Authors: Brandon Rozek, Selmer Bringsjord,

Conference: International Workshop on Human-Aware and Explainable Planning (HAXP)

Publication Date: 2024/06/02

Abstract: Techniques in automated planning under uncertainty capture whether an agent believes that a ground atomic formula is true, false, or uncertain; and, in some cases, the exact probability that it’s true at a given state. Sometimes, however, an agent does not have access to exact probabilistic information, but is instead able to judge the uncertainty qualitatively. We take initial but substantial steps towards characterizing a variant of conformant planning based on qualitative uncertainty. Our framework, QU-STRIPS, introduces levels of belief about ground atomic formulae which stratify uncertainty ranging on the negative side from certainly not, to agnostic, and then on the positive side up to certainly. In order to efficiently find plans, we present a sound compilation into classical STRIPS. We provide preliminary results on a new escape domain and show that state-of-the-art planners can effectively find plans that achieve the goal at a high positive belief level, while considering the trade-off between the strength of a plan and its cost.

PDF Link: https://openreview.net/pdf?id=soH9BIp0pL