A parameter control method inspired from neuromodulators in reinforcement learning

摘要

The brain gains appropriate behaviors, which get rewards and escapes punishments by trial-and-error. Reinforcement learning models such a system by an engineering approach. Neuromodulators, which project widely in the brain and adjust functions in each brain part, are matched with parameters of reinforcement learning. We propose a reinforcement learning algorithm, which can follow sudden changes in environment by considering how neuromodulators affect behaviors. This algorithm improves actions by controlling the parameters of reinforcement learning after the obtained reward decreased as compared with the past. Computer simulation shows that the robots with the proposed algorithm are able to respond flexibly to sudden environmental changes.