神经网络课4-45教案3 研究生课件教学教材.ppt

4.4 Introduction to special neural network architectures;0.0;3. Objective function;Example 1;Iterations;4. Structure 2-5-1;Weights deleted; ;Training process;;;Example 2;Weights deleted for four structures;Changing of the weights connected to input node 3 (structure 3-5-1);Weights changing curve;Example ;W0=?0.001;?0 =0.005;Record for redundant weights deleted with different E0;B. Autoassociative Networks;The input signal has three distinct components:;A three-layer autoassociative network ;Typical ratios are normally set between 2 and 8, depending on the signal type and noise in the system;C. Hierarchical Neural Networks;Advantages:;C1: Moving-Window Networks;Key challenge problems:;A moving-window network with a process-trend scanning window including 4 time increments;C2: Input-Compression Networks;A input-compression network;* Attaching a Dimensionality-Reduction Network to a Backpropagation Network;Developing an input-compression network involving two steps.;Out(1);D. Recurrent Networks;A recurrent network combines the feedback and feedforward connections of neural networks. It is simply a neural network with loops connecting the output responses to the input layer.;E. Radial-Basis-Function Networks;(1) Input Layer;(c) A transfer function which transforms the Euclidean summation Ik (k=1 to L) to give an output for each node.;2. Network Development;Using K-means clustering algorithm; Example of K-means clustering algorithm;(a) Initialization step;(2) Gaussian Function, Sigma ?k;(4) Weight factors, wkj;For example, after the 2000th iteration, the weight factors, cik, are fixed for the remainder of the training procedure. The weight factors wkj are then trained using a conventional backpropagation algorithm.;XOR Problem;Input;F. Input Training Neural Network ;The aim of training IT-NN is  to minimize the following objective function: ;The steepest descent direction for training network is;Effectiveness of IT-NN ;Nonlinear System Modeling with IT-NN RB