Robust forecasting using predictive generalized synchronization in reservoir computing

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Complex Systems

Background:

• Reservoir computers (RCs), a type of recurrent neural network (RNN), excel at time series forecasting.
• Hyperparameter selection for RCs is challenging, impacting forecasting accuracy.
• Predictive generalized synchronization (PGS) offers a potential solution for optimizing RC design.

Purpose of the Study:

• To analyze a PGS-based method for guiding the design and hyperparameter selection of RCs.
• To introduce an efficient pre-training test for determining PGS occurrences.
• To establish a robust evaluation metric for RC prediction capabilities.

Main Methods:

• Utilizing predictive generalized synchronization (PGS) to inform RC architecture and hyperparameter choices.
• Employing an auxiliary method for computationally efficient pre-training tests to identify PGS.
• Evaluating RCs by measuring the reproduction of input system's Lyapunov exponents.

Main Results:

• The PGS method provides a clear direction for designing and evaluating RCs.
• The auxiliary pre-training test efficiently guides hyperparameter selection.
• The Lyapunov exponent reproduction metric demonstrates robust prediction capabilities.

Conclusions:

• The PGS-based approach significantly improves the design and evaluation of reservoir computers for time series forecasting.
• This method offers a computationally efficient and robust framework for optimizing RNN hyperparameters.
• The proposed evaluation metric ensures reliable prediction performance in RCs.