Policy Gradient Methods for Reinforcement Learning with Function Approximation

本文探讨了强化学习中的一种替代方法,即直接通过策略的函数逼近并利用梯度更新来优化期望奖励。研究显示,通过使用近似的动作价值或优势函数辅助,策略的梯度可以被写成一种适合从经验中估计的形式。基于此结果,首次证明了一种带有任意可微函数逼近的策略迭代方法能够收敛到局部最优策略。

摘要生成于 C知道 ,由 DeepSeek-R1 满血版支持, 前往体验 >

Function approximation is essential to reinforcement learning, but the standard approach of approximating a value function and determining a policy from it has so far proven theoretically intractable. In this paper we explore an alternative approach in which the policy is explicitly represented by its own function approximator, independent of the value function, and is updated according to the gradient of expected reward with respect to the policy parameters. Williams’s REINFORCE method and actor critic methods are examples of this approach. Our main new result is to show that the gradient can be written in a form suitable for estimation from experience aided by an approximate action-value or advantage function. Using this result, we prove for the first time that a version of policy iteration with arbitrary differentiable function approximation is convergent to a locally optimal policy.
评论 1
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值