Enhancing Genetic Algorithm With Explainable Artificial Intelligence for Last-Mile Routing

Traditional evolutionary optimization algorithms often presuppose straightforward constraints and objective functions. However, in many real-world optimization problems, a clear-cut objective may be absent or hard to evaluate. Given these challenges, surrogate models have gained attention as proxies for evaluating objective functions or constraints. In this research, we leverage Machine Learning (ML) based surrogate modeling and Explainable Artificial Intelligence (XAI) to improve the performance of Genetic Algorithm (GA) for route sequencing problem. Our framework utilizes ML to capture the nuanced interplay within environmental data while using XAI to identify their relative importance, to ultimately uncover tacit knowledge embedded in desired solutions. The goal is to use ML and XAI to guide the GA search by identifying and utilizing significant genes, i.e., genes that produce high-quality solutions in a more efficient manner. The proposed approach is data-driven, focusing on enhancing core GA components, including (1) enhanced chromosome generation and initialization, (2) informed genetic operators, and (3) refined evaluation of fitness function. We demonstrate our framework using real-world data from the Amazon 2021 Last Mile Challenge as a case study. Our methodology extracts drivers’ preferred routing characteristics and applies these insights to generate new routes. Our method is shown to generate high-quality routes when compared to other approaches from the literature, demonstrating improvements in the convergence and effectiveness of GA enhanced by XAI. The code for this publication can be found at https://zenodo.org/records/15227301.