Authors: Brandon Rozek, Selmer Bringsjord,

Conference: International Conference on Agents and Artificial Intelligence

Publication Date: 2026/03/05

Abstract: Rationally inferring the goal of an agent from observations of their actions is challenging. In the goal-recognition-as-planning literature, it is often assumed that the initial state of the environment is known. However, actors and observers do not always operate with complete knowledge. Instead, agents may be working with little information regarding the uncertainty of the environment. In this light, we revisit goal recognition in the context of qualitative possibilistic planning (QPP). Agents in this setting do not know the exact probability of an event occurring but are able to determine whether one event is more likely than another. More specifically, agents describe the uncertainty regarding the initial state and the outcome of actions qualitatively. We show that for rational actors, the observer should not filter goals solely based on necessity thresholds and instead propose a technique that takes into account whether the actor followed a necessity-optimal plan. Using our novel compilation CQPR, we find those necessity-optimal plans that additionally satisfy the observed action sequence by casting the overall problem as a QPP problem. Our formal results and experiments show that this approach is sound and may narrow down the potential goals that a necessity-optimal agent is pursuing.

PDF Link: https://www.scitepress.org/Papers/2026/143479/143479.pdf