Guided Exploration in Sparse Reward Environments using Causal Prior Knowledge

In reinforcement learning (RL), sparse reward environments pose significant challenges for exploration and learning. Using prior causal knowledge, we propose a novel approach to guide RL agents during exploration. Our method starts with strong guidance, gradually tapering it off to encourage independent policy convergence. Our experiments demonstrate that this non-uniform guidance strategy aids in achieving better performance while allowing agents to eventually surpass the initial guided model. Our work sheds light on the delicate balance between exploration and convergence in RL tasks with sparse rewards.