Evaluation of Termination Criteria for Single-Objective Evolutionary Optimization Algorithms
Many improvements concerning the convergence behavior of evolutionary algorithms have been suggested since they have been invented. However, detection of convergence or stagnation is still largely done via heuristic rules. Recently, the online convergence detection (OCD) approach based on statistical testing has been introduced for finding adequate stopping generations for multiobjective evolutionary algorithms.
The aim of the thesis is the (straightforward) generalization of the OCD concept to the single-objective domain, especially the application to the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) and the experimental analysis whether the internal algorithm termination criteria coincide with the OCD behaviour.