AlphaGo Round 2: Google’s AI To Compete Against Top Go Players

Last year, Google’s AlphaGo AI, which was based on the core DeepMind AI technology, played a historic series of matches against Lee Sedol, an 18-time world champion at the Go board game. This year, after many requests from the Go community, Google plans to pit the AlphaGo AI against Ke Jie, the world’s current no. 1 player.

AI Wins At Professional Go For The First Time

Last year, AlphaGo beat Sedol 4-1 in a five-match game. Sedol held his own throughout most of the five matches, but ultimately he couldn’t defeat the AI more than once. This was a historic moment, similar to the one when IBM’s Deep Blue computer beat Garry Kasparov at chess about two decades ago.

Since then, computers and machine learning has evolved significantly, which is what made it possible for an artificial intelligence to defeat a human pro player at Go. However, chess had a manageable number of possible moves on the board. This eventually allowed computers to quickly test all the possible moves in a given scenario on the board and pick the “best” one.

Go, on the other hand, has many orders of magnitude more possible moves, which can’t be calculated by a traditional computer or program by trying them all out ahead of time. Therefore, for an AI to beat a human pro player at Go, it needs the same type of “intuition” as humans. AlphaGo will “consider” moving in multiple places on the board, and then pick a move that it believes has a higher chance of “winning,” but it can’t simulate the entire remaining match.

AlphaGo’s Upcoming Matches

Google’s DeepMind team announced that from May 23-27, it will collaborate with the China Go Association and the Chinese government to create another event in which AlphaGo will play against multiple top Go players, including Jie, the world’s #1 Go player at the moment, who has also beaten Sedol several times in the past.

The new matches will be set up as follows:

Pair Go - a game of Go between two human players who will be assisted by two different AlphaGo instances. The humans will alternate their moves with their AlphaGo assistants. The idea is to make the game more interesting, as well as to show that Go doesn’t have to disappear as a result of AIs beating humans, but it could evolve to include AI assistants, too.

Team Go - this should be another interesting match, which will test AlphaGo’s creativity against the creativity of a team of five top Go players. The idea here is to test AlphaGo’s weak points (if it has any) by coming at it with different styles of play.

Ke Jie vs AlphaGo - the final games will consist of three 1:1 matches between Ke Jie and AlphaGo. That’s when we’ll see if any single human has any chance of beating AlphaGo anymore.  

The chances should be quite low considering AlphaGo has had a whole year to improve by "watching" millions of recorded game plays from top Go players. The AI can also evolve by playing against a version of itself, as it did before the game with Sedol. This time, AlphaGo may also take full advantage of Google’s TPU chips, so it probably won’t be limited by hardware resources too much.

A Rapid Improvement In AI Technology

Google’s core DeepMind technology has continued to improve, not just for playing Go and other games, but also for solving real-world problems. The company revealed last year that it was using DeepMind machine learning to cut its cooling bill for a data center by 40%, for example, and the DeepMind team has also collaborated with hospitals and medical research teams to study various diseases and types of cancer.

Other machine learning teams inside Google have also made breakthroughs in object recognition in photographs, allowing Google Photos users to search for rather specific items within a photo, as well as in significantly improving the quality of Google Translate.

It’s still early days for machine learning, but we’re already seeing a strong focus on improving machine learning hardware as well as software from large industry players. This should lead to a quick evolution of machine learning over the next few years, making it increasingly more useful in solving more types of real-world problems.

Lucian Armasu
Lucian Armasu is a Contributing Writer for Tom's Hardware US. He covers software news and the issues surrounding privacy and security.
  • dstarr3
    I didn't know there was a CS:GO board game.
  • Virtual_Singularity
    "the DeepMind team has also collaborated with hospitals and medical research teams to study various diseases and types of cancer."

    I wonder, how many sophisticated techs have moved forward by proponents selling the mere prospect of their possible application to improve the human condition in some way? As though, surely, their potentially benevolent applications outweigh any other considerations...

    Deep Blue's AI win vs Kasparov, (despite his claim IBM violated the rules in the 2nd game) was a significant leap for AI. Seeing google's AI beat a reigning Go champion was far more significant. Both were joyless occasions imo. Both are complex, strategic, artistic, potentially passionate games; though Go is far more so. Seeing these champions lose to cold, calculating machines didn't seem a fair outcome, given the circumstances.

    The EU's vote to begin a framework to bestow "electronic personhood" for some existing & future robots & AI earlier this year was notable. Mr. Musk warning of the dangers of AI superintelligence, concluding AI would "treat humans like pets" without some sort of neural lace, seems to indicate AI-SI, with all its implications for better or worse, has progressed far sooner than most could've imagined. Many accept the above as the inevitable, sooner or later. Very interesting times. Though, fwiw(even if not much for most), I've faith the human mind, given a chance to evolve its true, natural potential in the right circumstances, could beat any AI & ensure a far more balanced, gratifying future for all concerned. ;)