Related Experiment Videos
Minimum description length neural networks for time series prediction.
1Department of Electronic and Information Engineering, Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong. ensmall@polyu.edu.hk
Summary
Artificial neural networks (ANN) training can be computationally intensive and lead to overfitting. This study introduces an optimal fitting routine to avoid overfitting and create smaller, effective models for chaotic time series prediction.
Area of Science:
- Computational neuroscience
- Machine learning
- Nonlinear dynamics
Background:
- Artificial neural networks (ANN) involve extensive training to fit parameters, risking overfitting.
- Traditional ANN training is computationally demanding and can lead to suboptimal data fitting.
Purpose of the Study:
- To adapt an optimal fitting scheme, previously used for RBF, to ANN for improved time series modeling.
- To develop a method that avoids overfitting in ANN by controlling neuron count and replacing intensive training with an optimal fitting routine.
Main Methods:
- Proposed an alternative scheme to traditional ANN training, replacing it with an optimal fitting routine.
- Controlled the number of neurons in the network to prevent overfitting.
- Applied the algorithm to chaotic differential equations, sunspot count data, and chaotic laser experimental data.
Main Results:
- Demonstrated that the proposed scheme leads to smaller ANN models (fewer neurons) that accurately capture system dynamics.
- Showed that the method effectively avoids overfitting in time series modeling and prediction.
- Found ANN particularly well-suited for modeling chaotic time series data due to structural differences compared to RBF.
Conclusions:
- The adapted optimal fitting scheme offers an efficient alternative to traditional ANN training for time series analysis.
- Controlling neuron count is crucial for preventing overfitting and achieving accurate models.
- ANNs exhibit strong potential for modeling complex chaotic time series data.