Design and Analysis of Competition-Based Neural Networks

A simple and systematic approach to the design of neural networks for combinato rial op t imization is present ed in this paper. This approach is based on compe t it ion among the neurons of a network. Our approach relies on the use of simple heuristics in network design . The act ivat ion rul e govern ing neuron be havior is derived by breaking down the global constraints for a problem into local constraints for individual neurons. The local const rai nts can be expressed as simple heuristic rules for network design. Using this approach, high performance neural networks for optimiza t ion can be readil y designed and the global properties can be eas ily verified. Simulation results indicate that this approach can outpe rform some sequential heu ristic algorit hms and simulated annealing algorit hms.