中国科学院自动化研究所机构知识库

Exploration is a major challenge in deep reinforcement learning, especially in cases where reward is sparse. Simple random exploration strategies, such as epsilon-greedy, struggle to solve the hard exploration problem in the sparse reward environment. A more effective approach to solve the hard exploration problem in the sparse reward environment is to use an exploration strategy based on intrinsic motivation, where the key point is to design reasonable and effective intrinsic reward to drive the agent to explore. This paper proposes a method called CEMP, which drives the agent to explore more effectively and continuously in the sparse reward environment. CEMP contributes a new framework for designing intrinsic reward from multiple perspectives, and can be easily integrated into various existing reinforcement learning algorithms. In addition, experimental results in a series of complex and sparse reward environments in MiniGrid demonstrate that our proposed CEMP method achieves better final performance and faster learning efficiency than ICM, RIDE, and TRPO-AE-Hash, which only calculate intrinsic reward from a single perspective.