【摘要】 本文提出一种用于多层前向网络的广义能量约束算法。新的能量函数不仅包括误差平方和项,而且还包括某些能使网络获得所期望的性能的约束项.讨论了某些典型的约束形式,施加在权重矢量或隐藏层输出或网络结构上.将本方法用于信号处理中广泛采用的奇异值分解(SVD)问题,取得较好效果.此方法还可以引申到主分量分析(PCA)。方法具有高度灵活性,并使矩阵的并行运算成为可能.能量约束中还可以引入有关信号特征的先验知识,从而使作信源分解时对信号和传输阵的正交要求有所降低.初步仿真结果表明本方法有良好应用前景.更多还原

【Abstract】 A generalized energy constrained method of multi-layer paceptron (MLP) is proposed. The new en-ergy function is composed of not only the sum of squared-error term but also some constrained terms whichimpose some desired features on network performance. Some typical constrain forms (on weight vectors orhidden-unit outputs or network structure) are discussed.This method is applied to solve the SVD problem whichis extensively used in signal processing. The method can also be extended to Primcipal Component Analysis (PCA).It is a highly flexible method which makes parallel processing possible. Apriori knowledge about the signal featurecan be incorporated into the energy function constraint. The orthogonality requirement on signal and transfer ma-trix for signal separation can be somewhat relaxed. Initial simulation results are encouraging.更多还原