Abstract:Although deep reinforcement learning can solve many complex control problems, it needs to pay the cost of a large number of interactions with the environment, which is a major challenge for deep reinforcement learning. One of the reasons for this problem is that it is difficult for an agent to extract effective features from a high-dimensional complex input only by relying on the loss of value function. As a result, the agent has an insufficient understanding of the state and cannot correctly assign value to the state. Therefore, this study proposes a regularized predictive representation learning (RPRL) method combining forward state prediction and latent space constraint to make agents know the environment and improve the sample efficiency of reinforcement learning. The method helps agents to learn and extract state features from high-dimensional visual observations to improve the sample efficiency of reinforcement learning. The forward state transfer loss is used as the auxiliary loss so that the features learned by agents contain dynamic information related to environmental transition. At the same time, the state representation of latent space is regularized on the basis of forward prediction, which further helps the agent to learn the smooth and regular representation of the high-dimensional input. In DeepMind Control (DMControl) environment, the proposed method achieves better performance than other model-based methods and model-free methods with representation learning.