Enhancing Abstract Argumentation Solvers with Machine Learning-Guided Heuristics: A Feasibility Study

Abstract

AbstractAbstract argumentation frameworks model arguments and their relationships as directed graphs, often with the goal of identifying sets of arguments capable of defending themselves against external attacks. The determination of such admissible sets, depending on specific semantics, is known to be an NP-hard problem. Recent research has demonstrated the efficacy of machine learning methods in approximating solutions compared to exact methods. In this study, we leverage machine learning to enhance the performance of an exact solver for credulous reasoning under admissibility in abstract argumentation.More precisely, we first apply a random forest to predict acceptability, and subsequently use those predictions to form a heuristic that guides a search-based solver. Additionally, we propose a strategy for handling varying prediction qualities. Our approach significantly reduces both the number of backtracking steps and the overall runtime, compared to standard existing heuristics for search-based solvers, while still providing a correct solution.