神经网络学习算法的过拟合问题及解决方法

Over-Fitting in Neural Network Learning Algorithms and Its Solving Strategies

针对反向传播学习算法及其改进算法中出现的过拟合问题 ,探讨了三种解决方法 :调整法、提前停止法和隐层节点自生成法 ,并用实例对三种方法进行了验证和比较。其中 ,调整法和提前停止法针对一个较大的网络可以解决过拟合问题 ,而隐层节点自生成法的提出既能避免过拟合问题 ,又能获得最少神经元网络结构。这三种方法有效地解决了在神经网络学习过程中的过拟合问题 ,提高了网络的适应性。它们不仅适合于函数逼近 。

To counter the over fitting problem in BP algorithm and its improvements, we proposed three solving strategies, called regularization, early stopping and hidden node self generating. The first two methods could get rid of the over fitting for large networks. The last one could not only avoid the over fitting, but also get the most appropriate network. The paper also gave the validation and comparison of the three methods by an actual example.