The Study of Crossover Mechanism of Genetic Algorithm for Diffractive Optical Elements

• Main Authors: Tsai-Chun Hou, 侯采君
• Other Authors: 徐巍峰
• Format: Others
• Language: zh-TW
• Published: 2007
• Online Access: http://ndltd.ncl.edu.tw/handle/aa446c

碩士 === 國立臺北科技大學 === 光電工程系所 === 95 === In the genetic algorithm, it is generally acknowledged the crossover mechanism is even more important than the mutation mechanism. However, the crossover mechanism did not increase the performance of phase-only diffractive optical elements(DOEs) as expected in our early study. In order to identify the effect of the crossover mechanism. we removed the mutation mechanism and only used only the crossover mechanism to simulate the evolution to design the DOEs. The simulation results of evolution for the crossover mechanism were poor compared with the mutation mechanism, which was different from this algorithm applying to other problems. We believe the main reason of the poor result by only using the crossover mechanism is the population-purity. Therefore, in this thesis, we analyzed the population-purification situation and different crossover mechanism for designing the diffractive optical elements in using the genetic algorithm. In this research, by using a regular crossover pattern and by adding a seed individual of best performance in the population, we found the populations were completely purified (individuals of the population all become the same) in about 20 generations. By dynamically changing the crossover patterns, the population was purified in hundreds to thousands generations. Therefore, we used a method of the population reload every 1000 generations in order to analyze the effect of different crossover patterns. In the population-reload method, the crossover pattern was changed every 10 generations, the initial population replaced the purified population every 1000 generations, and a total 100,000 generations of the GA revolution. Then, 5 different designs of crossover pattern of random parameters were designed to compare the performance of the binary-phase DOEs. No significant improvement of performance was observed in the simulation results when the number of exchanged pixels was about a half of the total pixels. However, the performance of the DOEs was significantly increased when the number of exchanged pixels decreased, or even increased. Consequently, according to the simulation results obtained in this study, we believe that the spatial location of the exchanged pixels of the crossover mechanism influenced the DOE performance more significantly than the number and pattern of the exchanged pixels.