Google’s AI will battle Go world champion

[Skanky Vamp]

Earlier this month, Google's AI company DeepMind announced that it had built a system of neural networks capable of beating a champion Go player. This ancient Chinese board game is considered an extremely difficult challenge for a computer — far harder than chess — and DeepMind's success was hailed as coup for AI research. However, the computer only beat the game's European champion, and in March, DeepMind's AlphaGo AI will take on the world champion, with YouTube live streaming the series of games.

A loss for AlphaGo would hardly be catastrophic though. DeepMind's win against Hui has been described as a "decade earlier than expected," and any setback would likely be temporary — just as it was with Deep Blue.

DeepMind's founder Demis Hassabis announced details about the competition on Twitter: the five-game matchup will take place in Seoul on March 9th to the 15th (there'll be a game a day apart from on the 11th and 14th), with AlphaGo facing off against the world's top-ranked player, South Korea's Lee Sedol, for a $1 million prize.

 

[vellavu]

It's nothing like Deep Blue. It's far more advanced and sophisticated. It's two main algorithms: MCTS and Deep Neural Network, are both new. They were explored and developed after 2000. The first one simulates human's "reading ahead" while the second simulates the whole visual cortex. Compared to the (basically) brute-force based Deep Blue, it's a totally different beast.

I heard the news in January 27 almost immediately after it's released, and I'm still reeling from it. As a go player, I'm both awestruck and terrified (in a good way).

I bet that Lee Sedol will crush this software in a 5:0 victory, but it will eventually become world's top in 1-2 years, far faster than the previously estimated 1-2 decades.

 

[Skanky Vamp]

So.. if it can be explained, why is go so much more difficult than chess?

 

[vellavu]

Because you have much more choices at each step and the game turns are usually much longer. Therefore human masters rely on more intuition and visualization (pictures in mind) and less tactical calculation to play it.

For computers, they also need something like "intuition" and "visualization" to play Go reasonably well. In Chess, computers can brute-search several dozen steps, but with Go, it's not an option, even if you have a computer the size of the Earth. So the Go AI was lagging behind for 2 decades until the software research recently had the breakthrough of how to simulate (part of) human's brain properly.

 

[Skanky Vamp]

rely on more intuition and visualization (pictures in mind) and less tactical calculation to play it.

Oh wow. Okay. Yeah, so a computer doing chess would seem like checkers in comparison. It would almost be a true 'thinking' calculations then.

 

[vellavu]

There's another fact: If you split all Go players into many levels, so that each level has a 80% win-rate against the level below, then there are about roughly 35-40 levels from a beginner to a pro. For chess, there are only about 15-20. This shit is inherently more difficult than Chess even for human, not for computers.