همایش بین المللی هوش مصنوعی

صفحه اصلی / دومین همایش بین المللی هوش مصنوعی

A novel idea for Abductive Planning on Temporal Knowledge Graphs

نویسندگان :

Amirhossein Sharafi¹ Alireza Shahbazi² Behrouz Minaei Bidgoli³

1- موسسه نجم/دانشگاه تفرش 2- موسسه نجم 3- موسسه نجم/ دانشگاه علم و صنعت ایران

کلمات کلیدی :

AI Planning،Temporal Knowledge Graph،Abductive Reasoning،Dynamic Description Logic،A* Search،Inconsistency Detection،Plan Validation

چکیده :

Reasoning and planning over dynamic knowledge graphs are critical for intelligent systems, but real-world applications are often hindered by incomplete information. While theoretical frameworks exist for integrating temporal and action formalisms, they often fall short in providing robust planning strategies for sparse-knowledge environments. This paper introduces APT (Abductive Planning for Temporal KGs), a comprehensive, practical, and implemented algorithmic framework that addresses this gap. Its architecture is founded on a novel "Generate-and-Test" paradigm. APT provides two primary, independent reasoning services: a Goal-Oriented Planner for generating logically consistent plans, and a Contradiction Hunter that proactively discovers hidden inconsistencies triggered by action sequences. Both services are powered by a shared backend composed of a raw A* search engine, a unified, simulation-based validation engine (temporal projection), and a multi-level Abductive Reasoner. This sophisticated abductive reasoner, operating on a level-by-level search strategy, serves a dual purpose: first, as a fallback mechanism to generate conditional plans by hypothesizing missing knowledge, and second, as a post-processing tool to find minimal, grounded, and consistent explanations for these hypotheses. We present the formal algorithms for this framework and demonstrate its power in handling knowledge gaps and discovering contradictions through detailed case studies.