Combination of Meta-Heuristics, Ensemble and Data Stream to Increase the Effectiveness of Memory-Based Reasoning

Memory-based Reasoning (MBR) has been used to capture past similarity experience and to solve new situations from precious past solutions. The MBR approach is particularly useful for solving problem, where we lack sufficient knowledge. MBR does not rely on statistical assumptions and its justifications are human-understand-able because they are based on the principle of analogy-based reasoning that humans frequently use during problem-solving. Before applying MBR to problems, parameter values must set in advanced. If the parameter values are not set appropriately, worse result may be obtained. On the other hand, a dataset may contain many features; however, not all features are beneficial for classification. The features may contain false correlations, which hinder the process of detecting process. Without feature selection, the prediction/classification accuracy rate may be worse due to the noises or too many dirty data. In most research either parameters turning or feature selection is used to improve classification accuracy rate. Some researches consider parameters turning and feature selection simultaneously to specific problems. However, they used specific data which can not be used to further comparison with other approaches. Therefore, three population-based meta-heuristics (SA, GA, PSO, SS and AIS), are proposed to feature selection and to obtain the better parameters for various public datasets, which may result in a better classification. Different predictive model has its own advantages and disadvantages and the suitability will influenced by the characteristic of problem. If these predictive models can work together, it is expected that the better result can be obtained, which is called ensemble. The purpose of ensemble is to integrate the option of many experts instead of only one expert to obtain better result. Therefore, this proposal is plan to use the ensemble architecture to further enhance the prediction/classification accuracy rate. On the other hand, data is increasing with time progress. The near unlimited data and the quick response become the challenge of current decision support system. Unlike the traditional steady database, data stream has multiple sources (multiple attributes), concept drifting, unlimited data, and model should provide the characteristic of quick response. Therefore, the proposed algorithms should be taken it into account to solve this problem effectively. In order to evaluate the proposed approaches, datasets in UCI (University of California, Irvine) are planned to evaluate the performance of the proposed approaches. It is expected that the approaches which apply the parameter turning and feature selection simultaneously, will obtain better classification rates and decrease the computational time, than approaches which apply either parameter turning or feature selection. Therefore, five meta-heuristics (SA, GA, PSO, SS and AIS) can be used to help MBR solve different problems with different parameter values and feature subset when facing various problems. In addition this proposal will add constraint of the maximum number of models in the data stream, in order to delete outmoded and non-reference case from ensemble and to provide the better prediction result.

Project ID:PF10001-0904
External Project ID:NSC99-2410-H182-023-MY2