聚类多Agent强化学习认知无线电资源分配  

Cognitive Radio Resource Allocation by Clustering Multi-Agent Enforcement Learning

作　　者：伍春[1,2] 江虹[2] 易克初[1] 

机构地区：[1]西安电子科技大学综合业务网理论及关键技术国家重点实验室,西安710071 [2]西南科技大学国防科技学院,四川绵阳621000

出　　处：《北京邮电大学学报》2014年第1期80-84,共5页Journal of Beijing University of Posts and Telecommunications

基　　金：国家自然科学基金项目(61379005);国家重点基础研究发展计划项目(2009CB320403);国家科技重大专项基金项目(2009ZX03007-004);西安电子科技大学ISN实验室开放课题(ISN10-09)

摘　　要：针对认知无线电多用户的信道和功率资源分配问题,提出一种基于用户聚类和可变学习速率的多Agent强化学习方法.首先使用分层处理分离信道选择与功率控制,采用快速最优搜索结合用户数均衡调节实现信道分配;其次,使用随机博弈框架对多用户功率控制问题进行建模,通过K均值用户聚类减少博弈参与用户数量和降低单个用户的环境复杂度,并使用可变Q学习速率和策略学习速率的方法进一步促进多Agent强化学习的收敛.仿真结果表明,该方法能使多个用户的功率状态和总收益有效收敛,并且使整体性能达到次优.A multi-agent enforcement learning method based on user clustering as well as a variable learning rate was proposed for solving the problem of channel allocation and power control within multi cognitive radio users. Firstly, a hierarchy processing method was used to separate channel selection and power control. The channel allocation was implemented by fast optimal search combined with user-num- ber balance. Secondly, stochastic game framework was adopted to model the muhiuser power control is- sue. In subsequent multi-agent enforcement learning, K-means user clustering method was employed to reduce the user number in game and single user＇ s environment complexity, and a variable learning rate scheme for Q learning and policy learning was proposed to promote the convergence of muhiuser learning. Simulation shows that the method can make multiuser＇ s power status and global reward converging effec- tively, moreover the whole performance can reach sub-optimal.

关 键 词：认知无线电 多AGENT强化学习 聚类 功率控制 可变学习速率 

分 类 号：TN929.5[电子电信—通信与信息系统]