Localized active learning of Gaussian process state space models