Emotional Agents in a serious game of Cooperation and
Competition
Bernardo Jacobetty de Abreu Teles Grilo
Thesis to obtain the Master of Science Degree in
Information Systems and Computer Engineering
Supervisor: Prof. César Figueiredo Pimentel
Examination Committee
Chairperson: Prof. José Luı́s Brinquete Borbinha
Supervisor: Prof. César Figueiredo Pimentel
Member of the Committee: Prof. Rui Filipe Fernandes Prada
November, 2015
Abstract
Emotions influence the way humans think, make decisions and act, determining their beliefs, motivations
and intentions. It is a generally accepted fact that these influences are, most of the time, beneficial. In this
work, we developed agents capable of replicating these emotion-driven decisions and that are also able to
recognize emotion in other agent’s actions.
These agents exist in the context of a serious multi-agent vs player game with cooperative and competitive components. In this game, agents are able to display behaviours associated with emotions of gratitude
that can promote cooperation between the community the agents belong to. On the other hand, agents also
display behaviours associated with anger that can help them avoid adverse behaviour from other agents.
In this work we also evaluate the developed agents according to believability of their behaviours, their
accuracy to human behaviour emulation and if these agents can actually get an advantage over other agents
through emotional action.
The solution that supports this work was developed with the INVITE Game in order to build our agents’
affective model.
Key words: Agents, Emotion, Sentiment, Influence, Memory, Decision-making, Cooperation, Competition,
Appraisal, Personality
2
Contents
1 Introduction
8
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8
1.2 Problem Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9
1.3 Document Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10
2 State of the Art
11
2.1 Problem Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11
2.1.1 Relevant Game Theory Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11
2.1.2 Cooperation and Competition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
14
2.2 Solution Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
16
2.2.1 The OCC Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
16
2.2.2 Sentiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
19
2.2.3 Big Five (OCEAN) Personality Theory . . . . . . . . . . . . . . . . . . . . . . . . . . .
20
2.3 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
22
2.3.1 Affective Cognitive Learning and Decision Making . . . . . . . . . . . . . . . . . . . .
22
2.3.2 Emergent Affective and Personality Model . . . . . . . . . . . . . . . . . . . . . . . . .
23
2.3.3 Theories on circumventing ”Tragedy of the Commons” . . . . . . . . . . . . . . . . . .
25
2.3.4 Cooperative Coevolution of Multi-Agent Systems . . . . . . . . . . . . . . . . . . . . .
26
2.3.5 Metrics for Character Believability . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
28
2.3.6 INVITE Game . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
29
2.3.7 Starting Point - Grateful Agents and Agents that Hold a Grudge . . . . . . . . . . . . .
30
3 Solution
33
3.1 INVITE Game configurations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
34
3.2 New Rules for Emotional Engagement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
34
3.2.1 Holding a Grudge: Nominations Mechanic . . . . . . . . . . . . . . . . . . . . . . . . .
35
3.2.2 Gratitude: Alliance Mechanic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
39
3.3 Scenario Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
44
3.4 Emotional Agent’s Day-to-day cycle
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
45
3.5 Agent Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
46
3.5.1 Agent Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
46
3.5.2 Emotional Agent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
46
3.5.3 Perceptions Manager . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
47
3.5.4 Personality Manager . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
48
3.5.5 Emotions Manager . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
48
3
3.5.6 Sentiments Manager . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
48
3.5.7 Nominations Manager . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
49
3.5.8 Alliances Manager . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
49
3.5.9 Logger Manager . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
49
3.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
50
4 Data Analysis
51
4.1 Agent-only Scenarios . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
51
4.1.1 Scenario 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
51
4.1.2 Scenario 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
53
4.1.3 Scenario 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
56
4.1.4 Scenario 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
58
4.1.5 Scenario 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
60
4.1.6 Scenario 6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
62
4.1.7 Scenario 7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
64
4.1.8 Scenario 8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
66
4.1.9 Scenario 9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
69
4.1.10 Scenario 10 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
72
4.1.11 Scenario 11 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
75
4.1.12 Scenario 12 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
77
4.1.13 Scenario 13 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
80
4.1.14 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
84
4.2 User Tests Observations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
85
4.2.1 Players vs Agents - Resource Collection Efforts . . . . . . . . . . . . . . . . . . . . . .
85
4.2.2 Player’s Nominations management . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
86
4.2.3 Player’s Alliances management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
87
4.2.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
88
4.3 User Feedback . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
89
5 Conclusions
96
5.1 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4
98
List of Figures
1
Emotion elicitation process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
16
2
Original OCC Model Structure of emotions . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
17
3
Revised Structure of emotions of the OCC Model . . . . . . . . . . . . . . . . . . . . . . . . .
19
4
Affective Cognitive Learning and Decision Making . . . . . . . . . . . . . . . . . . . . . . . .
23
5
Guide Personality Model Dimensions Mapping . . . . . . . . . . . . . . . . . . . . . . . . . .
25
6
Prey vs Predators simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
27
7
Neural Networks - Central vs Independent . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
27
8
Communication - With vs Without . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
27
9
Emotional agent’s day-to-day cycle diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . .
45
10
Emotional Agent Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
46
11
Resource collection for each agent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
52
12
Total wood collection the team. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
52
13
Nominations results for scenario execution. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
53
14
Tolerance for gold collection. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
53
15
Resource collection for each agent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
54
16
Nominations results for scenario execution. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
55
17
Resource collection for each agent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
56
18
Nominations results for scenario execution. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
57
19
Resource collection for each agent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
58
20
Nominations results for scenario execution. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
59
21
Resource collection for each agent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
61
22
Total wood collection the team. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
61
23
Nominations results for scenario execution. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
62
24
Resource collection for each agent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
63
25
Nominations results for scenario execution. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
63
26
Resource collection for each agent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
64
27
Resource collection for each agent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
65
28
Gold collected by each agent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
65
29
Resource collection for each agent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
66
30
Resource collection for each agent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
66
31
Gold collection source for each agent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
68
32
Number of Alliances for each agent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
68
33
Resource collection for each agent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
69
34
Gold collection source for each agent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
71
35
Number of Alliances for each agent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
71
5
36
Nominations results for scenario execution. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
72
37
Resource collection for each agent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
73
38
Gold collection source for each agent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
73
39
Number of Alliances for each agent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
73
40
Nominations results for scenario execution. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
74
41
Resource collection for each agent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
75
42
Gold collection source for each agent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
76
43
Number of Alliances for each agent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
76
44
Nominations results for scenario execution. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
76
45
Resource collection for each agent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
77
46
Gold collection source for each agent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
79
47
Number of Alliances for each agent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
80
48
Nominations for each agent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
80
49
Resource collection for each agent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
81
50
Gold collected from alliances for each agent. . . . . . . . . . . . . . . . . . . . . . . . . . . .
81
51
Resource collection for each agent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
82
52
Nominations for each agent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
83
53
Agent behaviour coherence evaluation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
90
54
Agent change w/ experience evaluation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
91
55
Agent emotion perception evaluation.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
91
56
Agent personality uniqueness evaluation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
92
57
Agent awareness evaluation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
93
58
Agent emotional expressiveness evaluation. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
93
59
Agent unpredictability evaluation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
94
6
List of Tables
1
Prisoner’s Dilemma pay-off matrix. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
13
2
Possible Emotional Agent configurations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
44
3
Agent configurations for scenario 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
51
4
Agent configurations for scenario 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
54
5
Nominations for scenario 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
55
6
Agent configurations for scenario 3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
56
7
Nominations for scenario 3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
57
8
Agent configurations for scenario 4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
58
9
Nominations for scenario 4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
59
10
Agent configurations for scenario 5. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
61
11
Nominations for scenario 5. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
62
12
Agent configurations for scenario 6. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
63
13
Agent configurations for scenario 7. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
64
14
Agent configurations for scenario 8. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
66
15
Alliance cooperative actions and reciprocations for scenario 8. . . . . . . . . . . . . . . . . .
67
16
Agent configurations for scenario 9. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
69
17
Alliance cooperative actions and reciprocations for scenario 9. . . . . . . . . . . . . . . . . .
70
18
Nominations for scenario 9. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
72
19
Agent configurations for scenario 10. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
72
20
Agent configurations for scenario 11. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
75
21
Agent configurations for scenario 12. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
77
22
Alliance cooperative actions and reciprocations for scenario 12. . . . . . . . . . . . . . . . . .
78
23
Nominations for scenario 12. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
78
24
Agent configurations for scenario 13. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
80
25
Alliance cooperative actions and reciprocations for scenario 13. . . . . . . . . . . . . . . . . .
82
26
Nominations for scenario 13. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
83
7
1
Introduction
1.1
Motivation
In recent years there has been a substantial increase in the study of Affective Computing, leading to very
important breakthroughs relevant to this field. We call Affective Computing the field of study and development of systems that can recognize, interpret, process, and simulate human emotions. The field’s increased
relevance comes from a growing need from society to create technology capable of interacting on an emotional level with humans, whether by systems being able to recognise user-generated emotion or by those
systems being able to express their own emotions. Our work’s focus is to design an agent-driven solution
for intelligent systems that enables emotion expression and allows for decisions to be made with emotional
reasoning, in a way that is recognizable to a human audience.
It is stated in Pimentel’s work [14] that the influence of emotions on rational thinking has long been
recognized since the times of ancient Greece, when Aristotle stated that the judgements we make depend
on whether we are influenced by joy or sorrow, love or hate. Spinoza considered emotions as states that
influence the mind to choose one option over another.
Pimentel [14] also specifies that until recently, this notion of the influence of emotion on rational thought
was undermined since this influence was seen as an impairment of reasoning. Kant coined the expression
”emotions are an illness of the mind”, and Descartes said that passions must be tamed/subjugated by reason, if we don’t want them to prevent rational thought. Emotions were seen as an impairment of rationality,
and for rational processes to happen, emotions had to be taken out of the equation.
When computer science began the development of ”intelligent” computer systems, it did not require
emotions to achieve its purpose. At the time, they were not seen as a relevant aspect of intelligence.
These computer systems focused on reasoning and the performing of logical processes, aspects considered
decisive of whether a system is intelligent or not.
This view of what defined intelligence was heavily supported. However there were different views that
considered the factor of emotion as the one deciding factor that actually defined intelligence. One such view
is the one of Minsky [12] who says that ”the question is not whether intelligent machines can have emotions,
but whether machines can be intelligent without emotions.”
We believe that this mechanism of emotions that we use on our daily lives, in order to improve our
decision-making process, can be replicated to intelligent systems and this is the main effort for the solution
that we developed and discuss throughout this document. Also, another purpose for our work was to measure our solution for affective agents, concerning their believability towards being capable to emulate human
behaviour both by enabling them to act emotionally towards others and also in generating and processing
emotions in a way that is recognisable to a human audience.
8
1.2
Problem Description
The problem that motivated our work, is the development of a multi-agent system capable of generating
and processing emotions, as well as, making decisions guided by those emotions. As we’ve stated on the
previous section, this work had two main objectives, the first one was to find evidence that affective agents
(agents that can emulate emotional processes) can benefit from a decision-making process that is guided
by emotions which will enable those agents to gain more utility then agents without affective behaviour;
the second was to test our solution against a human audience and also find evidence that our solution for
affective behaviour is in fact more believable, to a human audience in terms of how affective agents interact,
when compared with agents with non-affective behaviour. We developed this solution for a multi-agent
system that is able to make emotion-based decisions and process emotions in the context of a Prisoner’s
Dilemma-type of scenario.
This work’s relevance is scoped in the context of the Prisoner’s Dilemma-type of scenarios where agents
interact through mechanics that give them a choice of either cooperating or defecting. This dilemma has
a fundamental problem which consists of self-interested agents always taking the logical decision that
guarantees a small amount of utility instead of risking an illogical decision that could cost utility
but would possibly maximize the agent’s future utility; we developed a solution for affective agents that
through emotion enables them to act emotionally towards each other and either benefit individually or collectively because they act towards their emotions towards a certain entity. To do this, we developed mechanics
that allow for emotional actions and an architecture for affective agents that enables those agents to harness
these mechanics to act emotionally towards each other. This solution focuses on two fundamental purposes,
to allow agents to influence others to act in a beneficial way towards them or to influence others in order
to avoid adverse behaviours from them. This will enables us to analyse the differences in utility gained by
an emotional and a non-emotional agent-driven system in the same scenario. Our agent solution was also
developed to enable agents to analyse others agents’ actions and perceive if those actions are guided by
an emotion related with gratefulness or anger towards the agent that is target of that action.
We concluded our work by taking data that resulted from several test scenarios for these affective agentdriven systems’ with different personalities and different configurations for possible emotional responses,
and elaborated on our conclusions relating to our hypothesis which states that emotions provide a way for
our affective agents to make decisions in a more efficient way, which will give them an edge against nonaffective agents. We also validated with a human audience how our solution stands in terms of the agent’s
behaviour believability and how those emotional actions translate to a human audience that interacts with
these agents.
9
1.3
Document Structure
This document is divided into four sections.
The State of the Art chapter establishes the research behind what inspired our work or helped on decisions concerning the architecture of our solution. The subjects tackled in this document are extensively
interdisciplinary, so there was a need to fragment and structure our research into two distinct subsections.
The subsection of Background where we present the most prevalent and contextually relevant theories that
relate to the fields of game theory, emotion, sentiment, personality, cooperation, competition, memory and
motivation. The subsection of Related Work, where we will refer to work that applied successfully these
theories and are insightful guide lines for the work we developed. We will conclude this chapter with the
Starting Point subsection where we introduce the work of Pimentel[9], which set a road map for what we
wanted to accomplish with the work developed and that is described throughout this document.
In the Solution chapter we describe in detail every element related with the development of our proposed
solution to the problem mentioned on the previous section. We will detail possible agent configurations for
scenario development when building the team of agents that will interact during the INVITE Game execution.
We will also go into new mechanics that we extended from the INVITE Game that were developed in order
to fit our needs to test our hypothesis and the agent behaviour algorithms for affective behaviour that go
along with each mechanic. To conclude this chapter the agents day-to-day cycle will be described along
with an overview of the agents underlying module architecture.
The following chapter will be Data Analysis where we document our test case that will allow analysis
and discussion of results and elaborate on our hypothesis. This chapter will start by documenting the tests
done with agent-only scenarios where each team will be only comprised of agents running our solution with
different configurations. The second part of this chapter will go into user tests in order to understand the
validity of the agent’s behaviour towards believability from a human audience standpoint.
Finally, the Conclusions chapter will sum up the important steps and decisions taken throughout the
development of our solution as well as discuss the validity of our hypothesis. To conclude, we propose
opportunities for future work based on the work we developed.
10
2
State of the Art
Because the scope of our work is very interdisciplinary, it was developed with an extensive study of many
theories from diverse fields like Game Theory, Affective Computing, Theories of Cooperation and Competition, Memory Mechanisms and Personality Theories.
Since these theories are themselves very extensive, having many aspects that are not relevant for this
work, the following section will delve into the relevant theories and detail the most important and pertinent aspects of each one. This research allowed the development of an architecture for a multi-agent
system with the specifications needed for validating our hypothesis and take conclusions concerning if our
emotion-driven agents can have an edge over non-emotional agents when influencing their decision-making
processes with emotion, as well as validate if these mechanisms of emotion were possible to replicate in
Agent-Oriented software system in a way that is believable to a human audience.
2.1
Problem Background
The following section introduces relevant work which acted as background for introducing the problem that
motivated the work we developed and which is described in the following sections.
2.1.1
Relevant Game Theory Concepts
As previously mentioned, fundamental concepts of Game Theory are the base on which our work stands.
Concepts like utility, encounters, dominance, equilibria and the prisoner’s dilemma, which are the concepts
that set a basis for the work we developed. In this section we will clarify these concepts for future reference.
To explain them we will resort to the work of [1] from Wooldridge, only referencing those concepts that
actually are in the scope of this project. This work starts by assuming a universe where there are only two
self-interested agents (i and j) interacting in a certain scenario. By self-interested we mean, agents that
have their own preferences and desires about how the world should be. It will also be assumed that there is
a set of ”outcomes” or states that the agents have preference over:
Ω = {ω1 , ω2 , ...}
The preferences of these two agents will be defined by means of utility functions, one for each agent,
which assign to every outcome a real number. This number serves as an indicator of amount of benefit that
an outcome brings towards that agent. The larger the number, the more beneficial that outcome is, thus
agent’s i and j utility functions are defined as follows:
ui : Ω → R, uj : Ω → R
11
Before defining agent encounters, there needs to be a model for the environment in which these agents
will act. These agents will simultaneously choose an action to preform in the environment, and as a result
of the actions they select, an outcome will occur. This outcome depends on the combination of actions
preformed by the two agents involved in the encounter.
It is assumed that these agents have to decide on an action and execute that action. It is also assumed
that the agents can’t see the action preformed by the other agent. The agents can only take one of two
possible actions Cooperate (C) or Defect (D), let Ac = {C,D} be the set of these actions. The way the
environment behaves is determined by the following function:
T : Ac(agent i’s action) × Ac(agent j’s action) → Ω
After defining the concept of action, immediately follow the concepts of dominance and equilibrium. As
in [1] we will adhere to the game theory literature and start calling actions by strategies.
Dominant Strategy: A strategy s1 dominates a strategy s2 if the set of outcomes possible by playing s1
dominates the set possible by playing s2. Dominance occurs when one strategy is more beneficial than any
other, regardless of how others might act.
Nash Equilibrium: In general, two strategies s1 and s2 are said to be in Nash Equilibrium if:
• under the assumption that agent i plays s1, agent j can do no better than play s2; and
• under the assumption that agent j plays s2, agent i can do no better than play s1.
We will go on to describe the multi-agent scenario that motivated the work that will be described across
this document.
The Prisoner’s Dilemma
Consider the following scenario, two men are collectively charged with a crime and imprisoned in separate cells. They have no way of communicating with each other or making any kind of arrangement. The
two men are told that:
• if one of them confesses to the crime and the other does not, the one who did not confess will be freed,
and the other will be jailed for three years; and
• it both confess to the crime, then each will be jailed for two years.
Both prisoners know that if neither confesses, then they will each be jailed for one year. It will be assumed
that confessing is a defection (D) and not confessing is a cooperation (C). Therefore, there are four possible
12
outcomes to the prisoner’s dilemma, depending on whether the agents cooperate of defect and so the
environment is of type, the scenario is described by Table 1:
j defects
j cooperates
i defects
2/2
0/5
i cooperates
5/0
3/3
Table 1: Prisoner’s Dilemma pay-off matrix.
Note that the numbers in the pay-off matrix do not refer to years in prison but how good an outcome is
for the agents, in other words the pay-off matrix represents utilities corresponding to each outcome.
Since the scenario is symmetric, where both agents reason the same way, then the rational outcome
that will emerge is that both agents will defect, giving them a pay-off of 2.
Notice that neither strategy strongly dominates in this scenario, so the first route to finding a choice of
strategy is not going to work. Turning to Nash equilibria, there is a single Nash equilibrium of D, D. Thus
under the assumption that i will play D, j can do no better than play D, and under the assumption that j will
play D, i can also do no better than play D. Obviously this is not the best utility that the agents can achieve.
If they had cooperated, then they would receive a pay-off of 3 each. But if you assume the other agent
will cooperate, then the rational thing to do, that maximizes your utility, is to defect. The conclusion seems
inevitable: the rational thing to do in the prisoner’s dilemma is to defect, even though this appears to ”waste”
some utility.
There have been several attempts to respond to this analysis of the prisoner’s dilemma, in order to recover/emerge cooperation between both agents. We will now dive into one such attempt that actually tackles
the key problem with the prisoner’s dilemma that motivates this work of creating emotional agents that engage in a prisoner’s dilemma-type scenario with the purpose of emerging cooperation between agents.
Iterated Prisoner’s Dilemma
As described in [1] this particular scenario is played an unknown number of times. Each play is referred
as a round and each agent is able to see what the other played on the previous round. Assuming that the
game is played forever, every round is followed by another round, this means that the agents will be meeting
the same opponent in future rounds, making the incentive to defect a lot smaller for two reasons:
• If an agent defects, the opponent can punish that action by also defecting. Take into account that
punishment is not possible in the traditional prisoner’s dilemma because it only has one round.
• Despite the loss of utility that results from getting the ’suckers pay-off’ from cooperating when the other
agent defects, that amount of utility lost is inconsequential compared with the utility that can be gained
in future rounds as a consequence of starting a cooperative relationship.
