Extracting Decision Tree from Trained Deep Reinforcement Learning in Traffic Signal Control

Deep reinforcement learning (DRL) has achieved promising results on traffic signal control systems. However, due to the complexity of the decisions of deep neural networks, it is a great challenge to explain and visualize the policy of reinforcement learning (RL) agents. The decision tree can provide useful information for experts responsible for making reliable decisions. In this paper, we employ decision trees to extract models with readable interpretations from expert policy achieved by DRL methods. We evaluate our methods via a single-intersection traffic signal control task on the simulation platform of Urban MObility (SUMO). The experimental results demonstrate that the extracted decision trees can be used to understand the learning process and the learned optimal policy of the DRL methods.