Abstract
Recently, combinatorial optimization problems have aroused a great deal of interest in Machine Learning, leading to interesting advances in Neural Combinatorial Optimization (NCO): the study of data-driven solvers for NP-Hard problems based on neural networks. This paper studies the benefit of Transfer Learning for NCO by evaluating how model training can be improved taking advantage of knowledge learned while solving similar tasks. We focus, in particular, on two famous routing problems: the Traveling Salesman Problem (TSP) and the Vehicle Routing Problem (VRP). The latter being a generalization of the former, we study the effect of applying Transfer Learning from a model trained to solve TSP while training a model learning to solve the Capacitated VRP (CVRP). We present adaptations of a state-of-the-art NCO model for implementing Transfer Learning. Our results based on extensive empirical experiments in different settings show that Transfer Learning may help to speed up the training process while being more sample efficient.