Modular reinforcement learning with adaptive networks

Abstract

A modular reinforcement learning system (modular RL) with adaptive networks was proposed for applying the reinforcement learning into control tasks with numerous inputs. It is composed of several control modules and a selection module. All its modules are calculated by using the incremental normalised Gaussian networks (INGnet). The modular RL (INGnet) showed a better learning ability in all three control tasks performed in this study than the modular reinforcement learning whose all modules are calculated by CMAC [modular RL (CMAC)]. It showed a better or similar learning ability to the reinforcement learning using INGnet [RL (INGnet)]. From the simulation results obtained in this study, the modular RL (INGnet) is considered to have a better learning ability in the control tasks with a large number of inputs (8–10) than the modular RL (CMAC) and the RL (INGnet).