Multiagent models1

1
Multi-agent models1

• Introduction
• Agents What are they?
• Agent Architecture
• Examples of multi-agent modeling
• Building multi-agent simulations

1. Simulation for the Social Scientist, Nigel
Gilbert and Klaus Klaus G. Troitzsch, Open
University Press1999
2
IntroductionMulti-agent models

• Agents are designed to intelligently interact
  with their environment
• Distributed Artificial Intelligence (DAI) a
  network of interacting agents
• Relevance to social simulation
• Geographic possibilities?

3
Agents

• Related to free will and concept of goal
  achievement you need a good agent to get
• Computer agents have the following properties
• Autonomy operate without direct control of
  actions and state
• Social ability agents interact with other agents
  using some type of mechanism
• Reactivity agents perceive and respond to
  environment
• Pro-activity they take initiative
• Intentionality simulation of anthropomorphic
  characteristics (I.e belief, desire, motives)

4
Attributes to model in agents

• Knowledge and belief
• Action based on what is know about environment,
  accurate vs. erroneous knowledge.
• Inference
• Give set of beliefs agents infer additional
  information
• If find food at location a other food is nearby
  a
• Social models
• Learn about interrelationships between other
  agents in their world
• Create picture of social world (model)
• Learn about geography of environment
• Knowledge representation
• Use predicate logic to store declarative
  statements If agent is eating at location x
  then there is food at location x
• Goals
• Design sub-goals relevant to situation at hand
  and try to achieve them
• Language
• Multi-agent models include some interaction
  between each other and their environment
• Interaction may involve passing of information

5
Agent Architecture

• Physical-symbol system hypothesis AI approach to
  building agents with cognitive abilities
• System which manipulates symbols according to
  symbolically coed sets of instruction is capable
  of generating intelligent action.
• Problems
• fragility system situation specific
• Complexity dealing with it
• Commonsense knowledge easy for humans to do it
• Solutions
• Production systems
• Object-orientation
• Language parsing and generation
• Machine learning techniques

6
Agent Architecture (cont)

• Production systems
• Rules systems
• Set of rules
• Condition when rule fired
• Action what happens
• Working memory saves state of agent
• Rule interpreter
• Check whether conditions for rule are met
• Determine which rules fire if their conditions
  are met
• Object orientation
• Data stored in slots within object
• Agents instantiated from classes
• Rules may be same for agents but memory content
  differs
• Same parameters but different values
• Encapsulation each agent a distinct entity with
  characteristics and behavior in one package
• Modeling environment
• Multi-agent simulation agents located in
  environment
• Agents need sensors to perceive local
  neighborhood and means to affect environment
• Order of agent firing

7
Examples of multi-agent modeling

• Sugarscape
• Modified CA
• MANTA
• True agent based
• Evolution of Organized Society
• The final frontier

8
Sugarscape

• Multi-agent model simple yet results in emergence
  of social network, trade markets, cultural
  difference
• Model
• Agents can look in castle directions
• See distances based on genetic endowment
• Have differing metabolic rates
• survival of the fittest
• Run
• Results in strongly skewed wealth distribution
  with well endowed agents accumulate more wealth
• Characteristic of model goal-oriented and rule
  driven
• Modification add spice
• Simulate inter-agent trade
• Metabolic rate for sugar and spice different
• Agents can barter sugar for spice and vice versa
• Agents need welfare function to compare needs
  of sugar versus spice, negotiating a price and
  determining quantity of each commodity to
  exchange
• Resulting model adhered to neo-classical economic
  theory with price reaching an equilibrium

9
MANTA

• Modeled after the Ectatomma ruidum ant which
  begins with a lone queen who then gives birth to
  workers.
• Agents move around simulated ant nest environment
• Picking up objects, eating and looking after
  brood
• Task triggered by stimuli from environment or
  from ant goals
• Ants dont interact directly but by dropping
  stimuli
• One activity at a time
• Trigger dependent on weight, threshold and
  activity level
• Weight indicates relative importance
• Threshold decreases as time advance without the
  task being done
• Repeatedly doing a task increases its weight
• Ants switches task when threshold raised to a
  given level or the weight of a new task is
  greater (eating and attacked?)
• Environment also is represented by agents
• Food agents, humidity agents, light agents, and
  dead ant agents
• MANTA is goal driven with a repertoire of
  possible actions depending on the type of agent
  (larvae, cocoons and worker ants)
• It doesnt attempt to model cognitive or even
  symbolic processing

10
Evolution of Organized Society

• Simulate human population in the upper
  Palaeolithic period
• Focus on formation of relationships between
  agents either hierarchical (leader-follow) or
  alliances (group of agents with reciprocal
  relationships)
• Intension was to simulate environmental
  conditions an investigate consequence for
  patterns of interaction between agents
• Simulation
• Landscape with population of mobile agents and
  scattered resources
• Agents could act alone in obtaining resources or
  act collectively
• Model runs hierarchical relationships once
  established were persistent but not optimal for
  survival
• EOS experiment illustrative of a multi-agent
  simulation which includes simplified models of
  human cognition they perceive their environment
  and other agents, formulate beliefs of the world,
  plan, decide course of action and reap the
  results.

11
Swarm

• Software platform designed for building
  multi-agent simulations
• Scheduler
• Order events
• probes
• Monitoring simulation
• Reporting function
• Each agent is modeled as an object
• Agents organized in swarms
• Scheduler to determine who is next
• Can nest swarms (components can be swarms!)
• Agents possess swarms which represent the agents
  own social models of other agents
• Agents model environment in which agents act
• Observer agents
• Facilitates model building environment for
  simulation