I will consider the problem of probabilistic model checking of temporal logic formulae on stochastic processes with parametric uncertainty. This is computationally challenging, as in principle one needs an expensive model checking step for each value of the model parameters. We show that the dependence of truth probabilities on parameters is differentiable, which allows us to construct an efficient statistical learning approach to solve the problem: we collect a sparse sample of parameter/ truth probability pairs, and treat them as a training set for a machine learning predictor based on Gaussian Processes. Due to the smoothness result, this approach converges to the true value in the large sample size. I will further discuss how these ideas can be generalised in a system design/ identification context.
Guido Sanguinetti is a Reader in Informatics at the University of Edinburgh, UK. His interests focus on machine learning methodologies for complex systems, in particular dynamical biological systems. He has published >60 research articles in leading international journals including Science, PNAS, Nature Communications. He is in receipt of an ERC Starting Grant and was awarded the 2012 PNAS Cozzarelli Prize in Engineering and Applied Science.
This talk is part of the lecture series on research talks by the visiting professors of the Vienna PhD School of Informatics.