多Agent动态影响图的近似计算方法

Approximate Computation of Multi-Agent Dynamic Influence Diagrams

由于复杂系统具有高维性和不确定性常难以表示处理,因而知识表示和计算方法是复杂系统研究中的公开难题.当前,多Agent影响图不能建模动态环境和多Agent,马尔可夫决策过程难以表示A-gents之间结构关系的问题,因而提出一种用局部概率因式表示动态环境中多Agent之间关系的新决策模型——多Agent动态影响图(MADIDs).针对MADIDs模型的联合概率分布和联合效用函数在计算上的高维问题,研究该模型的近似计算方法.给出MADIDs概率结构部分的一种分层分解的分布近似方法,并通过对该近似方法的误差和复杂性的分析,给出一个可对近似分布的精度和复杂性进行均衡的函数δ(k);给出一种BP神经网络通过局部效用的学习来近似计算MADIDs的联合效用.在模型实例上的实验结果显示了MADIDs模型近似计算方法的有效性.

Due to high dimension and uncertainty of the complex system, the complexity system is often hard to represent and process, and the knowledge representation and computation methods of complex systems are open hard problems in complex system research. At present, MAIDs can not model dynamic environment and it is difficult for multi-agent MDPs to represent structural relations among agents; so a multi-agent dynamic influence diagrams （MADIDs） model is given to representation relations among multiagents in dynamic environment by local factor probability form. The computation of joint probability distribution and joint utility function of MADIDs are a high dimension problem, so the approximate computation methods are researched. A distribution approximation method of hierarchical decomposition of probability structural MADIDs is studied; based on analysis of the complexity and the error of the distribution approximation method, a function δ（k） is introduced to establish equilibrium between precision and complexity of approximate distribution. Then a BP neural network is given to approximately compute utility structural MADIDs by learning local utility. Finally, given model instances, the experiment results show the validity of the approximation computation method of the MADIDs model.