基于值等价的交互式动态影响图的求解方法研究与应用

Multiagent system is one of the most important areas in the development of artificial intelligence research, and plays a significant role in many applications. Interactive dynamic influence diagram (I-DID) is well recognized as a general technique for solving multiagent sequential decision making problems under uncertainty. Based on a large amount of I-DID research in the past years, this project aims to complement the I-DID solutions by proposing a new type of solutions that is developed from the principle of value equivalence. The new solutions, which are different from the previous I-DID solutions that are entirely based on behavioral equivalence, can significantly reduce the solution complexity. More importantly, they can provide a theoretical guarantee to the quality of agents’ optimal policies, which will therefore improve the reliability of applying I-DID in a practical setting. This project will develop the solutions by employing new advances of artificial intelligence techniques including sub-modular function optimization, Monte Carlo tree search, active learning, generative adversarial networks and so on. It aims to guarantee correctness of new I-DID solutions in a theoretical way and ensure their effectiveness in practice. The research outcomes will benefit the further development of artificial intelligence research and applications in a new horizon.

多智能体系统是人工智能技术研究的一个重要发展领域，在众多应用领域起到了不可估量的作用。交互式动态影响图是求解在不确定环境下多智能体序贯决策问题的一个普遍适用技术，在多智能体系统研究领域得到了高度的认可。基于前期大量的交互式动态影响图研究工作基础上，本项目继续完善交互式动态影响图的求解方法，提出一个基于值等价的求解体系。有别于传统的基于行为等价准则的交互式动态影响图求解方法，该崭新的求解技术不仅能够极大地降低模型求解的复杂度，而且更为重要的是能够在理论上对智能体的最优决策质量给出一个严格的保证。这将提高交互式动态影响图在实际应用中的可靠性。本项目将采用子模函数优化方法、蒙特卡洛树搜索方法、主动学习方法、生成对抗网络等人工智能最新发展技术，开发一套理论上正确、实际有效的交互式动态影响图求解算法。项目的研究成果将对人工智能技术的进一步发展和可靠应用有一定的借鉴作用。

多智能体系统是人工智能技术研究的一个重要发展领域，在众多应用领域起到了不可估量的作用。交互式动态影响图是求解在不确定环境下多智能体序贯决策问题的一个普遍适用技术，在多智能体系统研究领域得到了高度的认可。基于前期大量的交互式动态影响图研究工作基础上，本项目继续完善交互式动态影响图的求解方法，提出了一个基于值等价的求解体系。有别于传统的基于行为等价准则的交互式动态影响图求解方法，该崭新的求解技术不仅极大地降低模型求解的复杂度，而且能够在理论上对智能体的最优决策质量给出一个严格的保证，提高了交互式动态影响图在实际应用中的可靠性。本项目采用了子模函数优化方法、蒙特卡洛树搜索方法、生成对抗网络等人工智能最新发展技术，开发了一套理论上正确、实际有效的交互式动态影响图求解算法。