Computers are now learning and adapting through “reinforcement learning”. Without being taught by programmers, a computer can now learn behaviours through experimentation and adaptation.

Self-driving vehicles, for example, learn to merge and drive safely through “practicing”. They test maneuvering repeatedly, while adapting little by little with each new attempt. Every time a merge or a new move allows the vehicles to run smoothly, the system identifies and memorizes the behavior that is responsible for the desirable result. This approach is what we call reinforcement learning.

Thanks to reinforcement learning, machines can match human players in some of the most complex games known to man and were previously thought to be unbeatable by AI, such as the boardgame GO. In its essence, reinforcement learning copies natural behavioral patterns. Take animals for example – some of the most intelligent animals learn how to associate their behavior with the desirable outcome after a number of successful attempts.

Due to such, some AI pioneers began to believe that this process can be successfully applied in machine learning processes. The machines began using simple forms of reinforcement learning in order to mimic some of the natural behavioral patterns. Although the machines became skilled enough to match some of the best human players, the reinforcement learning failed to solve more complex problems.

One of the solutions for overcoming this problem and limitations of reinforced learning was to combine it with deep learning. Deep learning is a powerful learning technique that relies on an extremely huge simulated neural network in order to identify and apply the patterns in data. Each of the values are stored in an appropriate large table. As a result, the computer updates all of these new values as it learns along the way. Over the years, deep learning has helped improve the pattern recognition in data.

It was only a matter of time before deep learning made a name for itself. It was back in 2013 when the programs became capable of learning how to play and win against human players in various Atari video games. What were practical results of this impressive achievement? Well, Google acquired one of the most promising startups in this field named DeepMind in 2014 for more than $500 million. This was a turning point for other companies and researchers to focus their full attention on reinforcement learning.

There are numerous industrial-robot manufacturers who are testing this approach in order to improve their machines’ performance without the need to rely on or use the manual programming. Some of the researchers at Google used DeepMind and reinforcement learning to improve the energy efficiency of their data centers. It is definitely not an easy thing figuring out how all available elements in specific data center affect the current energy usage. However, with the help of a reinforcement-learning algorithms, it is possible to learn when and how to successfully operate the cooling systems.

The business field where you will notice the most significant changes is currently reserved for self-driving cars. So keep an eye on Google and Uber while they perform testing and improving the reinforcement learning for the next generation of the self-driving vehicles.