Self-Generation of Reward by Inputs from Multi Sensors -Integration of Evaluations for Inputs to Avoid Danger-

摘要

We proposed a method which generates a reward from sensor inputs without setting a reward when a robot does reinforcement learning in a complex environment. In previous research, this method generated a reward by weighted average of evaluations for each sensor. But previous method could ignore some small evaluations. Small evaluation means danger inputs. So, previous research had a problem of ignoring danger inputs by average. Therefore, we propose a method which generates a reward to avoid danger using inputs from multiple sensor in this research. This method evaluates sensor inputs from an environment and generates a reward. To evaluate inputs, we define the indicator based on organisms which is independent of an environment. This method can generate a suitable reward and make a robot learn an action independently an environment. Proposal method places importance on danger inputs and generates a reward to avoid danger. Therefore, proposal method makes a robot be able to survive. Then, we confirm that an agent can learn an action to avoid danger by an experiment using simulation environment.