TradingStone: Benchmark for Multi-Action and Adversarial Games

This paper introduces TradingStone, a simplification of the zero-sum game trading card game HearthStone that preserves the core elements of the game that challenge existing agents, namely: multi-action turns that require the management of a pool of resources; partially observable boards; Highly dimensional card spaces. However, it removes elements that are specific only to certain cards that render many existing approaches unusable, such as, effects that generate new arbitrary cards that lead to very sparse search spaces. We show that the core elements that can be studied with TradingStone are still unsolved by existing search and RL approaches and as such, it is a milestone that must precede solving Hearthstone.