分类回归树多吸引子细胞自动机分类方法及过拟合研究

Multiple Attractor Cellular Automata Classification Method and Over-Fitting Problem with CART

基于多吸引子细胞自动机的分类方法多是二分类算法,难以克服过度拟合问题,在生成多吸引子细胞自动机时如何有效地处理多分类及过度拟合问题还缺乏可行的方法.从细胞空间角度对模式空间进行分割是一种均匀分割,难以适应空间非均匀分割的需要.将CART算法同多吸引子细胞自动机相结合构造树型结构的分类器,以解决空间的非均匀分割及过度拟合问题,并基于粒子群优化方法提出树节点的最优多吸引子细胞自动机特征矩阵的构造方法.基于该方法构造的多吸引子细胞自动机分类器能够以较少的伪穷举域比特数获得好的分类性能,减少了分类器中的空盆数量,在保证分类正确率的同时改善了过拟合问题,缩短了分类时间.实验分析证明了所提出方法的可行性和有效性.

The classification methods based on multiple attractor cellular automata can process the classification of two classes, and they are difficult to overcome overfitting problem. There are not yet effective methods for constructing a multiple attractor cellular automata which can process multi- classification and overfitting problem. The pattern space partition in the view of cell space is a kind of uniform partition which is difficult to adapt to the needs of spatial non-uniform partition. By combining the CART algorithm with the multiple attractor cellular automata, a kind of classifier with tree structure is constructed to solve the non-uniform partition problem and overfitting problem. The multiple attractor cellular automata characteristic matrix is defined, and the learning method of classifiers as a node in a tree is studied based on particle swarm optimization algorithm. The multiple attractor cellular automata classifiers built on this approach are able to obtain good classification performance by using less number of bits of pseudo-exhaustive field. The classifier with tree frame of multiple attractor cellular automata reduces the number of empty basin and restrains overfitting problem without lost accurate rate, and shorts the classification time. The feasibility and the effectiveness of the proposed method have been verified by experiments.