| Life In 19x19 http://www.lifein19x19.com/ |
|
| AlphaGo vs X http://www.lifein19x19.com/viewtopic.php?f=10&t=12784 |
Page 1 of 1 |
| Author: | longshanks [ Tue Mar 08, 2016 3:03 am ] |
| Post subject: | AlphaGo vs X |
I read (forget where now) that Lee Se-dol can be considered the best Go player of the decade. Assuming the unlikely event that AlphaGo does win the series, is there someone stronger it could play against or is it Game Over for humans? And a related question: If humans could choose a player from history to play Google's bot, who would we choose? |
|
| Author: | pookpooi [ Tue Mar 08, 2016 3:13 am ] |
| Post subject: | Re: AlphaGo vs X |
Ke Jie is a little bit stronger than Lee Sedol |
|
| Author: | Bonobo [ Tue Mar 08, 2016 7:20 am ] |
| Post subject: | Re: AlphaGo vs X |
I can imagine that next could be AlphaGo (or other AIs) playing against a team of pros. |
|
| Author: | Kirby [ Tue Mar 08, 2016 7:53 am ] |
| Post subject: | Re: AlphaGo vs X |
Lee Sedol isn't ranked #1 in the world right now, but if AlphaGo wins, I would think that Google can say that computers have overcome humans. |
|
| Author: | hyperpape [ Tue Mar 08, 2016 7:59 am ] |
| Post subject: | Re: AlphaGo vs X |
Kirby wrote: Lee Sedol isn't ranked #1 in the world right now, but if AlphaGo wins, I would think that Google can say that computers have overcome humans. Yeah even if they win by a hair (3-2) it would be bizarre if they can beat Fan Hui in October, improve enough to beat Lee in March, but then stagnate between Lee and Ke Jie. You figure that any approach that can do that much in a year or two can eke out the remaining gap given some time. But I suspect they would play a second or third match for the sake of having a non-speculative victory. |
|
| Author: | gowan [ Tue Mar 08, 2016 8:02 am ] |
| Post subject: | Re: AlphaGo vs X |
It doesn't really matter which of Lee or AlphaGo wins. Here is a good article from Science Magazine on this topic: http://www.sciencemag.org/news/2016/03/ ... -what-does |
|
| Author: | Solomon [ Tue Mar 08, 2016 8:05 am ] |
| Post subject: | Re: AlphaGo vs X |
I think the most interesting way they could go about it after that point is to take the top 3 pros from Korea, Japan, and China and let AlphaGo play a 9-person simul. If AlphaGo wins more than 4, that would certainly put the nail in the coffin, and the country that wins the most games would have bragging rights like the Nongshim Cup. |
|
| Author: | EdLee [ Tue Mar 08, 2016 11:57 am ] |
| Post subject: | |
I await the day the computer takes W and beats the top pros at 3 stones. |
|
| Author: | Arms Longfellow [ Tue Mar 08, 2016 9:01 pm ] |
| Post subject: | Re: AlphaGo vs X |
I don't actually think that's possible. God, possessing the solved formula for Go, couldn't win after giving a top pro 3 stones. |
|
| Author: | EdLee [ Wed Mar 09, 2016 2:24 am ] |
| Post subject: | |
Arms Longfellow wrote: I don't actually think that's possible. Garry Kasparov on Letterman -- I wish I had recorded that interview -- Letterman asked Kasparov whether he thought the computer could ever beat the best human chess players, and Kasparov said No, he didn't think that's EVER possible.( This was some time before Deep Blue, obviously. ) Now the best chess apps can give the top chess pros a handicap. Just a few months ago, many people also thought the computer would never beat the best human pros, at least not for 10-50 years. I'm still hoping, as I did before this match started, that AG would clean up 5-0. We find out in a few days. |
|
| Author: | Krama [ Wed Mar 09, 2016 9:37 am ] |
| Post subject: | Re: |
EdLee wrote: Arms Longfellow wrote: I don't actually think that's possible. Garry Kasparov on Letterman -- I wish I had recorded that interview -- Letterman asked Kasparov whether he thought the computer could ever beat the best human chess players, and Kasparov said No, he didn't think that's EVER possible.( This was some time before Deep Blue, obviously. ) Now the best chess apps can give the top chess pros a handicap. Just a few months ago, many people also thought the computer would never beat the best human pros, at least not for 10-50 years. I'm still hoping, as I did before this match started, that AG would clean up 5-0. We find out in a few days. This is probably true, however we have no idea how advanced the theory and knowledge of both games are. Let's create an arbitrary scale from 0-100 Let's say humans in chess can get to 65 (with theory) and then computers appear and are able to get to 85-90. What if the humans(which I don't actually think is true) are around that level in go and the most this Ai can accomplish is a 80-90% victory in even conditions but could never actually give handicap to pros (or at most 2 stones)? We have no idea how advanced our theory of go is.. it could be that the top pros combined understand only 5% of the game and AI will advance to godlike strength and play moves which no pro would be able to comprehend. We live in great times and I just hope we will get a tool like chess players already have in no more than 5 years. |
|
| Author: | Charles Matthews [ Wed Mar 09, 2016 10:33 am ] |
| Post subject: | Re: Re: |
Krama wrote: We have no idea how advanced our theory of go is.. it could be that the top pros combined understand only 5% of the game and AI will advance to godlike strength and play moves which no pro would be able to comprehend. I sort of agree with you. Formulating a fairly standard view of pro strength, in a non-standard way: take ten plays selected randomly from the games of top pros, excluding later endgame. Then, logically, each play either loses a point or more to best play, or it doesn't (or go is a bizarre game where you really do need a ruleset to avoid triple ko and other long loops ... let's not go there). So, how many pro plays out of do we expect to fail the test, i.e. not to be best play? We are talking about two to three, I believe, on the consensus. Call this proportion F, say 20% or so. But there's a twist: what if it is true that there are more failures of pro play of this kind (dropping a point); but much computation is needed to verify that. I mean, serious server farm stuff. Then to get move 10 absolutely correct in High Chinese versus Sanrensei variations, say, may be like the computational underpinnings of bitcoin: you have to do the work. If there aren't general principles involved. Conclusion: one possible application of advanced go AI would be to take samples from pro games and try to work out the figure F empirically. Interestingly, this approach could reveal the systematic weaknesses of the software, instead. Over time, though, I suppose, we'd get more like contemporary backgammon, where the software is a key training tool. |
|
| Page 1 of 1 | All times are UTC - 8 hours [ DST ] |
| Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group http://www.phpbb.com/ |
|