基于知识工程的大规模多智能体系统自主规划方法的研究

As a key challenge in the design of the multi-agent systems, distributed automatic planning is required with the capabilities of high real time control, scalability as well as handling complex tasks. However, when the system scales up, the uncertainties of the system as well as its environments increase. When those uncertainties cannot be completely observable to agents, traditional static planning approaches are not suitable to the large dynamic multi-agent systems. On the other hand, solving dynamically decentralized partial observable multi-agent decision models are classified as NEXP-Complete. To address those challenges, in this project, we have proposed a novel hybrid decentralized automatic planning approach based on knowledge engineering, which is suitable for coordinating large-scale heterogeneous multi-agent systems. This approach is built based on Decentralized Partially Observable Markov Decision Process (DEC-POMDP), which is proven to model the uncertainties of the environment as well as the agents' behaviors well. The team-oriented plan schema based on ontology is used to significantly reduce the planning complexity and avoid the tedious low-level coordination controls between agents. In addition, we will design efficient heuristic algorithms for each decentralized agents. By making use of the proactive information sharing and states abstraction, decentralized agents can reduce their computation complexity and independently search for the optimal policies according to their own local POMDPs. As another key contributions, the agents' communication network which is used for agents' local interactions is taking into account for evaluating current tasks to produce agents' local joint optimized coordination so that the planning time as well as the algorithm complexity can be greatly reduced to improve the efficiency of automatic planning and system robustness.

分布式自主规划是多智能体系统设计的核心难点，它要求具有高实时性，可扩展性和复杂任务的处理能力，然而由于系统的扩大引入系统和环境大量的不确定性因素，当这些因素不可被智能体完全所观测时，原有的静态规划方法将不能再被大规模系统所采用；而动态非集中式方案是Nexp-Complete的。我们针对这一问题引入基于知识工程的分布式智能方法将设计一种创新的混合式团队导向型非集中式自主规划方法，该方案建立在分布式部分可观测马尔可夫决策模型（DEC-POMDP）基础上能很好建模环境和行为的不确定性。我们通过制定基于本体的团队导向方案大大简化计划复杂度并避免冗余的底层协调控制。通过设计高效的启发式搜索算法，智能体通过信息共享和状态抽取降低算法复杂度并独立寻找各自POMDP下的最优行为，多智能体通过网络局部交互在评价当前任务的前提下搜寻局部最优联合机制从而减少自主规划的时间和算法复杂度并达到规划的高效和鲁棒性。

本项目针对多智能体系统设计中构建高实时性、可扩展性和面向复杂任务的分布式自主规划问题，通过引入基于知识工程的多智能体协同语义表示方法，制定基于本体的团队导向方案并通过知识图搜索、智能信息共享等多项技术构建面向不同应用领域的异构型多智能体不确定性规划算法，研究结果表明这一方案通过引入人类经验知识能大大简化原有多智能体规划计算复杂度并避免冗余的底层协调控制。我们同时依据物联网、无人系统协同等多个应用领域知识本体描述，构建面向大规模智能体协同规划的验证系统，验证结果表明智能体可以通过独立寻找局部决策模型下的最优行为，在网络局部交互前提下搜寻局部最优联合机制以减少自主规划的时间和算法复杂度，并能保持原有集中式规划的高效和鲁棒性等优势。依托本项目的研究，项目组形成高水平研究成果40项，其中SCI检索论文15篇，EI或会议论文13篇，申请或获得国家发明专利10项，并获得国际会议最优论文1次。依托本项目研究成果，项目负责人获得多项国家级项目支持，包括国家科技重大专项项目1项，装备预研项目5项，对我国无人集群技术发展具有很强的促进作用。人才培养方面，依托本项目培养博士生5人，硕士生10余人，项目负责人同时获得美国匹兹堡大学信息科学青年校友奖。