Reinforcement Learning for Comfort Conditions in Buildings
Simulation-based reinforcement-learning experiments for control problems in building technology, using PPO with stable-baselines3, Gymnasium and Radiance-based simulation tooling.
Overview
Control problems in building technology involve competing objectives that have to be traded off against each other continuously, which makes them a natural setting for sequential decision-making. This work explored reinforcement learning as a research method for that class of problem.
My role: project lead of my employer’s workstream in a cooperative research project.
Problem
Rule-based control does not adapt to changing conditions on its own, and reinforcement learning offers a framework for learning a policy instead of specifying one. Validating that in a physical building is slow and expensive, so the work was simulation-based by design.
Approach
- Simulation: Radiance, run through its command-line tools and cached, provided the physical environment the agent observed
- What it controlled: the electric lighting, switched or dimmed, and the position of the shading
- Algorithm: Proximal Policy Optimisation, trained with stable-baselines3 on a custom Gymnasium environment
- Experiment tracking: TensorBoard across training runs, so results could be compared rather than recalled
What it built
Practical experience designing, training and evaluating reinforcement-learning experiments end to end, and, more usefully, the habit of keeping simulation assumptions explicit rather than letting a result in simulation imply a result in a building.
Limitations
Results are simulation-based only. Physical deployment would require a real-time sensor interface and safety constraints well beyond the scope of this research.