Searching for Balance: Understanding Self-adaptation on Ridge Functions

Abstract

The progress rate of a self-adaptive evolution strategy is sub-optimal on ridge functions because the global step-size, denoted σ, becomes too small. On the parabolic ridge we conjecture that σ will stabilize when selection is unbiased towards larger or smaller step-sizes. On the sharp ridge, where the bias in selection is constant, σ will continue to decrease. We show that this is of practical interest because ridges can cause even the best solutions found by self-adaptation to be of little value on ridge problems where spatially close parameters tend to have similar values.KeywordsSearch SpaceEvolution StrategyStrategy ParameterProgress RateObject ParameterThese 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.