A year after AlphaGo defeated Lee Sedol, an 18-time world champion at Go, the AI faced Ke Jie, who is currently considered the world’s best Go player. AlphaGo beat Ke Jie with only half a point difference--the smallest possible--but that may be due to the AI’s “safer” winning strategy.
AlphaGo AI
The AlphaGo AI is built on the core DeepMind technology that that has already been used to cut Google’s data center cooling costs and to discover better treatments for some diseases.
AlphaGo is a special version of DeepMind AI that only knows how to play Go. Google has “shown” it (as it learns through computer vision) millions of matches between professional (human) Go players, and the company also made AlphaGo play against itself, so it could learn from its own mistakes. Basically, AlphaGo learns similarly to how a human would learn, by watching others do a task well and then repeating that task over and over until its skill at the task improves.
This machine learning technique was not just an innovative way to teach AlphaGo how to play Go well, but it was quite necessary for AlphaGo’s successful learning. That’s because it wasn’t possible to program an artificial intelligence to play and win at Go in the same way chess-focused artificial intelligence was programmed before it.
There are more possible Go moves than atoms in the universe, according to Google, which would have made it impossible for the AI to “play ahead” all the moves until it found the winning ones. This is also why experts thought having an AI win at Go was going to take at least another decade, before AlphaGo was created.
By learning how humans play best, AlphaGo was able to not only replicate those winning moves in various Go-playing scenarios, but also create its own patterns for what a winning move would look like.
Ultimately, this allowed AlphaGo to beat even the best Go players in the world, such as Lee Sedol, and now Ke Jie, too. However, Ke Jie still has two more tries left before a final winner is declared.
AlphaGo Wins One Out Of Three
Although AlphaGo and Lee Sedol played five Go matches against each other, there will be only three matches between the AI and Ke Jie this time around.
In the first natch, Ke Jie lost to AlphaGo by the smallest margin possible: half a point. However, as is often the case in sports, the final score doesn't always clearly indicate how close a competition really was. According to Demis Hassabis, DeepMind’s founder, AlphaGo is not as interested in winning by large margins, as it’s interested in winning period. At the end of the day, winning is what matters in Go, and this seems to be what AlphaGo cares about, too.
Strategy Against AlphaGo
Unlike Lee Sedol, who didn’t really know what to expect from AlphaGo, Ke Jie was more prepared for this match. He’d seen the matches against Sedol, as well as other matches played by AlphaGo online under the nickname of “Master,” so he could understand a little better how the Google AI likes to play.
Ke Jie tried to use a strategy he’s seen AlphaGo use online before, but that didn’t work out for him in the end. Jie should’ve probably known that AlphaGo must have already played such moves against itself when training, which should also mean that it should know how to “defeat itself” in such scenarios.
A more successful strategy against AlphaGo may be one that AlphaGo hasn’t seen before. However, considering Google has shown it millions of matches from top players, coming up with such “unseen moves” may be difficult, especially for a human player who can’t watch millions of hours of video to train.
However, according to Hassabis, the AlphaGo AI also seems to have “liberated” Go players when thinking about Go strategies, by making them think that no move is impossible. This could lead to Go players trying out more innovative moves in the future, but it remains to be seen if Ke Jie will try that strategy in future matches against AlphaGo.
Although Google hasn’t mentioned anything about this yet, it’s likely that both AlphaGo’s neural networks as well as the hardware doing all the computations have received significant upgrades from last year. Google recently introduced the Cloud TPU, its second-generation “Tensor Processing Unit,” which should have not only have much faster inference performance, but now it comes with high training performance, too. As Google previously used the TPUs to power AlphaGo, it may have also used the next-gen versions to power AlphaGo in the match against Ke Jie.
Next AlphaGo Matches
The next match between Ke Jie and AlphaGo will happen on Thursday, and then the final one will be streamed on Saturday. At the “Future of Go Summit” in Wuzhen, China, where these matches take place, there will also be a match between five human players and one AlphaGo AI, as well as a match between two humans who are both assisted by AlphaGo AI instances.
