Related Experiment Videos
Short-term prediction of chaotic time series by using RBF network with regression weights
1Department of Architecture and Computer Technology, University of Granada, Spain. irojas@atc.ugr.es
International Journal of Neural Systems
|February 24, 2001
Summary
This study introduces a novel pseudo-Gaussian basis function (PG-BF) neural network for improved function approximation. The PG-BF network demonstrates superior performance in short-term chaotic time series prediction compared to standard RBF networks.
Area of Science:
- Artificial Intelligence
- Machine Learning
- Computational Neuroscience
Background:
- Radial Basis Function (RBF) neural networks are widely used for function approximation.
- Standard RBF networks utilize symmetric Gaussian functions, limiting hidden layer flexibility.
- Constant weights in the output layer can restrict the learning capacity of RBF networks.
Purpose of the Study:
- To propose a novel framework for constructing and training a modified Radial Basis Function (RBF) neural network using pseudo-Gaussian basis functions (PG-BF).
- To enhance the flexibility and function approximation capabilities of RBF networks by introducing asymmetry and adaptive learning.
- To demonstrate the effectiveness of the proposed PG-BF network in complex prediction tasks like chaotic time series forecasting.
Main Methods:
- Modification of Gaussian basis functions into pseudo-Gaussian (PG) functions with two scaling parameters (sigma) to remove symmetry restrictions.
- Implementation of regression weights instead of constant weights in the output layer of the PG-BF network.
- Development of a sequential learning algorithm for adaptive network structure modification, including adding new hidden units and removing inactive ones.
- Utilizing a weighted average of outputs from receptive fields for overall network output calculation.
Main Results:
- The proposed PG-BF network exhibits greater flexibility in hidden layer neurons due to the modified pseudo-Gaussian functions.
- The sequential learning algorithm effectively adapts the network structure by creating and removing hidden units as needed.
- The PG-BF system demonstrated superior performance compared to standard RBF networks in the task of short-term chaotic time series prediction.
Conclusions:
- The novel PG-BF network architecture offers enhanced function approximation capabilities over traditional RBF networks.
- The adaptive learning algorithm and modified basis functions contribute to improved performance in complex prediction tasks.
- The PG-BF system represents a promising advancement for applications requiring accurate short-term forecasting of chaotic dynamics.