TLDR: A new study introduces the “Intuitive Gamer” model, explaining how people quickly understand and play novel games for the first time. The model suggests humans use fast, shallow, goal-directed, and probabilistic mental simulations, outperforming complex AI models in capturing novice behavior across game evaluation, action selection, and prediction. This research offers insights into human-like AI for novel problem-solving.
For decades, games have served as a crucial testing ground for understanding intelligence, both in humans and artificial intelligence. Much of this research has focused on how experts achieve superhuman levels of play in complex games like chess or Go. However, a new study introduces a fresh perspective, exploring how people tackle decision-making problems they’ve never encountered before, specifically through the lens of playing novel games for the very first time.
The research, titled “People use fast, flat goal-directed simulation to reason about novel problems,” by Katherine M. Collins, Cedegao E. Zhang, Lionel Wong, Mauricio Barba da Costa, Graham Todd, Adrian Weller, Samuel J. Cheyette, Thomas L. Griffiths, and Joshua B. Tenenbaum, delves into the cognitive processes of novice game players. They found that even without prior experience, people are remarkably systematic and adaptively rational in how they approach new games, whether playing them or simply evaluating their potential fairness or fun.
Introducing the Intuitive Gamer Model
To explain these abilities, the researchers developed a computational cognitive model called the “Intuitive Gamer.” This model is designed to mimic how humans reason about new problems with limited computational resources. Unlike expert AI models that perform deep, extensive searches through countless possibilities, the Intuitive Gamer operates on principles of fast, flat (depth-limited), goal-directed, and probabilistic simulation.
Imagine you’re opening a board game for the first time. You don’t analyze every possible move to the end of the game. Instead, you quickly grasp the basic goals, consider a few immediate moves that seem to advance your objective or block your opponent, and make a decision based on these quick, mental simulations. This is the essence of the Intuitive Gamer. It uses very few stochastic samples, simple goal heuristics for evaluating actions, and no deep search, making it far more computationally efficient than expert-level models.
Experiments with Novel Games and Novice Players
The study involved over 1000 participants and 121 two-player strategic board games, most of which were entirely new to the participants. The researchers assessed human judgments and decisions across various scenarios:
- Game Evaluation: Participants judged how fair, challenging, or fun a game was likely to be, purely based on its description, without playing it.
- Action Selection: Participants played a game for the first time, choosing their initial moves against another novice player.
- Action Prediction: Participants watched videos of others playing and predicted what moves other first-time players would make.
The Intuitive Gamer model quantitatively captured human judgments and decisions across all these conditions, performing significantly better than much more compute-intensive expert-level models. It even accounted for how people evaluate subjective aspects like a game’s “funness” by considering factors like game balance, how much a game rewards strategic thinking, and its expected length.
Also Read:
- Beyond Imitation: How Large Language Models Develop Strategic Thinking and Unique Heuristics
- Dyna-Mind: Teaching AI Agents to Think Ahead Through Experience and Simulation
Key Findings and Implications
The research highlights that human novice reasoning occupies a valuable middle ground between purely random behavior and computationally expensive deep search. People are not random; they employ sophisticated, yet resource-limited, simulations to make sense of new situations. The Intuitive Gamer model was found to be approximately 700 times faster than the “Expert Gamer” model (a depth-5 variant) and nearly 40,000 times faster than Monte Carlo Tree Search (MCTS), while still aligning well with human behavior.
This work offers new insights into how people rapidly evaluate, act, and make suggestions when encountering novel problems. It suggests that the mechanisms of fast, flat, goal-directed probabilistic simulation are a highly efficient and resource-rational approach to problem-solving. These findings could inform the design of more flexible and human-like AI systems that can not only solve new tasks but also determine whether a task is worth engaging with at all.
The study’s focus on “pre-expertise”—the critical first moments before repeated experience builds expertise—is a significant departure from traditional AI and cognitive science research. It opens doors for understanding how humans navigate a vast range of novel situations, from economic mechanisms to interactions with unfamiliar AI systems. For more details, you can read the full research paper here.
