You are hereMotivational scenario for parameter setting in evolutionary algorithms

Motivational scenario for parameter setting in evolutionary algorithms


Evolutionary algorithms are more sensitive for their parameters than the pioneers of evolutionary computing have thought in the 1990ies. This means that finding appropriate parameter values, e.g., population size, mutation rate, selection pressure, is important for good performance. In the motivational scenario we have a problem or a set of problems to be solved, an EA to be used as problem solver, and we need to find parameter values that maximize EA performance. In general, there are two approaches for doing this:

  • before the run of the EA (called parameter tuning), or
  • during the run of the EA (called parameter control).

The second approach amounts to augmenting EAs by a parameter control mechanism that (re-)specifies the parameter values on-the-fly. Such EAs have the potential to a) calibrate their own parameters, thus liberating the users from doing it, b) using parameter values suited for the given problem, c) using parameter values appropriate in the different stages of the search process. Obviously, we want control mechanisms that are applicable for many problem-EA combinations. Therefore, a typical investigation addresses a set of problems. Furthermore, we want mechanisms that can handle different performance measures, e.g., algorithm speed or solution quality. Possible research questions include issues about the information to be used by the parameter control mechanism, the frequency of updates, the control mechanisms themselves.

Example publications of CI group members and students:

Background and more information:

  • Classic paper that identified the issue: here.
  • Recent state of the art overview: here.
  • A whole book about parameter setting in EAs: here.
  • Another excellent book on related issues: here.