AlphaGo, the first computer program to beat the world champion of the game of Go, recently found itself losing dozens of matches.
Who was the program's new challenger? A more powerful version of itself.
This new and improved AlphaGo can actually learn to play the game on its own, without any feedback from humans, according to DeepMind, the Alphabet-owned company behind the computer program.
On Wednesday, DeepMind detailed this latest evolution of AlphaGo, which it calls Zero, in a new research paper published in Nature.
What sets Zero apart from the older versions of AlphaGo is how the program learns. Previous iterations did so by competing with human players, both amateur and professional.
Zero is different. This version learned by playing the game against itself, DeepMind wrote in a blog post.
To accomplish this, the company used a machine learning technique called "reinforcement learning" to push Zero to optimize its gameplay. The program's algorithms were then fine-tuned to predict future moves and the eventual winner of each match.
"This technique is more powerful than previous versions of AlphaGo because it is no longer constrained by the limits of human knowledge," the company said.
That change helped Zero become an even stronger Go player than the earlier iterations.
After only three days of self-training, the new version was pitted against an earlier AlphaGo program that defeated 18-time world champion Lee Sedol last year. Zero performed so well that it won all 100 matches played.
"The system progressively learned the game of Go from scratch, accumulating thousands of years of human knowledge during a period of just a few days," DeepMind said.
After 40 days of self-training, Zero was then pitted against the AlphaGo program that defeated the current world champion Ke Jie earlier this year. It went on to win 89 of the 100 games played.
How any of this research might apply to other fields outside of an ancient board game still isn't clear. But according to DeepMind, the new version of AlphaGo shows that A.I. programs don't always have to rely on human-created data to become smart.