Resource rationality of fast, goal-directed simulation for novel game reasoning

Determine to what extent the intuitive game theory model that evaluates novel two-player grid games via fast, search-limited, goal-directed agents and partial game simulations is resource rational, and characterize the extent and specific ways naive human reasoners approximate answers to game outcome and fun judgments using resource-rational representations.

Background

The paper introduces an intuitive game theory model that simulates bounded, goal-directed play to make rapid probabilistic judgments about novel grid-based games (e.g., variants of m-in-a-row on different boards). The agent uses shallow lookahead and utility components that capture center placement preferences, progress toward one’s own win condition, and blocking the opponent’s progress. Outcome judgments are derived from a small number of partial simulations rather than exhaustive search to terminal states.

In the discussion, the authors connect their approach to the broader framework of resource rationality, which studies how people make near-optimal decisions given limited cognitive resources. They explicitly raise the question of whether their model itself meets the criteria of resource rationality and how naive human reasoning about these intuitive game evaluations aligns with resource-rational representations. This establishes a concrete open problem concerning the normative status of the model and the cognitive mechanisms underlying human judgments in such tasks.

References

It is an open question to what extent our current model can properly be called resource rational", or indeed, to what extent and in which ways naive humans approximate answers to these kinds of intuitive questions usingresource rational" representations.

— People use fast, goal-directed simulation to reason about novel games  (2407.14095 - Zhang et al., 2024) in Conclusion and future directions

The unresolved issue is when the cost of another experiment is justified during play.

— AI for Games in the Foundation Model Era  (2609.16679 - Luo et al., 15 Sep 2026) in Section 9.1, “Play and Act: Transfer Across Interfaces and Rules”

Although neural networks, in conjunction with classical methods, currently represent the state of the art in gameplaying AI, achieving victory is not the only goal. There are many other open questions in game theory, such as the relation between player performance and complexity of his/her behavior and underlying decision-making processes , the effects of diversity of playing strategies in multiagent gaming environments , , , or the mechanisms of appearance of power laws and memory effects in game-history databases , and in models' training and inference datasets .

— Turn-Based Combat Arena: A New Framework for Multiagent Training and Game Balancing  (2609.03122 - Vasyuta et al., 2 Sep 2026) in Introduction