Action-Conditional Video Prediction using Deep Networks in Atari Games