A Modular Deep-learning Environment for Rogue

摘要

Rogue is a famous dungeon-crawling video-game of the 80ies, the ancestor of its gender. Due to their nature, and in particular to the necessity to explore partially observable and always different labyrinths (no level replay), roguelike games are a very natural and challenging task for reinforcement learning and Q-learning, requiring the acquisition of complex, non-reactive behaviours involving memory and planning. In this article we present Rogueinabox: an environment allowing a simple interaction with the Rogue game, especially designed for the definition of automatic agents and their training via deep-learning techniques. We also show a few initial examples of agents, discuss their architecture and illustrate their behaviour.