Control of Memory, Active Perception, and Action in 3D world
In this work, we introduce a set of reinforcement learning tasks in Minecraft. We use these tasks to systematically compare existing deep reinforcement learning (DRL) architectures with our new memory-based architectures. These tasks are designed to emphasize issues that pose challenges including partial observability, delayed rewards, high-dimensional visual observations, and the need to use active perception so as to perform well in the tasks. We evaluate the generalization performance of the architectures on environments not used during training. The experimental results show that our new architectures generalize to unseen environments better than existing DRL architectures.
Junhyuk is a 2rd-year PhD student in Computer Science and Engineering at University of Michigan. He is working on the intersection between deep learning and reinforcement learning under the supervision of Professor Honglak Lee and Professor Satinder Singh.