Emergence of Information Processor Using Real World--Real-Time Learning of Pursuit Problem
Hiroyuki Fujii, Kazuyuki Ito, Akio Gofuku · International Conference Hybrid Intelligent Systems · 2006
Real-time reinforcement learning is difficult because number of trials is too much to complete learning within limited time. To solve the problem, we consider reduction of action-state space by information processor using real world without prior knowledge. We obtain the information processor in evolution by setting the fitness as ease of learning. As a typical example, we address pursuit problem in which dynamics is regarded. As a result, the processor has been obtained in evolution and agent has learned in real-time.