DeepMind Company, that became a Google subsidiary in 2014, created a neural network AlphaGo Zero, which trains itself without relying on human experience. This is a huge breakthrough, since previously AI-systems, including the ones of DeepMind’s production, analyzed data on human achievements for improvement.
For example, AlphaGo Lee is company’s first neural network that defeated the world champion in Go (traditional Chinese strategy game, more intuitive than chess), learned by analyzing moves of professional players.
Next version, AlphaGo Master, did not need developers adding information to it, the AI system independently searched data online. Nevertheless, it also took human experience as a basis. The main achievement of this neural network is victory over 60 strongest Go players among people.
AlphaGo Zero was only given basics of this game and the system perfected the skill fighting itself. In 3 days, the latest version of neural network reached the level of AlphaGo Lee, after 21 days, it matched the Master version, and after 40 days it completely destroyed all predecessors.