Ensemble Learning Method Based On Ant Colony Optimization Algorithm And Reasearch Of Its Parallel Characteristic

Ensemble learning can significantly improve generalization and precision of learn-ing system by combining weak classifiers together. What’s more, because of its flexi-bility and adaptability it has been successfully applied to various domains. Ensemble learning is a broad concept, consist of a set of different realization models, including: Bagging, Boosting and random forest, etc. Although these models have achieved sound results, there are still exist some defects. For example, Boosting can’t be deserialized for parallelization, random forest will overfitting in some noisy environment. In order to solve these problems, new ensemble learning models are constantly put forward.Ensemble learning model has two steps, including generation and combining of weak classifiers. Recently, researches on ensemble learning are focused on combining step. According to detail analysis on combining methods, we find it can be view as continuous optimization problem. To solve continuous optimization problem, heuris-tic algorithms usually be used. The reason is that heuristic algorithms can significantly reduce computation time and achieve sound percision. Ant colony optimization(ACO) is a good one of that which has been applied in many domains, including:combina-tion optimization, system identification and data mining. However, general ACO can’t be introduced directly, some special improvements are need to solve continuous op-timization problem. Improvements intorduced in this thesis are realized by changing the pheromone form. Pheromone is expressed as distribution instead of real value to control weights of classifier changing continuously. To test ACO-Ensemble learning algorithm, we apply it to drug-like prediction domain. Comparing to Bagging and Adaboost algorithm, it achieves ideal results. At last, we do some research on parallelization of ACO-Ensemble algorithm. With constantly generation of data, scale of data to process is more and more huge. There-fore, time consuming becomes a standard of the performance of algorithm. Many classical algorithms have been parallelized recently. With detail study, we find that ACO-Ensemble has parallel characteristics inherently. What’s more, we introduce a realization based on MapReduce cutting time consuming down observably.