Related Experiment Videos
A late-stopping method for optimal aggregation of neural networks
P M Granitto1, P F Verdes, H D Navone
1Instituto de Física Rosario, Consejo Nacional de Investigaciones Cientí ficas y Técnicas and Universidad Nacional de Rosario, Blvd. 27 de Febrero 210 Bis, 2000 Rosario, Argentina. granitto@ifir.edu.ar
International Journal of Neural Systems
|September 28, 2001
Summary
This study introduces a novel method for building artificial neural network ensembles, balancing accuracy and diversity for improved performance. The approach enhances predictions, including solar activity forecasting.
Area of Science:
- Artificial Intelligence
- Machine Learning
- Computational Statistics
Background:
- Artificial neural networks (ANNs) are used for classification and regression.
- Ensembles of ANNs offer better generalization than single networks.
- Effective ensembles require individual networks to be both accurate and diverse.
Purpose of the Study:
- To propose a simple method for constructing ANN ensembles.
- To achieve an optimal balance between accuracy and diversity in ensemble members.
- To improve the performance of regression and classification tasks.
Main Methods:
- Developing a method to create overtrained aggregate members for ensembles.
- Ensuring an adequate balance between accuracy and diversity among ensemble members.
- Benchmarking the proposed method against existing ensemble techniques.
Main Results:
- The proposed method outperforms other recently published techniques.
- Demonstrated improved performance on standard statistical databases.
- Successfully applied to predict sunspot time series for solar cycle 23.
Conclusions:
- The developed method effectively balances accuracy and diversity in neural network ensembles.
- The approach offers superior performance compared to existing methods.
- The technique has practical applications, such as in solar activity prediction.