Direct Encoding Evolutionary Learning Algorithm for Multilayer Morphological Perceptron

This paper presents a method based on evolutionary computation to train multilayer morphological perceptron (MLMP). The algorithm calculates network parameters such as its connection weights, pre-synaptic and postsynaptic values for a given network topology. Morphological perceptron are a new type of feed-forward artificial neural network based on lattice algebra which can be used for pattern classification. The representation scheme is based on a tree data structure which the algorithm to perform operations such as crossover by replacing or switching whole nodes between parent. Adaptive mutation is used as the genetic algorithm approaches convergence to fine tune network parameters final values. The algorithm uses a special fitness function based on the mean square error of the pattern classification and introduces a penalty function to reduce the number of redundant neurons in the resulting neural network.