Blog

The driving simulation at the AI conference in December 2016 in Barcelona showed software of Mobileye company governing the virtual cars’ behaviour programmed only to learn how to move, slickly and safely and to learn by practicing. The software performed the manoeuvres over and over, altering them a little with each attempt. Whenever the merge went smoothly, the system would learn to favour the particular behaviour.

Learning through experiments and simulations is known as reinforcement learning. A computer calculates the value that should be assigned to, say, each right or wrong turn. Each value is stored in a large table updated as system learns.

In 1992 IBM demonstrated a program that used the technique to play backgammon. It became skilled enough to rival the best human players.

In 2013 DeepMind company developed a program able to learn to play Atari video games at a superhuman level, leading Google to acquire the company year after.

In March 2016 AlphaGo, a system of DeepMind, mastered board game Go and beat one of the best players in the world Lee Sedol. It is virtually impossible to build a good Go-playing program with conventional programming. The game is extremely complex and even accomplished players may struggle to say why certain moves are good or bad.

Reinforcement learning of AI is well suited to driving and games because it enables good sequences of decisions. Progress would proceed much more slowly if programmers had to encode all such decisions in advance. Mobileye plans to test its software on vehicles of BMW and Intel. Also Google and Uber are using their own reinforcement learning for their self-driving vehicles, perhaps on a highway near you.

 

Source: https://www.technologyreview.com/s/603501/10-breakthrough-technologies-2017-reinforcement-learning