sparse-city

What should an agent do when the useful answer is hidden, looking costs money, and guessing confidently is worse than saying “I don't know”?

sparse-city is a tiny, deterministic environment for that question. An agent can buy a cell or a whole information layer, answer, or abstain. The score charges for information, penalizes wrong answers, and makes abstention visible instead of quietly treating it as failure.

fogged grid → QUERY (priced) / ANSWER / ABSTAIN → scorecard.json

Run the small argument

python -m venv .venv
source .venv/bin/activate
python -m pip install -e ".[dev]"
sparse-city demo --seed 42

The demo puts three deliberately simple baselines beside each other:

• wasteful_buyer purchases an irrelevant hazard layer and still guesses wrong;
• greedy_abstain purchases the target land-use cell and answers;
• threshold_abstain has too little budget to learn, so it abstains.

For machine-readable output:

sparse-city eval --agent greedy_abstain --seed 42
sparse-city leaderboard --seed 42
pytest -q

The agents are baselines, not trained models. The environment uses explicit dictionary actions rather than pretending to implement a Gymnasium contract it cannot satisfy cleanly.

A design choice that matters

Early versions made abstention easy to overlook. Here it has its own outcome, cost, scorecard field, and regression test. A timeout is also scored as abstention, so an agent cannot improve its apparent accuracy by wandering forever.

The default world is a generated 3×3 grid. It is intentionally small enough that a reviewer can inspect a complete trajectory. Scaling the grid is less interesting than adding agents that decide whether another query is worth its price.

Limits

• All places, prices, and layers are synthetic.
• This is a POMDP-like evaluation toy, not a city-planning model.
• Four hand-written baselines are included; no model leaderboard is claimed.
• Reward weights encode one particular preference about the cost of error.

Research lineage

I worked as a research assistant at SCI-Arc Research from May 2024 to January 2025. This is an independent implementation inspired by questions about partial urban knowledge in Views of Planet City, How Cities See, Backyard Home Data Explorer, and Multifamily Housing in Somaliland. It does not reproduce those projects or represent their communities. The original contributors are credited in ATTRIBUTION.md, with implementation provenance in PROVENANCE.md.

MIT — see LICENSE.