Computational Urbanism
We let algorithms design a neighborhood. Then we simulated it end-to-end.
How do you design a neighborhood?
Traditional masterplanning relies on intuition. The architect draws a plan based on experience and precedent. Maybe it works. Maybe it doesn't. You won't know until people move in.
We tried something different, standing on two well-tested methods. NSGA-II, the standard evolutionary algorithm for multi-objective search, proposes building arrangements. An agent-based crowd model in the lineage of Helbing's social force work walks 127,000 virtual pedestrians through each one. We tested 5,000 arrangements against conflicting goals: maximize density, maximize walkability, maximize sunlight, minimize wind.
The algorithm found something we wouldn't have drawn ourselves. Instead of spreading buildings evenly, it clustered them into five mini-centers. That pattern turns out to match what Jane Jacobs observed in thriving neighborhoods. But we discovered it computationally, and then checked the engine against reality: simulating an existing Istanbul district, our flows landed within 13 percent of published pedestrian counts.
127,000 Agents: Each colored dot is a simulated person. The heatmap shows where they congregate.
Theoretical Framework
Evolutionary Optimization
NSGA-II algorithm explores the design space. It finds solutions that balance competing objectives without human bias.
Agent-Based Simulation
127,000 virtual pedestrians with realistic behavior. They walk, shop, commute. Their patterns reveal design flaws.
Multi-Objective Balance
Density, walkability, sunlight, wind comfort. The algorithm finds trade-offs humans struggle to see.
Existing-Conditions Validation
We simulated an existing district and compared it against measured pedestrian counts; matched within 13%. The method works.
Research Process
Collect Site Data
GIS, traffic counts, demographics from municipal sources
Build Agent Model
127,000 agents calibrated to observed behavior
Run Evolution
NSGA-II tests 5,000+ variants overnight
Validate and Iterate
Compare predictions to real-world measurements
Research Phases
Data Gathering
GIS mapping, municipal traffic counts, published demographic data. The unglamorous groundwork before we touched the optimizer.
Agent Calibration
Walking speeds, route choices, and stopping patterns calibrated from published pedestrian-behavior studies, and sanity-checked the honest way, on foot in Kadikoy.
Evolution Runs
5,000+ design variants tested overnight. Each one simulated, scored, and ranked against objectives.
Validation
The Phase 1 scenario is simulated end-to-end. We compare its predictions against published existing-conditions counts. So far: 87% match.
Evolve, Then Let Them Walk: NSGA-II proposes, 127,000 agents judge, and one existing district keeps the model honest. 5,000 variants overnight.
Key Metrics
Key Thinkers
Jane Jacobs
Jacobs watched New York sidewalks for years. She saw that thriving streets need mixed uses, short blocks, and eyes on the street. Our algorithm independently discovered similar patterns.
Christopher Alexander
Alexander's Pattern Language proposed design rules that emerge from human behavior. Our agent-based approach generates those patterns computationally.
Jan Gehl
Gehl invented pedestrian counting as urban research. Our walkability metrics directly extend his methods.
Kevin Lynch
Lynch identified the elements that make cities legible: paths, edges, nodes. Our algorithms optimize for his criteria.
Case Studies
Campus Masterplan Concept
Concept study, not commissionedA 50-hectare university-campus scenario, run on our own concept study. Buildings 'found' their positions through 5,000 iterations. Average walking distance between academic buildings dropped 23% in the model.
Kadikoy Existing-Conditions Study
Istanbul, TurkeyWe simulated an existing district and compared the output against published pedestrian counts. 87% correlation. Encouraging; one district is a calibration, not proof.
Comparative Analysis
Traditional Masterplan
Architect Draws, City BuildsBased on precedent and intuition. Sometimes brilliant, sometimes disastrous. No way to test before construction.
Parametric Urbanism
Rules Generate FormGrasshopper definitions produce variants. Better than manual, but still designer-driven.
Agent-Based Modeling
Simulate Before BuildingVirtual pedestrians test the design. Problems show up before ground is broken.
Our Approach
Evolve + SimulateEvolutionary algorithms generate options. Agent simulations test them. Only validated designs survive.
Optimization Results
Where should buildings be tallest? The algorithm figured it out.
Scenario model: values from our own simulation runs
Key Findings
Polycentric beats uniform. When we optimized for both density AND walkability, the algorithm consistently produced 5 mini-centers, not an even spread.
5 clusters optimalSmall blocks emerge naturally. When walkability is weighted high enough, the algorithm produces blocks under 100m perimeter. Jane Jacobs was right.
<100m blocksPrediction looks possible. 87% correlation between our simulated flows and published pedestrian counts for one district. Urban design can be evidence-based.
87% accuracySpeed matters. Testing 5,000 variants in 24 hours means we can explore design spaces humans never could.
5,000 in 24 hoursHonest Limitations
Garbage in, garbage out. If our GIS data is wrong, so are our predictions.
Wind simulation is simplified. Full CFD at district scale would take weeks.
Agents calibrated to Istanbul. Different cities might behave differently.
Static optimization. Cities change over decades. Our model captures one moment.
Validation leans on published pedestrian counts for a single district. One district is a calibration, not proof.
References & Data Sources
Literature
Deb et al., A Fast and Elitist Multiobjective Genetic Algorithm: NSGA-II (2002)
IEEE Trans. Evolutionary Computation. The optimizer this study runs on.
↗Helbing and Molnar, Social Force Model for Pedestrian Dynamics (1995)
Physical Review E. The lineage of our agent behavior model.
↗Michael Batty, The New Science of Cities (2013)
Cities as flows and networks: the theoretical frame for simulating them.
Jane Jacobs, The Death and Life of Great American Cities (1961)
The observations our algorithm keeps rediscovering: mixed use, short blocks, mini-centers.
Kevin Lynch, The Image of the City (1960)
Paths, edges, nodes, landmarks: the legibility criteria we score against.
Jan Gehl and Birgitte Svarre, How to Study Public Life (2013)
The counting methods behind the pedestrian data we validate against.
Data Sources
Conclusion
Urban design doesn't have to be guesswork. With 127,000 simulated pedestrians and an 87% match against published counts, neighborhoods can be tested before anyone builds them. The algorithm found patterns Jane Jacobs described, but found them computationally. That's the future of planning.
Limitations
- Best for greenfield sites
- Static model for now
Future Directions
- Metropolitan scale expansion
- Real-time monitoring integration