Abstract: In order to solve the classifier problem of learning new sample knowledge, an incremental learning algorithm based on distance is proposed. The algorithm is based on the nearest neighbor algorithm to increase the function of incremental learning. Calculating the matching degree between new input sample and model samples, the algorithm finds the best and second best matching degrees and compares them with the threshold. The comparative results decide whether it increases samples amount or adds a new category to realize incremental learning. The algorithm is applied to the UCI standard database and experiment of vehicle type recognition. The experimental results prove the algorithm’s efficient. Also, more experiments are conducted to analyze and verify how standard sample amount and matching degree threshold affect the final results.

Key words: incremental learning, nearest neighbor algorithm, matching degree

摘要： 为解决分类器学习新样本知识的问题，提出一种基于近邻算法的增量学习算法。该算法以最近邻算法为基础，首先计算新样本与标准样本之间的匹配度，找到最佳匹配样本和次佳匹配样本，然后通过与匹配度阈值进行比较来决定是类内学习还是类别学习。算法采用UCI中的标准数据集进行实验并应用于车辆识别仿真，其结果验证了该算法的有效性。实验进一步研究了匹配度阈值的选择和初始化样本数量选取对分类正确率影响。

关键词: 增量学习, 最近邻算法, 匹配度