Semi-Autonomous Avatars in World of Minds
A Case Study of AI-based Game Design
ACE 2008
Yokohama, Japan
Thursday, Dec 4, D1, 9.00
Mirjam Palosaari Eladhari
Gotland University, Sweden
Michael Mateas
University of California, Santa Cruz,
USA
Introduction
Round Characters
Characters in MMOs
• Majority of characters in MMORPGs are player
characters/avatars.
• Most character interaction take place between
avatars.
• NPC’s in MMORPGs mainly give quests, sell
items or give information about the world.
Round and Flat
• A flat character consist of only a
few traits and usually have a single,
static aim, transparent to the
audience.
• A round character is complex and
capable of contradiction, surprise
and change. Round characters have
a multitude of contradictory traits.
(Forster, Aspects of the Novel, 1927)
Chatman on Round Characters
“Round characters […] possess a variety of
traits, some of them conflicting or even
contradictory [...] We remember them as real
people. They seem strangely familiar. Like
real-life friends and enemies it is hard to
describe what they are exactly like. “
(Chatman,Story and Discourse, 1978)
In MMORPGs, the player characters ARE
real
Round and Flat
• A flat character consist of only a
few traits and usually have a single,
static aim, transparent to the
audience.
• A round character is complex and
capable of contradiction, surprise
and change. Round characters have
a multitude of contradictory traits.
(Forster, Aspects of the Novel, 1927)
NPCs
Avatars
Where MMO’s fail
Despite the focus on player-to-player
interaction, current MMORPG
designs fail to provide explicit
support for helping players to
roleplay and fail to incorporate such
social interaction and roleplay into
the game mechanics themselves.
Our goal
• To develop an AI-based game mechanic that
brings round characters into MMORPGs to
support the player in roleplaying through their
avatar.
• Find ways to utilize the richness of the present
human intelligence in virtual game worlds and
via, the right triggers, enrich the game experience
in terms of characterdriven drama woven into the
virtual world through its rulesystem.
Outline and summary
One of the most exciting possibilities in game AI research is for AI to
open up new game design possibilities. New AI will suggest new
design possibilities while design will push back on AI requirements.
This paper provides:
• A concrete case study of AI-centric game design, specifically
describing how AI can be powerfully and effectively used in an
MMORPG design.
• The Mind Module, a technical framework for modeling personality
and emotion for both player and non-player characters.
• The World of Minds (WoM), an MMORPG that employs an AI-based
game mechanic designed around the MM.
• Lessons learned from a paper prototype playtest of the mechanic.
Mind Module
overview
Mind Module (MM) Overview
• a semi-autonomous agent architecture
• built to be used in a multiplayer environment as a part of the
player's avatar.
• The MM models the avatar's personality as a collection of traits
inspired by the FFM
• maintains dynamic emotion state as a function of interactions with
objects in the environment and trait values, and summarizes the
avatar's current emotional state as an inner and outer mood
• Spreading Activation Network (SAN)
Deconstruction and reconstruction
• In psychlogy and neuropsychology the
complexity of the human mind is made more
comprehensible by deconstruction into
subsystems.
• As builders of virtual humans we need to
reconstruct agents that can function in way
that is believable and useful to the users in the
contexts and systems we devise.
Node types
Weighted network of interconnected nodes
Two-dimensional affect plane
• Emotions: brief and focused (ie. directed at an intentional object)
disposition
• Sentiments: permanent and focused disposition
• Mood: brief and global disposition
• Personality: permanent and global disposition
Personality
• Plethora of personality theories;
psychoanalytic, behaviorist, cognitive and
biophysiological theories.
• Trait theory pioneered by Allport (1930ies)
where the most prominent personality
assessment is called the Five Factor Model.
• (also used by Guoliang, Klesen, Khirsagar, El
Jed, etc)
Five Factor Model
NEO PI-R (also called OCEAN)
Personality Trait Nodes
Defines how you ”are”.
Traits: OCEAN model (Big five)
Emotions in MM
Emotion
Weight to
Inner Mood
Amusement
+2
Interest Excitement
Enjoyment/Joy
Relief
Satisfaction
+1.5
+2
+2
+1.5
+1.5
+2
Surprise
+1.5
Confusion
-1.5
Distress/Anguish
Weight to
Outer Mood
-2
-1.5
Fear/Terror
-2
Anger
-2
Shame/Humiliation
Sadness
Guilt
-1.5
-2
-1.5
13 ”basic
emotions”,
common to
primates, as
described by
Ekman, Tomkins,
Nathansson
Emotion Nodes
13
Define how you
feel right now
Mood
In everyday sense:
An overall state or quality of feeling at a particular time
Changes slower than emotions
Temporary
Highly contextual
Lingers even if the context changes
Becomes part of the next context
Individual – ”same” type of event result in differents
moods for different people.
