摘要
研究了基于贝叶斯推理的多层前向神经网络训练算法,以提高网络的泛化性能。在网络目标函数中引入表示网络结构复杂性的惩罚项,以便能够在训练优化过程中降低网络结构的复杂性,达到避免网络过拟合的目的。训练过程中使用显式的概率分布假设对模型进行分析和推断,根据融入先验分布的假设和依据,获取网络参数和正则化参数的后验条件概率,并基于后验分布的贝叶斯推理得出最优化参数。利用上述算法训练前向网络,对一个微型锅炉对象进行了模型辨识,通过测试,证明所辨识出的对象模型能够较好地表现出对象的动态行为,且具有较好的泛化性能。
Bayesian inference methods are studied for adapting network and regularization parameters during the training of feed-forward neural networks in order to improve their generalization capabilities.The performance index includes a term interpreted as a log prior probability distribution over the network parameters and a sum-squared error term interpreted as the log likelihood for a noise model.The former is used for penalizing the network complexity.Prior knowledge about the model parameters is incorporated within analysis and combined with training data.The optimal network parameters can be obtained with the maximum posterior probabilities of the models.The regularizing constants are set by maximizing the evidence.And models can be ranked by evaluating the evidence.The resulting algorithm is applied to a non -linear mini -boiler system to construct a predictive model.A set of data are used to test the performance of the model,the result indicates that the neural network can behave with the dynamic characteristics of the boiler and good generalization abilities.
出处
《计算机工程与应用》
CSCD
北大核心
2005年第11期5-8,11,共5页
Computer Engineering and Applications
基金
国家自然科学基金资助项目(编号:60074021)
关键词
系统辨识
非线性
神经网络
泛化
system identification,nonlinearity,neural networks,generalization ability
