Machine and bit precision influence on the Hopfield-Tank model for the TSP

The influence of different bit precision and number representations with different sigmoid factors on the traveling salesman problem (TSP) modeled by the Hopfield-Tank neural network is presented. In order to simulate this model, a number of previous studies are used to determine some required parameters and a set of 10-city problems is generated by random number generator or obtained from existing research paper. To investigate the influence of the number representation and different ways of performing machine operations for the same precision, we simulate the TSP problem in two different architectures, namely, the IBM S/370 and the MIPS R3000. We have considered five different bit precision (8-16- 24- 32- and double precision mantissas), three sigmoid generation functions and different setting of the model parameters. The results of 14,160 simulations, convergence, average distance, number of cycles, and different precision performance of the network, are discussed.