Impact of weak generalized synchronization on time series forecasting using reservoir computers

Area of Science:

• Complex Systems
• Nonlinear Dynamics
• Machine Learning

Background:

• Echo state networks (ESNs) are recurrent neural networks effective for time series forecasting.
• Generalized synchronization (GS) describes a state where two coupled chaotic systems exhibit synchronized dynamics.
• Understanding the interplay between ESN dynamics and GS is key to optimizing forecasting performance.

Purpose of the Study:

• To investigate the relationship between ESN forecasting performance and generalized synchronization dynamics.
• To analyze the transversal stability of GS in ESNs driven by chaotic time series.
• To determine how noise affects the optimal operating point for ESN forecasting.

Main Methods:

• Treated ESNs as response systems driven by chaotic input.
• Analyzed transversal stability of GS using conditional Lyapunov exponents.
• Employed a replica synchronization error-based method to distinguish between strong and weak GS.

Main Results:

• Forecasting performance does not peak at the edge of conditional stability.
• Optimal performance for noise-free data occurs near the breakdown of strong GS conditions.
• With observational noise, optimal performance shifts towards the bubbling transition, characterized by reduced transversal contraction.

Conclusions:

• Transversal stability is a critical factor in determining ESN forecasting performance.
• The presence of noise shifts the optimal ESN operating regime.
• Performance degrades significantly when bubbling dynamics become dominant.