深層強化学習のハイパーパラメータと報酬関数のベイズ最適化～移動ロボットへの適用～

抄録

Deep reinforcement learning is a machine learning method that combines deep learning and reinforcement learning. Deep Q-network (DQN) is one of the typical methods of deep reinforcement learning. DQN uses Convolutional Neural Network (CNN) which can extract features from the input images. We have applied DQN method to the mobile robot navigation problem. The values of hyper-parameters, including the network structure of DQN, and the reward function used in the DQN algorithm, have been determined empirically. In this study, we attempt to optimize both of the values of hyper-parameters and reward function of deep reinforcement learning by using Bayesian optimization. We realized to optimize the values of hyper-parameters including the network structure of DQN, and the reward function by using Optuna, a framework of Bayesian optimization. We confirmed that the values of hyper-parameters and reward function obtained by Optuna have higher learning performance than that by empirical method.