Laplacian Echo State Network for Multivariate Time Series Prediction.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Echo State Networks (ESNs) are recurrent neural networks effective for time series prediction.
• A common issue with ESNs is the ill-posed problem arising from insufficient training samples relative to hidden layer size.
• This limitation hinders the accurate mapping of complex real-world data dynamics.

Purpose of the Study:

• To propose a novel Echo State Network architecture, the Laplacian Echo State Network (LAESN).
• To overcome the ill-posed problem in ESNs caused by limited training data.
• To achieve low-dimensional output weights for improved time series analysis.

Main Methods:

• An ESN maps multivariate time series data into a high-dimensional reservoir.
• Laplacian eigenmaps are employed to estimate the underlying manifold within the reservoir by constructing an adjacency graph.
• Output weights are subsequently calculated using this estimated low-dimensional manifold.

Main Results:

• The LAESN model effectively addresses the ill-posed problem inherent in traditional ESNs.
• The method successfully obtains low-dimensional output weights.
• Experimental validation on two real-world datasets confirms the LAESN's effectiveness and unique characteristics.

Conclusions:

• The Laplacian Echo State Network (LAESN) offers a robust solution for time series prediction with limited data.
• LAESN enhances the capability of ESNs to handle complex dynamics by leveraging manifold estimation.
• The model demonstrates practical applicability and improved performance in real-world scenarios.