Learning Combat in NetHack

Combat in roguelikes involves careful strategy to best match a large variety of items and abilities to a given opponent, and the significant scripting effort involved can be a major barrier to automation. This paper presents a machine learning approach for a subset of combat in the game of NetHack. We describe a custom learning approach intended to deal with the large action space typical of this genre, and show that it is able to develop and apply reasonable strategies, outperforming a simpler baseline approach. These results point towards better automation of such complex game environments, facilitating automated testing and design exploration.