Asynchronous Evolutionary Multi-Objective Algorithms with heterogeneous evaluation costs

Abstract

Abstract-Master-slave parallelization of Evolutionary Algorithms (EAs) is straightforward, by distributing all fitness computations to slaves. The benefits of asynchronous steadystate approaches are well-known when facing a possible heterogeneity among the evaluation costs in term of runtime, be they due to heterogeneous hardware or non-linear numerical simulations. However, when this heterogeneity depends on some characteristics of the individuals being evaluated, the search might be biased, and some regions of the search space poorly explored. Motivated by a real-world case study of multi-objective optimization problem -the optimization of the combustion in a Diesel Engine -the consequences of different components of heterogeneity in the evaluation costs on the convergence of two Evolutionary Multi-objective Optimization Algorithms are investigated on artificially-heterogeneous benchmark problems. In some cases, better spread of the population on the Pareto front seem to result from the interplay between the heterogeneity at hand and the evolutionary search.

Understanding Simple Asynchronous Evolutionary Algorithms

“…This raises the concern that simple asynchronous EAs may be biased away from slow-evaluating regions of the search space. As mentioned in Section 1.1, Yagoubi et al observe evidence that this does in fact occur in a multi-objective context [30]. They found that on at least one test problem, an evaluationtime bias helped to prevent premature convergence.…”

“…It's not clear how the relationship between the fitness landscape and heritable evaluation time traits may affect EA performance. The only prior work we know that has raised this question is Yagoubi et al's empirical analysis of a multi-objective test suite in [30], which found that an asynchronous MOEA had a harder time finding good solutions when they were located in a region of the search space that was artificially configured to have slower evaluation times. This raises the concern that asynchronous EAs may in general have a bias toward fast-evaluating individuals.…”

Evaluation-Time Bias in Asynchronous Evolutionary Algorithms

“…Yagoubi et al mention that an asynchronous EA was indeed slower in discovering solutions to a multi-objective problem that were located in a region of the search space that was artificially configured to have longer evaluation times than the rest of the landscape [6]. In another experiment, they found that penalizing high-quality solutions by giving them long evaluation times actually improved performance by helping to prevent premature convergence on a test problem.…”

Asynchronous Parallel Evolutionary Algorithms

“…Asynchronous Parallel EAs (APEAs) address this issue by performing survivor selection and all other evolutionary processes, without waiting for all individuals to be evaluated in the customary batch process. Since parallelizing EAs can be very time-consuming, many papers -including this one -rely on adding artificial evaluation time [2,5,6], which as a side-effect makes actual parallel execution unnecessary. Previous published work has mainly focused on the speed-up caused by the removal of the synchronization step in terms of less wasted clock cycles.…”