In this project we developed different stochastic, as well as, deterministic RL algorithms with variable transition periods (based on SMDPs rather than MDPs) and applied them to the problem of micro-climate control in smart buildings. The developed RL agents learn how to operate smart thermostats so as to improve energy consumption as well as the occupants’ comfort.
There are 4 main Jupyter Notebook files:
- SmartBuildings_EnergyControl_Main.ipynb: This is the main file to run.
- Environment.ipynb: This is the class file that contains our developed environment classes for the building.
- ReinforcementLearning.ipynb: This is the class file that contains our developed RL algorithms.
- Training.ipynb: This is the file that contains different training functions required for our algorithms in the RL class
Reference:
• Hosseinloo A. H. et al. “Data-driven control of micro-climate in buildings: An event-triggered reinforcement learning approach”. Applied Energy 277 (2020) (featured on MIT News at: https://news.mit.edu/2020/making-smart-thermostats-more-efficient-1218): https://arxiv.org/pdf/2001.10505.pdf
• Hosseinloo A. H. et al. “Deterministic policy gradient algorithms for semi-Markov decision processes”. International Journal of Intelligent Systems, 37 (2021): https://onlinelibrary.wiley.com/doi/abs/10.1002/int.22709
• Hosseinloo A. H. et al. “Event-Triggered Reinforcement Learning; An Application to Buildings’ Micro-Climate Control”. AAAI Spring Symposium: MLPS. (2020), Stanford, CA, United States. https://ceur-ws.org/Vol-2587/article_6.pdf