摘要
从理论上分析了隐含层激励函数满足Mercer条件的前向神经网络的数学本质,给出了网络学习的指导方向.提出3种网络在线学习算法,它们通过动态调整网络结构和权值来提高网络在线预测性能.算法完全符合统计学习理论提出的结构风险最小化原则,具有较快的学习收敛速度和良好的抗噪声能力.最后通过具体数值实验验证了上述算法的可行性和优越性.
The mathematic essence of feedforward neural networks whose activation function of hidden neurons satisfies Mercer condition is analyzed. And the guideline for feedforward neural networks learning is given. Then three online learning algorithms are proposed, which can improve the online prediction performance of the networks by adjusting its architecture and connection weights dynamically. Those algorithms with global convergence and good anti-noise performance correspond to the principle of structural risk minimization. Their reliability and advantage are illustrated through concretely test.
出处
《控制与决策》
EI
CSCD
北大核心
2005年第3期303-307,共5页
Control and Decision
基金
航空科学基金项目(01C52015).
关键词
前向神经网络
在线学习
统计学习理论
机器学习
Convergence of numerical methods
Feedforward neural networks
Neural networks
Robot learning
Signal filtering and prediction
Statistical methods
