arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2406.09563cs.LG
收藏

e-COP:策略的情景约束优化

e-COP : Episodic Constrained Optimization of Policies

University of Southern California(南加州大学);Google DeepMind(谷歌DeepMind)

Akhil Agnihotri, Rahul Jain, Deepak Ramachandran, Sahil Singla

AI总结:

本文提出首个用于情景(有限时域)设定下约束强化学习的策略优化算法 e-COP,通过建立策略差异引理并结合多种求解思路,实现了易于实现、数值稳定且具理论保证的优化,其实验性能优于适配的最先进非情景算法,并可扩展至大模型的安全约束人类反馈强化学习。

摘要:

在本文中,我们提出了 e-COP 算法,这是首个用于情景(有限时域)设定下约束强化学习(RL)的策略优化算法。当系统的行为具有独立的优化准则集合和约束条件时,这种建模方法是适用的。我们通过首先为情景设定建立策略差异引理来解决这一问题,该引理为算法提供了理论基础。然后,我们提出将一系列已有的和新颖的求解思路结合起来,以生成易于实现且数值稳定的 e-COP 算法,并在特定的缩放假设下提供了关于最优性的理论保证。通过使用 Safety Gym 套件中的基准进行广泛的实证分析,我们表明我们的算法与适配于情景设定的最先进(非情景)算法相比,具有相似或更好的性能。该算法的可扩展性为其在大型语言模型或扩散模型的安全约束人类反馈强化学习中的应用打开了大门。

英文摘要:

In this paper, we present the $\texttt{e-COP}$ algorithm, the first policy optimization algorithm for constrained Reinforcement Learning (RL) in episodic (finite horizon) settings. Such formulations are applicable when there are separate sets of optimization criteria and constraints on a system's behavior. We approach this problem by first establishing a policy difference lemma for the episodic setting, which provides the theoretical foundation for the algorithm. Then, we propose to combine a set of established and novel solution ideas to yield the $\texttt{e-COP}$ algorithm that is easy to implement and numerically stable, and provide a theoretical guarantee on optimality under certain scaling assumptions. Through extensive empirical analysis using benchmarks in the Safety Gym suite, we show that our algorithm has similar or better performance than SoTA (non-episodic) algorithms adapted for the episodic setting. The scalability of the algorithm opens the door to its application in safety-constrained Reinforcement Learning from Human Feedback for Large Language or Diffusion Models.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