Self-sampling strategies for multimemetic algorithms in unstable computational environments

Abstract

Optimization algorithms deployed on unstable computational environments must be resilient to the volatility of computing nodes. Different fault-tolerance mechanisms have been proposed for this purpose. We focus on the use of dynamic population sizes in the context of island-based multimemetic algorithms, namely memetic algorithms which explicitly represent and evolve memes alongside solutions. These strategies require the eventual creation of new solutions in order to enlarge island populations, aiming to compensate the loss of information taking place when neighboring computing nodes go down. We study the influence that the mechanism used to create these new individuals has on the performance of the algorithm. To be precise, we consider the use of probabilistic models of the current population which are subsequently sampled in order to produce diverse solutions without distorting the convergence of the population and the progress of the search. We perform an extensive empirical assessment of those strategies on three different problems, considering a simulated computational environment featuring diverse degrees of instability. It is shown that these self-sampling strategies provide a performance improvement with respect to random reinitialization, and that a model using bivariate probabilistic dependencies is more effective in scenarios with large volatility.