Solving a Combinatorial Optimization Problem with Feedforward Neural Networks

A new approach is proposed to solve a typical combinatorial optimization problem using artificial neural networks. Unlike the popular idea of using a Hopfield network for optimization, we design a new network architecture which consists of two parts: a feedforward network for optimization, and a feedback network for meeting the constraints. Radial-based functions are adopted in the feedforward network in order to utilize its spatial locality and facilitate selection of the numbers of hidden layers and nodes. The convergence of the proposed scheme is proved and a vector-form training algorithm is developed.