Curriculum Learning for agents in pixel based 3D Environments

Abstract

This thesis explores Curriculum Learning in Deep-RL. The focus is on VizDoom, an 3D environment with pixels as state representation. Two new curriculum methods are proposed. One simplifies the frames by using image processing techniques, to create an easier task as source task. The other use K-means on the frames to try to find clusters with distinct visual qualities. These clusters could potentially be used as sub-goals.