A Study on the Emergent Strategy Acquisition in the Massively Parallel Graph-Reduction

点击翻译按钮获取中文摘要

We have developed an interactive visual simulator for observing complex behavior of the emergent computation model. As a preliminary experiment using this simulator, we have tried to generate cooperative strategies in a multiagent game defined on a reward matrix. It is not so easy to identify emergent behavior in the multiagent game because of the accidental property that is the essence of the emergent computation. Moreover the parameters specifying the agents are very critical.We have tried to build a massively parallel graph-reduction system for lambda expressions on the visual simulator described above. To design the local transition rules for each agent is a harder work than we expected. Consequently, within the practical simulation steps, we could not find out interesting emergent behavior in the graph-reduction process by using the cooperative agents. Hence, we have partially changed our original research plan, and developed an emergent load balancing system for the massively parallel graph-reduction. In a massively parallel computation field, an autonomous agent (computation node) searches lightly-loaded nodes among its neighbors to transfer extra tasks. To apply a stochastic learning automaton to the fitness evaluation in a genetic algorithm, we have achieved highly adaptive load balancing with a dynamically changing load environment.We have implemented a prototype of the emergent load balancing system on the massively parallel computer SR2201. By using independent virtual tasks with no file access, we have found well load balancing emerged in the parallel computation field. Clusters of nodes for the task transfer is built up in a self-organized way. This is a distinguished feature of the emergent computation model. The load balancing in the massively parallel computation field is one of the promising applications of the emergent computation.