RL for Comfort-Based Shading Control
Proof-of-concept reinforcement-learning pipeline using PPO for solar-shading control, validated on synthetic Radiance simulation data.
Overview
Designed and implemented a proof-of-concept reinforcement-learning pipeline for automated solar-shading control, with the goal of optimising user comfort and visual quality in building environments.
Problem
Manual or rule-based shading control does not adapt to occupant comfort dynamically. RL provides a framework for learning optimal shading policies from feedback signals — but validating this in a physical environment is expensive.
Approach
- Simulation data: Radiance was used to compute illuminance levels and daylight-glare probability values as environment observations.
- RL algorithm: Proximal Policy Optimisation (PPO) was selected for its stability on continuous action spaces.
- Reward design: Reward functions were built around comfort thresholds (irradiance limits, daylight-glare probability) and visual quality criteria.
Comparisons were made with simpler approaches (Hill Climbing, classical rule-based control) to contextualise the RL results.
Results
The pipeline demonstrated feasibility on synthetic simulation data. RL-based control reliably satisfied the comfort constraints in the simulated environment. The work served as a research proof of concept presented in the bachelor thesis.
Limitations
Results are on synthetic data only. Physical deployment would require a real-time sensor interface and safety constraints that go beyond the current scope.