摘要
ＩＨＭＣＡＰ（ｉｎｃｒｅｍｅｎｔａｌｈｙｂｒｉｄｍｕｌｔｉ－ｃｏｎｃｅｐｔｓａｃｑｕｉｓｉｔｉｏｎｐｒｏｃｅｄｕｒｅ）算法将基于概率论的符号学习与神经网络学习相结合，通过引入ＦＴＡＲＴ（ｆｉｅｌｄｔｈｅｏｒｙ－ｂａｓｅｄａｄａｐｔｉｖｅｒｅｓｏｎａｎｃｅｔｈｅｏｒｙ）神经网络，成功地解决了符号学习与神经网络学习精度之间的均衡性问题，实现了两种不同思维层次的靠近．该算法采用一种独特的增量学习机制，当增加新的实例时，只需进行一遍增量学习，调整原结构，不必重新生成判定树和神经网络，即可提高学习精度，速度快，效率高．同时，这种增量学习机制还可以降低算法对噪音数据的敏感度，从而使ＩＨＭＣＡＰ可以应用于实时在线学习任务．
IHMCAP (incremental hybrid multi concepts acquisition) algorithm combines the probability based symbolic learning with neural learning. The balance of learning accuracy between the symbolic and the neural parts are proportioned successfully, and the two different levels of thought are aboard laid by adhibiting FTART (field theory based adaptive resonance theory) neural network. A unique incremental learning mechanism is employed with this algorithm, which can adjust the former structure to improve learning accuracy by learning once instead of rebuilding the decision tree and the neural networks when the new examples are provided. It has higher speed, and is efficient. Moreover, the noisy sensibility of the system is depressed by the incremental learning mechanism, which enables IHMCAP can be applied to the tasks that require real time online learning.
出处
《软件学报》
EI
CSCD
北大核心
1999年第5期511-515,共5页
Journal of Software
基金
国家自然科学基金
江苏省自然科学基金
关键词
增量学习
神经网络
多概念获取
算法
抗噪音
Hybrid model, incremental learning, neural network, noise disposal.
