Local Forward Model Learning for Sample-Efficient Sequential Decision Making in Open World 3D Games

视频游戏作为最先进的机器学习算法的平台已经有一段时间了。拥有一个定制的虚拟环境为我们提供了开发越来越复杂的算法来一步一个脚印地解决实际生活问题的可能性。这项拟议的研究将使用Malmo作为平台，在像《我的世界》这样的复杂3D部分可观察游戏中制作样本高效学习模型，并实现统计提前规划模型。一个高效的本地学习模型，与在各种2D游戏中成功的统计前向规划方法相结合，将允许过渡到3D环境，这将打破目前的障碍，使我们离创造通用人工智能的目标更近一步。

Video games have been a platform for state of the art machine learning algorithms for some time now. Having a custom virtual environment provides us with the possibility to develop more and more complex algorithms to tackle practical life problems one step at a time. The proposed study will use MALMO as a platform to make sample efficient learning models and implement statistical forward planning models in a complex 3D partially observable game like Minecraft. An efficient learned local model, combined with Statistical Forward Planning methods that have been successful in various 2D games would allow the transition to the 3D environment which would breach the current barrier and bring us one step closer to the goal of creating general artificial intelligence.