Improving SVM For Learning Multi-Class Domains With Multi-Objective Algorithm

Multi-Objective Genetic Algorithm based on Pareto (MOGA) is one of the most important algorithms to solve Multi-Objective problems. Genetic Algorithm is very suitable for Multi-Objective problems, because at each iteration of Genetic Algorithm, several best Pareto results can be selected, and by using fitness function the Genetic Algorithm can converge to the Pareto-front. But because during the iterations of the Genetic Algorithm, new generation's items are create randomly, and Genetic Algorithm excludes all the worse items of each generation, some better results of the Pareto Set are lost. This is the disadvantage of the Genetic algorithm. Because the Multi-Objective Genetic Algorithm uses the Genetic Algorithm as the search algorithm, the Multi-Objective Genetic Algorithm also has the disadvantage in search efficiency and precision.Compare with Genetic Algorithm, Tabu Search algorithm has the advantage in local search precision and intellective search ability. Tabu Search algorithm uses tabu list to save the local optimal results and the search routes. Tabu list can control how the route can be selected, and tabu some items to avoid the search plunge in some local search area. Tabu list can buildup the Tabu Search algorithm's ability of finding the global optimal results. The"Strategic Oscillation"strategy in the Tabu Search uses penalty factor to control whether using some wore result as the local optimal results to get better global results. These characters of the Tabu Search can remedy the disadvantage of the Genetic Algorithm in the local search. Base on the characteristic of the Tabu Search and Genetic Algorithm, this thesis combine Tabu Search and Genetic Algorithm to create a new Multi-Objective algorithm named as MOTSGA (Multi-Objective based on Genetic Algorithm and Tabu Search).In order to validate the algorithm MOTSGA is effective, this thesis uses the MOTSGA algorithm to find the best parameters for the Multi-Classes SVM problem. The method to solve the Multi-Classes SVM problem is change the Multi-Classes SVM problem to several Two-Classes SVM problems. The problem to find the best parameters of the group of the several Two-Classes SVM problems is a Multi-Objective problem. The experiment is used MOTSGA algorithm and MOGA algorithm respectively to solve the Multi-Classes SVM problem. The results of the experiment showed that the MOTSGA has the advantage in the local search ability, and is easy to find the global best result.