Research on the Performance of Noisy Chaotic Neural Network Based on Travelling Salesman Problem

Since Hopfield and Tank applied their neural networks to Travelling Salesman Problem (TSP), this classical NP-hard (non-deterministic polynomial) problem has been intensively studied in the field of artificial neurocomputing. Lipo Wang et al. proposed an efficient approach named noisy chaotic neural network (NCNN), which has been proved to be a powerful tool to solve combinatorial optimization problems in their literatures. However, its exact parameters choice is exquisitely sensitive and complicated under different scenarios. In order to improve the convergence performance, the characteristics of its parameters are investigated in detail again in this paper. We focus on further researching the effects parameters have on the performance of NCNN. Through a large quantity of analyses and numerical simulations, we present the modified scheme with a new parameter set which gives 1) less steep sigmoid function, 2) stronger synaptic weights, and 3) higher initial temperature for annealing. Simulation results show that the modified scheme has much faster convergence speed with a small amount of accuracy loss, compared with the original NCNN which used a traditional parameter set.