Multi-objective optimization of empty train allocation using a Fast Surrogate-Assist Evolutionary Algorithm based on incremental learning

Du, Zhigang

doi:10.1016/j.eswa.2025.127505

Return

Multi-objective optimization of empty train allocation using a Fast Surrogate-Assist Evolutionary Algorithm based on incremental learning

2025-06-01

PRE

Du, Zhigang

Pan, Jeng-Shyang

Chu, Shu-Chuan

DOI：10.1016/j.eswa.2025.127505

Original

Original request for help

Save

Abstract

En 中文

This research addresses the critical issue of empty train shortages, which significantly impact the efficiency of train dispatching within the railway transportation system. The objective is to develop an optimization model that minimizes both the economic costs associated with dispatching empty trains and enhances their punctuality. To this end, the study introduces a novel Fast Surrogate-Assisted Evolutionary Algorithm utilizing Incremental Learning to optimize empty train allocation. Traditional Surrogate-Assisted (SA) models suffer from high time complexity due to repeated training cycles; however, this research employs incremental learning techniques to alleviate these constraints, thereby streamlining the computational process. Additionally, the implementation of a regularized least squares radial basis function neural network within the SA framework enhances the model's generalization capability, making it suitable for complex multi-objective optimization scenarios. To further improve computational efficiency, the SA model substitutes the resource-intensive fitness function evaluations, reducing the overall number of evaluations needed. A strategically designed sample infilling criterion based on expected improvement selects samples that offer the highest informational gain for actual evaluation, thereby optimizing algorithmic efficiency. Throughout the optimization process, a Particle Swarm Optimization algorithm refines the SA model, with a non-dominated solution set sorting method employed to preserve population diversity and prevent convergence to local optima. This approach promises significant advancements in the scheduling and allocation of empty trains, potentially transforming operational practices within the railway sector.

Keywords:

Incremental learning

Regularized least squares

Fast Surrogate-Assist

Uncertain infilling criteria

Empty train allocation