Generative adversarial simulator

Knowledge distillation between machine learning models has opened many new avenues for parameter count reduction, performance improvements, or amortizing training time when changing architectures the teacher and student network. In case of reinforcement learning, this technique also been applied to distill policies students. Until now, policy required access a simulator real world trajectories. In paper we introduce simulator-free approach knowledge in context learning. A key challenge is having learn multiplicity cases that correspond given action. While prior work shown data-free possible with supervised by generating synthetic examples, these approaches are vulnerable only producing single prototype example each class. We propose an extension explicitly handle multiple observations per output class seeks find as exemplars reinitializing our data generator making use adversarial loss. To best knowledge, first demonstration policy. This improves over state art on networks benchmark datasets (MNIST, Fashion-MNIST, CIFAR-10), demonstrate it specifically tackles issues input modes. identify open problems distilling agents trained high dimensional environments such Pong, Breakout, Seaquest.