Aiming for a theoretically tractable CSA variant by means of empirical investigations

Evolution Strategies (ES) for black-box optimization of a function f : Rn → R are investigated. Namely, we consider the cumulative step-size adaptation (CSA) for the variance of multivariate zero-mean normal distributions, which are commonly used to sample new candidate solutions within Evolution Strategies (ES). Four simplications of CSA are proposed and investigated empirically and evaluated statistically. The background for these four new CSA-derivatives, however, is not performance tuning, but our aim to accomplish a probabilistic/theoretical runtime analysis of an ES using some kind of a CSA in the near future, and a better understanding of this step-size control mechanisms. Therefore, we consider two test problems, namely the Sphere function without and with Gaussian noise. Categories and Subject Descriptors F.2 [Analysis of Algorithms and Problem Complexity]: Experimental Analysis; G.3 [Probability and Statis

• Type: Conference
• Year: 2008
• Where: GECCO
• Authors: Jens Jägersküpper, Mike Preuss
• Keywords: Cumulative Step-size Adaptation, Evolution Strategies, GECCO 2008, Optimization, Zero-mean Normal Distributions