TransPath：a Deep Transfer Reinforcement Learning Method for Knowledge Reasoning
CUI Yuan-ning1，LI Jing1，CHEN Yan2，LU Zheng-jia2
1（College of Computer Science and Technology，Nanjing University of Aeronautics and Astronautics，Nanjing 211106，China）2（Information and Communication Company of State Grid Shanghai Electric Power Company，Shanghai 200000，China）
Abstract：Knowledge reasoning based on deep reinforcement learning（RL） aims to infer missing facts and complete the knowledge graph.The RL agent searches for paths on the knowledge graph and makes fact predictions and link predictions based on the paths.Due to their good performance and interpretability，the knowledge reasoning method based on deep RL has rapidly become a research hotspot in recent years.However，for a specific entity，there are a large number of invalid actions in the action space，and RL agents often stop walking due to choosing invalid actions，so the success rate of path search is low.In order to solve the problem of invalid actions，we propose a knowledge reasoning method based on deep transfer reinforcement learning-TransPath，which adds a one-step walking source task to select valid actions in addition to the target task.First，train the one-step walk on the source task to help the RL agent learn to choose valid actions，and then transfer to the target reasoning task for path search training to improve the success rate of path search.The experimental results on the data sets FB15K-237 and NELL-995 show that our method not only greatly improves the success rate of path search，but also achieves the best performance in most reasoning tasks.
崔员宁,李静,陈琰,陆正嘉. TransPath：一种基于深度迁移强化学习的知识推理方法[J]. 小型微型计算机系统, 2022, 43(3): 536-543.
CUI Yuan-ning,LI Jing,CHEN Yan,LU Zheng-jia. TransPath：a Deep Transfer Reinforcement Learning Method for Knowledge Reasoning. Journal of Chinese Computer Systems, 2022, 43(3): 536-543.