Control of memory, active perception, and action in minecraft

摘要: In this paper, we introduce a new set of reinforcement learning (RL) tasks in Minecraft (a flexible 3D world). We then use these to systematically compare and contrast existing deep (DRL) architectures with our memory-based DRL architectures. These are designed emphasize, controllable manner, issues that pose challenges for RL methods including partial observability (due first-person visual observations), delayed rewards, high-dimensional observations, the need active perception correct manner so as perform well tasks. While conceptually simple describe, by virtue having all simultaneously they difficult current Additionally, evaluate generalization performance on environments not used during training. The experimental results show generalize unseen better than