The algorithm was tested against Stockfish 8, an open source computer chess engine considered to be among the strongest available. In a paper on the research, Google’s DeepMind division reported that in 100 games played against Stockfish 8, AlphaZero won or drew all of them.
The AlphaZero algorithm is a generalized version of Google’s AlphaGo Zero, which itself was an evolution of AlphaGo – the first computer program to defeat a world champion at the ancient Chinese game of Go. The original AlphaGo trained on thousands of human amateur and professional games to learn how to play Go, while AlphaGo Zero was able to teach itself to play Go from scratch, starting from completely random play.
The AlphaZero algorithm takes this approach to a general level that, say the DeepMind researchers, can achieve superhuman performance in many challenging domains, including the games of chess and shogi (Japanese chess), as well as Go. “It replaces the handcrafted knowledge and domain-specific augmentations used in traditional game-playing programs,” they say in their paper, “with deep neural networks and a tabula rasa reinforcement learning algorithm.”
Presenting more details on the new AI program at an AI conference last week, CEO of DeepMind and expert chess player Demis Hassabis reportedly said, “It doesn’t play like a human, and it doesn’t play like a program. It plays in a third, almost alien, way.”
The DeepMind researchers note that the performance of AlphaZero would likely benefit further from using some of the techniques employed in other strong chess programs. However, they say, they had chosen to focus on a pure “self-play reinforcement learning” approach with AlphaZero and are leaving possible such extensions for future research.
NASA drone race pits human pro vs AI
Google’s Kaggle opens AI attack/defense competition
AI’s ‘human side’ is focus of new Google initiative
AI wars heat up with new Microsoft research lab