- Joined
- 08 Apr 09
- Moves
- 20026
10 Mar 16
Originally posted by sonhouseDo what exactly? I don't think I said that it's impossible for a computer to beat humans? My statement was only that Go is more complex than chess (for humans and computers), addressing Marinkatomb's confusion.
Well the fact AlphaGo defeated a sitting world champ says comps can in fact do just that.
10 Mar 16
Originally posted by tvochessSorry then. You are totally right, Go is almost infinitely more complex than go. It is said there are more positions in go than there are atoms in the universe.
Do what exactly? I don't think I said that it's impossible for a computer to beat humans? My statement was only that Go is more complex than chess (for humans and computers), addressing Marinkatomb's confusion.
It's no wonder conventional computers cannot get to high levels in Go, at least not THIS decade.
11 Mar 16
Originally posted by tvochessThe key is this: controlling more territory (and holding more prisoners) is the criterion for determining who has won the game, once they decide to stop playing. But that is not the criterion for deciding when to stop playing. The criterion for deciding when to stop playing is: when both sides believe they can no longer improve their positions. At that point, it is not necessarily known which side controls more territory; they then count up how much territory (+ prisoners) each side has. (Whereas in chess, if one side resigns or they both agree to a draw, that itself is the result--there is nothing else to be counted up to determine the winner.)
Ok, so the thing with Go is indeed: the game ends when both players decide to pass.
However, they couldn't continue eternally even if they wanted to. At some point, the game must end, because there are no more moves to make or the position would be repeated. So, that is not an issue.
However, you may be right that playing on is not necessarily the bes ...[text shortened]... the set of possible moves is still finite: putting a stone on one of the empty squares, or pass.
The thing that interests me about AlphaGo is this: how have the programmers dealt with the questions: a) "should the program now offer to stop playing (because the program cannot improve its position)?" and b) "Should the program accept an offer to stop playing (because the opponent cannot improve his position)?"--if the program has not already calculated that it does in fact hold more territory (+ prisoners) ?
- Joined
- 08 Apr 09
- Moves
- 20026
11 Mar 16
Originally posted by moonbusRegarding the ending of Go vs. chess: that's a good point you have there. The end result in Go is not clear until the score is counted. So when both players pass and end the game, the result is not yet known. However, at the point where both players pass, the end result is already determined, in the sense that the score follows from the final position in a deterministic way using the counting rules. Both players have perfect knowledge of the state of the game and could theoretically keep track of the score along the way. They should stop (pass) when they see no way to improve their score.
The key is this: controlling more territory (and holding more prisoners) is the criterion for determining who has won the game, once they decide to stop playing. But that is not the criterion for deciding when to stop playing. The criterion for deciding when to stop playing is: when both sides believe they can no longer improve their positions. ...[text shortened]... ogram has not already calculated that it does in fact hold more territory (+ prisoners) ?
So, for a theoretically unlimited computer, there is no bigger challenge to solve Go than to solve chess. But practically, given the current state of hardware and software, this is indeed a challenging task as you mentioned. However, I disagree with the statement that formulating it in a finite approach is more challenging.
About AlphaGo: as far as I understood the way they approached the problem from reading several articles a while back, they have mainly focused on training a neural network with a bunch of master games and playing against itself. This neural network is used as a position evaluation, giving a score for each position. Similar to most chess computers, this is then used in a tree-search with finite depth. In chess, the concept of position evaluation is much simpler, taking e.g. material evaluation, king safety, mobility, pawn structure into account. Such a heuristic model for a position in Go is much harder for the reasons I mentioned in a earlier post, i.e. much larger branching factor meaning much more possible future positions. This increases the uncertainty associated with a limited horizon. In addition, a game of Go has more moves than in chess on average. Instead of a heuristic evaluation function, the developers of AlphaGo have opted for a large neural network, which does not contain any heuristics but is just trained (tuned) by studying many games.
About your questions: since AlphaGo constantly runs its search algorithms in combination with its neural network position evaluation, it has an estimate of the score of the game at any time. It is then relatively straightforward to detect that making a move does not improve its (estimated) score, by just considering "pass" as an additional move option throughout its search.
- Joined
- 08 Apr 09
- Moves
- 20026
15 Mar 16
Originally posted by ketchuploverTurned out to be 4-1 for AlphaGo.
Final Score is 5-0
http://www.engadget.com/2016/03/14/the-final-lee-sedol-vs-alphago-match-is-about-to-start/
I wonder if Lee Sedol would have been able to win the match if it consisted of more games. It seems as if he managed to find AlphaGo's weak spots in the last two games, although just winning one of them.
16 Mar 16
Originally posted by plopzillaI suppose you can take that tack as long as you define 'superior' as, well, something other than English-speaking people do. 😛
Human still rules as it beat alphago. A human only needs to win 1 game to be superior to computers. As long as human can win one game computer wins are meaningless.
- Joined
- 08 Apr 09
- Moves
- 20026
16 Mar 16
We can definitely say Lee Sedol was playing competitively with AlphaGo, being able to win 1 of 5 games. It's unlikely a coincidence. I kind of understand why winning this one game makes humans still superior. It's a proof that AlphaGo is beatable, meaning there are weaknesses that humans can exploit.
Also, Lee Sedol only played 5 games against AlphaGo, where AlphaGo studied thousands of games of Go masters. I think if Lee Sedol could have prepared by practicing with AlphaGo and putting in some dedicated study of its style and strategy, he would have had a much better chance. The first games were definitely a surprise for Lee Sedol, having to adapt to a new style.
17 Mar 16
Originally posted by tvochessIt seems the neural net learns faster and deeper than humans. My question is, will Google let AlphaGo be used in the go world like the top chess programs.
We can definitely say Lee Sedol was playing competitively with AlphaGo, being able to win 1 of 5 games. It's unlikely a coincidence. I kind of understand why winning this one game makes humans still superior. It's a proof that AlphaGo is beatable, meaning there are weaknesses that humans can exploit.
Also, Lee Sedol only played 5 games against AlphaGo, ...[text shortened]... hance. The first games were definitely a surprise for Lee Sedol, having to adapt to a new style.
My guess is they will pack up and go home and reprogram the net to other more needed purposes.
I think it would be a boon for Go players to have a comp playing at higher than 9 Dan as a teaching aide to the top players. I bet new strategies would quickly emerge that would make all players stronger.