Automating and Optimizing Building Design with Deep Reinforcement Learning

Urban redevelopment requires smart, compliant, and economically viable building designs. In this project, we developed an end-to-end system that automates building design—following legal constraints and user preferences—while maximizing financial value through reinforcement learning.

---

Mission Statement

• Urban areas increasingly require reconstruction.
• Accurate valuation of new buildings is crucial for land investment decisions.
• If we can automatically estimate and generate optimal designs, identifying undervalued land becomes much easier and more scalable.

---

Problem Definition

Input Data

• Parcel polygons
• Legal boundary polygons
• Road network (as LineStrings)

• Numerical constraints:

  • Road width
  • Parking requirements (as equations)
  • BCR (Building Coverage Ratio) & FAR (Floor Area Ratio)
• User preferences (e.g., unit mix, aesthetics)

Objective

Generate a building design that:

• Complies with legal codes
• Maximizes total building value (unit area × revenue per area)
• Ensures livable and practical unit design
• Is visually appealing

---

Problem Structure

The problem is decomposed into three stages:

1. Parameterization: Represent the building design as a set of adjustable parameters.
2. Estimation: Evaluate the design value based on quantitative and qualitative factors.
3. Optimization: Search for the best parameter set to maximize the estimated value.

---

1. Parameterization

What We Did

• Defined core design components: massing, core, corridor, unit, parking

• Parameterized each component to make the design space computable

  • massing
    
  • core
    
  • corridor
    
  • unit
    
  • parking
    
• Prevented invalid designs to reduce search space complexity
  

My Contribution

• While my colleague Tzung-Kuan Hsu built the initial PoC in Rhino,
  I refined the logic and implemented the entire system as software.

Key Principles

• Compact parameter space: Easier to search and optimize.
• Intuitive mapping: Enables agent generalization and policy learning.

---

2. Estimation

What We Did

• Created metrics to evaluate both quantitative (e.g., unit area) and qualitative (e.g., livability, beauty) factors.
• Designed a weighted objective function combining multiple value dimensions.

My Contribution

• Integrated and extended the metric set.
• Developed the final scoring function used for optimization.

Key Challenges

• Conflicting goals: e.g., floor area vs. livability
• Need to quantify abstract values: beauty, openness, utility

---

3. Optimization

What We Did

• Framed the design process as a combinatorial optimization problem.
• Applied Deep Reinforcement Learning (DRL) to explore and generalize over large search spaces.

My Contribution

• Designed the DRL agent architecture and training pipeline.
• Selected REINFORCE as the learning algorithm for its simplicity and direct return optimization.
• Implemented masking and pointer networks to handle dynamic action/state spaces.

Why DRL?

• Traditional search methods (e.g., greedy, exhaustive) fail due to high interdependencies.
• DRL allows feature-based decision making, enabling generalization across parcels.

---

Technical Insights

• Used autoregressive policies to model joint action distributions.
  
• Incorporated pointer networks to remove index binding from action spaces.
  
• Inspired by AlphaStar’s architecture for structured state/action encoding.
  
• Designed Modular RL Agent Architecture for Adaptive Environments.
  

---

Real-World Deployment

• Launched as a mobile app service: Landbook Premium (Sep 2021)
• Generates viable building designs within 3 minutes

---

Retrospective

Strengths

• Successfully automated a multi-step architectural design process
• Achieved generalization and rapid generation over various parcels

Limitations

• Some action heads showed low entropy and weak training signals
• Requires further tuning and experimentation for better convergence

← To Profile