Decision-theoretic transformational planning

Abstract

In this paper we develop decision-theoretic transformational planning as a novel computational theory for planning reactive behavior under hard time constraints. The theory is based on three main paradigms: transformational planning, decision theory, and time-dependent computations. Knowledge about goals and the robot control language is accessed through transformation rules that define semantic relationships between constructs in the plan representation language and associations between goals and canned plans. The computational theory deals with uncertainty by applying decision-theoretic methods to control the planning process. The tradeoffs between planning and acting are weighed by applying time-dependent algorithms for testing the applicability and utility of transformation rules with respect to the current situation and the preferences of the robot.