Abstract
This investigation shows that a genetic algorithm seems to be able to search a solution space more effectively with the addition of a gradient-based local improvement scheme. The objective of this research was to determine if a new hybrid genetic algorithm, for optimization, is computationally more efficient than a standard genetic algorithm (G.A.). A direct method of comparison was used by adding the hybrid method onto the standard G.A. allowing for a direct comparison of the two algorithms. This is necessary due to the difficulties inherent in duplicating or comparing to existing algorithms. This new hybridization method applies a variable sized gradient procedure to a fixed percentage of the population during solution evolution. A series of tests were completed by using both the standard and hybrid genetic algorithms to evolve optimal solutions for three test problems. Comparisons of accuracy and efficiency between standard genetic algorithm and the standard G.A., with the addition of the gradient search ability, were completed. As expected, the gradient procedure generally improved the efficiency of the standard genetic algorithm for all of the test problems. This result is only applicable to the specific algorithm used in this research. However, similar algorithms, which use similar mutation rates, selection schemes, and crossover methods, should also benefit from the addition of the gradient search. The short-term efficiency, i.e. convergence rate, seems to be optimal at some initial step size range, which will be problem dependent. Long-term evolution, however, was found to reduce the effect of initial step size. Execution times were also found to be lower in almost all cases with the gap between the standard and hybrid algorithm narrowing as the population size increased.
Houser, Robert Charles (2000). Improving the searching abilities of a real-valued genetic algorithm. Master's thesis, Texas A&M University. Available electronically from
https : / /hdl .handle .net /1969 .1 /ETD -TAMU -2000 -THESIS -H674.