(IJCSIS) International Journal of Computer Science and Information Security,Vol. 8, No. 6, 2010
river regimes, it would appear to be appropriate thatalternative representations for flood forecasting should beconsidered.II.M
In this study different parameters like learning rule andactivation function are employed for rainfall-runoff modelingusing Radial basis function neural network model of artificialneural network.Radial basis functions networks have a very strongmathematical foundation rooted in regularization theory forsolving ill-conditioned problems.The mapping function of a radial basis functionnetwork, is built up of Gaussians rather than sigmoids as inMLP networks. Learning in RBF network is carried out in twophases: first for the hidden layer, and then for the output layer.The hidden layer is self-organising; its parameters depend onthe distribution of the inputs, not on the mapping from the inputto the output. The output layer, on the other hand, usessupervised learning (gradient or linear regression) to set itsparameters.
Figure 1. The Radial basis function neural network
In this study we applied different learning rules to theRBF neural network and studied the optimum performancewith different activation function. We applied Momentum,Deltabar Delta, Levenberg Marquardt , Conjugate Gradient,Quick prop learningrule with activation function Tanh, LinearTanh, Sigmoid and Linear Sigmoid.
The learning and generalization ability of the estimatedNNmodel is assessed on the basis of important performancemeasures such as MSE (Mean Square Error), NMSE(Normalized Mean Square Error) and r (Correlationcoefficient)
MSE (Mean Square Error):
The formula for the mean square error is:
PN ijij ji
dy MSE NP
… (1)WhereP = number of output PEs,N = number of exemplars in the data set,
= network output for exemplar i at PE j,
= desired output for exemplar i at PE j.
NMSE (Normalized Mean Square Error):
The normalized mean squared error is defined bythe following formula:
NN ijijPii j
PNMSE NMSE Ndd N
… (2)WhereP = number of output processing elements,N = number of exemplars in the data set,MSE = mean square error,
= desired output for exemplar i at processingelement j.
r (correlation coefficient):
The size of the mean square error (MSE) can be usedto determine how well the network output fits the desiredoutput, but it doesn't necessarily reflect whether the two sets of data move in the same direction. For instance, by simplyscaling the network output, the MSE can be changed withoutchanging the directionality of the data. The correlationcoefficient (r) solves this problem. By definition, thecorrelation coefficient between a network output x and adesired output d is:
xxdd N r ddxx NN