World champ human Go V Deep mind:

https://www.youtube.com/channel/UCP7jMXSY2xbc3KCAE0MHQ-A

This is the site that will show the games played starting at 11 PM EST.

Originally posted by sonhouse
https://www.youtube.com/channel/UCP7jMXSY2xbc3KCAE0MHQ-A

This is the site that will show the games played starting at 11 PM EST.

I've only played this game once or twice but it looks really good. One thing i was surprised about while reading a news paper article on this match was that the game is apparently considered to require a lot of intuition. This surprises me tbh, i would have thought a computer would easily beat a human as there appears to be less to calculate. In chess, the value of the pieces change depending on the position, that is why it is such a complicated game. With Go, all the pieces are the same value, at least on the surface it appears to be 100% calculation. I guess I just don't have a deep enough understanding of this game..

Originally posted by Marinkatomb
I've only played this game once or twice but it looks really good. One thing i was surprised about while reading a news paper article on this match was that the game is apparently considered to require a lot of intuition. This surprises me tbh, i would have thought a computer would easily beat a human as there appears to be less to calculate. In chess, t ...[text shortened]... ars to be 100% calculation. I guess I just don't have a deep enough understanding of this game..

Commentary by the only American to reach 9 Dan pro, Micheal Redmond:

https://en.wikipedia.org/wiki/Michael_Redmond_(Go_player)

Originally posted by Marinkatomb
I've only played this game once or twice but it looks really good. One thing i was surprised about while reading a news paper article on this match was that the game is apparently considered to require a lot of intuition. This surprises me tbh, i would have thought a computer would easily beat a human as there appears to be less to calculate. In chess, t ...[text shortened]... ars to be 100% calculation. I guess I just don't have a deep enough understanding of this game..

Here's my view on the topic. In principle. both chess and Go can be calculated entirely. But there are practical limitations on computation power.

The possible amount of positions in Go is much much larger than for chess. For example in chess, the amount of possible first moves for white is 20, then 20 for black, and on average about 30-40 moves throughout the game. Go is played on a 19x19 board, which means 361 possible first moves for white, then 360 for black and one less for each subsequent move. The amount of positions increases way faster with depth for Go than for chess. This is called the branching factor I believe, the ratio of nodes from generation to generation, i.e. half-move.

As a result of this, it is much harder to "calculate" future positions in Go, which is why it is harder to solve by computer. But also, it explains why humans have to rely more on experience and intuition, because it's impossible to calculate anything accurately. So Go is largely based on strategies that have proven to beat weaker opponents (with weaker strategies), in addition to as much calculation as possible. Such empirical strategies are effective until the next better strategy is developed, or something with more calculation power is taking up the game.

Go is also more abstract than chess. In chess, there is a formally defined terminus: namely, checkmate. Go ends by mutual agreement when neither player thinks he can improve his position; this means that go, unlike chess, is somewhat open-ended (the limiting case being that both players play until every node is occupied, but no human would). I don't see the level of complexity as the main problem for a go-program, given that the level of complexity of chess has already been adequately reduced to a finite-state engine. I see the bigger challenge this time to be how to program the ambiguity of go into a finite-state engine.

Go is not ambiguous. The rules are clear: the player with most territory wins (different rule sets for counting territory exist). However, it is almost impossible for humans (and computers) to determine whether they are winning or losing in a nearly even, but uncomplete position. The game continues until both players are convinced there is no more progress. But if they disagree, play is continued to the point where the outcome becomes clear.

This is similar to chess ending more often than not in an agreed draw or one player resigning. Rarely we see a checkmate at the GM level. If a draw offer is not accepted, play continues.

Originally posted by moonbus
I don't see the level of complexity as the main problem for a go-program, given that the level of complexity of chess has already been adequately reduced to a finite-state engine. I see the bigger challenge this time to be how to program the ambiguity of go into a finite-state engine.

The solvability of chess doesn't mean anything about the solvability of Go. Only recently we see a Go computer being able to beat the strongest humans. This has happened for chess a long time ago. Nor are the complexities of both games comparable.

