Javascript must be enabled for the correct page display

The Gated experts network and time series prediction

Bron, J. and Zijlstra, R.M. (1998) The Gated experts network and time series prediction. Master's Thesis / Essay, Computing Science.

Infor_Ma_1998_JBron.CV.pdf - Published Version

Download (2MB) | Preview


One of the main issues in the research on time series is its prediction. Artificial neural networks are suitable tools for this purpose. Most traditional models are global models,assuming stationary. This ignores the fact that most real—world time series are non—stationary. An important subclass of non—stationary is piecewise stationary, where the series switches between different stationary regimes. Using an artificial neural network for the prediction in each regime, solves the problem of non—stationarity. To predict the transitions between the regimes, an additional artificial neural network can be used,assuming that these transitions are unknown. The gated experts network combines these properties. Key elements are: non—linearity, predicting regime transitions and local predictors for each regime. The Expectation—Maximization learning algorithm is used to update the free parameters of the network. Our goal is to gain insight in the application of the gated experts network to real—world time series prediction. This is achieved by studying the gated experts network, implementing it in InterAct© and conducting several experiments. The gated experts network is a reasonable good choice for real—world time series prediction. It uses the experts as local predictors and the gate plausibly allocates the experts to local regions of the input space. The gate splits the input space, but not always as one might expect. This input space splitting depends on the initialization of the weights of the gate and experts. The choice of the free parameters depends on the kind of experiment conducted. This network is a useful tool for analyzing the underlying dynamics of a time series. The gated experts network can be modified by adding different density functions to individual experts, applying dynamic growth or pruning of the number of experts and hidden units of the experts. The implementation of this network can be extended to include separate tapped delay lines for the experts and the gate, to capture the periodicities in the time series.

Item Type: Thesis (Master's Thesis / Essay)
Degree programme: Computing Science
Thesis type: Master's Thesis / Essay
Language: English
Date Deposited: 15 Feb 2018 07:29
Last Modified: 15 Feb 2018 07:29

Actions (login required)

View Item View Item