Researchers from MIT and Technion, the Israel Institute of Technology, have developed an algorithm that combines imitation learning and reinforcement learning to train machines. The algorithm automatically determines when the machine should mimic the teacher and when it should explore on its own. This approach allows for more effective learning and better results. The researchers tested their method in simulations and found that it outperformed methods that used only one type of learning. The algorithm could be valuable in training machines for real-world situations where uncertainty is present, such as training a robot to navigate a new building. The researchers believe that their algorithm has the potential to improve performance in various applications, including language models and robotics. The research was supported by the MIT-IBM Watson AI Lab, Hyundai Motor Company, DARPA, and the Office of Naval Research.