An Adaptive Differential Algorithm
Vainioranta, Ville (2013)
Kuvaus
Opinnäytetyö kokotekstinä PDF-muodossa.
Tiivistelmä
Differential Evolution is an evolutionary algorithm designed for global optimization. Its main assets are simplicity, ease of use, and effectiveness, which has been demonstrated in different research papers and evolution computation competitions. Differential Evolution has three control parameters and each of them has a significant impact on the performance of the algorithm.
This paper introduces three modified algorithms which adapt the control parameters during the optimization process. The first modification VDE-1 adapts the mutation factor, the second modification VDE-2 adapts the crossover factor and the third modification adapts both factors. The idea in these modified algorithms is to use previously successful factor parameter values and calculate an exponential moving average value of these parameter values. This average value is then utilized to generate new control parameter values. In addition, the modifications apply Daniela Zaharie’s theory on population diversity in Differential evolution and on how control parameters affect the convergence properties.
The performances of the modified algorithms are compared with the original algorithm version with results on a benchmark of 25 functions provided by the “IEEE Congress on Evolutionary Computation 2005” evolution computation contest. The 25 functions are run on two different dimensions, and the results are gathered from all algorithms. The results are then analyzed and compared with each other. The main focus in the analysis is on the behavior of the control parameters for the modified algorithms.
The experimental results indicate that the modified algorithms are promisingly more effective and efficient when compared to the original Differential Evolution. Especially the convergence speeds of VDE-1 and VDE-3 are much faster compared to the original algorithm. The differences in performance are remarkably high for certain functions.
This paper introduces three modified algorithms which adapt the control parameters during the optimization process. The first modification VDE-1 adapts the mutation factor, the second modification VDE-2 adapts the crossover factor and the third modification adapts both factors. The idea in these modified algorithms is to use previously successful factor parameter values and calculate an exponential moving average value of these parameter values. This average value is then utilized to generate new control parameter values. In addition, the modifications apply Daniela Zaharie’s theory on population diversity in Differential evolution and on how control parameters affect the convergence properties.
The performances of the modified algorithms are compared with the original algorithm version with results on a benchmark of 25 functions provided by the “IEEE Congress on Evolutionary Computation 2005” evolution computation contest. The 25 functions are run on two different dimensions, and the results are gathered from all algorithms. The results are then analyzed and compared with each other. The main focus in the analysis is on the behavior of the control parameters for the modified algorithms.
The experimental results indicate that the modified algorithms are promisingly more effective and efficient when compared to the original Differential Evolution. Especially the convergence speeds of VDE-1 and VDE-3 are much faster compared to the original algorithm. The differences in performance are remarkably high for certain functions.