DeepLearningと曲がったパラメータ空間 (Scalable trust-region method for deep reinforcement learning using Kronecker-factored approximation)