Structure inference for networks with general non-parametric inter-object relationships

We present an algorithm for the estimation of the structure of a class of dynamic networks in which object interactions depend on some one-dimensional function of their joint state (e.g. inter-object distance). By using a non-parametric Gaussian process prior assumption on the inter-object relationship strength the algorithm is able to infer a wide range of relationship types. We demonstrate this on a physical object tracking problem. The algorithm is able to cope with a certain degree of noise and can deal with systems involving hundreds of objects on modest hardware.