Selecting the next action with constraints

Abstract

Traditional AI planning systems have focussed on batch planning, where an entire plan for achieving a goal is generated. An alternative approach is to select only the next action, a technique that has been used in situated planners, and, more recently, has been effectively applied to traditional AI planning domains. In this paper, we present an action selection framework sensitive to resource limits and based on constraint optimization. While the framework we present is very general, we are concerned with dynamic, time-pressured domains requiring reasoning under uncertainty. In such domains, batch planning is usually inappropriate or impossible to apply. We experimentally compare a number of local search algorithms, and give a detailed example of how action selection can be used to control the dialog of a course advising system, which allows for more flexible behaviour than in typical advice-giving systems.