多Agent交互式动态影响图的建模方法

The Exploration on Modeling Methods for Interactive Multi-agent Dynamic Influence Diagrams

交互式动态影响图是一种以动态影响图为基础,利用有向图构造Agent之间交互作用的决策概率模型,目前只能解决2个Agent的问题.根据概率图模型理论、交互式部分可观测马尔可夫决策过程性质、最大奖励期望值原理等以3个Agent为例建立多Agent交互式动态影响图(I-MADIDs)模型,探讨除建模Agent之外,其他非建模Agent之间存在稳定关系时,如何简化I-MADIDs模型.最后对老虎问题进行建模,利用HUGIN7.0对其进行求解,分别讨论了建模A-gent和其他Agent的决策情况,对比了精确方法和简化模型中贝叶斯参数学习近似方法中Agent的决策情况,证明了近似方法的有效性.

Interactive dynamic influence diagrams （I-DIDs） are a kind of probability graph models based on dynamic influence dia grams,using directed graph to construct decision-making models about interaction between agents. I-DIDs can only solve 2 agents＇ problems. Take 3 Agents for example,the paper tries to model interactive multi-agent dynamic influence diagrams （I-MADIDs） by means of probabilistic graph model theory,interactive partially observable Markov decision process nature and the principle of maxi- mum reward expectations,and explores how to simply I-MADIDs when there is the stable relationship between non modeling agents. Finally,we model the tiger problem, solve models using HUGIN7. 0, and discuss separately various decision-making cases for the modeling agent and other agents. Examples prove the validity of the approximate method based on Bayesian parameter learning through comparing the exact and approximate methods.