Because of these reasons, if the prisoner’s dilemma is played indefinitely, then cooperation becomes a
13
rational strategy. However, suppose the iterated prisoner’s dilemma is played a fixed number of rounds (lets
say 100), if we consider the final round, it basically is a one-shot classic prisoner’s dilemma encounter due
to the agents knowing that they won’t be interacting with each other after that encounter. As we have already
concluded, in a one-shot prisoner’s dilemma, the rational strategy is to defect, and both agents will defect in
the last round. So this makes the 99th round the final round. Counting this backwards induction, leads to
the inevitable conclusion that in a fixed iterated prisoner’s dilemma, defection is the dominant strategy.
This generates a phenomenon that is commonly called Tragedy of the Commons where each participant
in an encounter acts selfishly because they are self-interested, despite their understanding that they could
get better pay-off by cooperating. This is where our work comes in by making these agents interact in a fixed
iterated prisoner’s dilemma with the added feature that they reason about their strategies through personality
and emotion generation as a way of avoiding the tragedy of the commons.
2.1.2
Cooperation and Competition
Since the scenarios that we wish to simulate deal with the subjects of cooperation and competition, we
turned to the theory of cooperation and competition specified in [6] which was initially developed by Morton
Deutsch, and contemplates two main concepts, the concept of interdependence among goals of the people
involved in a given situation and the type of action taken by each individual. There are two main types
of interdependence, there is positive interdependence where the goals are linked in such a way that the
probability of an individual’s goal attainment is positively correlated with the probability of another individuals
attaining their own goals. On the other hand, negative interdependence occurs when the goals are linked so
that the probability of attaining the individual’s goals is negatively correlated with the probability of other individuals attaining their own goals. Putting it clearly, in a positive interdependence with another individual both
sink or swim together, when in a negative interdependence, only one swims when the other sinks. There are
also two types of action that can be taken by an individual, being effective actions, actions that improve the
individual’s chances of obtaining a goal, and bungling actions, actions that worsen the individual’s chances
of obtaining the goal.
People’s goals may be linked for various reasons. It is than logical that positive interdependence between
individuals can result from people liking one another, being rewarded in terms of their joint achievement,
needing to share a resource or overcome an obstacle together, holding common membership or identification with a group whose fate is important to them, being unable to achieve their task goals unless they
divide up the work or being influenced by personality. On the contrary, negative interdependence between
individuals will lead to people disliking one another, inequality of reward distribution among the group and
resource hogging.
There may also occur a situation of uneven interdependence relationships among individual, where one
participant of this relation is more dependent of it than the other is. In this type of relation, a master-slave
paradigm is usually verified, where the more independent individual will have control over the individual that
14
is more dependent and committed to that relationship.
Still in the scope of this theory of cooperation and competition, there are three basic social processes
inherent to every individual that must be mentioned and explained. The first process is substitutability and
defines how much an individual is accepting of other’s actions as a way of increasing the probability of
achieving his own goals. The second process is attitudes, this one defines, the attitude through which an
individual acts on a group context. A cooperative attitude is where an individual acts in favour of the group’s
interests with the knowledge that the group benefits from one another while a competitive attitude is where
an individual believes that everyone is against each other, and everyone is for themselves. Finally, the
third process is inducibility and refers to an individual’s readiness to accept another’s influence, in order to
do what the other wants. While positive inducibility can even induce the individual to harm himself for the
greater good of another, negative inducibility can induce an individual to obstruct the fulfilment of what the
other wants.
Now that cooperative and competitive processes have been explained, all that remains to be answered
is when do cooperative or competitive processes begin. The answer to this question is that, cooperation
induces and is induced by perceived similarity in beliefs and attitudes between individuals, readiness to be
helpful, openness in communication, trusting and friendly attitudes, sensitivity to common interests in instead
of opposed interests, orientation toward enhancing mutual power rather than power differences. In the same
way, competition induces and is induced by use of tactics of coercion, threat, or deception, attempts to
enhance the power differences between oneself and the other, poor communication, minimization of the
awareness of similarities in values and increased sensitivity to opposed interests, suspicious and hostile
attitudes.
Because of these facts, it can be concluded that the initiation of cooperative or competitive processes
is not a consequence of the individuals’ phenotype (traits perceived by others) but a consequence of their
genotype (traits inherent to an individual) of type of interdependence (positive or negative) and type of action.
15
2.2
Solution Background
The following section, much like the previous one, introduces relevant work which acted as background, this
time for the solution we developed to solve the problem of developing an architecture for affective behaviour
that occurs in human decision-making processes and replicate them for synthetic characters.
2.2.1
The OCC Model
The work we developed relies on the development of an agent-driven system that is capable of generating emotional responses to events. Appraisal theories are currently the most widely accepted theories of
emotion and are the ones that inspired most works developed in the field of Affective Computing. Appraisal
theories, according to the work of Clore [18], claim that the process of emotion elicitation, as depicted in
Figure 1 occurs in two steps. First, a situation or event must occur to become subject of appraisal processes,
this appraisal is done according to an agent’s concerns or goals and originates the appraisal outcome, which
can be viewed in terms of values of a set of variables. When these variables match a certain criteria, the
second step occurs where a specific emotion is elicited, these associations between variables and emotions
vary and are given by the appraisal theories.
Figure 1: Emotion elicitation process
There are many relevant appraisal theories in the field of Affective Computing. One approach to appraisal theory was through the Scherer’s Component Process Theory which defines continuous stimuli and
appraisal processes and also defines that the appraisal variables are continuous which leads to an infinite
universe of emotions. We did not progress further with this theory because what it established is beyond the
scope of what we want to accomplish with our work. Another promising approach was Roseman’s Theory
which defines emotion through the mapping of five variables. The variation of those variables originates one
of 16 emotions. Note that some emotions map to the same variable values. The fact that we needed to
implement this model with all five variables, just so we could use a portion of the emotions it is supposed to
deal with, meant that it was not ideal for our objectives.
The one theory that proved more appropriate for what we wished to accomplish was the OCC Theory,
because it is the most widely used and recognised appraisal theory in the field of Affective Computing which
16
is extensively documented in the work of Ortony [19]. Also because we based our work on the way the
OCC Model is approached in the revisions done by the work of Steunebrink [4], because of it’s ease of
partitioning the architecture only to the emotion hierarchies that we wish to develop and also because of it’s
intuitive structure in a software development point-of-view.
In this section we will go into some detail on what this theory is, how it works and how it can be interpreted
in order to facilitate the model’s implementation.
The work of [4] gives insight into the OCC Model stating that it describes a hierarchy that classifies 22
emotions as depicted by Figure 2. The hierarchy contains three branches, namely emotions concerning consequences of events, emotions concerning actions of agents and emotions concerning aspects of objects.
Some branches intertwine in order to cover compound emotions, those emotions are generated through
consequences of events caused by actions of agents. As the concepts of agent, event and object are well
known in agent-driven design, it makes the OCC model ideal for implementing affective agent systems.
Figure 2: Original OCC Model Structure of emotions
According to the emotion specification of the OCC Model, each emotion has three descriptive elements:
• Type Specification: provides the conditions that elicit an emotion of the type in question.
• Tokens: list of specific emotion words that are associated with the current emotion the agent is experiencing.
• Variables Affecting Intensity: these variables are local to the emotion in question. Typically, higher
variable values result in higher emotional intensities for the agent experiencing them.
17
The work developed in [4] aims to point out some ambiguities present in the OCC Model and present
possible interpretations and solutions to circumvent them. It is important to clarify that the solutions listed
below were designed from a computer science standpoint and are meant to solve ambiguities related with
the development of an accurate programmatic model for emotion generation through the OCC Model.
1. Desirable Event: In Figure 1. the phrase ”desirable event” is used many times. However, events are
always appraised with respect to their consequences hence, every instance of the phrase ”desirable
event” should actually be read as a shorthand for ”desirable consequence of an event”.
2. Generalization of emotion types: For example, ”being pleased” represents the situation where one
has appraised a consequence of an event as being desirable, but it says nothing about whether it is
hope or joy, or whether that consequence is presumed to be desirable or undesirable for someone
else, meaning one cannot say whether it is happy-for or gloating. When the structure as shown in
Figure 1. is regarded as an inheritance hierarchy, pleased/displeased, approving/disapproving and
liking/disliking then become generalized emotion types, from which all emotion types below them are
derived.
3. Joy vs Pleased: In the OCC Model, ”joy” is classified as an emotion type arising from positively appraising the consequences of an event where the focus is on consequences for the self and prospects
are irrelevant. However, the type specification of ’joy’ is given as “(pleased about) a desirable event”,
with no mention of a focus on the self or the irrelevance of prospects. This creates an ambiguity
between ”joy” and ”pleased”, just as ”joy”, ”pleased” is a valenced reaction to a desirable event.
4. Specification Crossing: According to the OCC Model, ’joy’ is specified as ”(pleased about) a desirable event” and ’distress’ is specified as ”(displeased about) an undesirable event”. Crossing these
specifications, one may wonder whether there is such a thing as ”being pleased about an undesirable
event” or ”being displeased about a desirable event”. In Figure 2. these duplications of specifications
are removed.
5. Hope/Fear Specializations: Satisfaction, Fears-confirmed, Relief, and Disappointment are not special kinds of hope or fear, but continuations of hope or fear, counting from the point when an event has
been perceived that signals the confirmation or denial of the thing hoped for or feared. In other words,
these four types are emotions in response to actual consequences of events, namely consequences
signalling the confirmation or not of a previously prospective consequence. In Figure 2., it is proposed
to move these four emotion types from under hope/fear to become specializations of joy/distress.
6. Fortunes-of-others emotions: Logically speaking, happy-for/gloating implies joy, and resentment/pity
implies distress. If the specifications of joy and distress are subsets of those of the fortunes-of-others
emotions, then joy and distress are generalizations of happy-for, resentment, gloating and pity should
be their parents in the hierarchy.
18
7. Aspects of objects: The branch of attraction-based emotions (liking/disliking and love/hate) is redundant, because there are no conditions to distinguish love/hate from its generalization liking/disliking. In
other words, love/hate does not seem to extend its parent type. According to OCC, the more familiar
one is with an appealing object, the more one will love it, and the more familiar one is with an unappealing object, the more one will hate it. To extend this branch of attraction-based emotions there needs
to be, besides love/hate, another branch that covers liking/disliking when confronted with unfamiliarity.
To do this, one could use the variable familiarity to differentiate between love/hate on the one hand for
familiar objects and interest/disgust on the other hand for unfamiliar objects.
In Figure 3 we can see the resulting revised model of the OCC Model that improves on the ambiguities
the work of [4] mentioned previously.
Figure 3: Revised Structure of emotions of the OCC Model
2.2.2
Sentiments
In this section we will mention briefly the topic of sentiments which is another relevant concept mentioned
several times in the work of Pimentel[9] making it a crucial part of the work we developed. According to [11],
a sentiment is a predisposition to trigger a certain emotion with respect to a certain object or individual. Such
predispositions can be acquired, for instance, due to past emotional experiences. Sentiments are referred
19
to as long-term emotional states, it is a long lasting predisposition that sporadically triggers emotions which,
in turn, may have long or short durations. A sentiment typically lasts long enough, to the point that it may
never fade away, but the individual only experiences the corresponding affective state when the object of the
sentiment comes to the individual’s attention.
2.2.3
Big Five (OCEAN) Personality Theory
To cover the topic of personality we resorted to the work of [3] which tells us that in psychology, the
Big Five personality traits are five dimensions of personality that are used to describe human personality.
The dimensions of the Big Five theory are openness, conscientiousness, extraversion, agreeableness and
neuroticism. For each of the five dimensions contemplated, a cluster of correlated specific traits is found. For
example, extraversion includes traits such as, gregariousness, assertiveness, excitement seeking, warmth,
activity and positive emotions tendency.
It is important to note that someone who scores highly on extraversion doesn’t necessarily exhibit all of
it’s specific traits, for example a person may be warm but not necessarily an excitement seeker. Each trait
has items that are used to attribute specific characteristics to each subject.
.Openness
Openness is a trait related with general appreciation for art, emotion, adventure, unusual ideas, imagination, curiosity, variety of experience and unconventional beliefs. People who are open tend to be, when
compared to closed people, more creative and more aware of their feelings.
Sample openness items: ”I have a vivid imagination.”; ”I am quick to understand things.”; ”I spend time
reflecting on things.”
.Conscientiousness
Conscientiousness is a tendency to show self-discipline, act dutifully and aim for achievement against
measures or outside expectations. High scores on conscientiousness indicate a preference for planned
rather than spontaneous behaviour.
Sample conscientiousness items: ”I pay attention to details.”; ”I get chores done right away.”; ”I like
order.”; ”I follow a schedule.”
.Extraversion
Extraversion is marked by a pronounced engagement with the external world, interaction with people,
enthusiasm, action-oriented behaviour. Subjects with high extraversion scores possess great group visibility,
20
like to talk and assert themselves.
Sample extraversion items: ”I am the life of the party.”; ”I feel comfortable around people.”; ”I start
conversations.”; ”I create movements.”
.Agreeableness
Agreeableness is a tendency to be compassionate and cooperative rather than suspicious and antagonistic towards others. The trait reflect a general concern for social harmony. Agreeable individuals value
getting along with others. They are generally considerate, friendly, generous, helpful and willing to compromise their interests for others. Agreeable subjects also have an optimistic view of human nature.
Sample agreeableness items: ”I am interested in people.”; ”I sympathize with others feelings.”; ”I have
a soft heart.”; ”I take time out for others.”; ”I feel others emotions.”
.Neuroticism
Neuroticism (or emotional instability) is a tendency to experience negative emotions, such as anger,
anxiety or depressions. Subjects that score highly in neuroticism often have low tolerance for stress or
adverse stimuli and are more likely to interpret ordinary situations as threatening. Their negative emotional
reactions tend to persist for unusually long periods of time, which means that they are often in a bad mood.
Neuroticism is connected to a pessimistic approach towards tasks and their completion.
Sample neuroticism items: ”I am easily disturbed.”; ”I change my mood a lot.”; ”I get irritated easily.”; ”I
get stressed out easily.”; ”I worry about things.”
21
2.3
Related Work
In the subsections that follow we will detail some prominent work done on the field of Affective Computing
that was relevant to our own work by defining a clear scope of the work we developed and also aided in
making decisions during the design of our solution’s architecture.
2.3.1
Affective Cognitive Learning and Decision Making
The work of Ahn and Picard in [15] goes into defining a clear mechanism for decision making applied to
affective agents which is a clear inspiration for what we achieved in our work when it comes to agents being
able to act towards their emotions.
When developing this mechanism they set some clear principles for their work, which are listed in their
work as follows:
• How previous emotional experiences assist the reasoning process (Bechara, Damasio).
• Most learning and decision-making models are still purely cognitive.
• The valenced affective feeling states provide fundamental values for guidance of behaviour.
What this work achieves is a mechanism that integrates these three inspirations and results on a decision
making algorithm for affective agents that takes into account cognition and emotion. In a very simplistic
overview the algorithm translates into the following formula:
Decision Value = external cognitive awareness + internal emotional valences
This formula tells us that agents when making their own decision on an emotional level they will take into
account the following:
• Their cognitive senses (perceptions) about the world that surrounds them in terms of all possible data
that goes into their emotional decision making process.
• Their internal emotional valences (affective state) and by this it means their own internal emotions
towards their situation and also their sentiments towards the other agents with whom they share the
world.
• Also as mentioned earlier, this algorithm also receives as input learned knowledge either from perceptions and continuous affective state.
To conclude the description of this important work we will leave the flow diagram in Figure 4 from the
described algorithm which illustrates in a summarized way everything that we’ve gone into throughout this
section.
22
Figure 4: Affective Cognitive Learning and Decision Making
2.3.2
Emergent Affective and Personality Model
The research documented in [5] describes the creation of an agent with ”attitude” to provide adaptive guidance and engaging interaction in an outdoor environment. To accomplish this, a personality model was
proposed to be implemented for the agent to be able to generate different personalities. The guide agent
will have a declarative memory model that will divide the agent’s memories into emotional memory and
semantic memory. It is important to note that this system does not generate emotions instead these are
implicit to the agent’s actions at a given moment in time. The basic architecture of the guide takes a lot from
the psi model, with the added feature of emotional memory.
Psi Model - Practical Use
In the work of [5] the agent perceives the environment continuously and generates intentions based
on external information and its own needs, meaning that the agent reads user-input, system feedback and
GPS information continuously. The agent then generates a goal (ex: story topic about a relevant monument)
based on all this information. These intentions together with the agent’s built in motivators - competence
level and uncertainty level - are stored in a memory of intentions, these intentions will then be selected by a
criteria of motivator strength. Some examples of contributions to these motivators are the user’s response,
for example, the degree to which he or she agrees with the guide’s argument, contributes to the guide’s
competence level, while the accuracy of the GPS reading contributes to the level of uncertainty. When
these motivators assume values that deviate from their desired values, they activate motivations for the
agent to take action in order to reset these motivators to their proper values. Besides these motivators,
there are three modulators that have critical roles when it comes to intention selection, they are arousal
23
level (speed of information processing), resolution level (carefulness and attentiveness of behaviour) and
selection threshold (how easy is it for another motive to take over) or in another word, the agent’s current
emotional state.
An agent with a higher arousal level will process information more quickly than a lower arousal level
agent. A careful agent will pay more attention to various circumstances and perform a more detailed planning before the execution of an intention compared to an agent with lower resolution level. While an agent
with a higher selection threshold will hold to its current intention more firmly than a lower selection threshold agent. Interaction between these modulators and built-in motivators results in complex emotional state.
There is no direct mapping of the high-level emotion labels to the different values of the modulators. In other
words, the resulting emotions are in the eye of the beholder.
Emotional Memory
Recent studies in neurology have provided evidence that memory files contain not only data or information but also contain emotions as well. Memory files thus consist of the information about an event and the
emotions we experience at the time when the event occurs. It is the emotional arousal, not the importance
of the information that organizes memory. The stronger the emotional factor associated with a memory, the
longer that memory remains. These emotionally arousing events are what comprises long-term memory,
being often responsible for behaviour control and mood setting.
The work developed in [5] adopts this idea, the guide possesses a long-term memory that is made up
of declarative memories, both semantic and emotional. Semantic memory is memory for facts, including
location-related information and user profile. While emotional memory is memory for experienced events
and episodes. Emotional information is characterized by two dimensions, arousal and valence, being that
arousal defines how exciting or calm an experience is and valence defines if an experience had a positive
or negative impact. This means that not only does the agent record information about what, when and how
an event occurred, but also attaches an ’arousal’ and ’valence’ tag to this information.
Personality Model
The guide’s personality model will affect the way it behaves and the way it presents stories. Take note
that the dimensions defined in this personality model is meant for the context of the guide’s purpose and it
is meant for tackling challenges specific to the tasks of guiding and telling stories. The dimensions defined
for the personality model are mapped in a 3D space and each point in that space maps to different combinations of these three dimensions, which are Impulsivity-Deliberateness which describe the resolution level,
Extraversion-Introversion which maps the arousal level, while the selection threshold is represented by the
Neuroticism-Stability dimension. A graphical representation of the model is shown in Figure 5.
In this model, rather than assigning different traits to the guide, personality emerges from varying the
weight of each dimension which ranges from 0 to 1. Different mappings of coordinates for each dimension
24
will result in different personality guides that when combined with the emerging emotions can produce a vast
range of expressions.
Figure 5: Guide Personality Model Dimensions Mapping
2.3.3
Theories on circumventing ”Tragedy of the Commons”
On this next section we will go into some well known theories formulated around a the dilemma associated
with a phenomenon that can be found frequently in many scenarios in the field of economy called ”Tragedy
of the Commons”. This phenomenon is based on a theory of economics which states that on a scenario
where a group of self-interested individuals are acting independently according to their own best interests,
will not act on the best interests of the group with the expectation that the others will. This propagates to all
individuals leading to the group ending up depleting some sort of common resource.
The following theories provide ways out of this dilemma by changing these individualistic behaviours
through factors which influence individuals into becoming altruistic towards each other.
Multi-level Selection
The first theory for avoiding the ”Tragedy of the Commons” is the work of Wynne-Edwards [16] on Multilevel Selection, proposing that in a society divided in groups that assumes that individuals are either
individualistic or altruistic, cooperative groups tend to be more successful than groups of defectors and
that what dictates success for an agent is if the agent is able to get to know not only it’s group but also
25
the others and with that information, pick the right group to belong to. These group selections are often
called Multi-level Selections as per the name of the theory. This is due to the social organization of the
individuals has a complexity that does not stop at interactions between individuals. Group selection
theories argue that an individual behaviour may spread in a population because of the consequences
of that behaviour influence the group’s performance in a visible way for the rest of the society.
Kin Selection
The second theory is related with the work of Hamilton [17] on Kin Selection which stipulates that
individuals who are more related to each other are more likely to cooperate. Hamilton’s rule is defined
by the following equation:
C
<r
B
r - represents the genetic relatedness of the recipient to the individual that performed the action.
