Reservoir Dynamic Interpretability for Time Series Prediction: A Permutation Entropy View.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Time Series Analysis

Background:

• Echo State Networks (ESNs) are efficient Recurrent Neural Networks (RNNs) for time series prediction.
• The
• black-box
• nature of ESN reservoirs limits their interpretability and development.
• Existing interpretability methods offer limited perspectives on reservoir modeling.

Purpose of the Study:

• To propose a novel framework for ESN reservoir interpretability.
• To establish a clear relationship between reservoir richness and projection capacity.
• To enhance the understanding of ESN dynamics for improved performance.

Main Methods:

• Developed a reservoir interpretability framework based on permutation entropy (PE) theory.
• Extracted instantaneous and time-varying reservoir states.
• Applied phase space reconstruction, sorting, and entropy calculation (ISE, GSE).
• Conducted multiscale complexity-entropy analysis and Pearson correlation for performance evaluation.

Main Results:

• Introduced Instantaneous State Entropy (ISE) and Global State Entropy (GSE) as measures of reservoir richness.
• Demonstrated that higher reservoir richness correlates with better projection capacity.
• Revealed detailed reservoir dynamics through multiscale complexity-entropy analysis.
• Established clear relationships between ESN performance and reservoir dynamics across different scales.

Conclusions:

• The proposed PE-based framework effectively enhances ESN reservoir interpretability.
• The framework provides a quantitative link between reservoir properties and predictive performance.
• This approach offers a superior method for understanding and optimizing ESNs for time series tasks.