Perceptron algorithms for the classification of non-separable populations
Robert Main Burton, Dehling Herold G., Venema Rienk S. · Stochastic Models · 1997
In this paper we study the behavior of the perceptron algorithm when the underlying populations cannot be linearly separated. Assuming that the inputs form an i.i.d. sequence, we can prove that the weights do not converge but that their distributions approach a steady state. We propose a modified algorithm, which appears to stabilize near an optimal weight setting