The roots of backpropagation: from ordered derivatives to neural networks and political forecasting
The roots of backpropagation: from ordered derivatives to neural networks and political forecasting
Nonlinear black-box modeling in system identification: a unified overview
Automatica (Journal of IFAC) - Special issue on trends in system identification
Nonlinear black-box models in system identification: mathematical foundations
Automatica (Journal of IFAC) - Special issue on trends in system identification
Time Series Analysis: Forecasting and Control
Time Series Analysis: Forecasting and Control
Artificial Neural Networks - Forecasting Time Series
Artificial Neural Networks - Forecasting Time Series
Non-linear Prediction of Vibration Series for Turbogenerator Unit
Proceedings of the 14th International conference on Industrial and engineering applications of artificial intelligence and expert systems: engineering of intelligent systems
Time series prediction with single multiplicative neuron model
Applied Soft Computing
Journal of Global Optimization
A time-dependent enhanced support vector machine for time series regression
Proceedings of the 19th ACM SIGKDD international conference on Knowledge discovery and data mining
| Hi-index | 0.00 |
Commonly, time series prediction problems are approached either from a stochastic perspective or, more recently, from a neural network perspective. Each of these approaches has advantages and disadvantages: stochastic methods are usually fast, but of limited applicability since they commonly employ only linear models. The NN methods, on the other hand, are powerful enough, but selecting an appropriate architecture and parameters is a time-consuming trial-and-error procedure.Combining the stochastic and NN methods may prove fruitful. This article explores the possibility of rapidly designing an appropriate neural network for time series prediction based on information obtained from stochastic modeling. Such an analysis could provide some initial knowledge regarding the choice of an NN architecture and parameters, as well as regarding an appropriate data-sampling rate.The motivation for this approach is that it is much more cost-effective to select an NN architecture with the help of linear stochastic modeling than by trial and error. The objective of this study is not to obtain "the optimal" NN architecture for a given problem, but to rapidly provide an architecture with close-to-optimal performance.Experiments on both a complex real-life prediction problem (an entertainment video-traffic series) and an artificially generated nonlinear time series on the verge of chaotic behavior (Mackey-Glass series) indicate that stochastic analysis can indeed provide some useful initial knowledge on which to base neural-network design. Although not necessarily optimal, such rapidly designed NN models performed comparably to or better than more elaborately designed NNs obtained through expensive trial-and-error procedures.