Reservoir computing, a machine learning tool for robust forecasting

Reservoir computers, networks of nodes that are trainable dynamical systems, are a relatively mature machine learning approach. Under certain circumstances, they can accurately forecast the future behavior of observed data sets.

Reservoir computers have been applied to predicting the evolution of dynamical systems that change over time. This could include geophysical systems such as weather and climate forecasting, possibly economic or financial trends, and other technological systems requiring processing and forecasting the future of observed data.

The training input signal presented to the network can be generated using a known dynamical system or observed data for which a model is unknown. Finding a set of reservoir hyperparameters that will permit robust predictions is a major challenge in reservoir computing and recurrent neural networks in general.

Platt et al. use a computationally efficient numerical test to guide hyperparameter selections in reservoir computing that result in good forecasting. They show that the key ingredient in good forecasting is generalized synchronization.

Other parameters, such as the connectivity among the dynamical nodes, are fixed during the selection of the reservoir, making the training of the reservoir computer often simpler and faster than usual machine learning networks. How to constructively estimate such parameters is the focus of the group’s work.

“It’s hard to speculate in physics, but the method of reservoir computing will be a useful sidelight in machine learning,” said author Henry D. I. Abarbanel. “Many of the salient ideas in the pursuit of reservoir computing are contained in the development of methods for unidirectionally driven dynamical systems from two decades ago, and many of the uses of those methods, in everything from communication to cryptography, have been uncovered.”

Source: “Robust forecasting using predictive generalized synchronization in reservoir computing,” by Jason A. Platt, Adrian Wong, Randall Clark, Stephen G. Penny, and Henry D. I. Abarbanel. Chaos (2021). The article can be accessed at https://doi.org/10.1063/5.0066013 .