Learning explainable task-relevant state representation for model-free deep reinforcement learning

State representations considerably accelerate learning speed and improve data efficiency for deep reinforcement learning (DRL), especially for visual tasks. Task-relevant state representations could focus on features relevant to the task, filter out irrelevant elements, and thus further improve performance. However, task-relevant representations are typically obtained through model-based DRL methods, which involves the challenging task of learning a transition function. Moreover, inaccuracies in the learned transition function can potentially lead to performance degradation and negatively impact the learning of the policy. In this paper, to address the above issue, we propose a novel method of explainable task-relevant state representation (ETrSR) for model-free DRL that is direct, robust, and without any requirement of learning of a transition model. More specifically, the proposed ETrSR first disentangles the features from the states based on the beta variational autoencoder (beta-VAE). Then, a reward prediction model is employed to bootstrap these features to be relevant to the task, and the explainable states can be obtained by decoding the task-related features. Finally, we validate our proposed method on the CarRacing environment and various tasks in the DeepMind control suite (DMC), which demonstrates the explainability for better understanding of the decision-making process and the outstanding performance of the proposed method even in environments with strong distractions.