Reinforcement Learning 

We are a team interested in IoT and the intersection of machine learning and reinforcement learning at the edge. Our work focuses on creating smart devices powered by on-device machine learning that work seamlessly in real-time environments. With our experience in embedded programming, we strive to develop solutions that make IoT smarter and more autonomous, able to adapt and respond effectively to dynamic environments. Our goal is to push the boundaries of what end devices can achieve independently with data analytics and advanced adaptive decision-making capabilities.

     

Here at REAL lab, we focus on developing reinforcement learning (RL) agents optimized to control HVAC systems autonomously. Using various RL algorithms, such as PPO, SAC, A2C and DDPG etc. We train these agents’ performances on simulation and real-world data, aiming to achieve high levels of efficiency and adaptability. A key part of the adaptability from simulation to real World data is to address the domain shift effect, which refers to performance gaps that arise when an agent trained in one environment (often a simulated one) encounters different dynamics in a real-world setting. Our research on addressing the domain shift effect is still ongoing. There we combine various methods and algorithms to optimize the performance gap, ultimately aiming it to be completely resilient. We use various approaches such as domain adaptation and transfer learning to achieve our aim.

Smart Buildings And Machine Learnings

Smart homes are environments where devices and systems interact with each other, enhancing user experience and quality of life.
Machine learning plays a critical role in analysing data and making predictions in these environments. Most machine learning models are trained on raw data, and this is also true for prediction algorithms used in IoT.
This means that we will lose information about the relationships between devices.

Machine Learning and Knowledge Graphs

Infographics are meaningful and structured data models that show the relationships and connections between data. They enable data to be used more effectively by creating semantic networks.

This is why it is important to create an information graph of the data coming from smart devices to train the prediction algorithms used to make predictions in IoT, and to make input to the algorithms without losing the relationships between the data. One of our work areas is to create graphs of the data, to specify the relationships between the data and to build more effective systems.