基于联合博弈的多Agent学习  

Joint Game-based Multi-agent Learning

作　　者：黄付亮[1] 张荣国[1] 陈大川[1] 刘焜[2] 

机构地区：[1]太原科技大学计算机科学与技术学院,太原030024 [2]合肥工业大学机械与汽车工程学院,合肥230009

出　　处：《计算机与数字工程》2011年第6期21-24,共4页Computer & Digital Engineering

基　　金：国家自然科学基金项目(编号:50775060)资助

摘　　要：在研究Q-Learning算法的基础上,将博弈论中的团队协作理论引入到强化学习中,提出了一种基于联合博弈的多Agent学习算法。该算法通过建立多个阶段博弈,根据回报矩阵对阶段博弈的结果进行评估,为其提供一种有效的A-gent行为决策策略,使每个Agent通过最优均衡解或观察协作Agent的历史动作和自身当前情况来预测其所要执行的动作。对任务调度问题进行仿真实验,验证了该算法的收敛性。This article discusses the joint learning process,propose a joint game based learning algorithm of multi-Agent.The method is through the establishment of multiple stage game,according to return matrix to assess the results of stage game,so that the optimal equilibrium solution for each Agent or observed through collaboration a historical action and self-Agent the current situation to predict their movements to be performed.Finally,the task scheduling problem by simulation results demonstrate the convergence of the algorithm,show that the learning algorithm is an efficient and fast way of learning.

关 键 词：AGENT 强化学习 联合博弈 MAS 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]