EnsembleDAgger: A Bayesian Approach to Safe Imitation Learning