Abstract reasoning for multiagent coordination and planning

As autonomous software and robotic systems (or agents) grow in complexity, they will increasingly need to communicate and coordinate with each other. These agents will need planned courses of action to achieve their goals while sharing limited resources. This dissertation addresses the problem of efficiently interleaving planning and coordination for multiple agents. As part of my approach, I represent agents as having hierarchies of tasks that can be decomposed into executable primitive actions. Using task hierarchies, an agent can reason efficiently about its own goals and tasks (and those of others) at multiple levels of abstraction. By exploiting hierarchy, these agents can make planning and coordination decisions while avoiding complex computation involving unnecessary details of their tasks. To reason at abstract levels, agents must be aware of the constraints an abstract task embodies in its potential decompositions. Thus, I provide algorithms that summarize these constraints (represented as propositional state conditions and metric resource us ages) for each abstract task in an agent's library of hierarchical plans. This summary information can be derived offline for a domain of problems and used for any instance of tasks (or plans) assigned to the agents during coordination and planning. I also provide algorithms for reasoning about the concurrent interactions of abstract tasks, for identifying conflicts, and for resolving them. I use these algorithms to build sound and complete refinement-based coordination and planning algorithms. I also integrate summary information with a local search planner/scheduler, showing how the benefits can be extended to different classes of planning algorithms. Complexity analyses and experiments show where abstract reasoning using summary information can reduce computation and communication exponentially along a number of dimensions for coordination, planning, and scheduling in finding a single agent's plan or in optimally coordinating the plans of multiple agents. In addition, I provide pruning techniques and heuristics for decomposition that can further dramatically reduce computation. Overall, the techniques developed in this thesis enable researchers and system designers to scale the capabilities of interleaved coordination, planning, and execution by providing agents with tools to reason efficiently about their plans at multiple levels of abstraction.