B - the benefit gained by the recipient as a consequence of the altruistic action.
C - the cost to the individual performing the altruistic action.
What this means is the following, the relatedness factor between the possibly altruistic individual and
the individual target of that altruistic act must be higher than the cost/benefit ratio of this altruistic act.
If this condition is true, then the individual performing the action should act altruistically towards the
other, else it will not.
2.3.4
Cooperative Coevolution of Multi-Agent Systems
According to [7] in our lives, on an every day basis, there are certain tasks which concern not only our selves
but also the immediate population involved in the context of that same task. Sometimes there is a need to
create cooperation processes within that population in order to achieve a common goal. The pertinent
question is how do these cooperative processes initiate and how do these processes evolve. The work of
Yong and Miikkulainen gives very interesting answers to these questions while at the same making ground
breaking conclusions on how agents strive to cooperate. The model implemented to back this research uses
the Multi-Agent ESP Architecture of agents with neural networks which we will avoid detailing since it is out
of the scope of this project. It is also important to note that this research of cooperative coevolution does
not go into affective computing and the effects of personality or emotion in the cooperative process between
agents.
The work discussed in this section focuses on a simulation where four agents interact, of these four,
three agents are predators and the other one is a prey. The objective of this simulation is for the prey to be
captured by the predators by means of cooperation (role assignment), since they are unable to catch the
prey individually as depicted in Figure 6. For this purpose, three types of encoding, evolving and coordinating
were implemented. The first was a central-controller neural network, responsible for controlling the entire
26
team of predators and another for the prey, the second implementation is based on independent neural
networks for each agent in the simulation and finally the third implementation is based on on independent
neural networks but without communication between predators.
Figure 6: Prey vs Predators simulation
The results of these simulations can be found in Figure 7 and Figure 8 so as to more clearly explain the
conclusions taken from them:
Figure 7: Neural Networks - Central vs Independent
Figure 8: Communication - With vs Without
The results listed in the figures above provide evidence for the following conclusions:
• Cooperative coevolution is more powerful than a standard centralized approach in the context of the
work developed.
• Evolution without communication produces teams that evolve specific and rigid roles for each team
member and utilize a single effective strategy in all cases. On the other hand, evolution with communication tends to produce teams with more flexible (although less effective) agents able to employ two or
more different strategies. This team that employs communication, can also use combinations of these
strategies, depending on the situation. For example by starting with one and finishing with the other
leading to each predator not having a specific role it has to perform rigidly, but modifies the strategy
27
depending on the situation. Each predator behaves not only according to the prey’s relative location,
but also observes the other predators in deciding how to act. This way, their strategy is more versatile, but also less efficient. Whereas the non-communicating teams resemble players in a well-trained
soccer team, where each player knows what to expect from the others in each play.
• The non-communicating teams are more robust against unpredictable preys than the communicating
ones. Apparently, the first two prey behaviours, which are noisy versions of the original behaviour, are
still familiar enough so that the rigid roles are effective: the teams still catch the prey about half of the
time. All the agents have to do is track the occasional erratic movement, otherwise their strategy can
remain the same.
• Communication is not necessary for success, but may actually make evolution less effective.
To conclude, we would like to make note of one fundamental concept that arises from the results of this
study, which is the concept of stigmergy applied to cooperative and competitive behaviour. As described in
[8], Stigmergy is a mechanism of indirect coordination between agents or actions. The principle is that the
trace left in the environment by an action stimulates the performance of a subsequent action, by the same
or a different agent. In that way, subsequent actions tend to reinforce and build on each other, leading to the
spontaneous emergence of coherent, apparently systematic activity. This definition implies that traces of
actions of agents that exist in the environment can actually influence and possibly manipulate other agents
to behave in a more cooperative manner that is in accordance with the goals of the manipulative agent.
These traces can be categorized as rewards for good behaviours or incentives for agents that are behaving
competitively, to become more cooperative, aligning their current goals with the goals of the rest of the group.
2.3.5
Metrics for Character Believability
The purpose of the work developed in [10] by Gomes was to define metrics for evaluating believable virtual
characters in interactive narratives. Specifically, in analysing how an audience unfamiliar with the subjective
concept of believability can provide feedback on the believability of characters whose behaviour is defined
by a computational system. The goal is to create metrics to help researchers assess how a computational
system is contributing to the final believability of characters in interactive narratives.
In the work of Gomes [10], dimensions of believability are identified, for an audience to be able to report
and describe the overall believability of a computer system with interactive narratives.
Directly asking an audience how believable a character is can be an ambiguous question. Unless, the
audience is familiar with the notion of illusion of life, the answer will probably not reflect this concept. Hence,
a form of measurement is proposed that uses several believability dimensions contributing to the overall
perception of believability. In this way, participants are asked about more objective aspects of the agent.
Here are the believability dimensions that are considered to be reportable by an audience:
28
• Behaviour coherence
• Change with experience
• Awareness
• Behaviour understandability
• Personality
• Emotional expressiveness
• Social
• Visual impact
• Predictability
By having believability measured through quantifiable factors, measurements can be more informative
as design feedback towards making agents more believable.
Likert scales are commonly used in assessing individual subjective perceptions, thus it is proposed that
they be used, by having a scale for each dimension defined previously. This scale ranges from ”totally
agreeing” to ”totally disagreeing” and is applied to the dimensions by associating a phrase for each one,
for example ”awareness: X perceives the world around him/her”, where X is the name of the agent being
considered.
The exception to this methodology is emotional expressiveness which is proposed to be evaluated during
the interaction of the user with the interactive story thorough the multiple choice methodology, where the user
tries to guess the emotion a certain agent is expressing at a given time. The criteria of this method is that too
much predictability and too much unpredictability are negative feedbacks, making a middle ground between
these too extremes, the desirable feedback.
Despite the ambiguity that remains inherent to this proposed method of evaluation due to the highly
subjective theme of emotion that it tackles, it still proves to be useful and is sure to enhance the way users
can express them selves to developers, about the believability of agent-driven interactive stories.
2.3.6
INVITE Game
On this section we will mention the work on which our own solution stands on and which serves as the basis
of development for our developed agent’s solution which is described in detail on the Solution chapter, that
work is the INVITE game[13].
The INVITE game sets scenarios which are populated with synthetic characters with a social identity
module that would increase the agent’s believability and at the same time explore the topics of cooperation
between human players in the context of a Prisoner’s Dilemma-type of scenario, observing how cooperation
29
and satisfaction are related to each other. The game’s scenario is a Prisoner’s Dilemma when it comes
to its pay-off matrix that strongly resembles the one described in section 2.1.1, but the game’s complexity
comes from the number of participating agents. Because it is not a system comprised of two agents, it is
a multi-agent system, the pay-off matrix is that much more complex than the classic Prisoner’s Dilemma
scenario.
The main objective of the game is to have the highest amount of gold at the end of the game. However,
to complete the game a team needs to escape from the island before the volcano erupts. To do that, every
team will try to build a raft with wood. Every team member has a fixed amount of time to spend in everyday
and they can decide to gather wood or gold during that day. Those resources (gold and wood) can be
acquired at different places on the Island. The Dilemma comes in because, to obtain points for the final
score and win the game individually, every element of a team should help to build the raft but also needs to
have more gold than everybody else. The wood at the end of the game is divided in equally to help build the
raft, but the gold that each element of the team gathers is only for their own benefit and is not shared.
When the game starts, every player is spawned in their team camp site. Each player should go to the
resource site where they can use their available hours to obtain wood or gold. If they choose the wood, that
wood is to be shared with the rest of the team in order to build a raft to escape from the island. However ff
the player chooses gold, than that gold is for their personal benefit. After this section plays out, each team
member should return to their team’s camp. When all players are in their team’s camps the day advances
and the players have a new day to spend gathering resources. The user interaction is mouse-based with a
point-and-click mechanic of exploration.
This project was meant to be played by real players and for them to interact and deal with the Dilemma
that was proposed. In our work we wish to extend the INVITE Project by developing a multi-agent system
simulation where only agents will participate and engage in emotional interactions. Due to technological
limitations each simulation will need to have at least one Human Player to host the game. Since this is
a Prisoner’s Dilemma, the rational decision will be to defect (acquire gold), we want to develop an agent
architecture that enables emotion generation, personality definition and emotional action as way of breaking
this tendency for defection and promote cooperation between the agents.
2.3.7
Starting Point - Grateful Agents and Agents that Hold a Grudge
We conclude the Related Work section by introducing the work of Pimentel [9] which establishes the groundwork for what we accomplished with our own work described throughout this document. This work proposes
mechanisms for promoting mutual cooperation between agents in a multi-agent scenario where each agent
is self-interested and each has no guarantee that other agents will be cooperative towards the community
that they are a part of. In [9] it is argued that specific traces of human behaviour related with emotion
can stimulate agents to more cooperative behaviours. Specifically, it is suggested that if an agent behaves
gratefully and another agent recognizes this emotional behaviour, mutual cooperation may occur as a result
30
of rational decisions taken in accordance to the recognition of grateful behaviours on another agent, where
otherwise they would not occur.
It is also suggested in Pimentel’s work [9] that if an agent behaves in accordance with the emotion of
anger, ”holding a grudge” against another agent that recognizes this emotional behaviour, future adverse
behaviours may be prevented, as a result of rational decisions taken in accordance to the recognition of
anger-driven behaviours from other agents.
A pertinent reference is also made in [9] to a real-life example where these emotional actions occur. The
example tells the story of a group of five friends that occasionally used to play a board game of military
strategy. According to [9] in this game, each player starts out with control over certain territories and may
attack the territories of other players to gain control over them. The first time that the game was played,
one of the players, David displayed a behaviour known as ”holding a grudge” meaning that whenever he
was attacked by another player he retaliated and continued attacking that player, repeatedly, until the end of
the game even in situations where such attacks did not seem to be advantageous for him. In the end, both
David and his initial attacker failed to win the game.
However as described in [9], in subsequent games David continued displaying behaviours consistent with
the personality of someone that ”holds a grudge” against whoever attacked him. The other players quickly
realized this and started being biased against attacking David because being the target of his constant
attacks makes the task of winning the game a difficult one. Players did not entirely stop attacking David, but
they started taking his personality into account when choosing which opponent to attack. This gave David a
great advantage in the game, because he was rarely attacked.”
”Holding a grudge” is an affective behaviour associated with the emotion of anger. Anger is the emotion
that results from an event that has undesired consequences for oneself and was caused by the actions
of another agent. On the other hand, when an event caused by the actions of another agent has desired
consequences for oneself, the resulting emotion, is gratitude.
According to [9], for a multi-agent system that cooperates and competes for resources to benefit from the
mechanisms mentioned earlier, it is proposed that this system implements an architecture that contemplates
the following three key features:
• Affective Behaviour: The ability to decide and act in accordance with an emotion. In particular, when
an agent’s situation is improved by the actions of another agent, that agent should act in congruence
with gratitude, by returning the favour. When an agent’s situation is harmed by the actions of another
agent, it should act in congruence with anger, by retaliating.
• Identification of Affective Behaviour: The ability to identify that another agent’s actions are the
result of an affective state, and associate, with that agent’s personality, the predisposition for such
affective behaviours. This ability is within the scope of a theory of mind (Baron-Cohen, 1995), which
describes the ability to attribute mental states to oneself and others. In particular a simplistic theory of
31
mind is required, to identify the gratitude and anger-congruent behaviours described in the previous
item.
• Reasoning about Personality: The ability to make decisions that account for the other agents’ personalities. Based on the aspects of personality that were identified according to the previous item, the
agent should be able to extrapolate and predict the behaviours of other agents under certain conditions. These predictions should be taken into account when deciding the best course of action.
The work of Pimentel [9] also revisits the challenges concerning cooperation and competitiveness, in
interactions that involve self-interested agents. It proposes an approach, to these challenges, inspired on
human affective behaviours, attempting to reproduce the beneficial roles that the emotions of gratitude and
anger play in human social interactions. In this approach it is proposed that agent architectures contemplate
simplistic approaches to producing affective behaviour, recognizing affective behaviour, and reasoning about
personality.
This approach was tested against four examples, to show how acting on gratitude can promote cooperation and help form alliances among agents, and acting on anger can also promote cooperation, as well as
dissuade other agents from having adverse behaviours toward the agent in question. After recognizing the
emotional behaviours, agents decided to cooperate or to avoid adverse behaviour, not as artificial decisions
designed to simulate human behaviour, but as the rational decisions that aimed at maximizing the overall
present and future pay-off. These decisions take into account the personality of other agents, to help predict
their future behaviours in specific situations.
32
3
Solution
This chapter will describe in detail the solution that was developed in an effort to respond to the objectives
that we set out to achieve and that were enumerated previously in this document. This solution is tightly
related with the inevitable problem posed by the prisoner’s dilemma-type of scenarios, as it was described
in section 2.1.1, where we explained that in these scenarios, although agents have the choice to cooperate
or defect, will ultimately defect because, according to Game Theory, logic dictates that assuming that the
other participant of a Prisoner’s Dilemma-type of scenario will defect then we must also defect to guarantee
a small amount of utility. With our solution for affective agents we aim to break this inevitability by allowing
agents to add emotion to their decision-making process and enable them to consider not only logic but also
emotion when considering to either cooperate or defect. This should result in our emotional agents being
able to increase their gain of utility either by influencing other agents to act in a more beneficial way towards
the affective agents or simply dissuade them from continued adverse behaviour that is harmful towards the
affective agents’ utility.
These actions to either motivate or dissuade through emotional action other agents, comes with an added
risk in the form of spent utility for the affective agent. However if the scenario plays out in the agent’s favour,
that should translate into more utility, turning an irrational decision into a rational decision, when considering
the long run. With this objective in consideration, we developed an algorithm for agent behaviour based on
the work of Pimentel [9] on the role of affective behaviours in sustained multi-Agent interactions that was
described in section 2.3.7. Our goal, by developing this solution for agent’s that display affective behaviour,
is to find evidence that in a prisoner’s dilemma-type of scenario agents who can display affective behaviour
and also can detect those same emotional behaviours in others, will increase their utility and overcome
agents whose behaviours do not have these affective characteristics.
Our affective agents were developed to be able to interact with each other, as well with human players,
as mentioned in section 2.3.6, in the context of the INVITE game.
We developed an algorithm which instantiates agents that can assume different personalities, are able
to develop emotions, display affective behaviour and also detect affective behaviour in other agents. For our
specific scenario, we will only develop affective behaviours related with the emotions of Anger and Gratitude
as specified by [9] Pimentel’s work.
As we already mentioned, our objective with this solution was to validate if we can replicate the assumptions detailed in Pimentel’s work [9] in the examples given for grateful behaviour and behaviour of agent’s
that hold grudges. This allowed to compare and validate how agent’s can benefit or not from acting with
emotional behaviour and without emotional behaviour.
33
3.1
INVITE Game configurations
Before going into specifics about the solution’s development, we will mention some of the setup configurations that were made to the INVITE game for our specific scenario of validating interactions between our
affective agents.
The setup configurations made to the INVITE game when running our affective agent solution are the
following:
• The game takes place for 7 days until the time for eruption.
• Agents can collect 15 resources per day.
• The raft requires 150 pieces of wood to be completed and for the agents to survive the eruption.
• The game will be comprised of two teams. Each team has a total of 4 members. Team 1’s
members will all run our solution for agents with affective behaviour, as for Team 2’s members,
all team members will run our affective agent except for one which will be used for players to
user test our solution.
3.2
New Rules for Emotional Engagement
In order to conform to the scenarios set up by Pimentel in [9], our solution was developed taking into account
that our agents would have to interact with each other on an emotional level as well as be able to identify
emotional behaviour on others. Since the original INVITE game does not support this type of interaction that
allows for displays of emotion, we had to extend it and develop mechanisms of our own that would enable
our agents to engage each other in affective behaviour.
The main specification of the scenarios of emotional behaviour found in Pimentel’s work [9] which played
a great part in the design choices of the emotional interactions we developed was that the participants of the
scenarios had mechanisms of emotional interaction, both for grateful and grudgy behaviours, that allowed
each participant to direct their emotions towards specific individuals. This design limited our use of the core
INVITE mechanics to replicate these emotional actions. For example, if an agent was holding a grudge
towards another and because of that specific grudge, that agent would stop collecting wood, this agent
would not only be punishing the target agent of it’s anger but would be also punishing all agents that belong
to the same team.
The mechanics that were developed for our solution take this very important requirement into account
so that agents can act on their sentiments towards other agents in a targeted way. Two mechanics were
developed to do this, one for agents to act on their grudges and another for them to act on their gratefulness.
Here follow detailed descriptions of each one of these mechanics:
34
3.2.1
Holding a Grudge: Nominations Mechanic
The first mechanic that we will describe is the one developed for agents to be able to express anger towards
each other. Note, as was mentioned, that these mechanics for expressing emotion towards other agents
through action must be directed at specific agents and not either punish or reward all agents at the same
time.
The first mechanic we will go into will be the one related with displaying grudges towards other agents.
This mechanic is based on a common nominations system where agents vote on others with the objective
of showing their displeasure towards them. An agent will show it’s displeasure towards another agent’s
action if those action are too adverse to them, for example if an agent is indulging in individualistic behaviour
by collecting too much gold. These agents are much more likely to get nominated by other agents whose
behaviours are more focused on the common good than their own.
When the game ends, if the island’s eruption takes place and if a team of agents was able to build the
raft in time, these nominations that took place during the game, will be used as a criteria to decide if one of
the agents must be kicked from the raft. If one of the agents has accumulated more than ten nominations,
that agent will then be kicked from the raft. If more than one agent has more than ten votes, the agent to be
kicked from the raft, will be the one with the most accumulated amount of nominations.
We will now move on to the description of this mechanic for nominations with a list of rules for these
nominations:
• Nomination issuing occurs at the beginning of the day for every agent.
• Nominations are checked by all agents and then cleared at the end of every day.
• Every agent is entitled to one nomination for each day.
• Agents are allowed to not issue a nomination if they do not have reason to do so.
• Nominations are permanent and are cumulative until the end of the game.
• At the end of the game, when the eruption takes place, agents will tally the number of votes
and decide if an agent should be kicked from the raft.
• All agents have access to the all nominations that were issued by all other agents.
Now that we have described with detail the nominations mechanic for the INVITE game, we can now
move on to clarify how these nominations will enable the development of the behaviour that we set out to
develop in the first place which was for agents to act on grudges towards other agents who nominated them.
To develop this behaviour the first thing we did was give agents a personality trait that tells if an agent is
either prone to act on grudges or not.
Taking this into account, we will now detail in pseudo-code the logic behind how agents issue regular
nominations. Note that for this part we will mention alliances which is a concept that is used for the gratitude
35
behaviour algorithm. The algorithm for regular nomination is as follows:
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
if (I am Distressed AND i don’t have any grudges){
if (I am able to hold grudges){
for (each agent in my team){
if (the agent is the one that I feel the most anger towards AND i am not in an alliance with
the agent){
Clear list of current possible names for nomination;
Add agent’s name to my list of possible nominations;
}
if (my anger valence towards current agent equals agent with greatest anger valence yet
AND we are not in alliance){
Add agent’s name to my list of possible nominations;
}
}
}
else{
for (each agent in my team){
if (i didn’t detect that the agent holds grudges AND the agent has the greatest anger
valence AND we are not in alliance){
Clear list of current possible names for nomination;
Add agent’s name to my list of possible nominations;
}
if (i didn’t detect that the agent holds grudges AND my anger valence towards current
agent equals agent with greatest anger valence yet AND we are not in alliance){
Add agent’s name to my list of possible nominations;
}
}
}
if (i don’t have possible names for nomination){
Do not issue any nominations;
}
if (number of possible nominations is greater than 1){
Choose one agent’s name for grudge nomination randomly;
}
if (agent selected for nomination surpasses gold collection tolerance){
Choose the agent that was selected for nomination;
}
if (I am feeling Anger towards agent AND anger valence surpasses my tolerance AND I collected
less gold than selected agent){
Choose the agent that was selected for nomination;
}
else{
Do not issue any nominations;
}
}
Algorithm 1: Default behaviour for agents issuing nominations.
Note how in this algorithm we trigger an emotional decision in the agents resorting to the work mentioned
in section 2.3.1 which mentioned that agents make emotional decisions by evaluating their cognitive abilities
36
which in our case is the way other agents spent their resource collection effort and also by evaluating their
inner emotions towards those agents.
Emotion valence plays a very important role and is used in this algorithm, inspired in the OCC Model
mentioned in section 2.2.1. Also from the OCC Model we developed how anger actions are triggered, which
is a combination of reproach towards other agents, as a consequence of their adverse behaviour towards
the agent’s goals, combined with the agent’s emotional state being in distress.
