Batch-enabled randomized parameter tuning for improved metaheuristic performance

Deepika Kaushik, Mohammad Nadeem

Abstract: Parameters are an important part of any metaheuristic algorithm. They play a pivotal role in deciding the results obtained from these algorithms. The problem of parameter tuning has become an optimization problem in itself, termed meta optimization. In the proposed work, a methodology for parameter tuning is proposed, in which the values of parameters vary randomly over an interval; hence, it is called random parameter tuning. It starts with dividing the whole population into sub-populations (called batches), and each batch is assigned a different range of parameter values. During each iteration, the value of a parameter is decided randomly according to the predefined interval of the corresponding batch. This approach is easy to understand and is general. It can be embedded with any of the metaheuristic algorithms. The proposed work has been embedded in the Genetic Algorithm, Particle Swarm Optimization, Firefly Algorithm, and Differential Evolution Algorithm. The approach has been tested over 15 benchmark functions, shifted rotated functions, and classical engineering problems. Moreover, the significance of the proposed approach is established by conducting a sensitivity analysis, a non-parametric Friedman test, and a Wilcoxon Rank Test. The proposed approach has been compared with the two state-of-the-art methods. The results show the superiority of the proposed approach.

External IDs:dblp:journals/evs/KaushikN25