Engineers are developing a new machine-learning methodology that lays the foundation for artificial intelligence to be utilize in applications that until recently were considered too sensitive. Working at the CSEM, the team has developed a new method that has been tested on a climate-control simulations for a 100-room building, and it’s looking like it will provide an energy savings of over 20%
A few years ago in 2016, a supercomputer beat the world’s best human Go champion. In case you aren’t familiar, Go is an fantasticly complicated board game that usually takes many years and a lot of experience to master. So, how did the computer manage to pull it off? By using machine learning. To be specific, reinforcement learning. Reinforcement learning is a kind of artificial intelligence system where a computer is able to train itself complicated matters after being programmed with only simple instructions. This way, the computer can learn from its mistakes and over time can become highly powerful.
The main disadvantage related to reinforcement learning is that its not very practice to apply in certain real-life scenarios. The reason why can be found in the training process itself. See, when computers first try to get something done, they will try just about anything before eventually finding themselves on the right path. This ‘trial-and-error’ phase can be problematic for things like climate-control systems, where abrupt swings in temperature wouldn’t be ideal.
The team of CSEM (Computer Science, Engineering and Mathematics) engineers were able to design such an approach that it sidesteps this issue. The research team was able to demonstrate a computer’s ability to be trained basic, simplified theoretical models before being exposed to real-world systems with live data. So, this way, when the computer begins the machine-learning process on the real-world systems, they will already have a little experience and knowledge to draw on, helping them acclimate. This allows computers to quickly get on the right track without having to go through the traditional periods of extreme volatility that you would expect. This study was just recently published in IEEE Transactions on Neural Networks and Learning Systems.
It’s like learning the driver’s manual before you start a car. With this pre-training step, computers build up a knowledge base they can draw on so they aren’t flying blind as they search for the right answer.Pierre-Jean Alet, head of smart energy systems research at CSEM
The team that performed the study tested their prototype system on an HVAC (Heating, Ventilation and Air Conditioning) system that controls a complicated 100-room virtual building. At first, the AI was trained on a ‘virtual model’ that was built from simple equations that describe a building’s behavior. Then, the computer was fed building data. Things like how long blinds were open, what the weather conditions were, the temperature, and other factors into the neutral network to increase its training accuracy. Finally, the AI ran its reinforcement-learning algorithms to find the best way to manage the HVAC system.
This discovery could open new opportunities machine learning and AI by expanding from its current limited use to applications where large deviations in operating parameters are important for either economic or security reasons.