Building a Control Method through Verbalizing the Internal Behavior of a Deep Reinforcement Learning Model

Due to its black-box nature, interpreting the internal behavior of a deep reinforcement learning model is a challenging task for humans. We therefore apply fuzzy modeling for the input-output relationships of a learned deep reinforcement learning model and express these relationships with fuzzy linguistic variables to make linguistic control rules. In this study, we use CartPole as an experiment subject; explain control rules of the model learned by a Deep Q-Network (DQN) with fuzzy linguistic variables; and attempt to control the CartPole using those control rules. We use two approaches to construct control rules. One is to construct control rules from input data and the other is to construct control rules from output data. As a result, through experiments, we have confirmed the trade-off between control performance and the number of rules. The learned DQN of CartPole model stands 200 steps, whereas the fuzzy controllers constructed in this study stand 171.9 steps on average. Though the accuracy of controling CartPole by constructed fuzzy controllers is lower than DQN, it is good that the internal behavior can be understood through the constucted rules.