A Model-Based Approach to Solve the Sparse Reward Problem
Shuailong Li, Xiaohui Wang, Wei Zhang, Xin Zhang · 2021
For reinforcement learning (RL) algorithms, the sparsity of reward has always been a problem to be solved. Because reinforcement learning cannot get effective feedback in most cases, the agent is difficult to learn effectively. We propose a model-based algorithm to create extra rewards and increase reward density to make it easy to learning. To reshape rewards, we use model error as an extra reward and add it to the return. We test our approach in the Google Research Football Environment, and our algorithm gets good results.