On the Problem of Training the Coulomb Energy Network

The Coulomb Energy network offers a unique perspective towards nonlinear transformations. However, its training as it was originally proposed by C. Scofield [1] presented difficulties that prevented its general use. We have investigated this model and we present here the reasons for its shortcomings. Further we propose refinements to the model and its training algorithm, and we present the study and results of various other modifications. We address these problems by constraining its architecture (topology) and present a derivation of the associated training algorithm. We also discuss further refinements of this algorithm. Existing genetic algorithms and simulated annealing are also evaluated as training techniques. Simulation results are also presented.