CRII: RI: Planning and learning with macro-actions in cooperative multiagent systems

该建议旨在研究分散规划的问题。一般的技术领域是分散的部分可观察马尔可夫决策过程（deco - pomdp）。PI提出了一种基于dec - pomdp有限状态控制器的宏观行为理论。宏操作使计划器能够在单个计算周期中执行多个计划步骤，而使用常规操作的计划器只能在每个计算周期中执行一个操作。因此，宏观行为有可能更有效地解决规划问题，从而实现(1)跨多个代理的分布式规划任务，(2)在代理对世界状态只有有限知识的环境中进行规划，以及(3)在不确定的环境中进行规划，其中操作可能有多种结果。这项研究有许多潜在的应用领域，包括监视网络安全漏洞的分布式代理、分布式军事规划和协调涉及灾难恢复任务的多个机器人。

The proposal aims at studying the problem of decentralized planning. The general technical area is decentralized partially observable Markov decision process (Dec-POMDP). The PI proposes a theory on macro-actions by using finite-state controllers of Dec-POMDPs. Macro-actions enable the planner to perform multiple planning steps in a single computation cycle whereas a planner using regular actions can only perform one action in each computation cycles. As a result, macro-actions have the potential to solve planning problems much more efficiently enabling (1) distributed planning tasks across multiple agents, (2) planning in environments where agents have only limited knowledge about the state of the world, and (3) planning in uncertain environments where actions might have multiple outcomes.There are many potential areas of application of this research including distributed agents monitoring a network for security breaches, distributed military planning, and coordinating multiple robots involved in disaster recovery tasks.