The Asian board game Go is considered particularly challenging. For a long time, it was very unlikely that an artificial intelligence in the difficult game would have many chances against a human player. Until Google’s algorithm AlphaGo proved the opposite. The software had to surrender to the world champion Lee Sedol , but the performance of the program was nevertheless impressive. Now AlphaGo developers went a step further: AlphaGo Zero is able to learn the complex board game within a few days without human teachers.
New AI does not have to learn from people
In the field of artificial intelligence, astounding progress has been made in recent years. Neural networks and machine learning ensure that AI can recognize our language, help doctors diagnose, make judgments, or chatbot human beings.
In 2016, when AlphaGo won against the European champion Fan Hui and then lost to the world champion, the program had to be trained before the start of the tournament with human experts and players for the board game. Only then was the program able to offer a virtual professional player a virtual forehead.
The new version of Alpha Go, AlphaGo Zero, is able to learn how to play Go without completely interacting with human game partners. After AlphaGo Zero is given the game rules, the AI can teach it independently. In doing so, she simulates several million games against herself. Thus, the program becomes virtually his own teacher. The researchers have made this possible by creating an algorithm that optimizes the neural network after each successful game. By and by, the game can assign a success to the individual moves.
AlphaGo Zero defeats his predecessor
AlphaGo Zero was so able to learn the game in just a few days. ” To our surprise, AlphaGo Zero was better than its predecessor AlphaGo Lee after 36 hours,” said David Silver from the Google research center DeepMind. After 72 hours, the researchers then had both programs compete against each other. The result was impressive: AlphaGo Zero defeated his predecessor with 100: 0 games.
The new version is not only better but also more economical than its predecessor. She is content with one computer and four special chips. AlphaGo Lee needs more than one computer and a total of 48 special offers.
Through the self-learning processes of the AI, especially innovative and powerful moves are developed. ” He acquired not only fundamental elements of the human go-knowledge, but also strategies that go far beyond the range of traditional go-knowledge,” the researchers said.
An AI that starts its own learning processes according to given rules has the great advantage that errors and weaknesses of human teachers can not be transferred to the software. The researchers also assume that they can transfer AlphaGo Zero from the Go game to any other domain. “The algorithm is so general that it can be used anywhere,” explains Silva.
Not for no reason do researchers see the future of AI in self-learning systems. In view of the rapid development in AI research, the next few years are likely to be tense in this area.