Deriving heuristics in evolutionary algorithms through families of focused experiments

Over the last decade (1990-2000), the field of evolutionary computation has experienced a very rapid growth. However, the existence of overly many parameters and choices in directing the optimization (search) process requires some additional heuristics. These heuristics may be derived from focused experiments that reveal an impact of the particular parameter on the efficiency of search for a global optimum. The paper concentrates on evolution strategies (ES), a subset of evolution programs developed as methods for numerical optimization, and in particular on multimembered (/spl mu/+/spl lambda/)-ES, where /spl mu/ individuals produce /spl lambda/ offspring. The paper analyzes the influence of various recombination techniques on the convergence rate of (/spl mu/+/spl lambda/)-ES. Experimental explorations of various recombination techniques are performed on the optimization task of nine standard, well-known objective functions.