论文部分内容阅读
对于传统的Ｋ 平均算法来说，如何选择适当类的数目是一个难以解决的问题．有人提出了次胜者受罚的竞争学习（ｒｉｖａｌｐｅｎａｌｉｚｅｄ ｃｏｍ ｐｅｔｉｔｉｖｅ ｌｅａｒｎｉｎｇ ： ＲＰＣＬ）算法试图来解决这一问题．但是，当数据类有重叠以及输入矢量含有非独立项时，ＲＰＣＬ算法的性能不能令人满意．本文提出了一种结合全协方差矩阵的ＲＰＣＬ算法，并逐步删除那些只包含少量训练数据的类．这种算法，我们称之为改进的ＲＰＣＬ算法．我们用改进的ＲＰＣＬ算法来确定高斯混合分布类的数目，并将其与原来的ＲＰＣＬ进行比较．实验证明，改进的ＲＰＣＬ算法比原来的ＲＰＣＬ算法能够更好地表征类
For the traditional K-means algorithm, how to choose the appropriate number of classes is a difficult problem to solve. Some proposed rivalpenalized pet petition learning (RPCL) algorithm to solve this problem. However, the performance of the RPCL algorithm is unsatisfactory when the data classes overlap and the input vector contains non-independent terms. This paper presents an RPCL algorithm that combines a full covariance matrix and gradually removes classes that contain only a small amount of training data. This algorithm, which we call the improved RPCL algorithm. We use the improved RPCL algorithm to determine the number of Gaussian mixture distribution classes and compare them with the original RPCL. Experiments show that the improved RPCL algorithm can better classify the original RPCL algorithm