arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2109.04033cs.LG
收藏

梯度时序差分学习的新版本

New Versions of Gradient Temporal Difference Learning

KAIST(韩国科学技术院);Shinhan Bank(新韩银行)

Donghwan Lee, Han-Dong Lim, Jihoon Park, Okyong Choi

AI总结:

本文基于凸-凹鞍点解释提出多种 GTD 新变体,将其统一到原始-对偶框架中并给出稳定性分析,同时通过数值实验比较各方法。

摘要:

Sutton、Szepesvári 和 Maei 提出了首批同时兼容线性函数逼近和离策略训练的梯度时序差分(GTD)学习算法。本文的目标是:(a) 提出若干 GTD 变体并进行广泛的比较分析;(b) 为 GTD 建立新的理论分析框架。这些变体基于 GTD 的凸-凹鞍点解释,该解释有效地将所有 GTD 统一到单一框架中,并依据原始-对偶梯度动力学的最新结果提供简洁的稳定性分析。最后,本文给出数值比较分析以评估这些方法。

英文摘要:

Sutton, Szepesvári and Maei introduced the first gradient temporal-difference (GTD) learning algorithms compatible with both linear function approximation and off-policy training. The goal of this paper is (a) to propose some variants of GTDs with extensive comparative analysis and (b) to establish new theoretical analysis frameworks for the GTDs. These variants are based on convex-concave saddle-point interpretations of GTDs, which effectively unify all the GTDs into a single framework, and provide simple stability analysis based on recent results on primal-dual gradient dynamics. Finally, numerical comparative analysis is given to evaluate these approaches.

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。

↑