Player Identification for Collectible Card Games with Dynamic Game States

Abstract

Collectible card games are a fruitful test space for studying resource allocation and battle strategy, given that their structures promote reactionary combat styles and allow players to obtain variable amounts of combat power by expending fixed resources. However, their large action spaces also allow for flexibility in play styles, thus facilitating behavioral analysis at the individual level rather than the aggregate level. When presented with the same options and the same amount of resources, a player's selection of cards and their choice of moves gives insight into their unique play style and decision-making tendencies. As such, we use the virtual collectible card game Legends of Code and Magic to determine whether we can identify a player from their actions and, conversely, predict the future actions of a known player. Our main contributions to this task are the creation of a comprehensive dataset of Legends of Code and Magic game states and actions, as well as the first use of large transformer-based language models to address this problem.