Pseudo Reward and Action Importance Classification for Sparse Reward Problem

Qingtong Wu, Dawei Feng, Yuanzhao Zhai, Bo Ding, Jie Luo

Abstract: Deep Reinforcement Learning(DRL) has witnessed great success in many fields like robotics, games, self-driving cars in recent years. However, the sparse reward problem where a meager amount of states in the state space that return a feedback signal hinders the widespread application of DRL in many real-world tasks. Reward shaping with carefully designed intrinsic rewards provides an effective way to relieve it. Nevertheless, useful intrinsic rewards need rich domain knowledge and extensive fine-tuning, which makes this approach unavailable in many cases. To solve this problem, we propose a framework called PRAIC which only utilizes roughly defined intrinsic rewards. Specifically, the PRAIC consists of a pseudo reward network to extract reward-related features and an action importance network to classify actions according to their importance in different scenarios. Experiments on the multi-agent particle environment and Google Research Football game demonstrate the effectiveness and superior performance of the proposed method.