A simple genetic algorithm applied to discontinuous regularization

Abstract

A simple genetic algorithm without mutation has been applied to discontinuous regularization. The relative slope of the energy-to-fitness function has been introduced as a measure of the rate of convergence. The intuitively better rate of convergence (slow in the beginning, faster in the end) has been shown to be superior to an exponential transformation-function in the present case. A probabilistic model of the performance of the algorithm has been introduced. From this model it has been found that a division into subpopulations decreases the performance, unless more than one computer is available.<>