Principles of neural network model selection with applications to time series analysis (36)

Theme Track: Methods of Spatial Analysis - Spatial Network Modelling

Authors:
Fischer, Manfred M. ; Koller, Wolfgang

Learning from examples, the problem for which neural networks were designed to solve, is one of the most important research topics in artificial intelligence. A possible way to formalize learning from examples is to assume the existence of a function representing the set of examples and, thus,enabling to generalize. This can be called a function reconstruction from sparse data (or in mathematical terms, depending on the required precision, approximation or interpolation, respectively). Within this general framework, the central issues of interest are the representational power of a given network model (or, in other words, the problem of model selection) and the procedures for obtaining the optimal network parameters. The tasks of parameter estimation and model selection are of crucial importance for the success of real world neural network applications. Model selection or the specification of a network topology is a key methodological issue (see, for example, Fischer, 2000) and the focus in this paper.

This issue is widely neglected in neural network modelling. The purpose of this paper is to suggest different model selection strategies for feedforward neural networks that are based on statistical concepts. The building blocks of the strategies rely on hypothesis tests, information criteria and crossvalidation. The efficacy of the strategies is demonstrated in the context of time series forecasting vis-a-vis to heuristic selection strategies such as regularisation and pruning. The testbed for the evaluation uses monthly unemployment rates in Austria.

[Home]    [Sitemap]