Using Enhanced Gaussian Cross-Entropy in Imitation Learning to Digging the First Diamond in Minecraft

Yingjie CAI, Xiao Zhang

Keywords: reinforcement learning ObtainDiamond

Abstract: Although state-ofthe-art reinforcement learning (RL) systems has led to breakthroughs in many difficult tasks, the sample inefficiency of standard reinforcement learning methods still precludes their application to more extremely complex tasks. Such limitation will make many reinforcement learning systems cannot be applied to real-world problem, in which environment samples are expensive. To solve this problem, MineRL (13) provide an ideal develop environment to facilitate the research that leveraging fewer human demonstrations with more efficient reinforcement learning systems. Based on the MineRL environmnet, we proposed an enhanced Gaussian cross entropy (EGCE) loss for imitation learnning problems to achieve ideal performance. In the ObtainDiamond task, our EGCE achieves about 7.7% improvement than a strong baseline imitation learning pipeline. The demo video is available at here.