Engineers Pre-Train AI Computers to Increase Their Applications

Back in 2016, Google‘s AlphaGo beat the world champion in the complicated board game Go.

To beat the Go world champion, Lee Sedol, the system had to learn from its mistakes through reinforcement learning. It’s a step by step process in which AI models train themselves to become powerful.

While in the process of training themselves, computers stumble around a bit before finding the right path. This trial-and-error approach, unfortunately, makes reinforcement learning not ideal for some real-life applications

For example, in climate-control systems, such abrupt swings in temperature could be a problem.

Now, the CSEM engineers have developed a new machine learning method to address this problem. It involves pre-training a computer on overly simplified models before setting to learn on the real-life system.

In a statement, the head of smart energy systems research at CSEM and co-author of the study, Pierre-Jean Alet, said:

“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.”

The engineers described the process in IEEE Transaction on Neural Networks and Learning Systems.

How Pre-Training AI Computers Could Make Them More Powerful

To test their approach, the engineers ran simulations on a climate control system for a 100-room building.

First, they built a virtual model using simple equations that describe the building’s behavior. After training the computer on the model, the engineers then fed actual building data into the computer.

These include the temperature, weather conditions, and how long the blinds were open. That way, the training process can be more accurate.

Finally, the engineers let the computer run its reinforcement-learning algorithms to identify the best way to manage the climate control system. As it turns out, the approach can slash energy use by up to 20 percent.

According to the team, the pre-training approach can open new applications for AI computers. It’s especially useful where large fluctuations in operating parameters would have financial or security costs.