在分布式体系结构的MAS（multi-agent system）中，Agent之间通过彼此的交互，协调完成共同的任务，但是由于没有中心化的管理权威可以依赖，导致对网络中Agent信誉信息进行判断存在一定的困难.传统的基于评价反馈的信誉评估方法存在反馈评价属性信息利用不足以及缺少确保反馈评价信息可信的可行机制等问题，为此，提出一种综合的信誉计算方法.该方法针对个别用户提交的恶意评价，采用CUSUM（cumulative sum）控制图理论对服务评价中的数据进行过滤；利用信息熵的方法对不同维度的评价数据进行整合；使用改进的PageRank算法对评价影响力进行度量，最终得到融合反馈评价真值与评价影响力的综合信誉.仿真结果表明，该方法在提高信誉计算收敛性和准确性、抵抗恶意攻击行为等方面表现出较好的效果.

In MAS with distributed architecture, Agents are employed to achieve task objectives through mutual coordination and collaboration. Since there is no centralized management authority to rely on, it is difficult to judge the reputation information of Agents. Beyond that, traditional reputation evaluation methods based on evaluation and feedback have some problems, such as the insufficient usage of feedback evaluation and the absence of feasible mechanism for credible feedback evaluation information, etc. To solve these problems, a comprehensive reputation calculation method is proposed in this study. Concerning about malicious evaluation submitted by individual users, the proposed algorithm first filters the service evaluation data by the CUSUM (cumulative sum) control chart theory, then integrates different dimension of evaluation data using information entropy method, after that, uses PageRank algorithm to measure the influence of individual. Finally, it gives the comprehensive reputation model incorporating feedback evaluation of truth value and individual influence. Simulation results show that the proposed method performs well in improving reputation computation accuracy and convergence as well as the resisting of malicious attacks.

TP18

国家自然科学基金（61572167，61502136）；科技部国际合作项目（2015DFA11450）