Large deviations, evolutionary computation and comparisons of algorithms

Abstract

This paper summarizes large deviations results for Markov chains with exponential transitions, and discusses their applications in comparing stochastic optimization algorithms. In the second half, two specific algorithms will be focussed on: Mutation/Selection vs parallel simulated annealing. Conditions which tell when an algorithm should be preferred to the other will be given. These conditions combine the parameters of both algorithms and a number of relevant geometric quantities.