From neurons connected by axons to Facebook profiles connected by friendships, interaction networks lie all around us. In new work recently published in Physical Review X, Amitava Banerjee, Joseph D. Hart, Rajarshi Roy and Edward Ott  applied machine learning tools to formulate and test a new approach to working out such interaction networks solely from the data of their observed behavior over time.

To do so, the researchers trained an artificial neural network to mimic the observed time evolution of the unknown system. They then tracked the spread of disturbances in that trained neural nSchematics of the RC trained for predicting the time series k time steps ahead. Lower: the four time series represent scalar components of X[t].etwork and used that information to infer the network structure of the original system. The method is particularly suited for the common but hard-to-solve situations—where the network dynamics are noisy, and the cause-and-effect interactions are time-lagged. The team also tested this technique on experimental and computer-simulated data from opto-electronic networks—an excellent testbed for complex dynamics—and showed that the technique is extremely effective. Determining the underlying interaction network is a key step towards understanding, predicting, and controlling the behavior of many complex dynamical systems. As such, this method offers the promise of widespread future impact for the study of networks and dynamics.

To read more, see the paper  "Machine Learning Link Inference of Noisy Delay-coupled Networks with Opto-Electronic Experimental Tests", in Phys. Rev. X 11, 031014.