The Application Of Multi-ant Colony Optimal Algorithm In The Optimal Design Of Chemical Pattern Classifier

Pattern classification is a kind of key technology used in a lot of project fields including automatic control monitor, image recognition, troubled diagnose, supplies compound, medical diagnosis, etc. The classical categorized method of pattern classification is mainly to count the analytical method based on pluralism. After that, artificial neural network technology becomes effective tool that pattern classified gradually too. These two kinds of methods have their own strong points. Pluralism counts the analytical method calculates normally, there are clear probability meanings, but need abundant samples, and should comply with certain distribution. The artificial neural network technology is strong in ability to express and suitable for extensive range, but the network is difficult in design, it is time-consuming to train, some extreme value are deficient.With the constant development of scientific calculation, search space has become widen and complicated. Traditional algorithm for pattern classification usually has something to do with population scale, the choice of parameters, the problems complexity, etc. It takes a long time to search the optimal value when the computing has a large scale population, the complicated parameters and large search space. It is even unable to receive the satisfactory result sometimes.In this thesis, the merits and disadvantages of ant colony optimization algorithm and other optimization algorithms are studied firstly. Then a novel algorithm, multi-ant colony optimization algorithm, MACO, is proposed to solve combinatorial optimization problems. To demonstrate the performance of this algorithm, we conduct several TSP benchmark problems and some real pattern classification problems. Obtained results demonstrate that MACO has well global optimization ability in combinatorial optimization problems. The major contributions of this work are summarized as follows.1. In this paper, basic ant colony optimization algorithm will be used in the design of chemical pattern classifier. The methods that often used for pattern classification, statistical analysis needs the sample to comply with certain distribution, the design and training of neural network is difficult, and the enactment of super parameter in support vector machine is difficult too. So in paper, we try to optimize the design of chemical pattern classifier by ant colony algorithm. The classification result of UCI database indicates that the chemical pattern classifier based ant colony algorithm has little number of classifier, concise form and good predication performance.2. Because of single colony and pheromone, the basic ant colony algorithm doesn't simulate the full message of real ant colony. So in this paper, we attempt to proposed a novel algorithm combination of local research and pheromone change strategy, multi-ant colony optimization algorithm, MACO. The ant colony algorithm has a fast convergence velocity, but it has the phenomenon of premature convergence and complexity of solution construction in ant colony optimization. In order to maintain the diversity of population and prevent the degradation of population, multi-colony collaboration with local research and pheromone change strategy was introduced to ant colony optimization. It demonstrated that MACO has well global and local optimization ability in combinatorial optimization problems.3.According to selection of appropriate rule expressions and rule evaluation function, rule induction problem was translated into corresponding combinatorial optimization problem, which was solved by MACO.4. Through analysis of the interrelationship among rules expression, evaluation function and training sample data, a method was proposed to extract candidate points from every continuous attribute, which can lead to scale reduction of rule leaning problem and improvement of optimization algorithm's performance. Finally, MACO was applied to extract classification rules, and classifier constructed by these rules showed well performance in practical pattern classification problems.In a word, this thesis has a comprehensive analysis of ant colony optimization algorithm and the rule leaning of pattern classification, it provide a new method for information mine of complex data in chemistry and chemical engineering and modeling.