Towards sample efficient deep reinforcement learning in collectible card games

Collectible card games (CCGs) are widely-played games in which players build a deck from a set of custom cards and use it to battle each other. They are notoriously more challenging than games such as Go and Texas Hold'em Poker, the protagonists of recent breakthroughs in game-playing AI. Deep reinforcement learning approaches have recently become state-of-the-art in CCGs, although requiring huge amounts of computational power to train. In this paper, we propose a collection of deep reinforcement learning approaches to battling in CCGs that are trainable on a single desktop computer. Each approach tries different mechanisms to increase sample efficiency. We use Legends of Code and Magic, a CCG designed for AI research, as a testbed and compare our approaches to each other, considering their win rate and other metrics. Then, we discuss the position of our approaches regarding the current literature, their limitations, directions of improvement, and extension to commercial CCGs.