The ability to act in multiple environments and transfer previous knowledgeto new situations can be considered a critical aspect of any intelligent agent.Towards this goal, we define a novel method of multitask and transfer learningthat enables an autonomous agent to learn how to behave in multiple taskssimultaneously, and then generalize its knowledge to new domains. This method,termed "Actor-Mimic", exploits the use of deep reinforcement learning and modelcompression techniques to train a single policy network that learns how to actin a set of distinct tasks by using the guidance of several expert teachers. Wethen show that the representations learnt by the deep policy network arecapable of generalizing to new tasks with no prior expert guidance, speeding uplearning in novel environments. Although our method can in general be appliedto a wide range of problems, we use Atari games as a testing environment todemonstrate these methods.
translated by 谷歌翻译