增量式双自然策略梯度的行动者评论家算法

发布时间：2018-08-13 19:02

【摘要】：针对强化学习中已有连续动作空间算法未能充分考虑最优动作的选取方法和利用动作空间的知识,提出一种对自然梯度进行改进的行动者评论家算法。该算法采用最大化期望回报作为目标函数,对动作区间上界和下界进行加权来求最优动作,然后通过线性函数逼近器来近似动作区间上下界的权值,将最优动作求解转换为对双策略参数向量的求解。为了加快上下界的参数向量学习速率,设计了增量的Fisher信息矩阵和动作上下界权值的资格迹,并定义了双策略梯度的增量式自然行动者评论家算法。为了证明该算法的有效性,将该算法与其他连续动作空间的经典强化学习算法在3个强化学习的经典测试实验中进行比较。实验结果表明,所提算法具有收敛速度快和收敛稳定性好的优点。
[Abstract]:In view of the fact that the continuous action space algorithm in reinforcement learning has not fully considered the selection method of optimal action and the knowledge of action space, an actor critic algorithm is proposed to improve the natural gradient. In this algorithm, the maximum expected return is used as the objective function, the upper and lower bounds of the action interval are weighted to obtain the optimal action, and then the weights of the upper and lower bounds of the action interval are approximated by a linear function approximator. The optimal action solution is converted to the solution of the dual strategy parameter vector. In order to speed up the learning rate of the upper and lower bound parameter vectors, the incremental Fisher information matrix and the qualification trace of the upper and lower bound weights of the action are designed, and the dual strategy gradient incremental natural action critic algorithm is defined. In order to prove the effectiveness of this algorithm, the algorithm is compared with other classical reinforcement learning algorithms in continuous action space in three classical test experiments of reinforcement learning. Experimental results show that the proposed algorithm has the advantages of fast convergence speed and good convergence stability.
【作者单位】：苏州大学计算机科学与技术学院;软件新技术与产业化协同创新中心;吉林大学符号计算与知识工程教育部重点实验室;
【基金】：国家自然科学基金资助项目(No.61272005,No.61303108,No.61373094,No.61472262,No.61502323,No.61502329) 江苏省自然科学基金资助项目(No.BK2012616) 江苏省高校自然科学研究基金资助项目(No.13KJB520020) 吉林大学符号计算与知识工程教育部重点实验室基金资助项目(No.93K172014K04) 苏州市应用基础研究计划工业部分基金资助项目(No.SYG201422,No.SYG201308)~~
【分类号】：TP181

【相似文献】