From the TSP to the Dynamic VRP: An Application of Neural Networks in Population Based Metaheuristic

Abstract

In this paper, we consider the standard dynamic and stochastic vehicle routing problem (dynamic VRP) where new requests are received over time and must be incorporated into an evolving schedule in real time.We identify the key features which make the dynamic problem different from the static problem. The approach presented to address the problem is a hybrid method which manipulates the self-organizing map (SOM) neural network similarly as a local search into a population based memetic algorithm, it is called memetic SOM. The approach illustrates how the concept of intermediate structure provided by the original SOM algorithm can naturally operate in a dynamic and real-time setting of vehicle routing. A set of operators derived from the SOM algorithm structure are customized in order to perform massive and distributed insertions of transport demands located in the plane. The goal is to simultaneously minimize the route lengths and the customer waiting time. The experiments show that the approach outperforms the operations research heuristics that were already applied to the Kilby et al. benchmark of 22 problems with up to 385 customers, which is one of the very few benchmark sets commonly shared on this dynamic problem. Our approach appears to be roughly 100 times faster than the ant colony algorithm MACS-VRPTW, and at least 10 times faster than a genetic algorithm also applied to the dynamic VRP, for a better solution quality.