Hybrid Genetic Search for Dynamic Vehicle Routing with Time Windows

TL;DR

This paper adapts the Hybrid Genetic Search algorithm to solve the dynamic vehicle routing problem with time windows, effectively handling real-time customer data and improving solution quality in an online setting.

Contribution

It introduces modifications to the HGS algorithm components for DVRPTW, enabling better balancing of solution quality and future constraints without prior training.

Findings

Significantly improved solution quality over baseline algorithms

Effective adaptation of HGS components for dynamic routing

Demonstrated on EURO meets NeurIPS 2022 data

Abstract

The dynamic vehicle routing problem with time windows (DVRPTW) is a generalization of the classical VRPTW to an online setting, where customer data arrives in batches and real-time routing solutions are required. In this paper we adapt the Hybrid Genetic Search (HGS) algorithm, a successful heuristic for VRPTW, to the dynamic variant. We discuss the affected components of the HGS algorithm including giant-tour representation, cost computation, initial population, crossover, and local search. Our approach modifies these components for DVRPTW, attempting to balance solution quality and constraints on future customer arrivals. To this end, we devise methods for comparing different-sized solutions, normalizing costs, and accounting for future epochs that do not require any prior training. Despite this limitation, computational results on data from the EURO meets NeurIPS Vehicle Routing Competition 2022 demonstrate significantly improved solution quality over the best-performing baseline algorithm.