Moods
private,
inner
mood
extroverted side of the mood, how the character
emotionally is relating to the game world and to
other characters.
Inner and Outer Mood
• The inner mood is the private sense of harmony that can
be present even if the character is in an environment where
events lead to a parallel mood of annoyance.
• The nature of the outer mood is social, and as such tied to
emotions that are typically not only directed towards
another entity but also often expressed toward an entity,
such as anger or amusement.
• The two scales for the mood nodes open up the possibility
of more complex states of mind than a single binary axis of
moods that cancel each other out.
Mood
Define how you
feel today
Sentiments – Emotional attachements
• MM allows several sentiments ie, of different
emotions, to be attached toward an other entity,
thus creating a compound set of sentiment.
• emergent sentiments origin from interactions
with other entities in the world, thus creating
emotional memories.
• authored sentiment sets have certain pre-set
combinations. For example ``infatuation" is a
combination of interest/excitement/amusement
and joy toward another character.
Player might realize it is a good idea to
RUN!
Sentiment
An emotional disposition towards a specific object or class of objects.
Define how you
generally feel about
specific things
Node types
Weighted network of interconnected nodes
MM Summary
• The MM thus provides the player with
information about the avatars feelings toward
other entities in the world.
• Proximity to objects or characters affect the
emotions, and thus the mood of an avatar,
functioning as information the player can use
to form an agenda for game play.
World of Minds
Game Play
World of Minds Summary
• Prototype mutiplayer game world where the
personalities of the inhabitants are the base
for the game mechanics.
• When interacting with other characters, the
action potential depend upon the player
character's current mood and personality
Basic Game Play of WoM
Players need to defeat physical manifestations of
negative mental states by using
- Spells
- Affective actions
The spells available to the character depends on
personality
The affective actions available to the character
depends on the current mood.
Spells and mind energy
• Spells cost Mind Energy to use (”mana”)
• Attacks reduce Mind Resistanse (”health”)
• The amount of Mind Energy is tied to Outer
Mood
• The amount of Mind Resistanse is tied to Inner
Mood
Fluctuations of Energy and Resistanse
Affective actions and Mood
• players can affect each other's moods by using
affective actions (AAs).
• AAs are actively chosen by the players
•the AA ``Comfort" can be used
successfully on targets that have an active
emotion node of Sadness, but only if the
player's own avatar is not in the area of
Furious on the mood matrix.
•If the AA Comfort is used successfully the
values of the emotion nodes Sadness and
Anguish of the target are diminished,
which in turn affects the mood of the
character.
Mood manipulation
• Current mood affects:
– Regeneration of Energy and Mana
– The range of availiable current actions
Thus, by affecting the mood of a another character
the action potential of that character is changed.
Mood, Spells and Resistanse
How a sentiment
and/or a manifestaion
is born in WoM
Playtest
Playtest Research Questions
The playtest was designed to elicit design feedback from players on
three different questions:
1) During gameplay, are players able to "reverse engineer" the MM to
build a mental model of how to manipulate emotional state to
achieve gameplay goals?
(For any successful game design, players must be able to build a model of the
mechanics that allow them to successfully interact with the game.)
2) Understand the player's theory of how traits relate to MM state and
gameplay in order to gather new design ideas for how traits might
influence emotional state and gameplay.
3) How do the players perceive the relationship between the
emotional state maintained by the MM and the game mechanics
and interaction options that depend on that state?
Approach
Evaluation of the game design via a paper prototype
During the play tests, the test leader walks individual
players through a paper simulation of several
scenarios.
Players are asked to think aloud while playing the
game; additionally, the test leader stops the game at
several points and conducts interviews.
(User-Centered Design, where the user's experience is a main
driver for design, Rapid prototype and playtesting)
• Ten players individually went through five
game mastered scenarios where her avatar
had a mind, represented by the character
sheet.
For each playtest, the participant:
• Filled in a short (less than ten questions) survey on demographic
data and previous gaming experience.
• Took the IPIP-NEO Personality test and emailed the results to the
test leader.
• Filled in a short survey about their experience taking the
personality test and their opinions about the use of personality
traits for avatar creation.
• Participated in the playtest, which took between 1 and 1.5 hours.
Each playtest session consisted of playing five scenarios, and
answering questions in two interviews, one in the middle, and one
at the end of the playtest. Each session was videotaped.
• Filled in two more short surveys, one focused on sentiment objects,
and the other on general impressions of the experience.
Playtest
•
During the five scenarios, the player was guided through using the main categories of actions in the
game including :
–
–
–
affective actions
navigation in a landscape of sentiment objects
mind magic spells.
