神经网络课4-45教案3 研究生课件.ppt

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 RBF-NN Combining IT-NN with RBF Results We can see that the IT-NN with 8 nonlinear PCs will be the best choice. Fewer PCs (7, 6 and 5) cause information lost, and more PCs (10 and 9) increase the input nodes of RBF. A classification network’s predictions moving from the class I decision region (y=[1,0,0]) to the II decision (y=[0,1,0]). The two most commonly used network architectures for classification problems are the BP network and the RBF network. Illustrative example of fault diagnosis of a chemical reactor low conversion [1,0,0] low catalyst selectivity [0,1,0] catalyst sintering [0,0,1] Hidden layer reactor inlet temperature, 0F reactor inlet pressure, psia feed flow rate, 1b/min Next, we use a example to compare the BP network with the RBF network. In addition, we compare different transfer functions(e.g. sigmoid and hyperbolic tangent) and network training rules(e.g., delta rule and normalized cumulative delta rule, nc-delta). Using BP network Using BP network with delta-learning rule and the hyperbolic tangent transfer function and 5 nodes in the hidden layer Using the RBF network Comparison of the BP network and RBF network RBF networks perform better than BP networks for classification problems The RBF network’s RMS error approaches 0 as the number of nodes increases, while the BP network has a much larger RMS error of 0.037. BP network has a much more compact system representation than the RBF network, and has significantly lower RMS error for smaller networks(e.g. less than 5 nodes). The RBF network also trains faster requiring only 7000 iterations, versus 30000 for the BP network. A input-compression network … Out(1) Out(2) Out(m) subgroup (1) subnet 1 subgroup (n) … hidden layer 1 hidden layer 2 hidden layer 3 sub