DARE Extended Architecture applied to a Multi-Agent World

1
DARE Extended Architecture applied to a
Multi-Agent World

• Márcia Maçãs

2
DARE Architecture

• Somatic Marker Theory (António Damásio)
• Decision as result of scenario anticipation based
  on past experience associated with body states
  that qualify them.

• Double Sensorial Processing (Joseph Ledoux)
• Fast, simple and reactive processing
• Slow, complex and informed processing

• Goal of DARE Extension
• Apply the architecture to a dynamic multi-agent
  world
• Obtain adequate individual and social behaviours

Desenvolvimento de Agentes Robóticos Emocionais.
3
New Concepts

• Sympathy
• Human ability to recognise others emotions,
  acquired by self-awareness of their own emotions
  Goleman, 1996.
• Evaluate others and predict their reactions based
  on a mental model which relates physical
  expressions with feelings and intentions,
  acquired by own experience or by observation.
• Non Verbal Communication
• Emotional Expression
• Implicit
• Involuntary
• Informative
• Verbal Communication

4
Extended DARE Architecture
5
Multi-Agent World - Market
Product Examples
Agents Expressions
Interface
6
Body or Internal State

• Agents Body or Internal State

Internal State at instant t
Ideal Internal State
Unbalance at instant t
7
Perceptual Layer
Perceptual Analysis

• Best Dv - Incentive Stimulus
• Reactive Selection given
• Ip, DVp e IS

8
Perceptual Layer

• Action Effects

All data is stored in memory, including
The action effects on Internal State are used to
adapt (temporarily or not) the meaning functions
from which results the DVp.
9
Perceptual Layer - Results
Mean of Maximum Unbalance
Number of Internal State Changes
20 Simulations, 500 cycles of 5 seconds
10
Cognitive Layer
Cognitive Analysis
11
Results
Mean of Maximum Unbalance
Number of Internal State Changes
20 Simulations, 500 cycles of 5 seconds
12
Symbolic Layer
Symbolic Analysis
Feature
Symbolic
Action
Actions
Is
Action
Body
Extraction
Evaluation
Selection
13
Symbolic Layer

• Ontology

• Maps features and symbols
• Descriptive symbols (set of features)
• Identifier symbols (name of a set of descriptive
  symbols)
• Syntactic knowledge about messages.

• Expanded Internal State
• Elements not directly related to survival

• Internal Representation of the Agent itself
• Necessary for sympathy and non-verbal
  communication
• There is an image of itself on each layer

14
Symbolic Layer

• Symbolic Evaluation and Action Selection
• Specific Actions in this Layer
• Communication.
• Memory includes
• Symbolic information
• Expanded Internal State
• Sequences include
• Own experience
• Other agents observed experience
• Dialogs that end with expression change
• Action is executed if belongs to sequence with an
  ending expression similar to the ideal one

15
Symbolic Layer - Results
16
Symbolic Layer - Results
17
Symbolic Layer - Results

• Order
• Long term goals and cooperation
• Ask for product
• If supplier has good recall of the product
  sympathises and starts negotiation.
• Supplier assumes a fictitious internal state
  identical to the past and searches the product.

18
Conclusion and Future Work
In the application of DARE architecture to a
multi-agent world

• The adaptation on the perceptual layer allows a
  best performance of the overall architecture
• The cognitive layer allows flexible behaviour and
  more accurate learning
• The symbolic layer allows communication and the
  emergence of cooperation/concurrence behaviours.

• Future Work
• Long term anticipation of the internal state.
• Inference mechanisms triggered by internal
  stimuli.
• Learning of Expression-Internal state relation.
• Improvements on memory managing