AgentOrganized Networks for Dynamic Team Formation

1
Agent-Organized Networks for Dynamic Team
Formation

• Gaton, M.E. and desJardins, M., In Proceedings of
  AAMAS-2005, pp. 230-237.

Seo, Young-Woo
2
Introduction

• Agent-Oriented Network (AON)
• An organizational network structure, or
  agent-to-agent interaction topology that is the
  result of local rewiring decisions made by the
  individual agents in a networked multi-agent
  system.
• Goal of the individual agents
• Increase the collective performance of the agent
  organization by rewiring of an initial
  (arbitrary) network topology instead of creating
  and removal of connections
• Strategy Design Issues
• Local estimation of global performance
• When/How to perform network adaptation (or
  rewiring)
• Effectiveness of a Strategy
• Learning rate, stability, global structure

3
Dynamic Multi-Agent Formation

• A Simple Model of Multi-Agent System
• Team of agents forms a structure (or a topology)
  on the fly in decentralized fashion
• Individual decision making is based on local
  information
• Tasks are generated periodically and broadcasted
  globally

4
uncommitted
committed
active

• si States of an individual agent ai on the
• team formation
• uncommitted available, but not assigned to any
  task
• committed assigned to a task,
• but the team works on the skill fulfillment on
  a task
• active the team is working on the task

5
A valid team is a set of agents that induce a
connected subgraph of the agent Social network
and whose skill set fulfills the skill
requirements for a given task.
An uncommitted agent is only eligible to a task
either initiate a team or join a team
1
2
i
Mk
6
Rewiring Strategies (1/2)

• Structure-based
• Preferential attachment
• Rich-Get-Richer A connectivity phenomenon
  observed in a scale-free network
• Probability of connecting to a given node in a
  network is proportional to that nodes degree

7
Rewiring Strategies (2/2)

• Performance-based
• Consider the local performance and referral
• Rewire if the local performance is below the
  average of its immediate neighbors performance
  measures
• Disconnect from the neighbor that has the lowest
  performance
• Establish connection to neighbor with the highest
  performance by requesting referrals

8
Experiments

• Setting
• Random geographic graph
• Randomly placing N agents in the unit square and
  connecting two agents if they are within a
  predefined distance d
• Actions of an individual agent
• Adapt the network / join teams
• Parameters
• N100, ? ? s T10, µ2

9
(No Transcript)
10
(No Transcript)
11
(No Transcript)
12
(No Transcript)