Detail buku

No BukuT.LN.04.05
UniversitasChiba University, JAPAN
PenulisWindhiarso Panco Adi Putranto
PembimbingTakashi Yahagi Professor
AbstrakABSTRAK INGGRIS : Time series prediction is takes an existing series of data Xt-n, .. Xt-2, Xt-1, Xt and predict the Xt+1, Xt+2, …….., data values. The goal is to observe or model the existing data series to enable future unknown data values to be predicated accurately. Examples of data series include financial data series, physically observed data series and mathematical generated data series. Many techniques have been implemented to perform time series prediction. This thesis will focus on wavelet networks. The idea of combining both wavelets and neural networks has resulted in the formulation of wavelet networks – a feed-forward neural network with one hidden layer of nodes, whose basis function are drawn from a family of orthonormal wavelets. The use of wavelet network have been applied to speech segmentation, speaker recognition, face tracking, forecasting, and prediction of chaotic signals. The problem is how to make an algorithm for determining the parameter values of the wavelet networks and training the network to adjust the parameters of the network to minimize some function (usually the square error between the output of the network and the desired aoutput). The usual method to do this in the domain of neural networks is the backpropagation. A large fraction of recent work in artificial neural network uses multilayer perceptrons training with the backpropagation algorithm converges slowly for large or complex problems such as speech recognition, where thousands of iterations may be needed for convergence even with small data sets. The Extended Kalman Filter (EKF) is a well known tool for recursive parameter estimation, machine learning applications, nonlinear system identifications and for the purpose of training networks. In particular, the EKF has been applied to the estimation and weights training of feedforward and recurrent neural network models. Wan and van der Merwe point out the underlying assumptions and dramback in the EKF, which can introduce large errors and may lead to suboptimal performance. They presented an alternative filter with performance superior than the EKF called the unscented Kalman filter (UKF). This new method improve the drawbacks in the EKF, and perform a better result on nonlinear estimation problem, including nonlinear system identification and training of neural networks. In this thesis, we show that training wavelet networks can be solved using the unscented Kalman filter. Simulation result are demonstrated to validate the ability and efficiency of the propose network.