A Preliminary Counterfactual Explanation Method for Genetic Programming-Evolved Rules: A Case Study on Uncertain Capacitated Arc Routing Problem

Shaolin Wang, Yi Mei, Mengjie Zhang

Abstract: In this study, we propose a novel method to enhance the interpretability of Genetic Programming Hyper-Heuristics (GPHH) by employing counterfactual explanations for Genetic Programming (GP) evolved rules in dynamic stochastic combinatorial optimisation problems. Focusing on GP-evolved rules, such as dispatching and routing policies, our research tackles the challenge of their complexity and improves user understanding. We illustrate our methodology using the Uncertain Capacitated Arc Routing Problem (UCARP) as a case study. The approach involves analyzing potential candidates in a scenario to understand why some are not selected. We introduce metrics to evaluate the quality of counterfactual explanations and adapt optimization methods to produce them. Additionally, we present a new attribute importance method based on these explanations. This research contributes to enhancing the transparency of GPHH in UCARP and may provide a reference point for future investigations into evolutionary computation and decision-making systems.