{"title":"Feasible Policy Iteration With Guaranteed Safe Exploration.","authors":"Yuhang Zhang, Yujie Yang, Shengbo Eben Li, Yao Lyu, Jingliang Duan, Zhilong Zheng, Dezhao Zhang","doi":"10.1109/TCYB.2025.3542223","DOIUrl":null,"url":null,"abstract":"<p><p>Safety guarantee is an important topic when training real-world tasks with reinforcement learning (RL). During online environmental exploration, any constraint violation can lead to significant property damage and risks to personnel. Existing safe RL methods either exclusively address safety concerns after reaching optimality or incorporate a certain degree of tolerance for constraint violations during training. This article proposes a feasible policy iteration framework that can guarantee absolute safety during online exploration, i.e., constraint violations never happen in real-world interactions. The key to maintaining absolute safety lies in confining the environmental exploration at each step always within the feasible region of the current policy. This feasible region is described by a newly defined constraint decay function with uncertainty, ensuring the forward invariance of the feasible region under the worst case. Within the proposed framework, the feasible region maintains its monotonic expanding property and converges to its maximum extent, even though only local samples are available, i.e., the agent only has access to samples within the feasible region. Meanwhile, the trained policy also improves monotonically within its corresponding feasible region if one can use different updating rules inside and outside the feasible region. Finally, practical algorithms are designed with the actor-critic-scenery architecture, consisting of three modules: 1) safe exploration; 2) model error estimation; and 3) network update. Experimental results indicate that our algorithms achieve performance comparable to baselines while maintaining zero constraint violation throughout the entire training process. In contrast, the baseline algorithm typically requires thousands of constraint violations to achieve the same performance. These findings suggest a substantial potential for applying feasible policy iteration in real-world tasks, enabling the online evolution of intricate systems.</p>","PeriodicalId":13112,"journal":{"name":"IEEE Transactions on Cybernetics","volume":"PP ","pages":""},"PeriodicalIF":9.4000,"publicationDate":"2025-03-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"IEEE Transactions on Cybernetics","FirstCategoryId":"94","ListUrlMain":"https://doi.org/10.1109/TCYB.2025.3542223","RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"AUTOMATION & CONTROL SYSTEMS","Score":null,"Total":0}
引用次数: 0
Abstract
Safety guarantee is an important topic when training real-world tasks with reinforcement learning (RL). During online environmental exploration, any constraint violation can lead to significant property damage and risks to personnel. Existing safe RL methods either exclusively address safety concerns after reaching optimality or incorporate a certain degree of tolerance for constraint violations during training. This article proposes a feasible policy iteration framework that can guarantee absolute safety during online exploration, i.e., constraint violations never happen in real-world interactions. The key to maintaining absolute safety lies in confining the environmental exploration at each step always within the feasible region of the current policy. This feasible region is described by a newly defined constraint decay function with uncertainty, ensuring the forward invariance of the feasible region under the worst case. Within the proposed framework, the feasible region maintains its monotonic expanding property and converges to its maximum extent, even though only local samples are available, i.e., the agent only has access to samples within the feasible region. Meanwhile, the trained policy also improves monotonically within its corresponding feasible region if one can use different updating rules inside and outside the feasible region. Finally, practical algorithms are designed with the actor-critic-scenery architecture, consisting of three modules: 1) safe exploration; 2) model error estimation; and 3) network update. Experimental results indicate that our algorithms achieve performance comparable to baselines while maintaining zero constraint violation throughout the entire training process. In contrast, the baseline algorithm typically requires thousands of constraint violations to achieve the same performance. These findings suggest a substantial potential for applying feasible policy iteration in real-world tasks, enabling the online evolution of intricate systems.
期刊介绍:
The scope of the IEEE Transactions on Cybernetics includes computational approaches to the field of cybernetics. Specifically, the transactions welcomes papers on communication and control across machines or machine, human, and organizations. The scope includes such areas as computational intelligence, computer vision, neural networks, genetic algorithms, machine learning, fuzzy systems, cognitive systems, decision making, and robotics, to the extent that they contribute to the theme of cybernetics or demonstrate an application of cybernetics principles.