不确定环境下基于HTN的应急任务规划方法研究

Emergency response task planning is essential in emergency response decision making process. This project will investigate the HTN based emergency response task planning method, taking into account typical nondeterminism in emergency response process. Firstly, aiming at the incompleteness of emergency response domain knowledge, the automated learning method will be proposed for HTN planning domain according to the partial plans in emergency response cases. Secondly, a hybrid reinforcement learning and HTN planning method will be proposed for emergency response task planning problem with nondeterministic action time duration and resource consumption quantity. The method combines the capabilities of reinforcement learning on handling nondeterminism and policy optimization with that of HTN planning on hierarchical decomposition and domain knowledge representation. Resource constraints network is designed, which is checked by the planning algorithm for the plan consistency to avoid resource conflicts among tasks. The searching space can also be pruned in time by resource constraints checking to improve the planning efficiency. The evaluation criterion, which is related with plan time duration, resource consumption, or other preferred performance of the decision maker, is adopted in the planning process to evaluate the task decomposition methods or operators. The evaluation process will eventually improve the quality of the obtained action plan. Thirdly, the method of action plan adjustment and replanning will be proposed to deal with the newly coming emergency response tasks and resource anormaly observed during the action plan execution process. This project is aiming at extending the HTN planning capability and improving the applicability of HTN planning in emergency response domain.

应急任务规划是应急响应决策过程的核心问题。本课题针对应急领域的典型不确定性，研究基于分层任务网络（HTN）的应急任务规划方法。首先，针对应急领域知识不完备的特点，研究以案例处置片段为依据、面向HTN的规划领域自动学习方法。第二，研究结合强化学习的HTN规划方法，将强化学习处理不确定性和策略优化的能力与HTN规划的分层分解思想和描述领域知识的能力相结合，解决行动执行时间和资源消耗不确定的应急任务规划问题。构建资源约束网络，在规划过程中检验方案一致性，避免任务间的资源冲突，并对搜索空间进行修剪，提高规划效率。在规划过程中采用方案执行时间、资源消耗或其他决策者偏好的评价指标，对任务分解方法或操作进行评价，从而提高行动方案的质量。第三，针对应急行动方案执行过程中新应急任务的产生和资源执行异常，研究方案调整和重规划方法。通过本课题的研究，可以扩展HTN的规划能力，提高HTN规划在应急领域的实用性。

应急任务规划是应急响应过程的核心问题。课题围绕基于分层任务网络（Hierarchical Task Network，HTN）的应急任务规划开展了一系列相关研究。第一，课题针对应急情景下任务间的复杂的时态、资源和协同关系，研究了面向HTN规划的规划领域表达，并提出了一系列相应的HTN规划算法，提高了HTN规划方法处理复杂约束的能力。第二，课题分析了应急任务规划中存在的典型不确定性，综述了考虑不确定性的HTN规划方法，并进一步提出了一系列针对应急情景不确定性的HTN规划算法，包括领域知识不完备、时态或资源不确定等，不仅提高了HTN规划处理不确定性的能力，而且还提高了HTN规划算法解决应急任务规划实际问题的实用性。第三，课题将理论研究成果应用于分析解决应急相关领域的实际问题，以北京市海淀区的突发事件应对为背景，设计并实现了应急资源调运任务规划系统。另外，课题在应用领域上进行了一定程度地拓展，以智能规划技术为核心，设计并实现了面向高层应急指挥人员的应急演练仿真平台，在平台中运用推理技术解决应急情景推演问题，并运用HTN规划技术解决应急决策问题。考虑到军事领域与应急领域在任务约束关系复杂和环境高度动态不确定方面具有高度的相似性，课题将HTN规划技术延伸应用于解决航空母舰甲板舰载机任务规划问题，设计并实现了舰载机任务规划系统，并在实验室环境取得了良好的应用效果。课题组按照申报书的内容认真执行了研究计划，实现了预期的研究目标，超额完成了预期研究成果指标。课题经费使用严格遵守财务制度及预算，无违规现象。