Recurrent Neural Networks with the Soft 'Winner Takes All' Principle Applied to the Traveling Salesman Problem

Abstract

This work shows the application of Wang’s Recurrent Neural Network with the Winner Takes All (WTA) principle to solve the classic Operational Research problem called the Traveling Salesman Problem. The upgrade version proposed in this work for the ‘Winner Takes All’ principle is called soft, because the winning neuron is updated with only part of the activation values of the other competing neurons. The problems in the TSPLIB (Traveling Salesman Problem Library Reinelt, 1991) were used to compare the soft version with the ‘Winner Takes All’ hard version and they show improvement in the results using the 'Winner Takes All' soft version in most of the problems tested. The implementation of the technique proposed in this paper uses the parameters of Wang‘s Neural Network for the Assignment problem (Wang, 1992; Hung & Wang, 2003) using the ’Winner Takes All' principle to form Hamiltonian circuits (Siqueira et al. 2007) and can be used for both symmetric and asymmetric Traveling Salesman Problems. The 2-opt technique is used to improve the routes found with the proposed Neural Network, thus becoming a technique that is competitive to other Neural Networks. Other heuristic techniques have been developed recently to solve the Traveling Salesman Problem, and the work of Misevicius et al. (2005) shows the use of the ITS (Iterated Tabu Search) technique with a combination of intensification and diversification of solutions for the TSP. This technique is combined with the 5-opt and errors are almost zero in almost all of the TSPLIB problems tested. The work of Wang et al. (2007) shows the use of Particle Swarm to solve the TSP with the use of the fraction (quantum) principle to better guide the search for solutions. The authors make comparisons with Hill Climbing, Simulated Annealing and Tabu Search, and show in a 14-cities case that the results are better than those of the other techniques. In the area of Artificial Neural Networks, an interesting technique can be found in the work of Massutti & Castro (2009), who use a mechanism to stabilize winning neurons and the centroids of the groups of cities for the growing and pruning the network. The authors show modifications in the Rabnet’s (real-valued antibody network) parameters for the Traveling Salesman Problem and comparisons made with TSPLIB problems solved by other techniques show that the Rabnet has better results.