Abstract

One of the central challenges of understanding complex systems—such as financial markets, neural circuits, and cellular information processing networks—is to identify which system components are causally related. This work introduces a probabilistic framework for learning the causal structure of sparsely coupled nonlinear dynamical systems from observed time series data. The proposed algorithm adopts a continuous time Gaussian Process model of the system dynamics and provides an estimated distribution over directed network topologies representing the latent interaction among system components. The method is shown to identify robustly the topological structure of a diverse class of synthetic gene regulatory networks. (Joint work with Sandy Klemm and Karsten Borgwardt.)