•
Using the character sheet the test leader updated the state of mind of the avatar and NPCs,
showing the player the effect of her actions in the game in terms of fluctuations in emotions,
mood, mind energy and mind resistance.
•
In order to best capture player's problem-solving processes within the game, and to best
understand potential areas for confusion, the players were given minimal explanations about how
and what to do.
•
At any point, players could access a ``help system" in order to ask any question.
•
The twenty-two interview questions focused on the player's understanding of the relationship
between values in the MM, effects of game actions, relationship between personality and
availability of actions such as spells, etc
Data analysis
• We used the video analysis tool Transana to
analyze the 15 hours of video of interviews
and play sessions.
• We developed a coding scheme for potentially
relevant phenomena and states of mind; this
provided us with the initial framework for
searching for patterns and regularities, as
advocated by Miles and Huberman
Scenario 1 - Sentiments
The avatar meets the character "Gate Keeper" (GK).
• Via a prewritten dialog script GK gives
information about the world the player has just
arrived to.
• The GK searches his bucket to give the player two
random sentiment objects.
(The player represented by the avatar Mastaya got sentiments of
anger toward mittens and amusement toward socks.)
• The GK asks the player to picture an unnerving
scenario where she can choose which of three
different objects would be most scary.
(Mastaya picked garden gnomes and got a fear sentiment towards them.)
Scenario 2 - Affective Actions
The avatar meets the character Teresa.
• Teresa says she is too sad to explain what affective actions (AAs)
are, and asks the avatar for a hug.
• A selection of affective actions is presented to the player.
• Teresa and the avatar exchange affective actions until threshold
values for emotions result in the generation of sentiment objects
between the characters.
(Mastaya chose to comfort Teresa instead of hugging her. Teresa's distress and sadness
decreased, and her mood improved. After a few exchanges of AAs, a threshold value for
Mataya's emotion Interest/Excitement was reached and the system generated a sentiment
for Mastaya of this emotion toward Teresa.)
Seven of the players
chose to ``hug'' Teresa,
while three of them
started the chain of
AA's in the scenario
with using ``Comfort".
The AA ``comfort"
would diminish the
emotions of sadness
and anguish in the
targeted character.
• More AA:s were
revealed as the scenario
continued
• Players were asked for
AA:s they would like in
such a scenario
• The first meeting with
Teresa resulted in
sentiment objects
between her and the
player’s character.
• Type of sentiment
depenent on the
exchange of affective
actions
Scenario 3 - Facing the Sentiments
• The player needs to guide the avatar through
an environment with sentiment objects in
order to successfully accomplish a quest.
• The state of mind of the avatar changes
according to which sentiments are
encountered in proximity of the avatar.
(Mastaya navigated the board successfully and spent some time on the
sock in order to gain amusement before moving on.)
Scenario 4 - Using Spells (and Affective Actions)
The player finds Teresa in a state of distress as she is
attacked by a manifestation of Confusion.
• The player finds a spell, Laser Pen of Clarity, which reduces
confusion and mental resistance in the target.
• The player is introduced to the concepts of mental energy
and resistance through seeing the mind values on Theresa,
the Colossus of Confusion and the players' own avatar.
• When the Colossus of Confusion is defeated, a new foe
enters the scene, the Sail of Sorrow.
• When this is defeated, Teresa explains that when an
emotion goes out of bounds a manifestation of that
emotion is created.
• The players has been doing another scenario and comes back to
Teresa.
• She needs help she is attacked by a manifestation of confusion.
• Here players act very in ways very different to each other for
manipulating the mood of themseleves and of Teresa
Casts spell:
Trumpet irrelevant
questions and
obscure answers
(++ Confusion)
Voulnerable to spell:
Laser Pen of Clarity
(-- Confusion)
In some of plays of this scenario Teresa, who’s
personality makes her prone to depression, got
so sad that a manifestation of sorrow spawned
Casts spell:
Wet Net of Tears
(++ Sadness)
Voulnerable to spell:
Accommodate sorrow
(-- Sadness)
Scenario 5 - Trait based spells
The Gate Keeper gives the avatar two spells that he claims
are based on the personality of the avatar.
Example:
• Mastaya earns the spell "Interest/Excitement Shower",
based on the fact that her highest factor except
Neuroticism is Openness.
• She also learns the "Soothing Hand", which lowers fear in
the target, based on the fact that the highest value of the
traits in the neuroticism factor is Anxiety.
• The Gate Keeper tells her that she will be particularly good
at defeating manifestations of fear, the Terror Trolls.
Lessons from Playtest
Effects of Personality
• All players built correct mental models of
at least one of the effects of personality
on the MM and the game mechanics.
