Learning to integrate reactivity and deliberation in uncertain planning and scheduling problems

摘要

This paper describes an approach to planning and scheduling in uncertain domains. In this approach, a system divides a task on a goal by goal basis into reactive and deliberative components. Initially, a task is handled entirely reactively. When failures occur, the system changes the reactive/deliverative goal division by moving goals into the deliberative component. Because our approach attempts to minimize the number of deliberative goals, we call our approach Minimal Deliberation (MD). Because MD allows goals to be treated reactively, it gains some of the advantages of reactive systems: computational efficiency, the ability to deal with noise and non-deterministic effects, and the ability to take advantage of unforseen opportunities. However, because MD can fall back upon deliberation, it can also provide some of the guarantees of classical planning, such as the ability to deal with complex goal interactions. This paper describes the Minimal Deliberation approach to integrating reactivity and deliberation and describe an ongoing application of the approach to an uncertain planning and scheduling domain.