Invited Paper Evolutionary Algorithms Software- Parametrisation Possibilities and Performance

Abstract:- Evolutionary algorithms (EAs) are popular general purpose optimisation methods. According to the No Free Lunch Theorem is no inherently superior EA parametrisation possible. Therefore the support of a wide range of parameters by EA software libraries is desirable. Firstly this allows the exact specification of an EA that proved to perform well at a certain problem type. Additionally, flexible parametrisation possibilities allow the testing of a wide range of EA variants on new problems. Despite a large number of EA libraries has been developed, exists no model for the generic description of parametrisation possibilities. As explained in this paper, also the published approaches of classifying EA parameters are unsatisfactory. Starting from a missing definition of the term parameter, an a new, explicit description of the parameter elements is introduced and several categories are proposed according to which parameters can be classified. The additionally performed benchmarking of EA software shows that the abandonment of good software design and use of low-level languages for the benefit of more efficient operators is not worthwhile