抄録
In this paper, we propose a new simulation-based Distributed Reinforcement Learning approach that solves large planning problems under uncertain environment. The proposed method is a distributed state-action representation for softening an interaction and reward design for making agents cooperate. We apply it to real sewerage control systems, as the problem with uncertainty. Simulation results show it finds good control rules, which can cope with various situations, by dealing with the uncertainty included in real data directly on a simulator based on real systems.
