Abstract
Traditional reinforcement learning (RL) methods mainly rely ontrial-and-error exploration, often lacking mechanisms to guide agents towardmore informative decision-making and struggling to leverage past experiences,resulting in low sample efficiency. To overcome this issue, we propose aninnovative framework inspired by cognitive principles: Cognitive Belief-DrivenReinforcement Learning (CBD-RL). By incorporating cognitive heuristics, CBD-RLtransforms conventional trial-and-error learning into a more structured andguided learning paradigm, simulating the human reasoning process. Thisframework's core is a belief system that optimizes action probabilities byintegrating feedback with prior experience, thus enhancing decision makingunder uncertainty. It also organizes state-action pairs into meaningfulcategories, promoting generalization and improving sample efficiency. Theconcrete implementations of this framework, CBDQ, CBDPPO, and CBDSAC,demonstrate superior performance in discrete and continuous action spaces indiverse environments such as Atari and MuJoCo. By bridging cognitive scienceand reinforcement learning, this research opens a new avenue for developing RLsystems that are more interpretable, efficient, and cognitively inspired.