Self learning algorithms are a regular part of the machine learning environment, and a number of high profile examples have showcased the possibilities that exist when software can learn from it’s experiences.
The latest of these comes from researchers at Lancaster University’s Data Science Institute, who have build software that can assemble itself in the most effective form and help researchers combat climate change.
The system, which they’ve called REx and which is documented in a recently published paper, is designed specifically to reduce the energy consumption required by the walls of servers in data centers around the world. By rapidly adjusting to the changing demands placed on it by the data center, it would allow servers under the control of REx to do less processing, and therefore consume less energy.
Micro variations
The process is underpinned by so called micro-variation, which involves a huge database of building blocks for software components that can be selected and assembled autonomously depending on the task at hand.
“Everything is learned by the live system, assembling the required components and continually assessing their effectiveness in the situations to which the system is subjected,” the researchers say. “Each component is sufficiently small that it is easy to create natural behavioural variation. By autonomously assembling systems from these micro-variations we then see REx create software designs that are automatically formed to deal with their task.”
“As we use connected devices on a more frequent basis, and as we move into the era of the Internet of Things, the volume of data that needs to be processed and distributed is rapidly growing. This is causing a significant demand for energy through millions of servers at data centres. An automated system like REx, able to find the best performance in any conditions, could offer a way to significantly reduce this energy demand,” they continue.
Dealing with complexity
Modern software is an increasingly complex beast. It typically consists of several million lines of code that requires an army of developers to maintain. The industry accepts that it’s a level of complexity that cannot continue, and so self-assembling software, such as REx, is not only capable of saving energy, but also of dealing with such complex environments.
The REx system consists of three distinct, yet complementary, layers. Firstly, a component based programming language sits at the base of the pyramid. This language, called Dana, allows REx to find, select and adapt the building blocks that form the software. Atop of this layer is a perception, assembly and learning framework (PAL) that is designed to perceive the behavior of selected components and configure them accordingly. Finally, an online learning layer gauges the best compositions of blocks in real-time via linear bandit models.
