Efficient Model Based Reinforcement Learning Control Using Relative Entropy Regularization

In this paper, we propose MBPO-CDPP, a novel model-based reinforcement learning algorithm that addresses performance limitations caused by model bias. Our approach integrates dynamic relative entropy regularization into Model-Based Policy Optimization (MBPO). By automatically adjusting the relative entropy constraint based on the predicted uncertainty, the proposed method successfully enhances the learning capability and stability based on MBPO while preventing the negative impacts caused by the biased model. Experiments across six control benchmarks demonstrate superiority in both learning efficiency and policy stability compared to related baselines. The results indicate the potential and effectiveness of adaptive relative entropy regularization in enhancing the efficiency of model-based Reinforcement Learning.