基于语言类任务的概念化强化学习框架OA北大核心CSTPCD

Language-assisted tasks are proposed to facilitate the generalization ability of reinforcement learning policy.The key question is to learn the general representation across different scenarios.Existing studies often implicitly learn the joint representation,which may include spurious correlation information and consequently compromise pol-icy's generalization performance and training efficiency.To address this issue,a conceptual reinforcement learning framework(CRL)is proposed,which exploits the motivation of human cognition that extracts similarits from nu-merous instances to generate conceptual abstraction,and incorporates a multi-level attention encoder and restricted loss functions to learn compact and invariant conceptual representation for the policy.Evaluated in challenging lan-guage-assisted tasks,the results demonstrate that CRL significantly improves the policy's training efficiency(up to 70%)and generalization ability(up to30%).Additionally,the conceptual representation also shows better inter-pretability than other representations.