Multiagent reinforcement learning with organizational-learningoriented classifier system

摘要

Organizational learning oriented classifier system (OCS) is a newarchitecture proposed by us for an evolutionary computational model. Wehave shown its effectiveness in large scale problems with printedcircuit board (PCB) redesign using computer aided design (CAD). Thepaper proposes a novel reinforcement learning method for multiagentswith OCS for more practical and engineering use. To validate theeffectiveness of our method, we have conducted experiments on real scalePCB design problems for electric appliances. The experimental resultshave suggested that: (1) our method has found feasible solutions withthe same quality of those by human experts; (2) the solutions areglobally better than those by the conventional reinforcement learningmethods with regard to both the total wiring length and the number ofiterations