Effect of Viewing Directions on Deep Reinforcement Learning in 3D Virtual Environment Minecraft

Abstract

In recent years, deep reinforcement learning has attracted interests from AI researchers. Deep reinforcement learning is a method combining a deep neural network (DNN) and reinforcement learning (RL). By approximating a function in RL with a DNN, it enables an agent to learn in complex environment represented by low-level features, such as pixels given by a 3D video game. However, learning from low-level features is sometimes problematic. For example, a small difference in input pixels results in completely different behaviors of an agent. In this study, as an example of such problems, we focus on viewing directions of an agent in 3D virtual environment Minecraft and analyze the effect of them on the efficiency of deep reinforcement learning.