AI Planning in Dynamic, Uncertain Domains

Extending Theories of Action: Formal Theory and Practical Applications

Volume

Papers from the 1995 AAAI Spring Symposium

Contents

A major reason for the success of the STRIPS planner and its derivatives was the use of a representation that succinctly captured the changes in the world due to different actions. As planning systems are extended to deal with uncertainty, we seek ways to model the world that maintain this advantage. In this paper I examine four somewhat representative planning systems designed for uncertain domains that can change independently of the actions performed by the planning agent. These are Buridan, anytime synthetic projection, Weaver and XFRM. I compare their positions on several design issues including the power of the languages used to represent plans and actions, and the use to which it is put in both plan synthesis and plan evaluation. I briefly explore the tradeoff that exists between the expressiveness of the language used to reason about the effects of action in the world and the extent to which the planning system can take advantage of the language to build plans efficiently.