This repository accompanies Goldfeder, J., Sipple, J., Real-World Data and Calibrated Simulation Suite for Offline Training of Reinforcement Learning Agents to Optimize Energy and Emission in Office Buildings, currently under review at Neurips 2024, and builds off of Goldfeder, J., Sipple, J., (2023). A Lightweight Calibrated Simulation Enabling Efficient Offline Learning for Optimal Control of Real Buildings, BuildSys '23, November 15–16, 2023, Istanbul, Turkey
In addition to our calibrated simulator, we have released six years of data from three buildings. This data can be used for further simulator calibration, and for training and evaluating reinforcement learning (RL) models.
The dataset is available for download from Tensorflow Datasets.
Alternatively, a smaller version of the dataset can be downloaded as a zip file from cloud storage.
View the official Documentation Site for a complete auto-generated API reference.
There is also a legacy unofficial Community-run Documentation Site containing more information about the project and the codebase. We plan to merge all this content into the official documentation site soon.
A great place to start is by reviewing the Soft Actor Critic Demo notebook. This notebook will walk you through:
-
Creating a gym-compatible Reinforcement Learning (RL) environment.
-
Visualizing the environment.
-
Training an agent using the Tensorflow Agents Library.
Alternatively, RL agents can be trained by running various scripts in the "smart_control/reinforcement_learning/scripts" directory.
Before running notebooks or scripts, make sure to complete the setup instructions linked below.
The Setup Guide provides all the information you need to run the code locally.
The Contributor's Guide provides more information on how to contribute to this repository.