In this algorithm we have two different occasions where an agent would issue a nomination. The first
is when the agent’s tolerance towards present adverse behaviour is broken, which means that the targeted
agent was too individualistic for the agent’s anger tolerance towards one particular resource collection effort.
The second occasion is when the agent’s tolerance towards continuous adverse behaviour is broken which
means that the targeted agent actually performed individually for too long the agent act on accumulated
anger from previous experience with the target agent for the nomination. These concepts of getting cognitive information, plus evaluating inner emotions towards other agents combined with maintaining previous
experience were all taken from the work mentioned in section 2.3.1.
Now we will move on to describe the algorithm behind issuing nominations for agents that are holding
grudges
1
2
3
4
5
6
7
if (I am able to hold grudges AND nominations were issued){
for (each nomination in nomination List){
if (agent A has nominated me AND he is not already in my list of grudges){
Add agent A to my list of grudges;
}
}
}
Algorithm 2: Agents that hold grudges checking nominations.
In this algorithm we are going into a specific feature for agents that are able to act emotionally towards
their grudges as described in the anger scenarios in Pimentel’s work [9]. Note how these particular agents
are able, much like agents that have this emotional behaviour disabled, to check nominations however the
agents that are able to act on grudges look through nominations issued towards them and how they keep the
names of the team mates responsible for those nominations in a list that references agents towards whom
sentiments of grudge are being felt by the agent.
Like with the algorithm for checking nominations for agents who hold grudges, we will also detail the
algorithm for selecting an agent to nominate when an agent holds grudges:
37
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
if (my personality holds grudges AND i am holding grudges towards other ){
for (each agent in my team){
if (agent is in my grudges list AND agent is the one that I feel the most anger towards){
Clear list of current possible names for nomination;
Add agent’s name to my list of possible nominations;
}
if (agent is in my grudges list AND the anger valence that I am feeling towards this agent
equals the anger towards whom I am feeling the most anger ){
Add agent’s name to my list of possible nominations;
}
}
if (no names selected for nomination){
for (each agent in my team){
if (agent is in my grudges list AND he is the one that I feel the least gratitude towards){
Add agent’s name to my list of possible nominations;
}
}
}
if (number of possible nominations is greater than 1){
Choose one agent’s name for grudge nomination randomly;
}
Choose the agent that was selected by the algorithm for grudge nomination;
}
else{
/* ...Default to regular nomination issuing algorithm...
*/
}
Algorithm 3: Agents that hold grudges selecting nominations.
This algorithm much like on the regular nominations issues an emotional decision of an agent nominating
another. But note how this time around the agent is not only looking for agents that are behaving adversely
towards them, it also looks for agents that are actually acting towards it’s goals (building the raft) but from the
moment the agent is ”attacked” and it keeps a grudge, that agent will issue an emotional nomination as a way
to dissuade others from ”attacking” it. This behaviour of emotional action towards grudges is documented in
Pimentel’s work [9] and the supporting mechanic for nominations issued from grudges, followed that same
specification.
Finally, we will describe the algorithm that enables agents to identify grudge related behaviour towards
other agents. This is useful to let the previous algorithm for regular nomination issuing to be able to know
which agents are capable of holding grudges and avoid nominating these agents. Note that since all agents
have access to every nomination issued during a day, this means that even if they weren’t the target of
a nomination, they will still be able to run this algorithm and identify if any agent from the team acted on
grudges.
38
1
2
3
4
5
Data: assuming that agent A is the one that issued the nomination and agent B is the agent that was
nominated.
for (each nomination in nomination List){
if (agent B collected more Wood than agent A AND agent A was nominated by agent B on the
previous round on nominations AND agent A is not yet identified as an agent that holds grudges){
Agent A is hereby considered an agent that holds grudges and is added to the list of grudgy
team mates;
}
}
Algorithm 4: Identifying agents who act on grudges.
Now that all algorithms are clarified for this mechanic, note the way that they were developed was to
enable agents to hold grudges towards other agents, act on these grudges and let other agents detect this
behaviour. Mind also how the algorithm for agents who don’t hold grudges is so that if an agent is about to
issue a nomination it will first check if that agent holds grudges and if it does then it will avoid nominating
that agent. This will give an advantage to emotional agents who hold grudges contrary to agents who do
not since, for example, an agent who holds grudges will be able to get away with an individualistic behaviour
without actually suffering the consequences of being nominated by it’s peers since they will fear the agent’s
grudge, while on the other hand, an individualistic agent who does not hold grudges will not be able to get
away with only collecting pieces of gold without getting constantly nominated.
3.2.2
Gratitude: Alliance Mechanic
The other mechanic that we will describe in this section is one that enables agents to act on gratitude
towards others. Like in the previous mechanic for issuing nominations, this one also was developed taking
into account that agents had to be able to act emotionally towards other agents in a targeted way to prevent
other agents from being rewarded from grateful behaviour that was not earned. This design, like for the
nomination mechanic, was inspired by Pimentel’s work [9] for scenarios for agents that display grateful
behaviours. To develop this mechanic we added to our agents another personality trait that defines if the
agents will display gratitude towards other agents or not.
The first instinct when designing this mechanic for grateful behaviour between agents was to enable
agents act gratefully towards another team mate by collecting more wood pieces to progress further the raft
completion, since this is the concept of being cooperative in the INVITE game. However this approach was
not adequate to the requirements scoped for our work since this way, every time an agent collected more
wood to reward a certain agent, it would actually improve every agent’s utility regardless of that agent having
earned that improvement or not. Because of this fact, we developed a different approach to this mechanic
that involves the resource of gold pieces instead of wood pieces, to try and design a reward system that
would allow targeted displays of grateful behaviour which is described across this section.
39
We enabled our agents with grateful behaviour by allowing them to create alliances among each other. If
an agent is enabled to display gratitude towards others it will try to find other agents who display this same
grateful behaviour with whom they can create alliances with. This process takes place at the beginning
of each day, where agents who are able to display gratitude start to probe other agents by offering them
pieces of gold as a gesture that represents a cooperative action. This is supposed to motivate the
targeted agents to also take cooperative action by returning that gold offer and through that action, create
an alliance with the other agents. Note that after an agent tries to create an alliance with another agent, that
agent becomes checked and won’t be probed again with cooperative actions in the future to create alliances
for the rest of the game.
Our emotional agents will risk a loss in utility when they issue a cooperative action that consists
of offering gold to initiate a potential alliance. This may seem like the agents are acting in a way that will
seem irrational, since agents are basically offering their utility to other agents which isn’t logical in a game
theory point of view. However if these agents are reciprocated by the targets of their cooperative incentive,
who also display gratitude, the utility lost with initiating the potential alliance will later be rewarded when
agents start cooperating and gaining extra resources that agents who do not display gratitude won’t be able
to enjoy since they are not able to create alliances.
The moment when an agent checks if it has any ”opportunities” to form new alliances by receiving gold
from other agents as a cooperative incentive, is when this mechanic allows what we want to accomplish to
come into focus, which is for grateful agents who are the targets of these cooperative actions, to be able to
reciprocate those actions and benefit from belonging to an alliance which is to earn extra pieces of gold for
each alliance that an agent belongs to.
Here follows the logic related with how agents earn extra gold for each alliance that they belong to:
Gold collected += 3 pieces of gold * Number of Alliances Members
In the example for grateful interaction between affective agents depicted in the work of Pimentel [9]
we can see how an agent takes the first step and risks utility in order to try and initiate a cooperative
interaction. This cooperative pattern of interaction will ensue only if the other agents gratefully reciprocate
that cooperative gesture. From this point, if both agents act cooperatively towards each other, then that
behaviour is prolonged throughout that scenario. Note however that although these agents will cooperate
and never betray that relationship on the specific scenarios depicted in the work of Pimentel [9], the agents
still keep the choice of defecting at any time instead of cooperating. This mechanic which enables agents to
acquire more gold in proportionate quantity to the number of alliances they belong to, aims to simulate that
same behaviour in the context of the INVITE Game’s concept.
Keep in mind that an alliance only takes place when an agent who was the target of a cooperative gesture
in order to start an alliance, reciprocates that same action, unless these agents decide to act cooperatively
40
towards each other on the same day which implicitly creates an alliance between them. From that point
forward, once the alliance is created, the mechanic for agents to collect extra pieces of gold during resource
collection is executed along with regular resource collection for the agents that belong to one alliance or
more.
To conclude the design specification of this mechanic we will refer a component related with how agents
can maintain their alliances, after them being established, across multiple rounds of interactions on each
day of an INVITE Game scenario.
This component of the alliances mechanic which tackles alliance maintenance, after it is established
between two agents, took influence from the typical scenarios for the prisoner’s dilemma described in section
2.1.1 where, after agents establish a cooperative behaviour pattern of interaction between each other, this
fact should not in any circumstance guarantee cooperation between both parts in perpetuity across multiple
rounds of interaction. Instead, agents must keep the ability to, at any time, choose to either keep cooperating
and maintain the alliance with that agent or defect which terminates the alliance with that particular agent.
Although we developed this mechanic to maintain the agent’s choice to either cooperate or defect at any time
during the course of an alliance. We refrained from developing agents that could display betrayal behaviours
towards agents with whom they established a cooperative behaviour pattern with. This is not in the scope
of the scenarios of gratitude specified in the work of Pimentel [9] where, once agents get in a pattern of
cooperative behaviours with another agent, although they have a theoretical choice to defect, they won’t.
Further details on how this mechanic works can be found on the following detailed descriptions developed
for our affective agents behaviours.
We will now move on to the description of this mechanic for alliances with a list of rules for how these
actually work:
• A cooperative action to initiate a possible alliance consist of an agent making on offering of
three pieces of gold to another agent.
• An agent has the option to reciprocate to a cooperative action or keep the three pieces of gold
that it received.
• An agent can respond to a cooperative action from an agent who tried to initiate an alliance by
sending three pieces of gold to that agent.
• After two agents exchange a cooperative action which is in turn reciprocated, the alliance
officially starts.
• Agents can’t vote on agents with whom they are keeping alliances with.
• If two agents issue cooperative actions towards other at the same time to initiate alliances,
then reciprocation is not required to form the alliance.
41
• When an agent belongs in an alliance, each alliance translates to extra three pieces of gold per
alliance to which the agent belongs to.
• An agent can form as many alliances as possible.
• Other agents are not aware of each others alliances.
• Agents can only issue one cooperative action and one reciprocation to a cooperative action
per day.
• At the end of each resource collection day, each agent will have a round where it must decide
whether to give three pieces of gold to each alliance member the agent has and maintain those
alliances or defect keep the gold that was given to it and terminate those alliances.
Once again, like we did for the algorithms for the mechanic that enables agents to display grudges, we
will describe in pseudo-code the relevant algorithms that are involved in the mechanic of alliance forging
between agents. We will start by detailing the algorithm for when grateful agents probe other agents and
start to issue cooperative actions to motivate the creation of alliances. This algorithm takes place every
beginning of day and it’s logic is as follows:
1
2
3
4
5
6
7
8
9
10
if (agent is able to act gratefully){
for (each of my team members){
if (i didn’t yet probe the current team member AND today i didn’t nominate the current team
member AND I have gold to donate){
Add team member to list of probed agents;
Decrement by three the number of pieces of gold;
Issue cooperative action towards agent;
Done with initiating alliances;
}
}
}
Algorithm 5: Grateful agents issuing cooperative actions to initiate alliances.
Note how the nominations mechanic has a part to play in this algorithm in the sense that if an agent
nominates another, it is because the behaviour displayed by the nominated agent was too adverse to the
agent’s goals by increasing it’s anger, which ”black-lists” the agent nominated from being able to form an
alliance with this agent. Here we can see how the concepts of both references from sections 2.2.1 for the
OCC Model and 2.3.1 for Affective Decision Making are used to define how grateful agents are able to start
the alliance creation process.
The next algorithm is executed right after the previous one for issuing cooperative actions and is responsible for enabling the agents to check if they were reciprocated by any of their issued cooperative actions
and the logic is the following:
42
1
2
3
4
5
6
7
8
9
10
if (agent is able to act gratefully){
/* ...Algorithm for issuing cooperative action...
for (each reciprocated actions issued by other team members){
if (any of those is targeted at me){
Increment by one my number of alliances;
Increment by three my amount of pieces of gold;
Add reciprocation issuer to my alliances members;
Remove processed reciprocation action from repository;
}
}
}
*/
Algorithm 6: Grateful agents checking alliance responses.
We will move on to detail the algorithm for when agents check for alliance opportunities issued through
cooperative action and also decide if they should reciprocate towards those actions. Mind that this takes
place when agents return to the camp site after a day of resource collection and the logic executed is this:
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
for (each cooperative action issued by agents){
if (i cooperative action was targeted at me){
Increment by three my amount of pieces of gold;
if (i am a grateful agent AND agent that issued cooperative action is the one I chose to
nominate today){
Do not reciprocate cooperative action;
Remove processed cooperative action from repository;
}
if (i am a grateful agent){
Reciprocate cooperative action for agent who initiated alliance;
Add agent to list of probed agents lists;
Increment by one my number of alliances;
Increment by three my number of collected pieces of gold;
Add agent to my list of alliances members;
Remove processed cooperative action from repository;
}
}
}
Algorithm 7: Agents checking cooperative actions issued and respective reciprocations.
Note how agents who do not display grateful behaviours also check for cooperative actions of gold
donations and will get the gold that was donated to them but these agents will not respond to them since
they are not grateful agents. This is how agents who initiate alliances will be able to flag which agents are
willing to create alliances and which agents are not.
Finally, we will go into the last mechanic for the alliances feature which is related with enabling agents
to choose whether to maintain or betray an alliance that they formed. Note that this was only developed
because the algorithm had to conform to game theory prisoner’s dilemma-type of scenarios where each
round must have a choice of cooperation and defection. However, our affective agent’s behaviour does not
43
take defection into consideration since the scope of Pimentel’s work [9] does not go into betrayal related
behaviours. For these reasons, our agents will never defect on their alliances, nevertheless, we kept this
choice available to conform to game theory requirements.
1
2
3
4
5
6
7
8
9
10
/* ...Start of day...
if (I have an alliance members){
for (each alliance member ){
if (alliance member did not give three gold pieces to me at the end of previous day){
Remove that team member as an alliance member;
}
}
}
/* ...End of day...
for (each alliance members){
Give three pieces of gold to that alliance member as a gesture to maintain alliance;
}
*/
*/
Algorithm 8: Grateful agents maintaining established alliances.
3.3
Scenario Design
As was already mentioned in the previous sections of the solution description our agents will need to have
different personalities and will also have to be able to hold grudges and/or show gratitude towards other
agents in order to be able to act on their emotions. Table 2 shows the possible setup configurations for our
emotional agents:
Agent attribute
Personality
Holds Grudges
Shows Gratitude
Possible configurations
Highly Cooperative
Cooperative
Lowly Cooperative
Lowly Competitive
Competitive
Highly Competitive
YES
NO
YES
NO
Table 2: Possible Emotional Agent configurations.
The personality attribute of our emotional agent is used as criteria to define the agent’s beliefs and
goals for the game so if an agent is on the interval of competitive agents, these will focus on individualistic
behaviour and their main goal will be to become the agents with the most gold collected with increasing
individualism as agents are more competitive. The opposite happens for agents that are set to a more
cooperative personality configuration, making their goal more related with building the raft as fast as possible.
44
Like with competitive behaviour, also as agents are more cooperative they will focus more on cooperative
behaviour.
As for the ”holds grudges” and ”shows gratitude” attributes they represent how the agent expresses it’s
emotions and these come into the algorithms that were explained for the mechanics of nominations and
alliances that were described in detail in section 3.2. It is also important to point out that on our test cases,
an agent with the absent abilities of both holding grudges or showing gratitude will be considered agents
without affective behaviour.
3.4
Emotional Agent’s Day-to-day cycle
In this next section we will go into how our emotional agents spend their day-to-day routines in our specific
scenario that we developed with the INVITE game. The diagram displayed in Figure 9 shows how the agent’s
routine takes place and the order in which every key moment happens. Note that this diagram includes our
own changes to the INVITE game so we can clearly understand when nominations related actions and
alliances related actions occur.
Figure 9: Emotional agent’s day-to-day cycle diagram
45
3.5
Agent Model
We will conclude the solution chapter of this document with an overview of the emotional agent’s underlying
architecture. In this section we will start by defining how each module that comprises our solution, structures
the emotional agent’s architecture and how these modules relate and interact with each other.
After structuring the architecture we will move on to describing each one of the defined modules that
make up our agent.
3.5.1
Agent Architecture
Our architecture is depicted in Figure 10 along with each module that makes up the emotional agent and
how these are related to each other.
Figure 10: Emotional Agent Architecture
3.5.2
Emotional Agent
The emotional agent module is the one responsible by insuring the main behaviour of our emotional agent. It
is in this module that the day-to-day activities are executed and where the agent makes decisions according
to several input data maintained by all the other modules.
We can see by the architecture diagram of the emotional agent, that this module is a centralized hub
through which all others can access each other and read data as needed. The Perceptions Manager,
Personality Manager, Emotions Manager and Sentiments Manager objects are instanced per agent which
means that each agent maintains their own references to these modules.
46
On the other hand, both the Alliances Manager and Nominations Manager are static objects only instanced once per game. These two modules are special in the sense that they need to be shared among
agents and their state had to be maintained across the whole game execution.
3.5.3
Perceptions Manager
This next module of our agent architecture is the one responsible for maintaining an updated state of what
the agent perceives about several data relevant about the world that the agent is populating. These perceptions are the point of entry for any relevant information that is external to the agent and this information will
than be used to enable the agent’s decision making processes.
Here follows a list of all data that is being kept by this module:
• Agent’s name.
• Current day.
• Days to eruption.
• Raft progression.
• Gold gathered from alliances.
• Number of alliances.
• Total wood currently gathered by the team.
• Amount of wood needed to complete raft.
• List of resource collection actions per agent for the current day.
• List of resource collection actions per agent for the previous day.
• List of total resource collection actions per agent for all the days that took place during the
game.
• List of agents already probed for alliance.
• List of alliances members.
47
3.5.4
Personality Manager
The Personality Manager module was developed conforming to a combination of influences, first the personality traits were defined from what was available on the OCEAN personality theory, specified in section 2.2.3,
and that made the most sense in the context of our particular scenarios; meanwhile the work mentioned in
section 2.3.2 gave as clear inspiration on how to quantify programatically the existence, non-existence or
amount of a personality trait that a certain agent displays. We gave agents personality traits that set clear
goals for them, which in turn will make agents have a behaviour that is consistent enough to enable them to
stir up emotions towards each other ultimately enabling them to act on these emotions which is the purpose
of our work. The agent’s state that is being kept by this module’s attributes are the following:
• The Agents Personality Trait (possible values for this attribute were disclosed on Table 2).
• Agent Holds Grudges? (Yes/No)
• Agent Is Grateful? (Yes/No)
3.5.5
Emotions Manager
This next module keeps the updated general agent’s emotion of how the it feels towards what it’s perceiving.
This emotion only assumes two possible values which are Joy or Distress as specified by the OCC Model in
section 2.2.1. Note that each emotion of Joy and Distress are valenced as a way of defining the intensity of
the emotion which will than mean different things for each agent depending on their Personality Traits. For
agents who are more competitive Joy and Distress depends on how the agent ranks on the amount of gold
collected, while for cooperative agents depends on how the raft is progressing.
3.5.6
Sentiments Manager
The Sentiments Manager is a very important module of our architecture since it handles the current sentiments (emotions associated with an object or individual) that the agents are experiencing towards each
other.
Agents keep a list of all agents that belong to their team and associate with them a sentiment as described in section 2.2.2 which goes into the general idea on how sentiments are managed. Our agents’ can
either assume sentiments of Anger or Gratitude towards others. Each sentiment towards an agent will also
have a valence associated with it, which sets the intensity of the sentiment as defined by the OCC Model in
section 2.2.1. These sentiments will be kept updated through the game’s execution as a reference from the
continuous interaction and bonding between the agents. These sentiments will then come into play for the
agent’s decision making when acting on their emotions towards other agents as defined on the work of [15]
in section 2.3.1.
48
3.5.7
Nominations Manager
Like it was mentioned before, the Nominations Manager like the Alliances Manager belong to a special group
of modules in the agent architecture because their are only instantiated once and that instance is shared
among all agents that participate in the game. Besides this, the object needs to be able to manage it’s
status along the whole game’s execution while also managing multi-threaded transactions since each agent
is managed by different processes.
This is required because this specific object keeps an updated list of the nominations that are issued by
each agent every day.
All agents access the static instance of this object to check nominations related information. The attributes defined for this object are the following:
• Dictionary of nominations issued where it’s keys represent nominators and it’s values represent nominees.
