Policy Improvement via Imitation of Multiple Oracles