Performance Assessment of Metaheuristic Algorithms for Structural Optimization Taking Into Account the Influence of Algorithmic Control Parameters

Abstract

Metaheuristic optimization algorithms are strongly present in the literature on discrete optimization. They typically 1) use stochastic operators, making each run unique, and 2) often have algorithmic control parameters that have an unpredictable impact on convergence. Although both 1) and 2) affect algorithm performance, the effect of the control parameters is mostly disregarded in the literature on structural optimization, making it difficult to formulate general conclusions. In this article, a new method is presented to assess the performance of a metaheuristic algorithm in relation to its control parameter values. A Monte Carlo simulation is conducted in which several independent runs of the algorithm are performed with random control parameter values. In each run, a measure of performance is recorded. The resulting dataset is limited to the runs that performed best. The frequency of each parameter value occurring in this subset reveals which values are responsible for good performance. Importance sampling techniques are used to ensure that inferences from the simulation are sufficiently accurate. The new performance assessment method is demonstrated for the genetic algorithm in matlab R2018b, applied to seven common structural optimization test problems, where it successfully detects unimportant parameters (for the problems at hand) while identifying well-performing values for the important parameters. For two of the test problems, a better solution is found than the best solution reported so far in the literature.