A leap forward in the accuracy of forecasting problems in chaotic time series has been recently obtained using Echo State Networks (ESN). This is a novel kind of recurrent neural network with a randomly connected hidden layer, called the reservoir, and an adaptable output layer trained using a simple regression algorithm. This architecture has the advantage of faster computation and presents less parameter tuning then a standard neural network. Since their inception there has been a lot of effort towards the improvement of the model and applications have been found in several fields of study: anomaly detections in geophysics , short term predictions in astrophysics and reconstruction of chaotic attractors among others.

This project aims to build a comprehensive Julia package in which are implemented the majority of the variations of the ESNs presented in the literature. Such an effort would represent the first attempt to create a library for this kind of models in any programming language, and would reap benefits not only for the consequent simplicity of application, but also for showing the actual state of the art in a family of models that has yet to go mainstream.





  • Ranjan Anantharaman
  • David P. Sanders
  • ChrisRackauckas