DGTRL: Deep graph transfer reinforcement learning method based on fusion of knowledge and data

Deep reinforcement learning has shown promising application effects in many fields. However, issues such as low sample efficiency and weak knowledge transfer and generalization capabilities constrain its in-depth development. This paper proposes a deep graph transfer reinforcement learning method based on knowledge and data fusion. Through knowledge transfer between tasks and data training in exploration, efficient learning driven by knowledge and data fusion is realized. First, we map the trajectories of agents in the source and target domains into graph structures to represent knowledge and environmental models and use a graph matching method to match the knowledge adaptation domain based on similarity. Second, real-time transfer of source domain knowledge is carried out in a similar mapping domain, and policy mapping, verification, and adaptive transfer mechanisms based on knowledge reuse frequency are designed to effectively improve the transfer effect. Finally, we evaluate and validate the proposed algorithm, mechanism, and their application effectiveness in diverse and multitask knowledge transfer scenarios using three indicators in environments such as maze. The results show that the method improves the target task time indicators by 84.7% and 70% under the conditions of no verification exploration and a standard transfer setting, respectively, achieving significant knowledge reuse effects.