Adaptive Elastic Echo State Network for Multivariate Time Series Prediction.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Time Series Analysis

Background:

• Echo State Networks (ESNs) are effective recurrent neural networks for time series prediction.
• High-dimensional reservoirs in ESNs can lead to collinearity problems, impacting multivariate time series prediction accuracy.

Purpose of the Study:

• To propose a new model-adaptive elastic ESN to address collinearity in high-dimensional reservoirs.
• To achieve sparse solutions and unbiased estimations in multivariate time series prediction.

Main Methods:

• Developed a model-adaptive elastic ESN incorporating an adaptive elastic net algorithm.
• Combined quadratic regularization with adaptively weighted lasso shrinkage for weight calculation.
• Evaluated the model on benchmark multivariate chaotic datasets and real-world applications.

Main Results:

• The proposed model effectively overcomes the collinearity problem in high-dimensional ESNs.
• The adaptive elastic ESN demonstrated superior performance in multivariate time series prediction tasks.
• Experimental results validated the model's effectiveness and desirable characteristics.

Conclusions:

• The model-adaptive elastic ESN offers a robust solution for collinearity in time series prediction.
• This approach enhances the reliability and accuracy of ESNs for complex datasets.
• The proposed method provides an unbiased estimation and enjoys the oracle property.