Parallel Implementations of the Complex-RF Algorithm
(English)Manuscript (preprint) (Other academic)
Even though direct-search optimization methods are more difficult to parallelize than population-based methods, there are many unexploited opportunities. Five methods for parallelizing the Complex-RF methods have been implemented and evaluated. Three methods are based on the unchanged original algorithm, while two require modifications. The methods have been tested on two test function and one real simulation model. An analysis of the algorithm has been performed. This provides a basis for parametrization of the parallel methods. Without changing the original algorithm, speed-up of 2.5-3 is achieved. With allowing modifications, a speed-up of up to 5 is obtained without significantly reducing the probability of finding the global minimum. Speed-up does not scale linear to the number of threads. When more threads are added, parallelization efficiency decreases. However, a comparison with a particle swarm method shows that Complex-RF performs better regardless of the number of threads, due to its fast convergence rate.
Parallel optimization, direct-search, simplex, Complex-RF
Electrical Engineering, Electronic Engineering, Information Engineering Fluid Mechanics and Acoustics
IdentifiersURN: urn:nbn:se:liu:diva-122751OAI: oai:DiVA.org:liu-122751DiVA: diva2:872642