A Study of an Indirect Reward on Multi-agent Environments
Kazuteru Miyazaki · Procedia Computer Science · 2016
In a multi-agent learning where multiple agents are learning, there is a problem about an indirect reward that is how to distribute a reward to an agent that does not obtain a reward directly. We have shown the theorem [3] about “negative effect” of an indirect reward. This paper focuses on the “positive effect” of an indirect reward such as an elimination of the perceptual aliasing pro blem [1] . First, we describe the relationship the theorem [3] and the “positive effect” of the indirect reward. Next, we propose a method to eliminate the perceptual aliasing problem and show the effectiveness of the proposed method by numerical examples.