多智能体对手建模及其真实模型的确定
Multi-agent opponent modeling and true model identification
Abstract
针对如何在竞争环境中更好地预测对手行为,并正确制定自身相应的对策进行研究.利用交互式动态影响图对环境中的对手智能体进行建模,并结合贝叶斯网络提出了一种判断对手真实模型的方法.首先,将对手智能体的候选模型保存在模型节点中并实时推理和更新对手的模型、信度和动作;然后,在每次交互中记录下观察到的对手的动作序列,以此作为训练动态贝叶斯网络的集合,得到网络参数后重新计算候选模型的权重,从而判断出对手的真实模型;最后,通过多智能体老虎问题和无人机侦查问题进行实验,并从对手候选模型的权重和我方智能体的收益值两方面验证了算法的有效性. How to better predict the behavior of the opponent in a competitive environment,in order to make one′s own corresponding strategy correctly was studied.The opponent agents were modeled using interactive dynamic influence diagram in the environment and a method was proposed to identify the true model of the opponent based on Bayesian network.First,the candidate models of the opponent were set in the model node,and models,beliefs and actions of the opponent were inferred and updated real-timely.Then in every interaction,the observed action sequences of the opponent were recorded as the training set of dynamic Bayesian network.The weights of the candidate models were recalculated using the parameters of the network so as to identify the true model of the opponent.Experiments on multi-agent tiger problem and unmanned aerial vehicle reconnaissance problems verify the effectiveness of this method from two aspects which are weights of the opponent′s candidate models and profits of our own agent.