Determination of parameters in a Hopfield/Tank computational network

Neural-like networks which minimize a global energy function have been proposed for solving computationally intensive optimization problems. These networks have several parameters that need to be selected and often carefully tuned for a network to produce a sensible computation. The authors examine the traveling salesperson problem (TSP) as a representative NP-complete optimization problem and present a cookbook approach to setting these parameters. There appears to be a linear relationship between two of the parameters. This relationship and the problem size lead to a simple understanding of why these networks are less and less useful for the TSP computation as the number of cities increases.<>