An AI city-building game explores the trade-offs of urban resilience

A planning problem turned into a game

Investing in urban resilience means making choices under constraints. Stakeholders may value different outcomes, and projects consume resources while changing the surrounding neighborhood. Jin Gao's study turns those tensions into a digital board game, then asks whether AI players can learn productive strategies through repeated play.

The source is an author version of a paper published in the proceedings of ACADIA 2025: Computing for Resilience, volume 2, pages 724–732. Its arXiv v1 submission date, November 7, 2026, is separate from the proceedings publication in 2025.

The game is non-zero-sum: one player's gain need not be another's loss. Players can benefit from improvements to shared conditions even while pursuing different priorities. That creates room for cooperation, without assuming that everyone's interests align.

The study examines behavior inside this simplified environment. It does not demonstrate that the resulting strategies would improve an actual neighborhood.

What the players learn to value

The board is a grid. Players take turns constructing projects or skipping a turn, subject to available space and resources. Their resources are money and reputation, representing financial and social capital.

Projects have immediate and continuing consequences. A park, for example, costs its builder money to maintain but earns reputation over subsequent turns. Construction also changes conditions in the surrounding three-by-three neighborhood.

These effects are prescribed by the game. Parks increase greenery but reduce vitality; houses increase density while reducing greenery. They are modeling choices, not measurements establishing that every real park or housing project has those effects.

Players receive rewards that combine personal resources with shared urban indicators. In the Tokyo experiment, those indicators include greenery, vitality and density, alongside adaptability, sustainability and flood resistance. Different weights represent stakeholders who emphasize self-interest, shared conditions or resilience.

The agents learn through multi-agent reinforcement learning: repeated interactions produce rewards that help update their decision-making strategies. Following initial comparisons, Gao uses multi-agent proximal policy optimization, or MAPPO. Its training process shares global information through a centralized evaluator, while trained agents make their own decisions during play.

Three boards, three kinds of evidence

The smallest experiment uses a four-by-four board, three players and three project types: parks, houses and shops. Across a 100-timestep episode, learned strategies approach a combined return of 7,000.

For comparison, a simulated annealing search—a method for exploring possible solutions—finds a return of 8,336. The learned result is roughly 84% of that benchmark. Crucially, the search does not establish the best possible score, so this is not evidence that the agents reached 84% of an optimum.

An eight-by-eight experiment adds terrain, existing infrastructure and resilience projects. In one illustrated self-play session, resilience-oriented projects occupy 82% of buildable cells. That example shows a learned preference under the game's incentives; it is not a reported average across all sessions.

The largest experiment uses a twelve-by-twelve grid initialized from historical maps of Tatsumi, Tokyo. Dominant land-cover or parcel features are assigned to the closest terrain or project type in the game. Four agents then explore development choices over 200 timesteps.

Across 10,000 independent simulation runs comparing trained and random policies, the trained agents achieve a mean total score of 452.1, versus 357.3 for random play. Their results are also less variable: reported standard deviations are 3.2 and 16.0, respectively.

But the higher total conceals a trade-off. Trained policies score better on adaptability, sustainability, flood resistance and greenery, while scoring lower on vitality and density. They have not improved every objective simultaneously.

A testbed, not a planning verdict

The scoring rules determine what counts as success. Beating random play demonstrates learned behavior within those rules, not superiority to professional planners or a validated forecast of urban development.

Gao identifies several limitations: square grids poorly represent irregular urban blocks, predefined indicators simplify stakeholder trade-offs, and trial-and-error training can be inefficient or unstable. Formal tests of whether the learned strategies reach a game-theoretic equilibrium remain future work.

An interface lets people replay simulations or take over a stakeholder's decisions. That offers a way to explore alternative choices, but full-scale human participant studies have not yet evaluated it.

The framework's demonstrated contribution is therefore a controlled setting for examining competing incentives. Its results make a central distinction visible: a higher combined score can still mean giving up some valued outcomes to improve others.