Abstract

Many real-world systems studied are governed by complex, nonlinear dynamics.By modeling these dynamics, we can gain insight into how these systems work,make predictions about how they will behave, and develop strategies forcontrolling them. While there are many methods for modeling nonlinear dynamicalsystems, existing techniques face a trade off between offering interpretabledescriptions and making accurate predictions. Here, we develop a class ofmodels that aims to achieve both simultaneously, smoothly interpolating betweensimple descriptions and more complex, yet also more accurate models. Ourprobabilistic model achieves this multi-scale property through a hierarchy oflocally linear dynamics that jointly approximate global nonlinear dynamics. Wecall it the tree-structured recurrent switching linear dynamical system. To fitthis model, we present a fully-Bayesian sampling procedure using Polya-Gammadata augmentation to allow for fast and conjugate Gibbs sampling. Through avariety of synthetic and real examples, we show how these models outperformexisting methods in both interpretability and predictive capability.

 

Introduction (beta)

None

 

Conclusion (beta)

None