Novel Methods For High-Dimensional Complex Patterns Recognition

The development of science and technology requires the new theory and method of pattern recognition. In the field of chemistry, along with the wider use of the precise analytical instruments and the needs of investigating objects in more details, the patterns employed to describe objectives become more complex and the dimension of the patterns becomes much higher. Thus, the method that can map the complex patterns from the high-dimensional space to the lower-dimensional space or that can model non-linear problem showed particular importance in the field of the chemical pattern recognition. Self-organizing map networks (SOM Networks) and non-linear mapping (NLM) are two mapping methods full of promise. In our work, a series of explorations for SOM networks and NLM to improve their performances had been implemented. They were applied to high-dimensional complex chemical pattern classification and satisfactory results were obtained. Based on SOM networks, a novel pattern-characteristic-preserving map algorithm was proposed. It was employed to high-dimensional complex chemical pattern classification and satisfactory results were obtained. To illustrate the performance of the pattern-characteristic-preserving map algorithm, its performance was compared with that by using SOM networks only, and the results showed the superior performances of the pattern-characteristic-preserving map algorithm over SOM networks. At the same time, an adaptive partial least square regression (APLSR) was proposed to obtain the model with high predicting correctness, and a typical example of modeling the quantitative structure-activity relationships (QSAR) of substituted aromatic sulfur derivatives was employed to verify the effectiveness of APLSR. Good performance of APLSR makes it an attractive alternative for the other model methods. To estimate the parameters of the mechanism model, a novel genetic algorithm, which is named as chaos genetic algorithm (CGA), was proposed. The major contributions of this work are summarized as follows.l)To improve the convergence rate of training SOM networks and to obtain a better topology-preserving map of the input patterns, the algorithms of dynamically adjusting the range of the neighborhood and the learning rate were designed and the algorithm of improving the utilization of the connection weights was proposed. Firstly, to distinguish the winning unit from the others, the winning unit should have larger degree of updating weight than other units and accordingly the learning rate is defined by a Duo-Bubble function.Secondly, to accelerate the network convergence, the learning rate is decreasing as a whole in the learning process. However, the decrease is managed to take a wavelike manner, which means the learning rate will take a little increase after some learning cycles. Thirdly, along with the decrease of learning rate, the range of the neighborhood is decreased in the learning process. Besides, the smaller the range of neighborhood is, the longer the learning time of the neighborhood will be. Finally, to improve the utilization of the connection weights, define the maximum number of times of winning for every unit of output layer.2) In general, complete recomputation is required in case new pattern is needed to be added to the original pattern set, since interpoint distances are dependent from each other by using the conventional NLM methods. To facilitate the adding of new pattern into the original pattern set, a new method was proposed. The pattern most similar to the new pattern in original training set is selected at first and the corresponding point on the original map denoted by o-point is found too. Then, a new initial map is formed by plotting the new point same as o-point on the original map. Due to the fact that the newly obtained map tend to be quite stable when the new pattern introduced is similar to the original training set, it will not take long time to obtain the new map from the old map.3) A new algorithm of adaptive mapping error was proposed to overcome the defici...