一种用于随机优化的连续动作学习自动机  

A Continuous- Action Learning Automaton for Stochastic Optimization

作　　者：刘晓[1] 

机构地区：[1]中航工业西安航空计算技术研究所,陕西西安710068

出　　处：《航空计算技术》2015年第5期73-77,共5页Aeronautical Computing Technique

摘　　要：学习自动机(LA)是一种随机优化技术,其中自动机的目标是从一个允许的动作集里选择最优的动作。现有LA的动作集大多都是有限集,因此对于连续参数学习问题,需要将动作空间离散化,并且解的精度取决于离散化的颗粒度。提出一种新的连续动作学习自动机(CALA),其动作集为一个可变区间,并依照均匀分布选择输出动作。算法通过一种类似提高拟退火的方式确定一个"最新且最好"的动作,并根据该动作对区间的两个端点进行更新。通过优化一个被噪声污损的多模态函数的仿真实验,演示了新算法相对于三种现有的CALA算法的优越性。Learning automaton（ LA） is a stochastic optimization technique where the goal of the automaton is to choose the optimal action out of a set of allowable actions.In most of the existing LA,the action set is finite.Hence,for continuous parameter learning problems,the action space needs to be discretized and the accuracy of the solutions depends on the level of discretization.A new continuous-action learning automaton（ CALA） is proposed in this paper.The action set of the automaton is a variable interval,actions are selected according to a uniform distribution over this interval.The two endpoints of the interval are updated by using a recently best action which is determined in a manner similar to the simulated annealing scheme.Simulation results are given to show the advantages of the proposed algorithm over three of the existing CALA algorithms for optimization a noise-corrupted,multi-modal function.

关 键 词：强化学习 在线学习 随机优化 学习自动机 连续动作学习自动机 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]