Deep Reinforcement Learning in a 3-D Blockworld Environment

Deep reinforcement learning is an effective method for training autonomous agents to a high level of performance on visual tasks. This work explores how reinforcement learning agents using deep Q-networks perform when visually processing 3-D virtual environments and how deeper network architectures can improve performance given the added difficulties of more complex environments. We provide results for tests on a variety of tasks in a virtual 3-D world and show that deeper convolutional neural networks lead to increased performance.