Oblivious Randomized Direct Search for Real-Parameter Optimization

Abstract

The focus is on black-box optimization of a function given as a black box, i.e. an oracle for f-evaluations. This is commonly called direct search, and in fact, most methods for direct search are heuristics. Theoretical results on the performance/behavior of such heuristics are still rare. One reason: Like classical optimization algorithms, also direct-search methods face the challenge of step-size control, and usually, the more sophisticated the step-size control, the harder the analysis. Obviously, when we want the search to actually converge to a stationary point (i.e., the distance from this point tends to zero) at a nearly constant rate, then step sizes must be adapted. In practice, however, obtaining an ε-approximation for a given ε> 0 is often sufficient, and usually all N parameters are bounded, so that the maximum distance from the optimum is bounded. Thus, in such cases reasonable step sizes lie in a predetermined bounded interval. Considering the minimization of the distance from a fixed point as the objective, we address the question, for randomized heuristics that use isotropic sampling to generate new candidate solutions, whether we might get rid of step-size control – namely of the problems connected to it, like so-called premature convergence – by choosing step sizes randomly according to some properly predefined distribution over this interval. As this choice of step sizes is oblivious to the course of the optimization, we gain robustness against a loss of step-size control. Naturally, the question is: What is the price w.r.t. local convergence speed? As we shall see, merely a factor of order ln (d/ε), where d is the diameter of the the decision space, an N-dimensional interval region.KeywordsIsotropic DistributionCovariance Matrix Adaptation Evolution StrategyIsotropic SamplingElitist SelectionRandomize Search HeuristicThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.