The Di culty of Coping With Tradeo s through Local Decision Making

In Distributed Problem Solving (DPS), Dynamic Reorganization is said to be e ective in order to cope with tradeo s between the global accuracy and the e ciency of the problem solving. In order to realize an architecture for the Dynamic Reorganization, it is important to consider a question \When should agents reorganize ?". Such architectures, where each agent has a Meta-level component and makes decisions locally to improve the e ciency, seem to be robust systems. There is, however, the possibility that such an architecture decreases its global accuracy because each agent pursues its own e ciency to the extreme. In this paper, we present several experiments on the pursuit game to discuss the di culty of implementing an architecture which aims to improve the e ciency of the problem solving through the local decision making. We show the di culty of controlling \When to reorganize" in such an system is a rather subtle problem and the \making plans of agents consistent" is di cult. Moreover, we show unsuitable \translation" of the global accuracy and e ciency into the individual agents' utility could arrow sel sh agents to exploit cooperative agents and co-evolve with other sel sh agents.