Stopping Criteria for Single-Objective Optimization

In most literature dealing with evolutionary algorithms the stopping criterion consists of reaching a certain number of objective function evaluations (or a number of generations, respectively). A disadvantage is that the number of function evaluations that is necessary for convergence is unknown a priori, so trialand-error methods have to be applied for finding a suitable number. By using other stopping criteria that include knowledge about the state of the optimization run this process can be avoided. In this work a promising new criterion is introduced and compared with criteria from literature. Examinations are realized using two relatively new algorithms, Differential Evolution and Particle Swarm Optimization. The study is performed on the basis of eight well-known single-objective unconstrained test functions. Depending on the applied stopping criterion considerable performance variations are observed. Recommendations concerning suitable stopping criteria for both algorithms are given.