Porovnání učících algoritmů neuronových sítí pro predikci reálných dat

Neural network learning algorithms comparison on numerical prediction of real data

conference paper

en

In this paper we concentrate on prediction of future values based on the past course of a variable, Traditionally this task is solved using statistical analysis – first a time-series model is constructed and then statistical prediction algorithms are applied to it in order to obtain future values. This paper describes two learning algorithms for training Multi-layer perceptron networks, widely known Back propagation learning algorithm and Levenberg- Marquardt algorithm. Both of these methods are applied to solve prediction of real numerical time series represented by Czech household consumption expenditures. Tested dataset includes twenty-eight observations between the years 2001 and 2007. The observations are represented by quarterly data and the goal is to predict three future values for first three quarters of 2008. Predicted values of both experiments are compared with measured values. In the next step, a comparison of neural network topology efficiency regarding to learning algorithms is made.

Práce je orientována na oblast predikce a optimalizace reálných dat pomocí metod umělých neuronových sítí v rámci rozhodovacího procesu. Jedná se o hledání různých vhodných typů, topologií a učících algoritmů umělých neuronových sítí na řešení predikce reálných datových souborů. Optimalizace je v kontextu řešení práce zaměřena na použité topologie neuronových sítí, optimalizace učících algoritmů a vlastního výpočtu neuronovými sítěmi.

In this paper we concentrate on prediction of future values based on the past course of a variable, Traditionally this task is solved using statistical analysis – first a time-series model is constructed and then statistical prediction algorithms are applied to it in order to obtain future values. This paper describes two learning algorithms for training Multi-layer perceptron networks, widely known Back propagation learning algorithm and Levenberg- Marquardt algorithm. Both of these methods are applied to solve prediction of real numerical time series represented by Czech household consumption expenditures. Tested dataset includes twenty-eight observations between the years 2001 and 2007. The observations are represented by quarterly data and the goal is to predict three future values for first three quarters of 2008. Predicted values of both experiments are compared with measured values. In the next step, a comparison of neural network topology efficiency regarding to learning algorithms is made.

Back Propagation, Levenberg-Marquardt, Prediction of Time Series, Neural Networks

2010

23.06.2010

978-80-214-4120-0

Mendel 2010

16

280–285

6