• All players demonstrated enough
understanding of the system to be able to
act in the world in such a way as to reach
an emotional state they found desirable.
Examples:
•Two players thought
that personality
affected the amount of
mind energy and mind
resistance.
•Three players thought
that the effect of AAs
on themselves and
others depended on
personality.
• One player thought
there were personalityspecific effects on
mood values,
hypothesizing a
mapping between
traits, emotions and
mood.
Effects of Personality Factors
• Strong trend towards building models about
extraversion. Eight of the ten players theorized
Kita:
about the factor.
Eric the Red:
"Depending on how extroverted
you are, sadness and guilt would
probably move more or less as
you are affected by them."
"So, like, if you are an extrovert you
might get interested and excited
more easily, but you also might get
distressed or anguished more easily,
so each one kind of ... you know ...
had an effect on your emotions."
• Agreeableness discussed by six players
• Conscientiousness discussed only by one
Dante: "If you are not conscientious at all [...], other people pick up on that, how [...]
are they going to ask you to do anything for them?"
Extraversion and the design risks of
FFM
•
•
•
Exraversion is a commonly described personality factor.
The GOOD thing with it is that players have an existing
mental model of it. The BAD thing is that it is often a
normative model.
Players expressed worries about how their avatars would
perform in social situations if their level of extraversion is
low.
Ancient Greek
philosophy, Jung’s
typology, etc.
Also the factor most
commonly studied in
tests of applications
for synthetic humans
and conversational
agents
The design goal of WoM is for personality traits to be non normative: we want a
game design in which each possible combination of personality traits allows a player
to success fully progress in the game.
Current design of the Mind DO weight connections between trait and emotion nodes
such that it may be more difficult for an introverted avatar to perform certain social
actions. However, this is balanced by having some mind magic spells only be available
when the avatar is in specific mood ranges.
Personality-based Spells
In the final scenario, each player was given two spells by the
Gate Keeper, based on their personality traits.
7 players were very positive to this while 3 players were
hesitant.
These 3 are all avid
players of single
player role playing
games, where a
character’s
ability is based on
it’s class.
Solemni:
"Getting spells from personality
is a different method of choosing how
your avatar interacts with the world - like
a class. Not having direct control over
your class may be a difficult pill to
swallow."
Trait system / class system
Since personality-based capabilities (such as mind
magic spells) are not organized under classes, but
instead rely on relationships between the different
traits, the combination space of the possible actions
for an individual avatar is larger than in a traditional
class-based RPG.
Experienced players of class-based RPGs may need
extra support to become comfortable with this system
– a clarification of what can be expected can be
necessary in future playtest, despite our principle of
minimal information.
Emergent Game Play
• The scenarios in the prototype are
purposefully tightly scripted, since a primary
purpose of the prototype is to explore the
player's understanding of the MM in the
context of WoM.
• Desprite this players leveraged the MM-based
game mechanics to discover alternative
strategies for completing the scenarios that
had not been foreseen by the designers.
Examples of emergent game play
In the fourth scenario, the players is asked to help Teresa battle the Colossus of Confusion (CoC). The
only other object in the environment is the Laser Pen of Clarity (LPC), which the player can pick up
to learn the spell of the same name. This spell reduces confusion and mental resistance.
As designers, we had assumed that players would use this spell on the CoC to help Teresa.
•
However, 70% of the players combined the use of AAs on Teresa with the use of their (only) spell on
the CoC. Players hoped to improve Teresa's emotional state through the AAs, and thus increase her
effectiveness at battling the CoC.
•
30% of the players used their LPC on Teresa, which decreases her confusion and again makes her
more efficient against the CoC.
•
20% percent of the players used AAs instead of the LPC on the CoC, damaging the CoC with a lesser
decrease in their mind energy than using the LPC.
•
20% ran off to the gate keeper and just so happened to learn another spell that they could go back
to the CoC to use.
• Encouraging for this case of AI based game
design:
• The fact that players discovered interesting,
alternative strategies even in very constrained
and simple scenarios validates the potential
for rich and emergent gameplay in MM-based
game mechanics.
Ending summary,
we have described:
• The Mind Module, a semi-autonomous agent architecture.
• An experimental MMORPG, World of Minds, in which the game
mechanics build upon the Mind Module's model of personality and
emotion.
• In a case study of AI-based game design, we have shared lessons
learned from a test of a paper prototype.
• The players were able to form and communicate mental models of
the mind module and game mechanics, validating the design and
giv ing valuable feedback for the future development of the project.
• Despite the constrained scenarios presented to test players, they
discovered interesting, alternative strategies, indicating that the
\mental physics" of the Mind Module may open up new game
design possibilities
Thank You for listening!
Questions?
[email protected], [email protected]