People use fast and flat simulation to reason about new games

Key Points:

• Researchers manually created 121 two-player competitive strategy game variants on M × N grids, varying board sizes, win conditions (K-in-a-row), and rule structures to ensure diversity and systematic variance, including infinite board variants.
• Human experiments involved 238–314 participants across multiple studies assessing zero-shot outcome predictions, funness evaluations, human–human gameplay, and watching/predicting moves from gameplay videos, using linguistic game descriptions and interactive scratchpads.
• The Intuitive Gamer cognitive model was developed to simulate human game reasoning and play, using heuristics for goal progress, blocking opponent progress, and center bias combined via a softmax action selection, enabling efficient approximate reasoning without deep search.
• Alternative models included an Expert Gamer with deeper search and more complex evaluation, a Random Gamer, and Monte Carlo Tree Search (MCTS) as a near-optimal gameplay oracle; Intuitive Gamer showed better alignment with human data and greater computational efficiency.
• Analyses involved comparing model predictions to human judgments on game outcomes and funness, assessing action selection likelihoods, fitting admixture models to player behavior, and exploring factors influencing draw acceptance, demonstrating the Intuitive Gamer’s effectiveness as a resource-rational model of human game reasoning.