2019 年 139 巻 12 号 p. 1494-1500
In applying reinforcement learning to a different environment, relearning is generally required. The relearning, however, is time-consuming, and therefore a method without the relearning should be developed. This paper proposes a reinforcement learning method with generalization ability for solving an optimal routing problem with a given set of multiple goal positions. The proposed method can rapidly find the optimal route for any set of the multiple goal positions once a reinforcement learning agent learns. In the proposed method, a graph search algorithm determines the visiting order of the goal positions, and an ordinary learning algorithm such as Q-learning determines each route between goal positions. The performance of the proposed method is evaluated through numerical experiments.
J-STAGEがリニューアルされました! https://www.jstage.jst.go.jp/browse/-char/ja/