KGRL: A Method of Reinforcement Learning Based on Knowledge Guidance

Reinforcement learning usually starts from scratch. Due to the characteristic of receiving feedback from interactions with the environment, starting from scratch often leads to excessive unnecessary exploration. In contrast, humans have a lot of common sense or prior knowledge when learning. The existence of prior knowledge accelerates the learning process and reduces unnecessary exploration. Inspired by this, in this article, we propose a knowledge-guided approach to learning reinforcement learning policies. We use prior knowledge to derive an initial policy and guide the subsequent learning process. Although prior knowledge may not be completely applicable to new tasks, the learning process is greatly accelerated because the initial policy ensures a fast start to learning and guidance in the middle allows for avoiding unnecessary exploration. This is a new framework that combines human prior suboptimal knowledge with reinforcement learning. We refer to it as KGRL - Knowledge-Guided Reinforcement Learning. The KGRL framework includes defining fuzzy rules based on prior knowledge, learning an initial policy based on fuzzy rules, and guiding the new policy through knowledge-guided encoding. We conducted experiments on four tasks in OpenAI Gym and PLE, and the results show that our algorithm combines prior knowledge and reinforcement learning effectively. By utilizing knowledge-guided encoding, it improves the performance of reinforcement learning, trains faster, and increases rewards by approximately 12.7 \(\%\) .