复杂环境下智能体通用慎思模型研究

Agent-based systems are a promising paradigm for building software that exhibits a high degree of autonomy and operates in dynamic, complex domains. Agents have been used to develop applications in a wide variety of domains, including military, transportation systems and health care. Arguably the dominant paradigm in agent development is the Belief-Desire-Intention (BDI) model, in which the behaviour of an agent is specified in terms of beliefs, goals, and plans. In state of the art BDI programming languages, there is no general approach that can avoid conflicts and exploit synergies between intentions in all problem domains at the level of individual actions. There are several approaches proposed in the literature to reason about interactions between intentions. However, they all have certain limitations, and most of their approaches are done at plan level. In this project, we propose to design and implement a new BDI programming language which is capable of scheduling agents’ intentions at action level, as well as to deal with rich element types including achievement goal, maintenance goal，repeat goal, deterministic and non-deterministic actions and finally the beliefs with uncertainty. The detailed research problem includes: 1) Redesign agents’ deliberation cycle to provide the capability for BDI agent to schedule a rich type of goals at the level of individual actions. 2) Provide a general approach for BDI agent to schedule agents’ intentions at action level such that the agent is able to avoid conflicts and to exploit synergies at the same time. 3) Design and implement a new BDI programming language based on the proposed deliberation cycle and deliberation strategy. The outcome of this project not only improves the ability of BDI agents for achieving goals. It also has an important significance to reduce programmer’s effort and secure a future for agent programming as a discipline.

随着分布式人工智能技术不断发展，多智能体系统已被广泛应用于军事、商业、医疗服务等诸多领域。BDI模型通过智能体的思维状态对其行为进行解释分析，是研究和实现多智能体系统时常用的智能体体系结构。然而，当前BDI模型存在着目标种类单一、无法处理不确定性信息、缺乏有效的多目标并行实现能力等问题，严重阻碍了其在主流人工智能领域的应用与发展。针对以上现状，本项目将对BDI智能体的慎思模型展开研究，通过重写智能体的慎思过程框架，并且制定智能体在多目标实现过程中采用的意图调度策略，最终建立支持智能体在复杂环境中多目标并行实现的通用BDI慎思模型并实现基于该模型的BDI语言。本课题的预期研究成果不仅能够完善BDI语言的理论模型，提高智能体慎思过程的普适性、鲁棒性与可拓展性，而且初步实现了通用智能体，重新定义了BDI语言的编程模式，具有重要的理论意义和工程应用价值。

多智能体系统是实现分布式人工智能的主要方法之一，其相关研究已被广泛应用于军事、商业、生产制造等多个重要领域。BDI模型通过智能体的心智状态对其行为进行解释和建模，是当前智能体领域研究最多的模型之一。然而，当前BDI模型存在着目标种类单一、无法处理不确定性信息以及缺乏有效多目标并行实现能力等问题。针对以上现状，本项目针对BDI模型中的慎思模型展开研究，采用基于随机模拟的蒙特卡洛树搜索、强化学习以及线性时序逻辑等方法，建立了支持BDI智能体在复杂多目标环境中使用的通用慎思模型。该模型基于目标计划树结构，可以应用于所有当前常见的BDI智能体语言。相关实验结果已经发表于顶级人工智能会议与期刊，实验结果表明相比于传统的BDI慎思模型，本项目的研究成果能够大幅提高智能体实现目标的效率，且该模型适用于动态、不确定性环境，并且能够应用于多智能体场景。对比计划书中的研究内容，本项目进展顺利，成功实现了通用的智能体慎思模型。该研究不仅促进了通用智能体的研究，还重新定义了BDI语言的编程范式，对BDI模型的推广和应用有着重要的意义。