3.5.8
Alliances Manager
Just like the previous module, also the Alliances Manager keeps information that is shared across all participating agents. In this particular object, all cooperative actions issued and respective reciprocations for
alliances are kept here.
All agents access the static instance of this object to check alliances related information. The attributes
defined for this object are the following:
• Dictionary of cooperative actions issued for alliances where it’s keys represent cooperators
and it’s values represent targets of cooperation.
• Dictionary of reciprocations for alliances where it’s keys represent reciprocators and it’s values
represent targets of reciprocations.
3.5.9
Logger Manager
This object is another one that is crucial to our architecture because it enables agents to write onto text
files every thing that they go through during the game’s execution. These files are written in the form of
diaries of the agents where, for each day, agents write down their current perceptions, their emotions, theirs
sentiments and also their decisions.
However, this object’s greatest value is for data gathering, since agents keep all information on their
perceptions available for reading. This is particularly useful for extracting relevant data for all manner of
tests, their documentation and finally the analysis of results.
49
3.6
Summary
During this chapter we have covered the setup configurations of the INVITE game, we described in detail the
new rules that we added to the INVITE game which were developed in the scope of our prototype namely the
nominations and alliances mechanic. After that we moved on to listing all possible emotional agent attributes
setup configurations. We followed with a description of the diagram depicting the day-to-day routine of the
emotional agents with the new processes for expressing emotion. Finally we went into some detail about
the underlying emotional agent’s architecture which plays the INVITE game with the new rules for emotional
expression that we defined in this chapter.
With the solution description concluded we will move on to the section that details all the results that
were able to gather from executing several tests to the emotional agent solution that was implemented.
50
4
Data Analysis
This next chapter will go into the analysis of all data extracted from the execution of several test scenarios on
the solution presented on the previous chapter. This analysis is divided in two different sections, the first one
documents all data related with scenarios that were executed with a team composed only of agents that are
running our solution for affective behaviour. On the other hand, the second section goes into the analysis of
data extracted from scenarios that were tested with teams composed of human players and agents.
We had two hypothesis to validate through the data that we were able to gather. First we wanted to
show evidence that, with our solution for affective behaviour, agents that are able to act emotionally towards others can benefit from an advantage against agents who are not enabled to act emotionally. The
second hypothesis can only be validated through user tests which was to get evidence that our solution for
agents with affective behaviour believability from a human player’s perspective increases when an agent
acts emotionally towards them.
4.1
Agent-only Scenarios
As was mentioned in the introduction to this chapter we will start by documenting the tests done with scenarios composed only of agents running our solution for affective behaviour. For each scenario we will detail
the configuration that is being tested for that specific scenario and than we will display graphs with several
relevant data gathered from the INVITE game which we analysed and commented on.
4.1.1
Scenario 1
The first scenario that we tested for teams composed only of agents that are running our solution for affective
behaviour is shown in Table 3:
Agent Name
Berna
Jamie
Pat
Simon
Personality Trait
Competitive
Cooperative
Cooperative
Cooperative
Holds Grudges
No
No
No
No
Is Grateful
No
No
No
No
Table 3: Agent configurations for scenario 1.
This first scenario, as described on Table 3, has a team of agents that are not enabled to act either
on grudges or on gratitude. This scenario is set as a basis for the next scenarios analysis. We want to
understand how agents that don’t act on their emotions behave and what utility they can actually harness
in terms of the INVITE Game’s concepts. With this data we can compare it against the scenarios that will
follow where affective agents will be a part of, and after careful analysis, reach conclusions.
We will now display all relevant data extracted from the execution of this particular scenario.
51
Figure 11: Resource collection for each agent.
We can see from Table 3 how this is a scenario for setting a basis for our analysis. We are analysing
a scenario where we have one competitive agent and three cooperative agents. Note that all these agents
are behaving with their emotional actions disabled which means that they won’t act either on grudges or on
gratitude.
From the graph shown in Figure 11 we can see that agents collected resources in accordance with their
own personalities. We can see how cooperative agents focused on the collection of wood until the raft
construction was completed and than they started to collect gold for the remainder of the time they had.
Meanwhile, the competitive agent focused clearly on collecting gold for the whole game but since the agent
wasn’t Highly Competitive we can see that some small effort was spent by this agent in collecting wood for
the construction of the raft.
Figure 12: Total wood collection the team.
We will move on to the data related to wood collected by the team, as shown in Figure 12 and for this
analysis keep in mind that the INVITE Game was configured so that for the raft to be considered built, the
number of wood pieces collected by the team needs to be over or equal to 150.
The graph in Figure 12 displays the total number of pieces of wood gathered by the team of agents
throughout the scenario execution. We can see how the agents have been able to collect enough wood
to successfully build the raft. Seeing as agents were able to save themselves from the eruption we should
52
conclude the analysis of this scenario by checking the results for the nominations.
Figure 13: Nominations results for scenario execution.
The graph in Figure 13 shows how agents have conducted their nominations throughout the scenario’s
execution. To understand these nominations and what triggered them, recall the tolerance feature that was
mentioned in the nominations algorithm. Here follows a concrete example related to the scenario currently
under analysis:
Figure 14: Tolerance for gold collection.
We can see from Figure 14 that, from the first round of the game, the competitive agent broke the
tolerance for adverse behaviour from the cooperative agents, which triggered the nominations from all of
them. Also note that the accumulation of all the nominations resulted in a total of 18 nominations which,
according to nominations rules, means that the competitive agent at the end was kicked from the raft which
means that although the agent collected more gold it ended up not legible to win the game.
4.1.2
Scenario 2
In this next scenario we will begin to go into some examples of agents enabled with the ability to act emotionally. On this specific scenario we will see a competitive agent that holds grudges and how the other
cooperative agents, who do not hold grudges, will behave when they perceive this ability to act on held
grudges towards others.
From this second scenario going forward, since we’ve set a basis for analysis, we will refrain from commenting every piece of data extracted from the scenario and instead comment only where it seems relevant
and compare these results with previous scenarios and take some conclusions about the data that was
53
gathered. Table 4 shows the relevant configurations for Scenario 2:
Agent Name
Berna
Jamie
Pat
Simon
Personality Trait
Competitive
Cooperative
Cooperative
Cooperative
Holds Grudges
Yes
No
No
No
Is Grateful
No
No
No
No
Table 4: Agent configurations for scenario 2.
Figure 15: Resource collection for each agent.
Since the agents were configured with the same personality traits from the previous scenario, it was
predictable that regarding resource collection effort we would see similar results between both scenarios.
As seen in Figure 15, once again we see how the competitive agent has focused mainly on collecting gold
and how the cooperative agents started out on wood collection and than they collected gold for the remainder
of the days. Like in the previous scenario the team was able to collect enough wood pieces to build the raft.
With the agents’ resource collection efforts data analysed, we will move on to the nominations results
displayed in Table 5 and understand the impact of the competitive agent now being able to hold grudges and
act on these grudges towards other agents.
Looking at data that was extracted on the nominations activity for this scenario in Figure 16, we can
see differences from the previous scenario. To analyse this data, we will begin by examining the log for the
nominations issued by each agent displayed on Table 5.
Starting with Round 1 we can see how no activity took place as expected since the first round is meant
for agents to make their first round of resource collection and start developing emotional bonds towards
other agents. Regarding Round 2 we can see that, like on the previous scenario, the cooperative agents
start by issuing their nominations on the competitive agent, since this one broke their tolerance for adverse
behaviour.
54
Round Number
Round 1
Round 2
Round 3
Round 4
Round 5
Round 6
Round 7
Activity
No Activity
Jamie nominated Berna.
Pat nominated Berna.
Simon nominated Berna.
Jamie nominated Berna.
Pat nominated Berna.
Simon nominated Berna.
Berna nominated Jamie.
Berna nominated Jamie.
Berna nominated Jamie.
Berna nominated Jamie.
Berna nominated Jamie.
Table 5: Nominations for scenario 2.
Figure 16: Nominations results for scenario execution.
Following to Round 3 we can see the same nominations from the cooperative agents towards the competitive one and we can also see a new nomination issued by the competitive agent Berna of agent Jamie.
This nomination from Berna was issued from a grudge that it felt towards all cooperative agents who issued
nominations on the previous round. The cooperative agents see this nomination and can perceive, through
the algorithm detailed in section 3.2.1, that this nomination was issued from a grudge and since these
agents don’t act on grudges they set Berna as an agent who acts on grudges and adapt their behaviour
to avoid Berna’s grudgy behaviour. From this point forward, all following rounds are only composed of one
nomination resulting from Berna acting on his grudge towards Jamie.
In the end, if we analyse the final results for nominations activity as displayed in Figure 16, we can see
that only two agents were nominated for the duration of the scenario’s execution first agent nominated was
Berna with 6 votes and we can see how Jamie was also nominated with 5 votes, all of them resulting from
the grudge that Berna felt towards Jamie. Seeing how neither agent was nominated more than 10 times, not
only did the agents build the raft successfully, no agents were kicked from the raft.
We can conclude from this scenario that, because the competitive agent also was able to act on grudges
55
towards others, it was able to get away with individualistic behaviour only focusing on collecting gold for it’s
personal gain. Taking into account the previous scenario we can see how, because the competitive agent
did not act on it’s grudges towards other agents, it was not able to get away with it’s individualistic behaviour
and was kicked from the raft at the end of the game. We can conclude from this scenario that, with our
solution, the affective agent was able to avoid adverse behaviours from other agents, not be kicked from the
raft and in the end be able to win the game as the agent that collected the most pieces of gold.
4.1.3
Scenario 3
The next scenario, like the previous one, will continue the analysis into behaviours related with agents
displaying grudge behaviours towards each other through the nominations mechanic. This time we wanted
to analyse two competitive agents in the same team with two other cooperative agents, where one of the
competitive agents has it’s affective behaviour enabled and the other one does not.
The configurations for this scenario are shown in Table 6:
Agent Name
Berna
Jamie
Pat
Simon
Personality Trait
Highly Competitive
Competitive
Cooperative
Cooperative
Holds Grudges
Yes
No
No
No
Is Grateful
No
No
No
No
Table 6: Agent configurations for scenario 3.
Figure 17: Resource collection for each agent.
The data related with this scenario shown in Figure 17 tells us that, because the team has replaced
one of it’s cooperative members with a competitive one, the other cooperative agents spent a lot more resource collection effort on collecting wood pieces to compensate the lack of agents focused on building the
raft. Meanwhile, the two agents who were competitive displayed different degrees of competitiveness, which
means that although both agents are individualistic, it was expected that the cooperative agents would converge their nominations on the most competitive agent. Keep in mind how the setup configuration tells us
56
that the most competitive agent is also able to act emotionally when holding grudges.
Since the agents were able to collect enough wood to build the raft we’ll move on to analyse the agent’s
nominations activity as displayed in Table 7.
Round Number
Round 1
Round 2
Round 3
Round 4
Round 5
Round 6
Round 7
Activity
No Activity
Simon nominated Berna.
Pat nominated Berna.
Simon nominated Berna.
Pat nominated Berna.
Berna nominated Pat.
Simon nominated Jamie.
Pat nominated Jamie.
Berna nominated Pat.
Simon nominated Jamie.
Pat nominated Jamie.
Berna nominated Pat.
Simon nominated Jamie.
Pat nominated Jamie.
Berna nominated Pat.
Simon nominated Jamie.
Pat nominated Jamie.
Berna nominated Pat.
Table 7: Nominations for scenario 3.
Figure 18: Nominations results for scenario execution.
This scenario is interesting in the aspect that it shows, in Figure 18, more clearly how the affective
behaviour when enabled gives the agents an edge over agents who otherwise do not possess this ability
to act emotionally. According to Table 7 we can see how the cooperative agents are the first ones to
issue their nominations and converge on the agent that is displaying the most individualistic behaviour as
expected. However on the third round, since agent Berna is enabled to act on it’s emotions, it issues a
nomination as a grudge action towards Pat. On the forth round we can see a change in behaviour by the
57
two cooperative agents which in perceiving that Berna acts on it’s grudges refrain from nominating Berna
and start nominating Jamie instead, since Jamie is another competitive agent that broke the tolerance for
individualistic behaviour for both cooperative agents. On the fifth round forward, we can see how all agents
maintained their nominations and also how Jamie did not nominate anyone since Jamie does not act out of
grudges towards the agent that issued Jamie’s nominations. In the end of the game this translates into every
agent being able to save itself without anyone being voted out of the raft and also with the most individualistic
agent, who won the game, not even being the highest voted agent on the team.
As we said earlier this is a clear scenario that shows how an agent can act emotionally and benefit
from it’s actions while also being able to exploit it’s competition by focusing all nominations on the other
competitive agent which is not enabled to act on grudges towards other agents, hence not being able to
avoid adverse behaviour taken from the other cooperative agents that are issuing their nominations.
4.1.4
Scenario 4
The following scenario takes a step in a different direction from the previous ones in that it now enables cooperative agents to hold grudges along with the competitive agents as we showed on the previous scenarios
that were analysed.
Configurations for this next scenario are as shown in Table 8:
Agent Name
Berna
Jamie
Pat
Simon
Personality Trait
Highly Competitive
Competitive
Cooperative
Cooperative
Holds Grudges
Yes
No
Yes
Yes
Is Grateful
No
No
No
No
Table 8: Agent configurations for scenario 4.
Figure 19: Resource collection for each agent.
Once again, the agents that took part in this scenario were configured, as shown in Table 8, with the
same personality traits that were configured for scenario 3, which means that for each agent, resource
58
collection effort results are the same. We can see, from Figure 19, how the competitive agents focused on
the collection of gold pieces while, because of this, the cooperative agent spent that extra effort on getting
wood pieces for building the raft. We will move on to analyse the nominations events, as seen in Table 9,
and understand the impact of the cooperative agents being able to hold grudges and act on these grudges
towards other agents.
Round Number
Round 1
Round 2
Round 3
Round 4
Round 5
Round 6
Round 7
Activity
No Activity
Simon nominated Berna.
Pat nominated Berna.
Simon nominated Berna.
Pat nominated Berna.
Berna nominated Pat.
Simon nominated Berna.
Pat nominated Berna.
Berna nominated Pat.
Simon nominated Berna.
Pat nominated Berna.
Berna nominated Pat.
Simon nominated Berna.
Pat nominated Berna.
Berna nominated Pat.
Simon nominated Berna.
Pat nominated Berna.
Berna nominated Pat.
Table 9: Nominations for scenario 4.
Figure 20: Nominations results for scenario execution.
When comparing the graph displayed in Figure 20, for the total number of nominations, with the same
graph from the previous scenario in Figure 18, we can see clearly a major shift in nominations issued by
the agents. When on the previous scenario the most competitive agent was nominated but also was the
one with less number of accumulated votes now, on a scenario where the cooperative agents also are able
to act on grudges, we see that this is not what happens and in fact the most competitive agent not only is
59
nominated but is now the one with the most number of votes, in this case 12, which means that although the
agent has the most amount of gold pieces collected it was also kicked from the raft at the end of the game.
Analysing the log of nominations we can understand how the agents behaved differently from the previously scenario and how the resulting nominations displayed the discrepancy displayed between Figure 18
and Figure 20. As we’ve seen on previous scenarios, on the second round we start to see agents issuing
their nominations because of the tolerance mechanic that detects for agents who are behaving cooperatively,
others who are behaving too individualistically. After the second round we can see the same nominations
plus the nomination of Berna that results from a grudge towards Pat. From this point the differences start to
become clear. Looking at the nominations from Round 4 forward, we can see that because both cooperative
agents act on their grudges both agents did not concede after the grudge nomination from Berna like on the
previous scenario. This continues along the rest of the scenario where the cooperative agents keep voting
on Berna and Berna keeps issuing grudge votes on Pat, which ultimately leads to the accumulation of a total
of 12 votes for Berna as it is kicked from the raft at the end of the game.
This leads us to conclude that at first, holding grudges allowed individualistic agents to intimidate other
agents on previous scenarios and exploit them, by allowing the individualistic agents to get away with collecting many of pieces of gold without getting nominated enough to be kicked from the raft. However, we can
see from this scenario that holding grudges does not enable individualistic agents to exploit the behaviour for
nomination issuing of the cooperative ones when these are also able to hold grudges and therefore do not
become fazed by grudge nominations from other agents, meaning that they do not let go of their nominations
towards the individualistic agents, making them lose the edge they had over the cooperative agents.
4.1.5
Scenario 5
In the next scenario no cooperative agents will participate on it’s execution, all the team’s members have
competitive personality traits of different degrees.
In Table 10 follows the configurations for the first scenario for a team of competitive-only agents.
In this scenario we can clearly see from the results, in Figure 21, for the resource collection effort of the
agents that this is actually a ”tragedy of the commons” type of scenarios where no agent contributes for the
general good of the society they belong to, expecting that the others will in their stead. We can also see from
Figure 22, how these actions led the team to only be able to collect 85 pieces of wood which is insufficient
for building the raft and getting off the island at the moment the volcano’s eruption takes place. We will move
on the analysis of the nominations as displayed in Table 11.
Through this information for the nominations activity in Table 11, we can see how they all converged on
the most competitive agent. On round 1 Simon nominates Berna because it actually triggers the tolerance
for current adverse behaviour since Simon was configured as an agent that possesses the competitive trait
that collects less gold pieces.
60
Agent Name
Berna
Jamie
Pat
Simon
Personality Trait
Highly Competitive
Competitive
Competitive
Lowly Competitive
Holds Grudges
No
No
No
No
Is Grateful
No
No
No
No
Table 10: Agent configurations for scenario 5.
Figure 21: Resource collection for each agent.
Figure 22: Total wood collection the team.
After that, on the next rounds the other agents Jamie and Pat start nominating Berna but these nominations are triggered by another algorithm that was detailed in section 3.2.1 which is tolerance for continuous
adverse behaviour. Since Jamie and Pat are collecting more wood than Berna, this entitles them to nominate
Berna if it reaches a level of adverse behaviour that the agent considered enough to issue a nomination.
However, despite all these details on the behaviours analysed for this scenario, as shown in Figure 23,
the final standings for the nominations are actually irrelevant for the game’s ending since the team was not
able to build the raft and no agent can be selected as victorious.
61
Round Number
Round 1
Round 2
Round 3
Round 4
Round 5
Round 6
Round 7
Activity
No Activity
Simon nominated Berna.
Simon nominated Berna.
Pat nominated Berna.
Jamie nominated Berna.
Simon nominated Berna.
Pat nominated Berna.
Jamie nominated Berna.
Simon nominated Berna.
Pat nominated Berna.
Jamie nominated Berna.
Simon nominated Berna.
Pat nominated Berna.
Jamie nominated Berna.
Simon nominated Berna.
Pat nominated Berna.
Jamie nominated Berna.
Table 11: Nominations for scenario 5.
Figure 23: Nominations results for scenario execution.
4.1.6
Scenario 6
The following scenario shares the same configurations for personality traits of the agents of the previous
scenario but now all agents are enabled with the ability to act on their grudges as displayed in Table 12.
In terms of resource collection effort we can see from Figure 24 that, as expected, the graph has similar
data to the one from the previous scenario which means that once again the agent were victims of the
tragedy of the commons since all agents were focused on their own well being expecting that the other
would work on building the raft.
62
Agent Name
Berna
Jamie
Pat
Simon
Personality Trait
Highly Competitive
Competitive
Competitive
Lowly Competitive
Holds Grudges
Yes
Yes
Yes
Yes
Is Grateful
No
No
No
No
Table 12: Agent configurations for scenario 6.
Figure 24: Resource collection for each agent.
Figure 25: Nominations results for scenario execution.
As for the nominations data, as shown in Figure 25, the only difference is that instead of the most
competitive agent being the only one nominated by it’s peers now we can see some nominations activity on
another agent which means that, in this scenario, Simon was the target of Berna’s grudge nominations. Note
that despite these grudge nominations there were no benefits or advantages to this emotional behaviour
since all team members also hold grudges and behave in individualistic manner. This scenario only enforces
the point that the previous scenario made, which is that the advantages of behaving emotionally towards
others gives a circumstantial advantage, meaning that this is not a winning strategy and advantages can
exist or not for this emotional behaviour depending on how the environment, in which the agent exists,
adapts to the agent’s emotional actions and if that is either beneficial of harmful for the affective agent.
63
4.1.7
Scenario 7
The previous scenarios tested in detail several different configurations for the agent’s team, where those
configurations aimed at testing the mechanic for agents holding grudges towards others. From this scenario
forward we will move away from the grudge mechanic and configure team configurations that, at first, allow
the testing of the alliances mechanic which involves enabling agents to behave on gratitude towards others.
After these scenarios, for testing the alliances mechanic and understand how agents interact with each other
through gratefulness, we will go into some hybrid scenarios meant to test both alliances and nominations
mechanics in the same test groups and how both these mechanics for emotional behaviour can influence
one another when an agent is able to interact on both levels.
