Retrieving Game States with Moment Vectors

Game scholars need to find moments in games that advance their arguments, and artificial intelligence algorithms need to recall states that are most promising for exploration. This paper considers the problem of engineering a representation for game states that is suitable for retrieval in the vector space model. Retrieving moments from gameplay traces for two popular Super Nintendo Entertainment System games, we evaluate several different representations including one derived from a deep embedding of screenshot pixels based on a supervised memory prediction task. The results suggest compact moment vectors may be a promising representation for building future systems that intend to build higher level knowledge about games.