A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning