Max-min propagation nets: learning by delta rule for the Chebyshev norm

This paper introduces max-min propagation nets, which are composed of multiple-input maximum (max) and minimum (min) operators, besides linear weighted sum units. A learning algorithm based on gradient descent is derived for these networks. Two different criteria of error measurement are tested: the Chebyshev norm and the least squares norm. Weight-update rules for both criteria are deduced and implemented together with acceleration techniques to speed up convergence. A comparison of results on the parity problem is presented.