Asynchronous Teams: Cooperation Schemes for Autonomous Agents

  • Authors:

  • Sarosh Talukdar;Lars Baerentzen;Andrew Gove;Pedro De Souza

  • Affiliations:

  • Carnegie Mellon University, 5000 Forbes Ave., Pittsburgh, PA 15213. talukdar@cmu.edu;Carnegie Mellon University, 5000 Forbes Ave., Pittsburgh, PA 15213;Carnegie Mellon University, 5000 Forbes Ave., Pittsburgh, PA 15213;Carnegie Mellon University, 5000 Forbes Ave., Pittsburgh, PA 15213

  • Venue:
  • Journal of Heuristics
  • Year:
  • 1998

Quantified Score

Hi-index 0.00

Visualization

Abstract

Experiments over a variety of optimization problems indicatethat scale-effective convergence is an emergent behavior ofcertain computer-based agents, provided these agents areorganized into an asynchronous team (A-Team). An A-Team is aproblem-solving architecture in which the agents are autonomousand cooperate by modifying one another‘s trial solutions. Thesesolutions circulate continually. Convergence is said to occur ifand when a persistent solution appears. Convergence is said tobe scale-effective if the quality of the persistent solutionincreases with the number of agents, and the speed of itsappearance increases with the number of computers. This paperuses a traveling salesman problem to illustrate scale-effectivebehavior and develops Markov models that explain its occurrencein A-Teams, particularly, how autonomous agents, withoutstrategic planning or centralized coordination, can converge tosolutions of arbitrarily high quality. The models also perdicttwo properties that remain to be experimentally confirmed:• construction and destruction are dual processes. In otherwords, adept destruction can compensate for inept constructionin an A-Team, and vice-versa. (Construction refers to theprocess of creating or changing solutions, destruction, to theprocess of erasing solutions.)• solution quality is independent of agent-phylum. In otherwords, A-Teams provide an organizational framework in whichhumans and autonomous mechanical agents can cooperate effectively.