Rule extraction from a multilayer perceptron with staircase activation functions

We tackle the problem of rule extraction from multilayer perceptrons. Our approach consists of characterising discriminant hyper-plane frontiers built by a special neural network model, denoted as a discretized interpretable multilayer perceptron (DIMLP). Rules are extracted in polynomial time with respect to the size of the problem. Further, the degree of matching between extracted rules and neural network responses is 100%. We apply DIMLP to five data sets of the public domain in which for some of them it gives better average predictive accuracy than standard multilayer perceptrons and C4.5 decision trees.