Parameter control in evolutionary algorithms

Deciding on the best performing parameter setting for evolutionary algorithms in a problem domain is a nontrivial task. For example, in the case of the deployment optimisation of software components in automotive systems, despite the wide range of evolutionary algorithms already published, it is still unknown which parameter settings of the algorithms are the optimal choices. The values of the parameters of an evolutionary algorithm are dependent from the given problem instance and greatly determine whether the algorithm will efficiently converge to the optimal solutions or not. The selection of the parameter values is, however, a difficult task. The existing approaches for setting the parameter values of evolutionary algorithms follow two main streams, those who set the parameter values in advance of the search process, i.e. parameter tuning, and those who change them during the search process, refered to as parameter control. Optimal parameter tuning is not only hard and time consuming but also impossible, since different values of the parameters are optimal at different stages of the search. On the other hand, although parameter control is more efficient than parameter tuning, chosing a well-working procedure among the different possibilities one can design is also a hard task. The approach proposed will use statistical information to design an efficient parameter control procedure which overcomes the drawbacks of previous works done in this field. To validate the approach, multiple runs of the optimisation algorithms on realistic case studies taken from the automotive industry will be considered