Neural network compensation of optimization circuit for minimax path problems

A neural network approach is proposed for error compensation of a class of optimization circuit which was previously derived based on the binary relation inference network for minimax path problems. In contrast to the direct calibration method which has been used in an earlier attempt to reduce the error, the neural network based calibration gives a significant improvement in accuracy. As there are many unknown and unmodeled errors in the circuit, we construct three different learning models for error correction. The basic architecture and the assumption of each model are described. A feedforward neural network (multilayer perceptron) with different learning algorithms and a radial basis function network have been investigated. Experimental results on a simple three nodes network show that significant reduction of error is possible. The comparative advantages of each model are presented