毕业生知识库

  本文围绕不完美信息、多智能体复杂对抗条件下，融入知识经验识别对手中高层计划意图的问题，探索了领域知识与数据融合的对手意图识别学习算法，提出了一套基于意图空间和事件图的对手意图识别框架，并以陆军战术兵棋推演为实验环境进行了实验验证。论文总结了描述意图语义的实体行为，构建了意图空间知识描述体系，并基于此意图空间标注、构建了融入知识的计划意图识别研究数据集；进而通过上游融入知识的对抗态势建模表示和下游意图识别，实现了对手计划意图识别模型的学习。 论文主要完成的工作包括：
  （1）构建了计划意图空间、实体行为知识描述体系和融入知识的计划意图识别数据集。本文针对策略层级计划意图识别的需求，基于陆战兵棋推演领域经验知识，定义计划意图为实体行为的遂行任务，构建了包括攻击、隐蔽、侦察、夺控和坚守等 10 种遂行任务在内的计划意图空间。此外，本文还针对计划意图形成过程的可解释问题，构建了从微观单实体动作、局部单实体行为到宏观多实体关联行为的层次化实体行为知识描述体系。基于上述计划意图空间和实体行为知识描述体系，本文通过人类数据标注和数据挖掘相结合的方式构建了首个融入知识的陆战兵棋推演研究数据集。
  （2）提出了一种基于事件图的对抗历史态势建模与表示方法。针对以往时序状态型态势建模方法存在的弱化动作信息、难以建模多智能体行为和难以引入领域知识的问题，提出一种基于事件图的对抗历史态势建模方法。该方法以单实体行为为节点，以实体行为间的顺承、协同和敌对关系为边构建事件图建模对抗演化过程，进而利用图神经网络获得图中各行为节点的深层特征表示。在模型学习过程中，通过将单实体行为语义作为数据标签赋予事件图节点，实现了领域知识与数据学习的融合。
  （3）构建了基于事件图的对手计划意图识别模型。针对下游对手计划意图识别问题，本文设计了端到端模型和两阶段模型两种意图识别模型，实现了上游对抗态势理解与下游意图识别的关联。并通过实验证明了端到端模型在实际应用时具有一定的优越性。除此之外，本文还设计了多种对比实验和消融实验，证明了所述事件图方法的有效性。实验结果表明基于事件图的意图识别模型相较基线模型在准确率指标上最高取得了 6.96%的提升。

  In complex confrontation environments with characteristics of imperfect information and multi-agent, recognizing opponent’s middle and high level planning intention is a difficult problem. It usually requires domain knowledge and experience to be incorporated into the recognition algorithm. This paper was mainly designed to study this problem. In this paper, the intention recognition algorithms integrated knowledge and data were explored. An intention recognition framework based on intention set and event graph was proposed. And all these ideas were verified by experiments in tactical land wargame environment. In this environment, typical behaviors and planning intentions were summarized into a knowledge system. Based on the knowledge system, an intention recognition dataset integrated with domain knowledge was constructed. Furthermore, the intention recognition models were realized through upstream situation modeling and representation, and downstream target entity’s intention recognition. The main work of this paper includes:
  (1) Constructed a strategy-level planning intention set, a hierarchical knowledge system, and a planning intention recognition dataset integrated with knowledge. Aiming to solve the problem of strategy-level planning intention recognition, this paper defined the planning intention as the target mission of entity behavior. A planning intention set was constructed based on the domain knowledge and experience, which includes ten kinds of target missions such as attack, concealment, scout, capture and hold. In addition, aiming to explain how planning intentions came into existence, this paper constructed a hierarchical knowledge system. It covers the micro-level actions of single entity, the local-level behaviors of single entity and the macro-level associated behaviors of multiple entities. Finally, based on the planning intention set and the knowledge system mentioned above, the first dataset integrated with knowledge for land wargame researches was constructed by human annotation and data mining methods.
  (2) Proposed a method of modeling and representing historical confrontation situation based on event graph. Previous situation modeling methods can’t fully describe action information and multi-agent behaviors, and also difficult to be combined with knowledge. Therefore, this paper proposed a new method to model historical confrontation situation by event graph. It takes behaviors of single entity as nodes, and the compliant, cooperative and aggressive relations among behaviors as edges. And then graph neural networks can be used on the event graph to learn deep features for each node. In this process, the semantics of behaviors are assigned to node data as labels, which realized the integration of knowledge and data learning method.
  (3) Constructed planning intention recognition models based on the event graph. This paper proposed an end-to-end model and a two-stage model to realize downstream intention recognition. They can predict the planning intention of the target entity based on the deep features obtained by event graph. And through some experiments, the end-to-end model was shown to be better than the two-stage model. In addition, a variety of comparison experiments and ablation experiments were designed to verify the effectiveness of the proposed method. The experimental results showed that the accuracy of intention recognition model based on event graph is up to 6.96% higher than the baseline model.