Go champion in just three days: Researchers have created an artificial intelligence that can teach highly complex skills like the Go game without any human teachers - and that amazingly fast and good. Already after three days AlphaGo Zero defeated its predecessor, which had hit the world champion in the Go 2016 tournament, as the scientists in the specialist magazine "Nature" report. In such self-learning programs you see the future of AI.
In the Go game, there are more than 10 to the power of 170 possible game constellations - but the AI AlphaGo Zero managed to master it in just three days. © DeepMind
Hardly any technology area has experienced such rapid advances as artificial intelligence in recent years. Thanks to neural networks, these programs are now adaptable and flexible enough to recognize our language, to maintain web pages as a web bot, or even to help doctors diagnose.
The victory of the Artificial Intelligence AlphaGo caused furore in one of the most difficult strategic games ever which took place in 2015: the Asian board game Go. After months of intensive training with human experts and games against itself, the program was able to defeat a human pro in this game for the first time.
Now AlphaGo's makers have gone a step further: they have changed AlphaGo so that the AI program can now learn the Go game completely without training with human game partners. Only the game rules are given, the rest is done by the AI independently - playing millions of times against itself.
AlphaGo becomes its own teacher, explains David Silver and his colleagues at Google's research center DeepMind. AlphaGo Zero is trained only by playing against itself and the reinforcing learning - without any supervision or human interventions.
This was made possible by an algorithm that adapts and optimizes the neural network after every successful game. The system learns how much the winning chances of a turn are.
The AlphaGo Zero's learning curve is extremely steep. © DeepMind
The surprising thing: AlphaGo Zero not only learned much faster than its predecessors taught by humans, it also surpassed them in its abilities.
To our surprise, AlphaGo Zero was better than its predecessor AlphaGo Lee after 36 hours, Silver and his colleagues report.
This would had required several months of training for victory against a human grandmaster.
After 72 hours, the researchers let the two artificial intelligences fight directly against each other. The result: AlphaGo Zero defeated its predecessor with 100 won games to zero. In addition, the new, self-learning version of the AI is significantly more economical. It only needs one computer and four special chips, while AlphaGo Lee still needed the performance of several computers and 48 special chips, as the scientists report.
Exciting: Self-learning helped AlphaGo to develop particularly innovative and thus powerful features.
It acquired not only fundamental elements of the human Go knowledge, but also strategies that go far beyond the range of traditional Go knowledge, says Silver and his colleagues.
According to the scientists, this is one of the decisive advantages of an AI that learns without human input: it is not hindered by the possible weaknesses or limitations of its human teachers. At the same time, it can be used virtually indefinitely:
We have a system that can be transferred from the game Go to any other domain, explains Silver. The algorithm is so general that it can be used anywhere.
Once again, the colleagues at DeepMind have made a real and great coup, commented KI researcher Klaus-Robert Müller from the TU Berlin. They were able to show that an intelligent Go program that plays against itself can learn even better than when it is learning from people's games, which is as if one were able to draw knowledge out of nothing - so to speak, a Munchausen-Trick of the AI. "
Source: Nature, 2017, doi: 10.1038 / nature24270