Effects of selection schemes and scoling strategies on the performance of genetic algorithms

number: 
490
إنجليزية
department: 
Degree: 
Imprint: 
Computer Science
Author: 
Azhar Mawlood Kadim
Supervisor: 
Dr. Mazin Z. Othman
Dr.Venus W. Samawi
year: 
2000
Abstract:

Genetic Algorithm (GA) has become a suitable searching or optimization method for solving many of today complex problems. Comparing with the traditional search techniques; GA has advantages of i) Fast convergence to reach near global optimal. ii) Effective global searching capability in complex or/and huge search space, iii) Capability of search in space where one can not use traditional techniques. However, GAs suffer from some disadvantages of i) Premature convergence near global optimal point. ii) Many parameters to be selected that affect on the solution. Therefore, certain manipulations are usually essential to speed up and improve the GA .• performance. Among those manipulations are the fitness scaling and selection operations. Fitness Scaling refers to modifications to the fitness values to overcome the problems related to the characteristic of the fitness being optimized. The goal of the scaling is to reduce the selection (of the best string) early in a run, thereby preventing domination of a population by a single superindividual, and increase the selection later in a run, therefore maintaining appropriate competition among highly fit. Scaling of fitness function values has been used widely in practice, and many researches were done using this strategy. The implementation of fitness scaling and selection methods, and the discussion of their effects on the performance of GA are presented in this work. A comparison between these methods and traditional GA are also illustrated in this thesis. Three problems (Knapsack, Traveling Salesman, and Substitution Ciphers) were solved in this work using GA. Many scaling methods were applied on these problems and the results were compared with each other for each problem. Also several selection schemes were implemented. Some analysis was performed to compare between these schemes with scaling methods. finally, it was found that the use of selection and scaling mechanisms are important to ': make GA less susceptible to premature convergence (convergence means that all individuals in the population are the same), and these mechanisms have important role in GA performance.