基于模因计算的多智能体迁移强化学习研究

The research on multi-agent systems (MAS) is currently one of the most popular topics in the field of artificial intelligence over the world. MAS have a wide array of applications in various important fields including military, economics, industrial engineering, etc. In conventional MAS, the behaviors of agents are usually pre-defined based on human expertise for the states that would be encountered in the environment. However, this approach becomes unreliable when the environment becomes complex and changes over time since the space of environmental states is enormous and hence manually defining these action-state mappings is infeasible. Thus, the interest of this project is placed on investigating the online learning strategy of multiple agents in a dynamic environment. Different from the existing methods, this project first investigates the expression of memetics in multi-agent systems based on the theory of memetic computing. Subsequently, we will derive the corresponding memetic evolution paradigms as well as the efficient knowledge transfer schemes for speeding up the learning process of MAS. In this manner, the real-time performance and effectiveness of agents’ behavior strategy on solving multiple complex problems in the dynamic environment would be guaranteed. This project is of great value in real-world applications for intelligent control and fast learning in MAS. The idea in this project provides a new choice for solving multi-agent online learning problems.

多智能体系统研究是当前国际人工智能领域的研究热点之一，在军事、经济、工业生产等重大领域均具有广泛的应用背景。传统的多智能体系统通常利用专家知识预先设计智能体的行为策略。然而，当处于动态、复杂的环境空间时，这种决策系统往往缺乏足够的完备性和灵活性。因此，本课题围绕多智能体在动态环境中的在线学习策略展开相关研究。不同于传统的多智能体学习方法，本课题将基于智能计算中的模因计算理论深入探讨模因在多智能体系统中的通用知识型表达，建立和推导其相应的知识演化基本模型，并利用多智能体之间有效的模因迁移策略加速个体及整体系统的学习效率，以满足多智能体系统在动态环境中行为策略的实时性和有效性，提高多智能体在解决多种复杂问题时的性能。本课题对于实现多智能体系统的智能控制及其快速学习具有较高的应用价值，理论研究上属于国际源头创新，为解决多智能体在线学习问题提供了一种全新的选择。

传统的多智能体系统通常利用专家知识预先设计智能体的行为策略，当处于动态、复杂的环境空间时，这种决策系统往往缺乏足够的完备性和灵活性。因此，本课题围绕多智能体在动态环境中的在线学习策略展开相关研究，通过提炼多智能体之间有用的知识信息，构建高效的在线知识学习及迁移模型，以加速多智能体学习方法，提高其在解决复杂问题时的性能。不同于传统的多智能体在线学习算法，本项目基于演化计算中的模因计算理论，结合现有多智能体学习以及智能优化方法，构建了多智能体模因知识型的通用定义、模因知识模块挖掘以及通用模因型知识迁移方法，并建立和推导了其相应的知识演化模型，通过利用智能体之间有效的模因迁移策略加速个体及整体系统的学习效率。.本项目主要研究工作包括：.（1）针对复杂对抗多智能体学习环境，研究了基于对手行为模型预测的多智能体迁移学习方法，集成了类人社交选择机制，通过迁移有益模因知识提高系统整体学习效率并降低行为预测的复杂性。.（2）研究多智能体复杂在线学习环境中群组的有效表征以及协同与对抗行为之间的关系，建立了群体协作关系的拓扑优化模型，对智能体行为与结果之间的关系进行推理，实现智能体在线学习行为的决策优化。.（3）提出了基于细粒度分类以及注意力机制的分层表征学习与类人认知理解模型，将人类认知行为的信息处理机制融入智能体知识生成与学习模型，提升复杂环境的智能体认知理解。.本项目的完成可以促进多智能体在线学习算法在实际多智能体系统应用中的推广，提高了多智能体系统在解决复杂问题时的实时性和精确性。理论研究可推广到分步式计算、迁移学习、智能计算等学术领域，具有重要的实际价值。相关研究成果有助于在多智能体智能控制、多智能体快速在线学习等研究方向提出具备自主知识产权的关键技术提案。.