Keep in mind that the alliance mechanic maintains the approach detailed in section 3.2.2 where an alliance requires both agents of the alliance to participate in daily rounds of either cooperating and maintaining
the alliance or defecting from it as conforming to Game Theory specifications of the prisioner’s dilemma and
also to replicate the behaviours specified in the work of Pimentel [9]. However, since our agent’s are not
programmed to have a behaviour of betrayal because it is not in the scope of our work, we will only make
reference to the moment where two agents form an alliance and won’t mentions the next rounds since all
agents will always cooperate from the moment they belong to an alliance.
To start these scenarios we will begin by setting a basis for our analysis by configuring all agents to be
cooperative but are neither enabled to act on the grudges or on gratitude, as shown in Table 13.
Agent Name
Berna
Jamie
Pat
Simon
Personality Trait
Cooperative
Cooperative
Cooperative
Cooperative
Holds Grudges
No
No
No
No
Is Grateful
No
No
No
No
Table 13: Agent configurations for scenario 7.
Figure 26: Resource collection for each agent.
64
Figure 27: Resource collection for each agent.
From this scenario, as shown in Figure 26, resource collection from all agents is primarily focused on
wood collection for building the raft as soon as possible. After that all resource collection effort went into
collecting as much gold as possible. We can also see that, as expected, the agents were able to collect
enough wood to build their raft and escape the island. Looking carefully at the resource collection effort for
the agents, we can see that although the agents have the same personality and behaviours the agents seem
to have a discrepancy for gold collected. This is due to the fact that on the round where the construction of
the raft was concluded, while the first agents to arrive at the campsite used their collected wood to build the
raft to completion as shown in Figure 27, the later agents to arrive with wood were not able to do so because
the raft was already built and since the INVITE Game’s agents only keep track of effective resources that
were collected and actually used this explains the difference of gold collected that can be seen in Figure 26.
Figure 28: Gold collected by each agent.
For these scenarios that involve alliances and grateful behaviours we will show another piece of relevant
data specific to this mechanic that displays the interactions between agents in an effort to create alliances
with others. It was mentioned in section 3.2.2 where we explained how agents belonging to alliances can
benefit from extra pieces of gold. Here we will detail the amount of gold that each agent collected and how
much was collected due to alliances that they belong to. For this particular scenario this data isn’t relevant
65
since no alliances were made and all the gold the agents collected was only related to their own resource
collection effort. However, Figure 28 sets a basis towards which we will be able to compare values and
reach conclusions on how agents that are able to act gratefully towards others and belong to alliances will
be able to thrive comparing to agent that are not able to do that.
4.1.8
Scenario 8
The following scenario introduces the alliance mechanic by configuring all agents to be able to act gratefully
towards others. The configurations for this scenario are displayed in Table 14.
Agent Name
Berna
Jamie
Pat
Simon
Personality Trait
Cooperative
Cooperative
Cooperative
Cooperative
Holds Grudges
No
No
No
No
Is Grateful
Yes
Yes
Yes
Yes
Table 14: Agent configurations for scenario 8.
Figure 29: Resource collection for each agent.
Figure 30: Resource collection for each agent.
66
In this scenario we can instantly perceive how agents who were enabled to act gratefully towards other
agents and began to initiate the process of creating alliances, were able to collect more gold when comparing
with the agents who were not able to act gratefully, as shown in Figure 29. Note how, although the agents
collected more gold pieces, they still collected the same amount of wood and that led, as shown in Figure
30, to them being able to collect enough wood to build their raft and escape the island and win the game.
Still we can see how there is a difference in amount of gold collected for each agent and to understand
how this is happening we need to look carefully to all the data related with the alliances mechanic which we
will do in Table 15.
Round Number
Round 1
Round 2
Round 9
Round 4
Round 5
Round 6
Round 7
Activity
No Activity.
Jamie acted cooperatively towards Simon on day 2.
Simon acted cooperatively towards Jamie on day 2.
Pat acted cooperatively towards Jamie on day 2.
Berna acted cooperatively towards Jamie on day 2.
Jamie reciprocated Simon.
Simon reciprocated Jamie.
Jamie acted cooperatively towards Pat on day 3.
Simon acted cooperatively towards Pat on day 3.
Pat acted cooperatively towards Jamie on day 2.
Berna acted cooperatively towards Jamie on day 2.
Pat acted cooperatively towards Simon on day 3.
Berna acted cooperatively towards Simon on day 3.
Simon reciprocated Pat.
Pat reciprocated Jamie.
Jamie reciprocated Pat.
Simon acted cooperatively towards Berna on day 4.
Simon acted cooperatively towards Pat on day 3.
Jamie acted cooperatively towards Berna on day 4.
Berna acted cooperatively towards Jamie on day 2.
Pat acted cooperatively towards Berna on day 4.
Berna acted cooperatively towards Simon on day 3.
Berna acted cooperatively towards Pat on day 4.
Simon reciprocated Berna.
Berna reciprocated Simon.
Jamie reciprocated Berna.
Jamie acted cooperatively towards Berna on day 4.
Pat acted cooperatively towards Berna on day 4.
Berna acted cooperatively towards Pat on day 4.
Pat reciprocated Berna.
No Activity
No Activity
Table 15: Alliance cooperative actions and reciprocations for scenario 8.
Table 15 represents a log for the activity concerning alliances and how agents issue cooperative actions
to start those alliances across rounds. Note that in this log we only show the first iteration of creation of these
alliances, we refrain from showing interactions from these agents related with maintaining these alliances
67
since all agents won’t defect from the alliances they belong to, during the execution of this scenario. This
log gives a good impression on how this mechanic takes place and how agents interact with each other.
Remember how each cooperative action and subsequent reciprocation costs 3 pieces of gold, which means
that the agents that are issuing a cooperative action are taking a risk by relinquishing their own pieces of
gold with the purpose of forming alliances and possibly benefit from grater utility from continuous interactions
with each other.
This is a very straightforward scenario in the sense that all agents are enabled to act gratefully and as
mentioned in the algorithms for the alliance mechanics, agents with grateful behaviour enabled will take
the initiative and start probing others by giving them pieces of gold as a gesture for acting cooperatively
and signal an interest for creating an alliance. We can see this behaviour in Round 2 where all agents
act gratefully and start issuing their cooperative actions for alliances at the beginning of the day. Still on
Round 2 we can see that, because the agents are able to act gratefully towards each other, they reciprocat
the agents that issued the cooperative actions by returning the same amount of pieces of gold and start
collecting resources together and getting extra pieces of gold. This process goes on for the rest of the
rounds as alliances are formed between all the agents of the team.
Figure 31: Gold collection source for each agent.
Figure 32: Number of Alliances for each agent.
68
Regarding the data that resulted from the alliances mechanic, we can see how all agents were able
to benefit from behaving gratefully towards other agents in this particular scenario where the gold collected
from all agents is boosted by almost times 1,5 comparing to the gold collected by the agents that participated
on the previous scenario as depicted in Figure 31.
Finally, according to Figure 32 all agents were able to get the same number of alliances, however there
is a discrepancy between the gold collected from alliances for each agent. This is because of another
specification for the algorithm detailed on section 3.2.2 where agents only are able to issue one cooperative
action and reciprocation per round which means that some agents will be able to form alliances sooner then
other agents. We can see from the log of alliances activity how Jamie and Simon were the first to forge
alliances, while Berna and Pat were the last ones to get their own alliances, which leads to the difference of
gold collected from alliances that we mentioned earlier.
4.1.9
Scenario 9
On the next scenario we will start to go into configurations where agents enabled with grateful behaviour will
interact with others without this affective behaviour and that also are competitive.
The configurations are as seen in Table 16.
Agent Name
Berna
Jamie
Pat
Simon
Personality Trait
Cooperative
Cooperative
Cooperative
Competitive
Holds Grudges
No
No
No
No
Is Grateful
Yes
Yes
Yes
No
Table 16: Agent configurations for scenario 9.
Figure 33: Resource collection for each agent.
We can see from the resource collection data in Figure 33, that as in previous scenarios which had similar
agent configurations, the team was able to collect enough wood pieces to build the raft successfully and that
the competitive agent was actually able to collect the most amount of gold comparing to the other cooperative
69
agents. Note however how the cooperative agents despite not being able to surpass the competitive agent,
they still collected a greater amount of gold pieces while having the grateful behaviour enabled. To get a
better understanding of this gold collection we will look to data related with the alliances mechanic in Table
17.
Round Number
Round 1
Round 2
Round 3
Round 4
Round 5
Round 6
Round 7
Activity
No Activity.
Jamie acted cooperatively towards Pat on day 2.
Pat acted cooperatively towards Jamie on day 2.
Berna acted cooperatively towards Jamie on day 2.
Jamie reciprocated Pat.
Pat reciprocated Jamie.
Jamie acted cooperatively towards Berna on day 3.
Pat acted cooperatively towards Berna on day 3.
Berna acted cooperatively towards Jamie on day 2.
Berna acted cooperatively towards Pat on day 2.
Jamie reciprocated Berna.
Pat reciprocated Berna.
Berna reciprocated Jamie.
Pat acted cooperatively towards Berna on day 3.
No Activity
Pat acted cooperatively towards Simon on day 6.
Jamie acted cooperatively towards Berna on day 6.
Berna acted cooperatively towards Jamie on day 6.
No Activity
Table 17: Alliance cooperative actions and reciprocations for scenario 9.
We can see from the log displayed in Table 17, how all grateful agents started to issue cooperative
actions towards each other and how Simon, which is a competitive agent without grateful behaviour, did not
issue a cooperative action to form an alliance. Also note how the grateful agents did not act cooperatively
towards Simon because of it’s competitive nature which they perceive as self-centred behaviour. The final
alliance standings can be seen in Figure 35.
Despite Simon being a competitive agent, still we can see how the grateful agents issued cooperative
actions for alliances towards Simon on Round 6. This happened because the cooperative agents finished
building the raft and since they are no longer experiencing distress from building the raft on time, they tried
to cooperate with Simon. Note however how Simon did not respond to the alliances cooperative actions,
which means that the grateful agents gave gold pieces to Simon which in turn exploited this and did not
return those pieces because it was configured to not act gratefully towards other agents. This means, in
terms of resource collection for Simon, that although it did not get as much as it could, it did benefit from a
small boost in gold collection as depicted in Figure 34.
70
Figure 34: Gold collection source for each agent.
Figure 35: Number of Alliances for each agent.
From the data for the nominations mechanic, as shown in Table 18, we can see how the cooperative
agents issued their nominations towards Simon, as expected, because Simon broke their tolerance for
adverse behaviour. The interesting piece of information that can be taken from the nominations log is how
the cooperative agents stopped issuing their nominations from Round 6 forward which also supports the
alliances mechanic log where agents issued cooperative actions of gold donations for alliances to Simon
since, from Round 6 forward, the cooperative agents were no longer distressed from completing their goal
of building the raft which led them to stop issuing nominations.
To conclude the analysis of this scenario we can see from Figure 36, how the cooperative agents were
able to issue enough nominations towards Simon to kick Simon from the raft. Note how, because Simon
doesn’t act on grudges, it was not able to retaliate the cooperative agents with nominations towards them.
This means that at the end of the game Simon was kicked from the raft which makes Jamie the winner of
the game.
This particular scenario shows a situation where emotional behaviour for acting on grudges could have
enabled Simon to avoid harmful behaviour from other agents and improve it’s situation.
71
Round Number
Round 1
Round 2
Round 3
Round 4
Round 5
Round 6
Round 7
Activity
No Activity
Jamie nominated Simon.
Pat nominated Simon.
Berna nominated Simon.
Jamie nominated Simon.
Pat nominated Simon.
Berna nominated Simon.
Jamie nominated Simon.
Pat nominated Simon.
Berna nominated Simon.
Jamie nominated Simon.
Pat nominated Simon.
Berna nominated Simon.
No Activity.
No Activity.
Table 18: Nominations for scenario 9.
Figure 36: Nominations results for scenario execution.
4.1.10
Scenario 10
For this scenario the configurations will enable grateful behaviour for both competitive and cooperative
agents to see how the competitive ones behave towards other agents and understand to what point they
can exploit any benefit from this behaviour.
The configurations are as shown in Table 19.
Agent Name
Berna
Jamie
Pat
Simon
Personality Trait
Cooperative
Cooperative
Competitive
Highly Competitive
Holds Grudges
No
No
No
No
Is Grateful
Yes
Yes
Yes
No
Table 19: Agent configurations for scenario 10.
72
Figure 37: Resource collection for each agent.
The resource collection graph, as displayed in Figure 37, for this scenario has a rather interesting output,
in the sense that Pat is a less competitive agent than Simon but still Pat was able to collect an overall total o
collected pieces of gold greater than the amount collected by Simon. Seeing as this hasn’t been registered
in any other of the previous scenarios, being this the first time a competitive agent has it’s grateful behaviour
enabled, we should move on to try and understand how Pat was able to achieve this result by analysing the
alliances data and the nominations data.
Figure 38: Gold collection source for each agent.
Figure 39: Number of Alliances for each agent.
73
By looking at the data for the alliances mechanic in Figure 38, we can start to understand how the
scenario played out and why Pat was able to come out on top of Simon which was the most competitive
agent and the expected winner of the game. We can see from Figure 39 how Simon was the only agent
that was not able to establish an alliance with another agent and also Simon did not acquire any gold from
alliances throughout the whole game.
Meanwhile, according to Figure 39, the alliance count for Berna, Jamie and Pat is of three which translates to an alliance between the three agents, excluding Simon due to being the agent that displayed the
most competitive behaviour. We can spot on Figure 37 how Pat actually dedicated some resource collection effort to collecting wood. It wasn’t much but it was enough to distinguish itself from Simon which in turn
made the cooperative agents focus their nominations against Simon. Most interesting of all is the data found
in Figure 38 where we discern that not only was Pat able to collect it’s rewards from being an agent able to
behave gratefully, it was also the amount of gold collected from that reward that led to Pat collecting more
gold than Simon and ultimately being able to win the game since enough amount of wood was collected to
fully build the raft.
Figure 40: Nominations results for scenario execution.
Looking at the nominations data in Figure 40 we can see consistency between this and the alliances
data. Since Simon was nominated persistently by it’s team mates and Pat was able to start alliances, all
nominations fell on Simon which means that the agent was, in the end, kicked from the raft. Since Simon
was not a grateful agent, even if the agent were able to issue alliances, it would never be able to establish
one since the algorithm for alliances does not allow agents to act cooperatively or reciprocate regarding
alliances towards agents that they are currently issuing nominations towards.
Pat ”exploited” the alliances mechanic by acting less individually than Simon. Pat was able to concentrate
all it’s team mates nominations on Simon. Meanwhile, by behaving gratefully, Pat was able to build alliances
and surpass Simon’s amount of collected gold pieces and winning the game.
However note that this scenario might have turned out differently were Simon less competitive or even if
it was an agent that held grudges which would enable it to avoid the adverse behaviour from it’s team mates.
We will go into agents that act with gratitude and grudges towards others in the following scenarios.
74
4.1.11
Scenario 11
Before going into the final scenarios where agents are enabled with emotional behaviour both for holding
grudges and displaying gratitude we will analyse another scenario very much similar to the last in configuration with the exception that the most competitive agent is also enabled to act gratefully and understand if
this agent can leverage any advantage from this emotional behaviour.
In Table 20 follows the configuration for this scenario.
Agent Name
Berna
Jamie
Pat
Simon
Personality Trait
Cooperative
Cooperative
Competitive
Highly Competitive
Holds Grudges
No
No
No
No
Is Grateful
Yes
Yes
Yes
Yes
Table 20: Agent configurations for scenario 11.
Figure 41: Resource collection for each agent.
Starting with the Figure 41 we can see as expected how resource collection was very much similar to
the previous section with the exception that the difference of the amount of gold collected between Pat and
Simon is even further apart than in the previous scenario. This shows us how enabling Simon with emotional
behaviour not only did not improve Simon’s situation from the previous scenario, it also aggravated that
situation. To understand why, we will take a look at the data related with both alliances and nominations
mechanics that were extracted from this scenario.
According to Figure 42 it is clear how this time around Simon was a grateful agent contrary to the previous
scenario. This time around Simon’s gold collected from alliances not only isn’t null it is actually negative.
Being a grateful agent, Simon tried to cooperate with others to create alliances by giving away pieces
of gold to it’s team mates but never saw the return from that investment. Also, we can discern from Figure
43 how Simon ended up in fact without alliances and as in the previous scenario, the other three agents
created alliances among themselves.
75
Figure 42: Gold collection source for each agent.
Figure 43: Number of Alliances for each agent.
Once again this was due to the fact that being Simon a highly competitive agent, since the beginning of
the game it broke the other agent’s tolerance for individualistic behaviour and was nominated again as we
can see from the following nominations data in Figure 44.
The fact that the Simon was enabled act gratefully did not improve it’s dire situation, which again goes to
the point that emotional behaviour does not consistently improve the agent’s chances for a sound strategy.
Figure 44: Nominations results for scenario execution.
76
4.1.12
Scenario 12
Before moving on to analysing the user tests results, we will complete our agent-only scenarios analysis
with two scenarios with teams that have both grudge and gratitude behaviours enabled.
The configuration for this scenario are as displayed in Table 21.
Agent Name
Berna
Jamie
Pat
Simon
Personality Trait
Cooperative
Cooperative
Competitive
Highly Competitive
Holds Grudges
No
No
No
Yes
Is Grateful
Yes
Yes
Yes
Yes
Table 21: Agent configurations for scenario 12.
Figure 45: Resource collection for each agent.
Regarding the collection resource graph, as shown in Figure 45, seeing as personality wise the configurations are the same as in previous scenarios, the data distribution for collected resources is similar to the
one on those scenarios. Once again the team has been able to collect enough resources to be able to build
the raft successfully. Seeing as Simon is now enabled to behave gratefully, as well as able to act towards
it’s grudges we can already see how Simon this time around has collected the same amount of gold as Pat
which means Simon and Pat won the game instead of just Pat like in previous scenarios. To understand
better how this came to be we’ll start by looking at the alliances data.
Through careful analysis of the alliances mechanic log, in Table 22, we can see on Round 2 how all
agents began to issue their nominations towards each other including Simon who was also configured as a
grateful agent. However no agents reciprocated Simon’s cooperative actions to start alliances because this
agent displayed a highly individualistic behaviour. Also note in the nominations mechanic log in Table 23,
how on Round 2 Jamie and Berna issued nominations towards Simon which explains why Jamie was not
responsive to Simon’s cooperative action to initiate an alliance.
77
Round Number
Round 1
Round 2
Round 3
Round 4
Round 5
Round 6
Round 7
Activity
No Activity.
Jamie acted cooperatively towards Pat on day 2.
Simon acted cooperatively towards Jamie on day 2.
Pat acted cooperatively towards Jamie on day 2.
Berna acted cooperatively towards Jamie on day 2.
Jamie reciprocated Pat.
Pat reciprocated Jamie.
Jamie acted cooperatively towards Berna on day 3.
Pat acted cooperatively towards Berna on day 3.
Simon acted cooperatively towards Pat on day 3.
Berna acted cooperatively towards Jamie on day 2.
Berna acted cooperatively towards Pat on day 3
Jamie reciprocated Berna.
Pat reciprocated Berna.
Berna reciprocated Jamie.
Jamie acted cooperatively towards Simon on day 4.
Pat acted cooperatively towards Berna on day 3.
Pat acted cooperatively towards Simon on day 4.
Berna acted cooperatively towards Simon on day 4.
Simon acted cooperatively towards Berna on day 4.
Berna reciprocated Simon.
Simon reciprocated Pat.
Berna acted cooperatively towards Simon on day 4.
No Activity.
No Activity.
Table 22: Alliance cooperative actions and reciprocations for scenario 12.
Round Number
Round 1
Round 2
Round 3
Round 4
Round 5
Round 6
Round 7
Activity
No Activity
Jamie nominated Simon.
Berna nominated Simon.
Jamie nominated Simon.
Pat nominated Simon.
Berna nominated Simon.
Simon nominated Jamie.
Simon nominated Jamie.
Berna nominated Pat.
Simon nominated Jamie.
Simon nominated Jamie.
Simon nominated Jamie.
Table 23: Nominations for scenario 12.
On Round 3, in the alliances log, in Table 22, we see how the agents that began creating alliances continued that activity with Simon issuing another cooperative action that did not find any response. Meanwhile
on the nominations log, seeing as Simon is now an agent who is retaliates when confronted by nominations
due to it’s ability to act towards it’s grudges. Simon held grudges towards both Jamie and Berna but chose
78
to act on it’s grudges towards Jamie as seen in Round 3 of the nominations mechanic. Finally we find that
Pat also nominated Simon this time around due to it’s tolerance towards continued adverse behaviour form
other agents.
