RL-Building Generator
Agent-based Reinforcement Learning to Increase Housing Density in London
Project Overview
This research project addresses London's critical housing crisis through innovative application of agent-based reinforcement learning. With housing prices rising 130% between 2005-2023 and the city meeting only a fraction of its housing expansion targets, our study proposes an AI-driven solution to optimize urban density while maintaining livability standards.
The project combines advanced machine learning techniques with urban planning principles, utilizing multi-agent systems to identify optimal densification strategies. Through comprehensive site digitalization and intelligent agent behavior modeling, we create automated tools that can support planners, architects, and policymakers in making data-driven decisions for sustainable urban development.
Housing Crisis in London
Price Surge
130%
Housing price increase between January 2005 and January 2023
Low Delivery Rate
< 50%
Of original housing expansion plan met in 2023
London faces the highest rents in the country, with particularly acute challenges in areas like Waltham Forest, which has the fourth highest overcrowding rate in Outer London at 18% of homes classified as overcrowded. This project targets these critical areas for intelligent densification solutions.
Methodology
Site Digitalization
2D/3D mapping, land use analysis, building indexing
Agent Modeling
Multi-agent systems with ground and roof agents
RL Training
120,000 step training runs with reward optimization
Optimization
Density increase while maintaining livability
Our approach integrates comprehensive site analysis with intelligent agent-based modeling. The digitalization process captures multiple data layers including land use patterns, building functions, solar exposure, and spatial relationships through quadtree algorithms for empty space identification.
The multi-agent reinforcement learning system employs distinct agent types - ground agents for horizontal expansion and roof agents for vertical densification - each trained through extensive simulation runs to optimize housing density while preserving environmental quality and regulatory compliance.
Site Digitalization & Agent Types
Land Use Analysis
- • Residential areas identification
- • Commercial zone mapping
- • Green space preservation
Building Index
- • Height and density mapping
- • Function classification
Solar Analysis
- • Ground radiation mapping
- • Shadow impact assessment
Agent
Move
- • 6 directions
- • up,down,front,back,left,right
Occupy
- • 2 options
- • Occupied or not
Coodinate
- • 6 directions
- • up,down,front,back,left,right
Neighbor
- • 25 + 8 + 1 positions
- • Also check celltypes
If agent has neightbors underneath
- ✅AddReward(0.04f * underbuildingCell)
- ✅AddReward(2.0);(Exactly above)
- ❌AddReward(-2.0);(No exactly above cube)
If agent has neightbors surrounding
- ✅AddReward(1.0)
- ❌AddReward(-1.0)
- • The first cube always gets reward
If agent's footprint in residential area
- ✅AddReward(1.0)
- ❌AddReward(-1.0)
If agent's footprint away from building
- ✅AddReward(distance - 2.0)
If agent on ground
- ✅AddReward(0.4 - (0.1 * floor));
- • Only the first six cubes
BoundingBox Reward
- Compactness = volumeRatio - diagonalRatio
- ✅AddReward(Compactness > 0)
- ❌AddReward(Compactness < 0)
Mean Solar Index Reward
- ☀️MeanSunIndex = Sum(sun_map) / (width * length)
- SolarDelta = (MeanSunIndex - InitialMeanSunIndex) * 50 (negative)
- From the ground cast a ray to sun and check obstacles, if not exists, the value of this position in the sun_map +1.
- select winter solstice’s sunlight vector per hour in London as our sunvector input
Training Results
Orthogonal Building Plot
Non-orthogonal Building Plot
Plot without Greenery
Key Performance Metrics
Training Parameters:
- • Training Steps: 120,000 per run
- • Training Duration: ~1.2 hours
- • Site Dimension: 400 x 400 x 40m
- • Voxel Resolution: 8 x 8 x 8
Optimization Targets:
- • Maximize housing density
- • Improve sunlight access
- • Maintain plot ratio compliance
- • Preserve green spaces
Case Study: Waltham Forest
Site Selection Criteria
- Mixed land use covering residential, commercial, and green spaces
- Well-connected transportation networks and road junctions
- Sufficient site dimension (400x400x40m) for comprehensive analysis
- Representative of broader London housing challenges
Application & Optimization
Agent Performance
Training duration: 1.2 hours per 120,000 steps
Optimization Results
Comparison with Traditional Urban Planning
| Process Stage | Traditional Planning | AI-Driven Approach |
|---|---|---|
| Site Analysis | Manual assessment of physical characteristics | Automated plot digitalization with 2D/3D mapping |
| Contextual Studies | Analyze surrounding architecture manually | Comprehensive building index and distance mapping |
| Program Development | Define use mix based on stakeholder input | Automated land use optimization |
| Initial Massing | Preliminary diagrams considering solar/wind | Ground solar radiation mapping integration |
| Training/Iteration | Manual refinement based on feedback | 120,000 steps automated training run |
| Final Production | Human designer creates final design | AI-generated optimized urban layout |
