From Active Objects to Autonomous Agents
Zahia Guessoum & Jean-Pierre Briot
In quest of the missing link between active
objects and autonomous agents ...
OASIS
( Objets et Agents pour Systèmes d'Information et de Simulation)
LIP6
(Laboratoire d'Informatique de Paris 6)
e-mail: {guessoum, briot}@ poleia.lip6.fr
http:\\www-poleia.lip6.fr\{~guessoum, ~briot}
2
Active-Object Models
Plan
Example of Active-Objects Model : Actalk (Briot 94)
recept ion
(receiveMessage:)
1 Active Objects
message
(a Message/Invocation)
2 A Generic Architecture of Agent
selec t ion
(nextMessage)
m ailBo x
(a MailBox)
process
3 From Actalk to DIMA
actual compu tat ion
(performMessage:)
ac t iv it y
(an Activity)
4 Experiments
behavior
(an ActiveObject)
5 Discussion and Future Work
address
(an Address)
In spite of their communicating subjects appearance, active objects do not
reason about their behavior, on their relations and their interactions with
other objects... (Ferber 89).
3
4
Some Extensions of Active Objects
Example : Intensive Care Patient Monitoring
VentilationObserver
➫ T. Maruichi
• a message interpreter
• the notion of environment
Classifier
TherapyPlanner
Persistent Apnea
➫T. Bouron (MAGES IV)
• speech actes
Frequency x 2
Clinician
CMV
(Controlled Mechanical
Ventilation)
SignalProcessorr
➫Y. Shoham (Agent-Oriented Programming)
• mental states to have an interaction-based behavior
Ventilator
➫S. Giroux (ReActalk)
• reflexive and adpative active objects
TherapeuticActor
Warning !
apnea alarm
Patient
5
6
A Generic Agent Architecture
Example : Intensive Care Patient Monitoring
Different behaviors
Various kinds of behaviors (reactive or deliberative)
➫ Different asynchronous and concurrent modules
➬ Each agent is an autonomous and active entity
Agent
➬ Each agent may have various kinds of behaviors
(communicates with other agents, scans its
environment, makes decision, ...).
➫ Each agent may have simple or complex reasoning
capacities
➬ Each agent owns its control knowledge, the control
is not achieved through separated architecture.
behavior 1
behavior 2
Module 1
Module 2
behavior 3
behavior n
Module 3
Module n
Examples :
Perception, communication, reasoning, action, learning...
7
8
A Generic Agent Architecture
A Generic Agent Architecture
A Meta-Behavior
➫ Supervision Module
Adaptive Control ==> Autonomy
A Meta-Behavior
Self-Control Mechanism
Supervision Module
ATN
Supervision Module
ATN (states, transitions)
behavior 1
behavior 2
Module 1
Module 2
Adaptative
Control
behavior 3
behavior n
Module 3
Module n
Intelligent
Control
Control
Objects
MetaRules
Rules
Reactive Module
Reactive Module
fire a rule
Process
Control
Deliberative Module
9
10
Implementation (Actalk --> DIMA)
Behavior Hierarchy
An Active Object
body
[true]whileTrue:
[self acceptNextMessage]
ReactiveBehavior
createProcess
^[self body] newProcess
Perception
An Agent Structure
anAgentActivity
createProcess
^[self body] newProcess
DeliberativeBehavior
(aKB)
aMetaBehavior
(AtnInterpreter)
Process
body
self atnInterpreter
SimpleCommunication
(anATN)
SpeechActesCommunication
Behavior 1
Behavior 2
...
Reasoning
Behavior n
RealTimeReasoning
11
12
The Developed Environment
Meta-Behavior Hierarchy
DIMA
ActiveObject
Smalltalk
BasicAgent
(name)
Reactive Agents
NéOpus
ReasoningAgent
(aReasoningModule, atn)
Actalk
Discrete
Event
Simulation
Agents
Deliberative Agents
CommunicatingReasoningAgent
(aCommunicationModule)
PerceivingReasoningAgent
(aPerceptionModule)
RPCTalk
CommunicatingPerceivingReasoningAgent
(aPerceptionModule)
Other machines
Hybrid Agents
13
14
Use of DIMA
Use of DIMA
➬ Several real-life applications
➬ to study multi-agent problems : real-time (anytime reasoning)
• NéoGanDi : Intensive Care Patient Monitoring (INSERM, Paris)
• Meveco : Economic Agents Evolution (HEC, Banque de France)
• ATM Congestion Simulation
• ...
Multi-Agent Systems
NéoGanDi
Meveco
Number of Agents
constant
variable
Nature of Agents
cooperative
competitive
Context of Agents
agents activity
relies on context
context relies on
agents activity
Temporal Constraints
strong
weak
Social Reasoning
absent
Model-based
Schematic structure
package 1
Execution cycle of the real-time metabase
select
the fist package
contexte
package 2
Solution 1
Solution 2
.
package j
.
Solution j
evaluate
the remaining time
package n
Solution n
15
execute the rules
of the current package
sufficient quality
or insufficient time
to add the next rule
base
add
the next package
16
Conclusion & Future Work
Characteristics of the proposed multi-agent platform :
• A generic and modular agent architecture
• Several paradigms (object, production rules, ATN, ...)
• Various kinds of agents (reactive, deliberative, hybrid)
• Control mechanisms at two levels
• At the agent level : a self-control mechanism
• At the multi-agent level : a coordination mechanism
• Several real-life applications
Future work :
• Real-time multi-agent systems
• Learning behavior
17