An Analysis of the Taguchi Method for Tuning a Memetic Algorithm with Reduced Computational Time Budget

Determining the best initial parameter values for an algorithm, called parameter tuning, is crucial to obtaining better algorithm performance; however, it is often a time-consuming task and needs to be performed under a restricted computational budget. In this study, the results from our previous work on using the Taguchi method to tune the parameters of a memetic algorithm for cross-domain search are further analysed and extended. Although the Taguchi method reduces the time spent finding a good parameter value combination by running a smaller size of experiments on the training instances from different domains as opposed to evaluating all combinations, the time budget is still larger than desired. This work investigates the degree to which it is possible to predict the same good parameter setting faster by using a reduced time budget. The results in this paper show that it was possible to predict good combinations of parameter settings with a much reduced time budget. The good final parameter values are predicted for three of the parameters, while for the fourth parameter there is no clear best value, so one of three similarly performing values is identified at each time instant.