R
Rosie
Text
Unsupervised Detection of Bifurcation Drivers with a VAE-Reservoir Framework for Anticipating Critical Transitions
Removes the requirement that parameter-driven reservoir computing know the bifurcation parameter explicitly, by prepending a variational autoencoder that detects the hidden driving factor from time series in an unsupervised fashion; the extracted parameter is fed to a reservoir computer to forecast an approaching critical transition. Demonstrated on prototypical systems including the spatiotemporal Kuramoto-Sivashinsky equation, and extended to multiple independent driving parameters and partial state observations.