The capacitated vehicle routing problem is an NP-complete combinatorial problem. In addition to its usefulness for delivery services, many other problems can be efficiently mapped to it. To solve the problem, I use the GRU recurrent neural network with the attention mechanism. After a learning phase that lasts as long as using stochastic optimization algorithms on thousands of cases, we get comparatively good results on small graphs. On larger graphs the neural models do not learn fast enough and produce worse results due to an increasing problem complexity. Among the tested graph embedding methods, node2vec and GraRep give the best results.
|
