Re: The success of Go AI’s
- So they’re not as dumb as the article makes them out to be. Most AI’s will look at the board, rule out stupid moves (playing into a ladder trap, illegal moves, suicide, etc). It depends on the stage of game, but in the end there will be a candidate list of “good moves” the AI has found based on heuristics, patterns, etc. Then each candidate is “tried” – make the move and run countless simulations to find out how the game ends up.
- The entire approach ease easily parallelized so having more computers with more storage will only make things much more effective.
- What is surprising to most researchers is how effective the ‘dumb playouts’ actually are. When a candidate next move is being played out, the successive moves in the playout are mostly random (for both colours). Doing this a large number of times and aggregating the results works surprisingly well.

Re: How the AI’s will improve from here:
- Already mentioned the increase in storage / computing power (which will directly have an impact on the ability of the AI’s).
- Another avenue I don’t believe is fully tapped yet is the vast amounts of gameplay to learn from. There are so many archives of great go games that AI systems can tap into for knowledge. I could definitely see Go AI’s that are produced to play with the style of .
- Beyond that, knowing your opponent is another possibility for the AI’s to tap into. Just like a sports team will review footage of their opponents, learning the pro player’s strengths / weaknesses could have a great impact on the performance of AI’s.

My guess is that the AI vs Pro tipping point will be somewhere in between both of your predictions (5 and 20). Its just starting to have an uptake and focus from a larger research community and with the annual contests in place, things will only accelerate from here .

Cheers

I actually find the idea of computer-go competing on even grounds with top professionals VERY exciting. Just don’t think it’ll happen in the next 20 years.

I don’t think that CPU speed will keep growing, since some “physical barriers” are met with the current technologies.

My understanding is that the recent success in computer-go is mostly due to software, not to hardware (well, it’s always good to have fast hardware, but in this case I think it’s about running lots of concurrent simulated games in parallel, as opposed to using some complicated rules-based heuristic for moves and position evaluation). I am sure Christian knows more about this and can shed some light

As for “try to imagine” what’ll happen in the future, hm…. From the human perspective, once one gets to a reasonably high amateur rank, getting one rank stronger become more and more difficult.
If we assume strongest programs to be 6 stones away from pro level… that means they are about 3-4 dan amateur.
Not sure if comparing the humans’ learning process with the advances in computer-go is relevant, but I believe it to be to some degree: take any 3-4 dan amateur from today: with 99.9999% probability, they’ll have more or less the same level in 20 years
I am sure some computer-go programs will make progress… but I am skeptical of the rate of the progress.

My bet: computer-go will beat top-pros at 5 stones handicap in 5 years from now, but they’ll not beat top pros at 3 stones in 20 years.

The article glosses over what’s been done, but rest assured there is lots of /new/ research that has gone into the modern Go AI’s. Since checkers fell last summer (complete solved by a research from Calgary I believe), Go has been on the radar for many Academics — its a great candidate for applying new research ideas on since the rules are so simple and yet humans still have the upper hand.

Of course there’s no definite way to say if/when the computer programs will best the world’s top players, but just look at what the last 20 years has achieved and try to imagine what will come in the next 20. At the very least its a wager I can’t lose in my lifetime .

The reason is that the complexity is not linear, but exponential.

That being said, I’m still amazed at the recent progress of the computer-go level.
My understanding, though, is that the methods used are same that have been used in Chess for quite some time; if my understanding is correct, that there is basically no “new development”, but rather a “sync-up”.

The use of computers for such things is so new, it’s hard to fathom that the task of beating the world’s best go players can’t be achieved in our time.