基于数据驱动的多智能体交互式动态影响图算法研究与应用

Individual decision making model provides a general framework for solving sequential multiagent decision making problems. By modeling other agents’ decision making process, a subject agent optimizes its decisions through solving the model. Since the true model of other agents is often unknown to the subject agent, most of the current research assumes a large number of candidate modes of other agents that can be developed by a domain expert. This leads to difficulty in solving complex decision models. This project aims to develop a data-driven multiagent decision model by exploiting available data of agents’ interactions. Instead of manually building models of other agents, this project learns policies of other agents by adapting probabilistic automata inference methods and extends the well-known decision model, namely Interactive Dynamic Influence Diagram (I-DID). It is the first time that this project focuses on online solutions to the extended I-DID models and proposes active learning based techniques to identify the true behavior of other agents. In addition, this project will investigate practical applications of the proposed solutions in multiple experimental platforms. In summary, this project will significantly improve the sequential multiagent decision making techniques and provide an example of integrating machine learning techniques into multiagent decision making research. The research outcomes can also be generalized to other multiagent decision models and inspire new solutions to complex sequential multiagent decision making problems.

从个体决策的角度研究多智能体序贯决策问题是一种普遍适用的方法。通过建模其他智能体的决策过程，主体智能体优化本身的决策。由于主体智能体不知道其他智能体的真实模型，研究主要依赖于其他智能体决策过程模型的建立，并假设存在数目众多的其他智能体候选模型。这造成了繁杂的相互建模过程，导致模型难于求解。本项目利用大量存在的多智能体交互数据，借助概率机器模型推理技术自动学习其他智能体决策行为，建立基于数据驱动的多智能体序贯决策模型。通过拓展交互式动态影响图（I-DID），本项目首次提出模型在线求解技术，采用主动学习策略迅速而有效地确定其他智能体的真实决策行为，并建立多个仿真试验平台以评估研究的实际应用价值。本项目的研究将全面提高多智能体序贯决策模型的求解技术，是将机器学习技术无缝嵌入到多智能体决策研究的典范。该研究成果也可以被广泛地应用到其他多智能体决策模型，为解决复杂多智能体决策问题提供新的思路。

从个体决策的角度研究多智能体序贯决策问题时，主体智能体通常通过建模其他智能体的决策过程优化本身的决策。由于主体智能体不知道其他智能体的真实模型， 研究主要依赖于其他智能体决策过程模型的建立并，假设存在数目众多的其他智能体候选模型。这造成了繁杂的相互建模过程，导致模型难于求解。本项目利用大量存在的多智能体交互数据，借助概率机器模型推理技术自动学习其他智能体决策行为，建立基于数据驱动的多智能体 序贯决策模型。通过拓展交互式动态影响图，本项目提出模型在线求解技术， 采用主动学习策略迅速而有效地确定其他智能体的真实决策行为，并建立多个仿真试验平台以评估研究的实际应用价值。