【作者】 王冲冲； 许佳； 李莉；

【Author】 Wang Chong Chong;Xu Jia;Li Li;College of Electronics and Information Engineering,Tongji University;

【机构】 同济大学电子与信息工程学院；

【摘要】 随着人工智能领域的迅速发展,强化学习与因果发现两个分支日益显现出深厚的交集和相互促进的潜力。本文通过分析两个领域当前面临的算法效率、数据利用率、模型泛化能力及可解释性等挑战,深入探讨了两个领域的交叉融合路径。首先分析了强化学习和因果发现的现有研究方法和困境,总结了通过融合这两个领域如何提升因果发现算法和强化学习算法的性能。其次讨论了利用强化学习作为因果发现搜索工具的现有应用,指出其在干预规划和泛化性上的优势以及在计算效率上的限制。最后探索了在强化学习算法中使用因果信息的研究进展,并针对如何提高数据利用率和模型泛化能力与可解释性等问题给出了合理建议。更多还原

【Abstract】 With the rapid development of the field of artificial intelligence,the branches of reinforcement learning and causal discovery are increasingly showing a deep intersection and potential for mutual enhancement.This paper delves into the paths of cross-fusion between the two fields by analyzing the current challenges faced in terms of algorithm efficiency,data utilization,model generalizability,and interpretability.Firstly,it analyzes the existing research methods and dilemmas in reinforcement learning and causal discovery,summarizing how the integration of these two fields can enhance the performance of causal discovery algorithms and reinforcement learning algorithms.Secondly,it discusses the current applications of using reinforcement learning as a search tool for causal discovery,pointing out its advantages in intervention planning and generalizability,as well as its limitations in computational efficiency.Lastly,it explores the research progress in using causal information in reinforcement learning tasks,and offers reasonable suggestions on how to improve data utilization,model generalizability,and interpretability,among other issues.更多还原