New (feature) paper published on Entropy
A special issue about “unconventional methods for particle swarm optimization” was organized on MDPI’s Entropy journal.
We exploited the opportunity of such a peculiar venue to showcase one project that we developed in the last months: how to smooth out the fitness landscapes by means of Fourier transforms.
Specifically, we created and investigated surF, a novel methodology for Fourier surrogate modeling: by collecting a bunch of samples in the search space, surF creates a surrogate model of the fitness landscape. Thanks to the upstream filtering of Fourier coefficients, such surrogate model is smooth and, hopefully, simpler to optimize. surF works under the assumption that high-frequency noise is often not relevant for the optimization and it just misleads the optimization algorithms.
In the paper, we coupled surF with Fuzzy Self-Tuning PSO, showing that an optimization on the surrogate model can already provide good results. By feeding such preliminary information to an actual optimization, the optimization could yield better results, notably using the same budget of fitness evaluations.
The methodology seems to be very effective in the case of noisy fitness function; however, it is memory intensive, preventing its naïve application to high-dimensional optimization problems. We will investigate possible solutions in future work.
You can read our paper about surF on MDPI’s website.