Verification of Gaussian Process Dynamical Models

Recent work has shown it to be possible to integrate model building and verification. This is a desirable approach which can yield systems that are correct by design, an elegant solution to common industry demands [1]. Traditionally, we can utilise Bayes’ rule to make predictions on confidence of model parameters given data that satisfy a given property of interest. The interest can be expressed formally in temporal logic. The approach was shown successful for linear time invariant systems [2], as well as systems expressed as parameterised Markov chains [3] (basic and linear) with consideration for extensions towards non-determinism using Markov decision processes. We now aim to generalise the same framework to a new type of nonparametric dynamical system model, namely a Gaussian process dynamical model (GPDM). In the absence of a parametric model representation, we propose a novel algorithm to extend the framework of system verification.