Asynchronous teams : cooperation schemes for autonomous, computer-based agents

An asynchronous team (A-Team) is a strongly cyclic network of autonomous agents (workers) and memories (workplaces). Results (trial solutions to computational problems) circulate continually through this network. Agents work in parallel and cooperate by modifying one another's results. We have accumulated a good deal of experience in making the circulating results converge to better solutions of optimization and constraint satisfaction problems than the agents can find when working independently. This paper does three things. First, it distills our experiences with A-Teams into a protocol for designing them. Second, it points out that a sufficient condition for the circulating results to converge is that the skills of the agents that construct new results be complementary to the skills of the agents that destroy old results. Third, it argues that this complemantarity is relatively easy to achieve. The practical implications are: a) the quality of solutions obtained by any problem-solving-algorithm, even the best one available, can invariably be improved by combining it with other available algorithms into an A-Team, b) the quality of solutions obtained by any A-Team can invariably be improved by expanding its size, and c) expansions are relatively easy to make.