On Round 4 we see a turn of events, since it became clear for all other agents that Simon is an agent
that holds grudges and so they stopped nominating Simon as a gesture to avoid Simon’s retaliations. Also
we can see how in the alliances log for Round 4, since all agents stopped nominating Simon they figured
that they might as well try to benefit by cooperating with Simon, hence all the cooperative actions of gold
donation sent towards Simon on Round 4. Note that Simon did not need to respond to all agents due
to already having acted cooperatively towards them on previous rounds. Another important detail on the
alliances log is how Jamie tried to act cooperatively towards Simon to start an alliance despite being grudge
nominated by Simon. However Simon did not join in alliance with Jamie precisely because it was acting on
grudges that it was feeling towards Jamie. This is why on Figure 47 we see Simon and Jamie with only two
alliances while both Berna and Pat were able to create three.
One final note to an unexpected nomination from Berna towards Pat since we can see from previous
rounds that there was already an intention from Berna to join in an alliance with Pat. But seeing as Pat
failed to reciprocate on time, when Simon showed that it held grudges, Berna switched it’s nomination to
Pat. Seeing as Berna already cooperated with Pat and Pat tried to cooperate with Berna on Round 3. This
means that on Round 5 Berna took the decision to stop nominating Pat and join in alliance with Pat.
Regarding nominations, in Figure 48, we can see how Simon was able this time around to avoid further
nominations and was not kicked from the raft which means that in this instance, being able to show gratitude
and hold grudges benefited Simon since not only was it not kicked from the raft it was also able to collect the
same amount of gold pieces as Pat while being able to get away with it’s individualistic behaviour, as seen
in Figure 46. In this case it is also relevant to add how, according to FIgure 47, Pat associating itself with
Simon may have led to Simon not winning the game which means that in this case Pat may not have totally
benefited from acting emotionally towards Simon.
Figure 46: Gold collection source for each agent.
79
Figure 47: Number of Alliances for each agent.
Figure 48: Nominations for each agent.
4.1.13
Scenario 13
To conclude the analysis of our series of agent centred scenarios, we will go into a scenario similar to
the previous one, only differing in the configuration of the cooperative agent’s personality trait for holding
grudges. The purpose of this test is to analyse how the most competitive agent that won the previous
scenario by taking advantage of being both grateful and a holder of grudges fairs when the cooperative
agents of the team are also able to act emotionally towards their grudges.
In Table 24 follows the configuration for this scenario.
Agent Name
Berna
Jamie
Pat
Simon
Personality Trait
Cooperative
Cooperative
Competitive
Highly Competitive
Holds Grudges
Yes
Yes
No
Yes
Is Grateful
Yes
Yes
Yes
Yes
Table 24: Agent configurations for scenario 13.
80
Figure 49: Resource collection for each agent.
If we begin our analysis on Figure 49 we can already understand that by changing the cooperative agents
personality trait for holding grudges, the most competitive agent Simon no longer was able to position itself
equally to Pat as in the previous scenario. Actually, Simon’s situation in this last scenario achieved some
of the worst results on all tested scenarios despite being enabled with both grateful and grudgy behaviours.
On the other hand we can see Pat achieving the best results from all scenarios. To understand the reasons
behind these results we will look into the data extracted from both the alliances and nominations mechanics.
Figure 50: Gold collected from alliances for each agent.
From Figure 50 we can see how the agents fair on gold collected from alliances and how these results
tell us that Simon lost gold while trying to belong to alliances as all other agents were able to profit from the
alliances that they were able to create for themselves.
As expected, we can understand from Table 25 how, much like on previous scenarios, all agents began issuing their cooperative actions for alliances and their respective reciprocations on Round 2 since all
are enabled to act gratefully. Note how all agents were able to be reciprocated in their alliances with the
exception of Simon who was only able to create one alliance that lasted for two rounds with Pat.
81
Round Number
Round 1
Round 2
Round 3
Round 4
Round 5
Round 6
Round 7
Activity
No Activity.
Jamie acted cooperatively towards Pat on day 2.
Simon acted cooperatively towards Jamie on day 2.
Pat acted cooperatively towards Jamie on day 2.
Berna acted cooperatively towards Jamie on day 2.
Jamie reciprocated Pat.
Pat reciprocated Jamie.
Jamie acted cooperatively towards Berna on day 3.
Pat acted cooperatively towards Berna on day 3.
Simon acted cooperatively towards Pat on day 3.
Berna acted cooperatively towards Jamie on day 2.
Berna acted cooperatively towards Pat on day 3
Jamie reciprocated Berna.
Pat reciprocated Berna.
Berna reciprocated Jamie.
Simon acted cooperatively towards Berna on day 4.
Pat acted cooperatively towards Berna on day 3.
Pat acted cooperatively towards Simon on day 4.
Simon reciprocated Pat.
No Activity.
No Activity.
No Activity.
Table 25: Alliance cooperative actions and reciprocations for scenario 13.
Figure 51: Resource collection for each agent.
From Figure 51 it is also interesting to note how from all agents, Pat who was a competitive agent was
the only one able to create three alliances throughout the game. To understand the reason behind this we
must move on to the nominations mechanic data.
Analysing the nominations log displayed in Table 26, allows us to connect all the dots and understand
clearly the reasons behind all the events that took place in this scenario. Note how on the second round
the cooperative agents nominate Simon to be kicked from the raft and how on the third round we see both
cooperative agents nominating Simon with the added nomination from Pat this time around as well as a
nomination from Simon towards Jamie as a clear act of grudge towards that team mate.
82
Round Number
Round 1
Round 2
Round 3
Round 4
Round 5
Round 6
Round 7
Activity
No Activity
Jamie nominated Simon.
Berna nominated Simon.
Jamie nominated Simon.
Pat nominated Simon.
Berna nominated Simon.
Simon nominated Jamie.
Simon nominated Jamie.
Jamie nominated Simon.
Berna nominated Simon.
Simon nominated Jamie.
Jamie nominated Simon.
Berna nominated Simon.
Simon nominated Jamie.
Jamie nominated Simon.
Berna nominated Simon.
Simon nominated Jamie.
Jamie nominated Simon.
Berna nominated Simon.
Table 26: Nominations for scenario 13.
Figure 52: Nominations for each agent.
On the forth round note how Pat, not being able to hold grudges, ceases to vote on Simon while Jamie
and Berna continue to nominate Simon as it nominates Jamie on it’s continued grudge. This situation
progresses until the end of the scenario’s execution.
Returning our attentions to the alliances log in Table 25 we can see how Pat and Simon start efforts to
create an alliance from the moment that Simon displays it’s emotional behaviour for holding grudges which
dissuades Pat from nominating Simon and considers joining an alliance with Simon. Because the grudgy
behaviour from Simon did not faze Berna and Jamie, which continued nominating Simon, alliances between
them would be impossible as designed by the alliances algorithm. This explains why on Figure 51 we see
that Pat was the only agent able to join alliances with all it’s team mates and benefit from these alliances
which ultimately allowed Pat to win this scenario and also how, as shown in Figure 52, Simon was not able
83
to stop nominations form other agents and finished with the most accumulated amount of nominations.
4.1.14
Summary
Our final scenario of agent-only teams, enforces the point that these mechanics for emotional behaviour
give an advantage. However, that advantage seems to be extremely dependent on the team that each
agent belongs to and the way the agent’s team mates react towards it’s affective behaviour. Another clear
example of this circumstantial advantage was presented when Simon, on scenario 12, was able to leverage
on it’s grudges to avoid nominations from it’s team mates as well as use it’s predisposition to act gratefully.
Meanwhile Simon set alliances with it’s team members which in the end led to Simon winning the game.
However, as we’ve seen from the analysis of scenario 13 Simon, with the same configuration as in scenario
12, was not able to perform at the same level. Simon actually scored some of the worst results achieved in
alliances, nominations and total gold pieces collected, simply because we enabled the cooperative agents
to act on their grudges towards Simon.
These facts lead us to the conclusion that these mechanics of emotional behaviour are useful but their
use seems to have a circumstantial effect in the sense that affective behaviour is in no stretch a bullet-proof
strategy against loss of utility since we were not able to see this in the results that we extracted and analysed
from our agent-only test scenarios.
We see a clear advantage of affective-driven strategies when agents are interacting with a team whose
member’s personalities allow the affective agent to leverage on it’s emotional actions and influence others
to behave in a way that is most beneficial. However if the agent belongs to a team in which it’s personality
traits do not allow it to influence the other team mates, we have evidence that this advantage of affective
behaviour becomes a harmful handicap.
Concerning this conclusion, the work of Wilson [16] on Group Selection which was described in section
2.3.3 gives great insight by stating that in social scenarios where individuals belong to certain groups, these
individuals should be allowed to maximize the utility that they can achieve with their behaviour by enabling
them to act on multiple levels. This consists on individuals being able to interact between themselves in the
same social group as well as between groups, allowing agents to bond not only with their own groups but
also with other individuals outside of the group with the objective of being able to select which group they
wish to interact with in order to maximize their utility. Throughout most of the scenarios that we tested, the
problem that kept persisting was that when agents were performing badly because of how their personalities
were not able to influence their team mates, they would become doomed to failure because agents would be
limited to interact with the same team mates throughout the whole game. We believe that allowing agents to
interact on multiple levels that do not limit the agents to bond only towards their team mates but allow them
to also bond with agents from other teams, could allow them to find individuals that would possibly maximize
the utility that they can achieve by surrounding themselves with agents that they could influence to act in
their best interests through both their personality traits and affective behaviours.
84
4.2
User Tests Observations
The following sections will go over the user testing sessions of our solution and the corresponding analysis
of all data that we were able to gather. On this particular section we will refer to our user testing scenarios
and report all relevant observations that we took during those sessions. The sample of testers that had the
opportunity to interact with our solution consists of fifteen players and all of these players were requested,
during these sessions, to play the game two times and in the end give us feedback about the solution by
filling out a survey with questions related with evaluating our agent’s believability through quantifiable data.
We won’t go into too much detail in terms of dispensing exhaustive data on every scenario that was
tested as done for the agent-only scenarios. Since the point of these tests was to allow human players to
evaluate our agents believability as well as to understand the suitability of our agents emotional behaviours
towards human interactions in the context of the INVITE Game.
Our user testing sessions were not set up like the agent-only scenarios, where configurations were
pre-configured before each scenario’s execution. Instead, the teams that the players belonged to were
procedurally generated to guarantee that each scenario would have a different configuration for the agents
that would comprise each team. This made each scenario become dynamic and unpredictable for each
player, so they would not be influenced in any way by their teams’ configurations and guarantee an unbiased
assessment, avoiding configurations that could be more convenient to our evaluations. The point of this was
to allow the players to bond with their agent team mates and try to figure out both the agents’ personalities
and the emotions they are going through by analysing the way they behaved.
After this section we will go into the player feedback that we gathered during these sessions, which were
used in an effort to evaluate with quantifiable data how believable the players felt their interactions were with
the agents that ran our solution for affective behaviour.
4.2.1
Players vs Agents - Resource Collection Efforts
The following section will go into what we were able to observe from the players’ collection resource efforts
and how they made their decisions during both playthroughs that each player went through.
Both Playthroughs
When players started to play the game on their first playthrough, the first round saw most players
not taking risks, acting in the most neutral way possible by collecting a balanced amount of gold
and wood. However on the following rounds, once the players were able to start studying the logs
of the agents activities for resource collection on previous rounds, players began to see how agents
maintained their collection efforts the same through each round, depending on how individualistic
or how cooperative their were configured to be during the scenario’s procedural configuration. This
predictability on the agents resource collection efforts turned out to be an issue because this actually
harmed the believability of our agent’s behaviour from the players’ perspective, we will go into more
85
detail on this feedback on section 4.3.
All players, after figuring out this behaviour, used it to their advantage to try and keep collecting enough
gold to be ahead of the other agents but not too much to be the one being nominated, which led the
most individualistic agent of the team that actually collected the most gold, to be kicked from the raft
and the players were able to win the game. However, in situations when a highly individualistic agent
acted on grudges, this would make players have to take more risks and try to collect even more gold
which would make the player a target for nominations, this will be explained in the next section.
4.2.2
Player’s Nominations management
We will now mention some observations taken from the human players interactions with agents that ran our
solution for affective behaviour, more specifically, interactions related with the nominations mechanic.
First Playthrough
In this playthrough most players acted in a more proactive way. They took initiative and started issuing nominations without giving thought to how their agent team mates might react to those actions.
These interactions branched into two possible outcomes, the outcome where players were not harmed
by their actions because the scenarios they were in did not have emotional agents that would hold
grudges towards them; and the outcome where agents turned on the players and started grudge nominating them. The latter outcome was the most common and resulted in players not being able to avoid
these grudges and keep being retaliated until the end of the scenario they were playing. Some players
tried to mitigate this by collecting less gold then in their first rounds but this did not change the outcome
and resulted on the players performing badly in terms of accumulated nominations and collected gold
pieces. Some players even accumulated enough nominations to be kicked off the raft which made
them lose the game. As for the players who did not get retaliated by the other agents, these were able
to kick from the raft an agent that would behave in a more individualistic way form the other team mates
point of view. The players were also able to collect enough gold to win the game without becoming
targets of nominations themselves. It was also interesting how most of the players, when nominated
by another team mate for collecting too much gold, would try to avoid further nominations from their
agent team mates, not by nominating out of grudge towards them in an attempt to dissuade further
nominations, but instead the players would simply change their resource collection efforts to become
less individualistic, which consequently led to the to the other team mates ceasing further nominations
towards the players.
Some players that focused on winning the game were confronted with competitive agents that were
also acting on their grudges which made the task of winning that much more difficult and forced the
player to take more risks and try to collect greater amounts of gold and expose themselves to the
cross-hairs of their agent team mates for nomination. The way players avoided further nominations
86
was to fold and decrease their individualistic behaviour which allowed the agent that collected the most
gold to win the game.
Second Playthrough
During this playthrough we observed players changing the way they approached their agent team
mates. This consisted on players becoming much more passive in their interactions by starting the
rounds without making waves and just focusing on collecting their resources and analysing carefully
the logs with each round’s information. The players mainly focused on understanding how much gold
and wood each agent was collecting and also how the agents were interacting towards each other
in terms of nominations issued. This allowed players to have a clearer picture of their team mates
personalities and emotions towards each other. With this information they started nominating the
agents they felt were being more individualistic and also allowed them to avoid issuing nominations
towards agents that were displaying grudgy behaviours which could harm them if they became their
targets. Interestingly enough we also witnessed some players acting in a purely emotional way by
issuing nominations towards some of the more grudgy and individualistic team mates in the final
rounds of the scenario as a way to harm them while knowing the grudgy agents would not have
enough rounds to nominate the players in a meaningful way that could jeopardize their safety on the
raft when the game ended.
On these playthroughs the players that were focused on winning the game during the first scenario,
tried to use the mechanic for showing grudges instead of just decreasing the amount of gold they
collected. We were able to witness mainly positive results to this approach which allowed players
to get away with collecting more gold without risking being kicked from the raft at the end of the
game. However, we also saw a situation where a player tried to show grudges towards an agent that
nominated him but that also acted on it’s grudges which led the player to end up kicked from the raft
because the other agent did not yield.
4.2.3
Player’s Alliances management
The following observations, as for the nominations mechanic, were taken from the interactions between
human players’ and the agent team mates related to the alliances mechanic during the players’ both
playthroughs.
First Playthrough
This playthrough played out for the alliances mechanic, much in the same way we described for the
nominations mechanic, in the sense that players started out by issuing an cooperative actions by
giving gold pieces to team mates on the first opportunity they got. They issued these cooperative
actions to initiate alliances, taking into account the resource collection efforts of their team mates
which they accessed through the scenario logs. Invariably players always issued cooperative actions
87
towards agents that collected the least amount of gold pieces. On the other hand, our agents tried
to initiate alliances with any team member that were available unless that team mate had already
been nominated by that agent, meanwhile the players chose only to create alliances with agents that
only collected more wood pieces than gold. As the alliances mechanic progressed throughout the
scenario’s execution the players ended up receiving their gold back from agents that displayed grateful
behaviour and also noticed that some agents actually kept the gold they gave away. It was also
interesting to witness how some players decided to issue a nomination because of this, on a purely
emotional action of frustration towards the agent that did not reciprocate their cooperative actions.
We also perceived that some players received cooperative actions for alliances which made them
understand that they could ”exploit” this by not risking their own gold pieces and wait for the agents
to give their own gold pieces first. Players decided to not create alliances sooner then they possibly
could which cost them pieces of gold for each round that passed without them forming an alliance
while waiting to receive gold pieces from their agent team mates.
Second Playthrough
On this playthrough some players continued to refrain from initializing cooperation with their team
mates waiting for other agents to take that risk. Meanwhile, others players actually risked once to cooperate by donating their pieces of gold to a team mate for to start an alliance. It was interesting how
this time around, after some of testers became targets of cooperative actions to initiate alliances with
other agents with individualistic behaviour, they actually reciprocated this, which did not happen on
the previous playthrough since the players tried to avoid alliances with individualistic behaviours to impede those agents from further increasing their amount of gold collected. Eventually players received
cooperative actions from all the team mates that displayed grateful behaviours and it was interesting
to note how no players ever refused an alliance, always reciprocating and forming alliances whenever
the opportunity presented it self regardless of that agent’s resource collection efforts. However, on
this playthrough all players that belonged to alliances, chose to not cooperate on the final round of the
game towards their respective alliances. This gave them an edge by receiving gold from all their alliances and not reciprocating with that same amount of gold pieces. In the end of the sessions, testers
recalled this opportunity for exploiting the agents and that it would have been interesting to not feel that
these alliances were a sure things. Instead the agents that act gratefully could have been enabled to
either betray or maintain their alliances. In doing so, it would keep things more unpredictable, adding
another opportunity of emotional interaction with the agents.
4.2.4
Summary
As we mentioned in the nominations and alliances observations, one of the most interesting aspects of the
players interactions with our agents was how some of the players’ approaches towards the agents changed
between the first and second times they played the game. Mostly all players started out, on the first time
88
they played, by interacting more ”aggressively” towards the agents by both issuing alliances and nominating
agents very early in the game. From these interactions we observed some cases where this aggressive
approach paid-off, however most of the times players were confronted very early on with agents either
retaliating the nominations or by not answering to cooperative actions to initiate alliances.
The scenarios where players were not harmed by their rash actions on their first round resulted in them
not changing their tactics in their second playthrough which most of the times resulted in them not being as
lucky on the second round and finally being confronted by their actions through emotional behaviour by our
affective agents. This led to players being overwhelmed by the agents new behaviours for emotional action
which harmed the players and made them not able to perform as well in their second round. We observed
that most of the times players would become victims of other agents’ anger and receive many nominations
due to grudges culminating in the players ending up kicked from the raft. We also witnessed a few situations
where players would be presented with agents that would not act emotionally also in the second round
meaning that these situations were not that meaningful in terms of bonding interactions between the agents
and the players.
Another type of interactions that was also interesting to observe were situations where the players were
confronted early on with emotional agents and how these could act in meaningful ways towards the players
that could dictate the outcome of their end results. It was interesting how, in these situations, the players
would become much more passive in their actions during the second playthrough. Players would cease
to nominate or cooperate for alliances towards other team mates and instead remained passive while the
agent team mates started testing the waters among themselves. This allowed the players to bond passively
with the agents by acting first as observers and only after they had a more clear picture of their team mates
they would begin to act towards them. This allowed for some players to do better then in the first time they
played, even those who also had on their second playthrough more emotional agents to interact with.
Interactions from these player tests actually resembled most of the scenarios that were tested with our
agent-only scenarios as we witnessed many interactions between the players towards our agents that we
had already seen before during interactions between our own agents. With this data he had a first lead that
our agent solution to affective behaviour in the context of the INVITE Game concept, could represent a valid
emulation to believable human interaction from the evidence gathered on our observations of these user
testing sessions.
4.3
User Feedback
In this next section we will detail another important source of data for the analysis of one of our hypothesis
which is to validate with a human audience the believability of our affective agents’ behaviours. This user
feedback gathered during user tests was given by the same sample of fifteen users that participated in the
user testing sessions described in the previous section.
To gather this feedback from the testers, we resorted to the methodology for evaluating believability of
89
synthetic characters through quantifiable data as defined by the work of Gomes [10], in section 2.3.5. We
developed a survey that can be found in Appendix A with questions related with the users experience
and their interactions towards our affective agents. We also approached questions concerning how those
interactions translated to them in terms of believability in the way they bonded with the user and the way
they acted emotionally towards them. Throughout this section we will go into each question of the survey,
display the data gathered from each one and analyse what this data means to our work. More details on the
actual survey can be found on the annex section of this document.
