Fig. 1: A landscape of games revealed by the proposed response graph-based workflow.

This landscape is generated by collecting features associated with the response graph of each game, and plotting the top two principal components. At a high level, games whose response graphs are characteristically similar are situated close to one another in this landscape. Notably, variations of games with related rules are well-clustered together, indicating strong similarity despite their widely-varying raw sizes. Instances of Blotto cluster together, despite their payoff table sizes ranging from 20 × 20 for Blotto(5,3) to 1000 × 1000 for Blotto(10,5). Games with strong transitive components (e.g., variations of Elo games, AlphaStar League, Random Game of Skill, and Normal Bernoulli Game) can be observed to be well separated from strongly cyclical games (Rock–Paper–Scissors and the Disc game). Closely-related real-world games (i.e., games often played by humans in the real world, such as Hex, Tic-Tac-Toe, Connect Four and each of their respective Misere counterparts) are also clustered together.