Agent动态交互信任预测与行为异常检测模型

Prediction and Abnormal Behavior Detection of Agent Dynamic Interaction Trust

在agent理论中,信任计算是一个有意义的研究方向.然而目前agent信任研究都是以平均交互成功率来计算,较少考虑信任动态变化,因而准确预测和行为异常检测的能力不能令人满意.针对上述问题,以概率论为工具,按时间分段交互历史给出agent交互信任计算模型CMAIT;结合信任的变化率,给出信任计算的置信度和异常检测机制.实验以网上电子商务为背景,实验结果显示该计算模型的预测误差为TRAVOS的0.5倍,计算量也较少;既可用于对手历史行为的异常检测,防止被欺骗,又可用于对手未来行为的预测.改进了Jennings等人关于agent信任的工作.

Computation of trust is an interesting research direction in agent and multi-agent systems theory. There have been a lot of researches on agent trust and reputation in the past few years, such as TRAVOS model presented by Teacy and FIRE model presented by Huynh. However, previous abnormal behavior of CMAIT is also given. It is proven that trust of average historical interaction is a particular case of CMAIT. Experiments are conducted on Web e-commerce at Taobao website. Experimental results demonstrate that computational error of CMAIT is half of that of TRAVOS and its computational complexity is also low. It can be applied in the detection of abnormal behavior and the prediction of future behavior. It improves the work of Jennings on agent trust.