To analyse the data gathered from the surveys given to our testers we will go through each question and
display both statistical data as well as mention any written feedback given by the testers on the pertinent
sections of our survey.
The first question from our survey was based on the methodology defined by the work of Gomes [10]
which set a basis of metrics for synthetic character believability that while it is aware of the highly subjective
data that it is dealing with, could still be able to output quantifiable data evaluates concretely our affective
agent’s believability. We did not evaluate all characteristics mentioned in that work, some like Social or
Visual Impact which we believe are not possible to evaluate in the scope of what we developed in our
solution. Note however that our survey has this question repeated three times, also recall that our testers
integrate teams with three other agents when playing the INVITE Game. The way these questions are
structured allows each tester to be able to evaluate each agent they interacted with, individually according
to their believability. It is also important to understand that when the users evaluated the each agent, they did
not know the affective settings of the agents they interacted with. We believed that this lack of knowledge
would work towards better validating our results, otherwise that knowledge would defeat the purpose of
getting unbiased data purely based on how the players perceived their interactions with the agents without
knowing their affective configurations.
We will now display, for each evaluated trait of believability, the results that we were able to extract from
our sample of testers.
Figure 53: Agent behaviour coherence evaluation.
Behaviour Coherent: According to Figure 53, users felt in a general way that agents were displaying
90
coherent behaviour in the sense that they were able, through interactions with the agents, to identify their
personality traits, allowing them to adapt correctly to the agents behaviours as well as begin to predict
outcomes of agents behaviours. However, some users felt that agents without emotional behaviour were
too apathetic towards what was happening which for them felt incoherent with the rest of the game. Note
how there is a small amount of Neutral feedback towards the Grateful agents which is due to some users
feeling that it didn’t make much sense that the agents that acted on Gratitude never tried to betray their
alliances and it was always certain that they would cooperate every day.
Figure 54: Agent change w/ experience evaluation.
Changes with experience: As seen in Figure 54, our testers were able to perceive how our agents
developed emotional attachments towards their team mates as the agents bonded with each other through
prolonged interaction between them. Users also appreciated how this changed the way they behaved and
acted on those emotions towards their team mates and also how agents reacted differently in different
scenarios with different team mates. When it comes to change with experience we can see users being able
to identify easily which agents were the emotional ones and which were the not and how the best evaluations
converge on the affective agents, while the worst evaluations belong to the agents without emotional setting.
However, there is some conflict on the evaluation of the agents without emotional setting since they are
capable of becoming intimidated when they are grudge nominated by a team mate.
Figure 55: Agent emotion perception evaluation.
91
Perceives emotional behaviour: In a general way, as shown in Figure 55, testers were satisfied by
the way our agents perceived the game world as well as their team mates actions. This allowed the testers
to detect emotional action from other agents and how every time an agent behaved emotionally, it would
change another agent’s behaviours and that would impact their perceptions and inherently change the game
every round. Once again we are able to see our testers strongly agreeing that our affective agents have
believable emotional perceptions, while mainly disagreeing towards agents without emotional setting.
Figure 56: Agent personality uniqueness evaluation.
Personality Uniqueness: From the data from Figure 56, we can conclude that testers were not generally in much agreement about the believability of our agents’ personality uniqueness. The main arguments
for this feedback was the fact that users felt like the personality traits that determined if an agent was more
competitive or more cooperative made the agents’ resource collection efforts too one-slated in the sense
that users were never surprised with the agents’ efforts since depending on their personality that would
determine the amount of resources they would collect from the beginning of the game until the end without
any unpredictability. Players felt that the way they chose how to act wasn’t consistent through the entirety
of the game, meaning that they were never just competitive or just cooperative through the games execution. Instead there were turns when the users were more cooperative and others where they were more
individualistic. Although there was some diversity of personality traits, the testers felt that the six settings
of personalities developed for our solution were still limited and harmed the agent’s believability. Common
opinion was that although the agents that acted on their emotions were the most interesting and also the
most challenging to play with or against, the agents who would not act on their emotions were also important
to maintain the overall believability of the scenarios they were in. The existence of agents with both affectivedriven and logic-driven behaviour would make for more varied types of agents, which made each intractable
agent more unique to the human players, in this case making for the most important aspect of believability of
the agents which is their uniqueness in terms of emotional behaviour. This meant that, for some players, the
fact that emotion was in the mix of possible behaviours was important for the believability of the agents, however the existence of these agents makes that much more relevant the existence of non-emotional agents
which use logic to make their decisions. This actually increased the game’s overall believability because the
92
existence of agents that rely only on logic to make decisions was actually conceivable to the testers.
Figure 57: Agent awareness evaluation.
Awareness: As displayed in Figure 57, users did not have any feedback to give for the awareness
characteristic of our agents since they felt that agents were correctly aware of the status of the world they
belonged to and they kept this status correctly up-to-date as this information impacted the way the agents
acted as the scenario played out.
Figure 58: Agent emotional expressiveness evaluation.
Emotionally expressive: The way our agents expressed emotion was the criteria of believability where
we got some of the best evaluations, according to Figure 58, regarding how agents displayed their grudges
through the nominations mechanic and how they expressed gratitude through the alliances mechanic. Users
also enjoyed how the fact that the agents expressed emotions through action not only influenced the other
team mates actions, but also impacted the way they perceived each agent and changed the way they also
acted towards each agent making the game more dynamic overall. Like on previous believability criteria
we see a focus on strongly agreeable feedback for the more affective agents, while the least favourable
feedback sided with the agents without affective setting. The users who answered in favour of Emotional
Agents argued that agents who acted on either their gratitude, anger or both were the most interesting and
most relatable. This was mainly because the agents with affective behaviour would, at first, be unpredictable
93
in their behaviour towards the testers which would make them go through rounds of getting to know each
agent and bond with them in order to understand how to best behave towards them. Testers that supported
the emotional agents as the most believable also made the case that these agents acted emotionally towards
them depending on how the testers would behave which made the consequences of these emotional actions
that much more valuable to the human players.
Figure 59: Agent unpredictability evaluation.
Unpredictable: In Figure 59, this characteristic much like the Personality Uniqueness were the targets
of most feedback from testers. These characteristics were criticised mainly for not giving enough behaviour
diversity which led the agents to act too predictably from the perspective of a human audience. For this
reason we can understand the reason behind the more negative results associated with this characteristic.
We concluded our feedback survey with the following two questions. Note that there are less answers
then the number of our sample but this is due to some of the feedback being the same and we just mentioned the answers that referred different aspects of the solution:
We asked the testers to list from our solution what they felt made our agent’s behaviours more
believable to them.
• ”I enjoyed how emotional agents were able to interpret our actions and react emotionally to them.”
• ”The fact that agents would act differently depending on their personality made me need to understand
them and get to know them, in order to define how to act towards them.”
• ”I enjoyed the risk associated with both making alliances and nominating agents in the sense that it
makes you actually think before you act towards each agent.”
• ”I like the way each agent felt like a different individual in the team.”
• ”The fact that it is not possible to predict what type of agents that are going to belong to our team,
made every scenario feel unique and forced an effort of getting to know each team mate.”
94
• ”Letting other agents interact towards each other was also useful to understand what to expect from
each one.”
We also asked to list what testers felt were improvement points to increase the believability of
the agent’s behaviours.
• ”Agents without emotional action did not have meaningful actions towards the player.”
• ”The agent’s personalities give some amount of variety to the agents you can interact with, however I
felt that the personality made the agents resource collection too predictable.”
• ”I believe that it would be nice to be able to switch teams if we find that our team collects too much
gold.”
• ”I felt that when an agent feels a grudge towards you, it was overkill that the agent kept nominating you
until the end of the game or until another agent nominated it.”
• ”I think that if I so chose, I should be able to request multiple alliances in the same round.”
• ”Agents should not be constricted to never being able to betray their alliances. It would make for a
more fun and tense interaction to never know how the agents will act on that front.”
95
5
Conclusions
It is accepted that humans are influenced by their emotions when it comes to the way they think, make
decisions, act, determine their beliefs, motivations and their intentions. These influences are considered,
most of the time, to be beneficial to the individuals that are able to use emotions on their decision making
processes. In this work we aimed at developing affective agents that would be able to bond emotionally
with others and act on those same emotions, enabling them to possibly replicate the same benefits that are
generally accepted to occur most of the time with humans.
Our work is very interdisciplinary, so we began by looking into the state of the art background of the
most prevalent theories on the fields that were most relevant to our work, fields like game theory, theories of emotion, personality theories and also theories of cooperation and competition among societies of
individuals.
After a careful analysis of the theoretical background, we moved on to investigating some of the related
work developed in the same fields that we investigated in the state of the art background. Here we introduced
some of the most important works that took part in the inspiration for architecture design for our solution’s
developments. Works like the one by Pimentel [9] which set the stage for what we wanted to accomplish and
also the work of Antunes [13] that consisted on the INVITE Game which allowed for the development and
testing of our affective agents solution in the context of the INVITE Game’s concept were both indispensable
for the the work we developed.
Following the research done on these fields we moved on to the solution development where we developed affective agents able to have emotional attachments towards others, act on those emotions and also
be able to detect emotional behaviour from other agents. This was all developed with a mindful approach of
the INVITE Game’s concept of the tropical island eruption escape where players, either human or synthetic,
were presented with two options, collecting wood for building a raft and survive or collecting gold for personal
gain and in the end possibly win the game. We had some issues when developing the emotional actions
mechanics, since the mechanics already available in INVITE Game did not meet the required specifications
set by the work of Pimentel [9] on the design of desired emotional interactions to be analysed. In other
words, the INVITE Game’s core mechanics did not allow for the development of emotional actions that were
able to target specific team members and enable displays of gratitude and grudges towards them. Because
of these limitations we focused also on the extension of the INVITE Game to have two new mechanics that
would enable team mates to act emotionally towards each other in a targeted way that would allow to either
reward or punish one team mate at a time.
To validate our solution we moved on to the data analysis of the results from two sources of tests to
the solution we developed for affective agents behaviour. The first source of data was to set configurations
for different teams of agents that were running our solution for affective behaviour as they interacted in
different INVITE Game scenarios. The second source data consisted on procedurally generating random
configurations of teams composed by our affective agents playing with human players, these play tests would
96
then translate to a survey with questions asking for feedback on how our solution’s behaviour believability
stands towards a human audience like our testers.
We believe we were able collect enough evidence from the data analysis towards validating our hypothesis. According to the data related to the agent-only scenarios we were able to confirm how the affective
agents have a clear advantage over the non-affective agents which confirms that emotional judgement can
be more beneficial against pure logical decision making. However we were also able to conclude that
when pinning affective agents against each other, these agents were not always able to benefit from these
emotional behaviours. Those benefits occurred in a circumstantial manner depending on the affective and
personality characteristics/configurations of the other team mates. This leads us to infer that emotionally
driven behaviour does not lead to guaranteed advantage, this is actually determined by the other agent’s
affective configurations and how those configurations react to that emotional behaviour and if they allow the
other agent to either sink or swim.
Regarding user tests, we were able to gather more evidence leading us to conclude that our developed
solution for emulating emotional behaviour is able to do it in a believable way that allows for emotional interaction between human players and agents running our solution for affective behaviour. We witnessed similar
interactions on scenarios of agents vs agents and agents vs human players. However, much like in scenarios
composed only of agents, once again any advantages from emotional behaviour proved to be circumstantial
and only took place depending on the type of player that was interacting with the agents, as well as the agent
configurations at a given scenario. We also witnessed scenarios where humans players were able to exploit
the emotional agents when their configurations were set in ways that were not compatible with the human
player and, consequently, did not allow the agents to influence the human player into behaving more beneficially towards them. The fact that these similarities took place for most of the scenarios gives us evidence
that our solution for affective agents is able to emulate emotional behaviours that occur in human individuals.
The conclusions we arrived at are actually coherent with the work of Deutsch and Coleman [6] on Conflict
Resolution Theory, more precisely on the chapter on Cooperation and Competition, which states that ”(...)
it can be concluded that the initiation of cooperative or competitive processes is not a consequence of the
individuals’ phenotype (traits perceived by others) but a consequence of their genotype (traits inherent to an
individual) of type of interdependence (positive or negative) and type of action.” which means that the way
agents interact with each other and how those interactions impact a particular agent or player in terms of
either harming them or giving them an advantage is highly dependent on the way each individual agent is
configured internally on their inherent characteristics, either of emotional behaviour or personality, and how
those configurations actually fit when agents start to interact and bond with each other.
Finally we analysed the agents believability feedback given by our testers, which allowed us to conclude
that some disapproval was noted towards the way the agent’s personality shaped their resource collection
efforts which made them look too predictable in the sense that agents were too consistent in their behaviour.
97
However in a general way our agents were classified as believable to a human audience in the way the
affective agents behaved in their actions that resulted from emotional responses. Actions like nominations
issuing and alliances building, according to the mainly positive results from the survey, were classified as
believable by our testers when comparing with results concerning the absence of these interactions.
5.1
Future Work
Throughout the Data Analysis section, a theme that came up frequently related to the conclusions taken
from most of the scenarios was that the agent’s successes or failures to benefit from advantages over other
agents and influencing them to act in a beneficial way towards them, was tightly related with the team
that surrounded these agents and how their own configurations for affective behaviour reacted to the other
team mates’ emotional actions. This fact raises an interesting question concerning how agents enabled
with affective behaviour could harness their ability to both act emotionally and recognise emotion on other
agents behaviour and do this to profile each agent on the society they belong to, keeping information on
their personalities, emotions that they are feeling and sentiments felt towards others.
In our work we have already developed a solution that allows for this type of profiling, however we believe
that an interesting next step would be to mitigate one of our solution’s limitations, which is the fact that an
agent which has a certain affective behaviour configuration that does not allow that agent to thrive among
it’s team mates, condemns the agent to failure since it has no choice but to keep interacting with the same
team mates until the end of the game.
An interesting future work would be based in the work of Wilson [16] on the theory of Group Selection,
and use this limitation from our work as a stepping stone in allowing the agents to not conform to one
group of agents to bond with through the full duration of the scenario they belong to. Instead allow them
to bond with all the agents within the scenario that belong to other social groups and with the information
gathered from those bonding interactions, enable the understanding of which agents would be the ideal ones
to interact with, in order to leverage on their particular set of affective behaviour configurations and allow
them to maximize their utility when interacting with these agents. This would enable the them to not become
doomed from day one to interact with agents that are not influenced by their emotional actions and allow
them to choose the agents they wish to interact with in order to make the most out of their time interacting
with other agents that could allow themselves to be influenced by emotional action to both motivate them to
act in a more beneficial way, as well as avoid those agent’s adverse behaviour.
Concerning the user test observations section 4.2, users complained that for grateful agents, they still
yearned for more emotional interaction by adding a betrayal mechanic to agents that belong to alliances.
This leads us to believe that a possible opportunity for improvement would be to develop a solution that
takes even further the OCC Model for more than only the two emotions of Gratitude and Anger which were
the ones developed for this particular work. It would be interesting to increase the complexity of this model
and see a system that could generate a wider range of emotions and actually apply this to a more generic
98
and complex scenario for agents to interact in. Ultimately, this more generic solution should reflect similar
conclusions to the ones we reached in our work which was to show how agents interacting on an emotional
level would eventually be able to make better decisions than agents who are acting only on a logical level,
while at the same time enforce this idea with the added feature that agents can profile the society they
belong to and choose which agents to interact in order to improve their amount of collected utility through
continuous interaction.
99
References
[1] Wooldridge M.: ”Multiagent Interactions”, in An Introduction to MultiAgent Systems, John Wiley & Sons
LTD (eds), pp 105-129, University of Liverpool, Liverpool, UK, 2009
[2] Carver J.M.: ”Emotional Memory Management: Positive control over your memory”, in Burn Survivors
Throughout the World Inc, available at http://www.burnsurvivorsttw.org/articles/memory.html,
2005
[3] De Raad B.: ”The Big Five Personality Factors”, in The psycholexical approach to personality, Hogrefe
& Huber Publishers (eds), Ashland, OH, US, 2000
[4] Steunebrink B.R., Dastani M. and Meyer J.C.: ”The OCC Model Revisited”, in Proc. of the 4th Workshop
on Emotion and Computing, Utrecht University, Utrecht, NLD, 2009
[5] Lim M.Y., Aylett R. and Jones C.M.: ”Emergent affective and personality model”, in Intelligent Virtual
Agents, Heidelberg, DEU: Springer Berlin Heidelberg, 2005
[6] Deutsch M., Coleman P.T. and Marcus E.: ”Cooperation and Competition”, in The Handbook of Conflict
Resolution Theory and Practice, John Wiley & Sons (eds), San Francisco, CA, 2011
[7] Yong C.H. and Miikkulainen R.: ”Cooperative Coevolution of Multi-Agent Systems”, University of Texas
Austin, TX, 2001
[8] Theraulaz G. and Bonabeau E.: ”A Brief History of Stigmergy”, in Artificial Life, pp 97-116, Institute of
Technology, MA: MIT Press Journals, 1999
[9] Pimentel C.: ”Grateful Agents and Agents that Hold a Grudge: The Role of Affective Behaviours in
Sustained Multi-Agent Interactions”, ICAART, Barcelona, 2013
[10] Gomes P., Paiva A. and Martinho C.: ”Metrics for Character Believability in Interactive Narrative”, Instituto Superior Técnico, Lisbon, PT: Springer International Publishing, 2013
[11] Oatley K.: ”The Sentiments and Beliefs of Distributed Cognition”, in Emotions and Beliefs - How Feelings Influence Thoughts, Frijda N.H., Manstead A.S.R., and Bem S. (eds), pp 78-107, Cambridge, UK:
Cambridge University Press, 2000
[12] Minsky M.: ”Emotion”, in The Society of Mind, Simon & Schuster Paperbacks (eds), pp 162-172, New
York, US, 1985
[13] Antunes B.V.: ”Believable Synthetic Characters with Social Identity”, MSc Thesis, Instituto Superior
Técnico, Lisbon, PT, available at http://gaips.inesc-id.pt:8081/invite/, 2012
[14] Pimentel C.: ”Emotional Reasoning in AI: Modelling Some of the Influences of Affects on Reasoning”,
PhD Thesis, Instituto Superior Técnico, Lisbon, PT, 2010
100
[15] Picard R. and Ahn H.: ”Affective-cognitive learning and decision making: A motivational reward framework for affective agents”, in Affective Computing and Intelligent Interaction, pp 866-873, Cambridge,
MA: MIT Media Lab, 2005
[16] Wilson D.: ”Multilevel selection theory and major evolutionary transitions implications for psychological
science”, in Current Directions in Psychological Science, Binghamton University, New York, US, 2008
[17] Brembs B.: ”Kin Selection”, in Preliminary draft for the Encyclopedia of Genetics, New York, US: Academic Press, available at http://brembs.net/hamilton/, 1996
[18] Clore G.L.: ”Appraisal theories: How cognition shapes affect into emotion”, in Handbook of Emotions
(3rd ed.), Lewis M., Jeannette M. and Barret L. (eds), pp. 628-642, New York, US: Guilford Press, 2010
[19] Ortony A., Clore G. and Collins A.: The Cognitive Structure of Emotions, Cambridge, UK: Cambridge
University Press, 1990
101
Appendix A
Emotional Agents in a serious game of Cooperation and Competition – Survey
102
Emotional Agents in a serious game of Cooperation and Competition – Survey
Please rate our agents according to the following characteristics of believability for
our agents’ behaviours. Please leave comments whenever it seems relevant.
Name of the Agent: _________________
Strongly
Disagree
Disagree
Neutral
Agree
Strongly
Agree
Comments
Strongly
Agree
Comments
Behaviour
Coherent
Changes
with
experience
Perceives
emotional
behaviour
Personality
uniqueness
Emotionally
Expressive
Predictable
*Please do not fill. Grateful behavior
Grudgy Behaviour:
Name of the Agent: _________________
Strongly
Disagree
Disagree
Neutral
Agree
Behaviour
Coherent
Changes
with
experience
Perceives
emotional
behavior
Personality
uniqueness
Emotionally
Expressive
Predictable
*Please do not fill. Grateful behavior
Grudgy Behaviour:
Emotional Agents in a serious game of Cooperation and Competition – Survey
Name of the Agent: _________________
Strongly
Disagree
Disagree
Neutral
Agree
Strongly
Agree
Comments
Behaviour
Coherent
Changes
with
experience
Perceives
emotional
behavior
Personality
uniqueness
Emotionally
Expressive
Predictable
*Please do not fill. Grateful behavior
Grudgy Behaviour:
List in a few words what you think are points of our solution that made our agent's
behaviours more believable.
List in a few words what you think are possible improvement points to our solution
in order to increase the believability of the agent's behaviours.