Two-Stage Reinforcement Learning based on Genetic Network Programming for mobile robot

This paper studies the adaptability of Two-Stage Reinforcement Learning based on Genetic Network Programming for a mobile robot to cope with sudden changes in the environments, i.e., sensors break suddenly in the implementation. Two-Stage Reinforcement Learning (TSRL) uses two kinds of learning, that is, (1) sub node selection proposed in the conventional Genetic Network Programming with Reinforcement Learning and (2) branch connection selection. As a result, when the sudden changes occur in the environments, the proposed method can determine the actions more appropriately.