Design of analog circuits via soft computing methods

This thesis presents a simple speciation procedure to improve local search capability of the genetic algorithm in analog circuit design. Performing an intense search around prospering solutions is the main idea of the procedure. There is no need additional circuit simulation in order to detect promising search regions. Application simplicity and requiring no extra computation are the main advantages of the proposed procedure. The procedure is tested in Gaussian, sigmoid, cube, and square function circuit design problems. Two sets of 125 simulations are performed for each problem. The simulation results show that the method is statistically better than the canonical genetic algorithm with at least 99 % confidence, and since speciation method requires no additional circuit simulation, the total simulation time of the algorithm does not change too much. Effect of the different parameters on performance are investigated. Gaussian and sigmoid circuit design problems are solved for different population sizes. It is found out that effectiveness of the method increases in small populations. Gaussian and sigmoid circuit design problems are experimented for different speciation threshold coefficients, and it is concluded that lower threshold coefficient impose stringent condition on speciation, and speciation algorithm can not take place in most cases. Therefore, the performance of the procedure deteriorates for the lower threshold coefficients. The Gaussian circuit is designed for various template individual number to see effect of it. It is observed that both few and high number of template individuals have negative effect on the performance.