DRLViz: Understanding Decisions and Memory in Deep Reinforcement\n Learning

Abstract

We present DRLViz, a visual analytics interface to interpret the internal\nmemory of an agent (e.g. a robot) trained using deep reinforcement learning.\nThis memory is composed of large temporal vectors updated when the agent moves\nin an environment and is not trivial to understand due to the number of\ndimensions, dependencies to past vectors, spatial/temporal correlations, and\nco-correlation between dimensions. It is often referred to as a black box as\nonly inputs (images) and outputs (actions) are intelligible for humans. Using\nDRLViz, experts are assisted to interpret decisions using memory reduction\ninteractions, and to investigate the role of parts of the memory when errors\nhave been made (e.g. wrong direction). We report on DRLViz applied in the\ncontext of video games simulators (ViZDoom) for a navigation scenario with item\ngathering tasks. We also report on experts evaluation using DRLViz, and\napplicability of DRLViz to other scenarios and navigation problems beyond\nsimulation games, as well as its contribution to black box models\ninterpretability and explainability in the field of visual analytics.\n