Assumption-based Decentralized HTN Planning

This paper describes our approach to decentralized planning via Hierarchical Task Networks (HTNs), which we call Autonomy and Rationale Coordination Architecture for Decentralized Environments (Arcade). Arcade is a decentralized AI planning framework that can incorporate a number of Shop2 HTN planner instances. Each Shop2 instance may have a different HTN planning domain definition than the others in the framework. Arcade does not assume full communications among the planners. For this reason, Arcade planners must make and manage assumptions about parts of the world state that are not visible to them, including the tasks and plans of other planners. The individual planners also must operate asynchronously, and may receive new tasks, either from outside, or from other planners in Arcade. In this paper, we describe our assumption-based planning approach and how Arcade coordinates multiple, asynchronously interacting HTN planners, using assumptions and task queues. We first present a formal framework, Assumption-based, Decentralized Total-order Simple Task Network (DTSTN) planning, based on Total-order Simple Task Network planning. This is necessary because of our use of Shop2-style task semantics, instead of goal semantics. Then we describe the Arcade framework, and how it implements the framework. Finally, we present preliminary experimental results in a simplified air operations planning domain, which shows that Arcade realizes the expected speed-up when applied to weakly coupled planning problems. We conclude with directions for future work.