Both chess and Go lack a complete and exact way of evaluating a position without playing it out in all variations. Calculations are therefore important (to computers and humans). Nevertheless, the human approach is to simplify calculation by using empiricism and positional judgement, but these only result from the actual moves that may follow. The more difficult positional judgement in Go is therefore a direct effect of the greater complexity of that game.

Originally posted by tvochess
Go is not ambiguous. The rules are clear: the player with most territory wins (different rule sets for counting territory exist). However, it is almost impossible for humans (and computers) to determine whether they are winning or losing in a nearly even, but uncomplete position. The game continues until both players are convinced there is no more progress. ...[text shortened]... ing. Rarely we see a checkmate at the GM level. If a draw offer is not accepted, play continues.

The computer wins game one. But the computer is using all the master games in memory, I would think if humans were to play more than this one match, humans should be able to consult the same data base to even things out. Given how powerful modern computers are, even laptops now, and the massive data base it contains, it is an unfair advantage by having built in millions of master games which even the best human could not have at all times on hand for a game.

Originally posted by sonhouse
The computer wins game one. But the computer is using all the master games in memory, I would think if humans were to play more than this one match, humans should be able to consult the same data base to even things out. Given how powerful modern computers are, even laptops now, and the massive data base it contains, it is an unfair advantage by having buil ...[text shortened]... llions of master games which even the best human could not have at all times on hand for a game.

I don't think AlphaGo has a database of master games in its memory. Have a look at

http://www.theverge.com/2016/3/8/11178462/google-deepmind-go-challenge-ai-vs-lee-sedol

AlphaGo uses a neurol network which was trained (!) with a large database of master games and playing against itself. This would mean that it's parameters were tuned using these games, but not that the games themselves are stored in the engine. This is similar to the human players who have studied many games (applies to chess and Go). Letting the humans consult the database during the game would not even things out.

Originally posted by sonhouse
The computer wins game one. But the computer is using all the master games in memory, I would think if humans were to play more than this one match, humans should be able to consult the same data base to even things out. Given how powerful modern computers are, even laptops now, and the massive data base it contains, it is an unfair advantage by having buil ...[text shortened]... llions of master games which even the best human could not have at all times on hand for a game.

When DeepBlue beat Kasparov, the next challenge should have been a team of GMs consulting against DeepBlue in correspondence-chess time periods. DeepBlue had analyzed tens of thousands of games and had thousands of openings programmed into it; no single human can match that. I'd wager that the right team of human GMs could still have beaten DeepBlue.

A similar situation will soon be reached in go; the google program has already played itself millions of times in preparation. Soon, only a team of go masters consulting will be able to hold their own.

Originally posted by tvochess
Go is not ambiguous. The rules are clear: the player with most territory wins (different rule sets for counting territory exist). However, it is almost impossible for humans (and computers) to determine whether they are winning or losing in a nearly even, but uncomplete position. The game continues until both players are convinced there is no more progress. ...[text shortened]... ing. Rarely we see a checkmate at the GM level. If a draw offer is not accepted, play continues.

The fact that chess games end more often than not by agreement is irrelevant. There is nothing analogous to checkmate in go; that is the salient point.

The rules of go are not ambiguous, nor is it ambiguous who has won (once the players have decided to stop); however, the point at which the game stops is not formally defined by the rules, as it is in chess. That was the point I was making.

Originally posted by tvochess
I don't think AlphaGo has a database of master games in its memory. Have a look at

http://www.theverge.com/2016/3/8/11178462/google-deepmind-go-challenge-ai-vs-lee-sedol

AlphaGo uses a neurol network which was trained (!) with a large database of master games and playing against itself. This would mean that it's parameters were tuned using these game ...[text shortened]... hess and Go). Letting the humans consult the database during the game would not even things out.

I wonder if the actual number of neural nodes is a secret? I didn't see anything but them saying they used two of them.

We will see if Lee modifies his game style in game 2. Otherwise he is cooked.

Originally posted by moonbus
The fact that chess games end more often than not by agreement is irrelevant. There is nothing analogous to checkmate in go; that is the salient point.

The rules of go are not ambiguous, nor is it ambiguous who has won (once the players have decided to stop); however, the point at which the game stops is not formally defined by the rules, as it is in chess. That was the point I was making.

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 best strategy. Maybe there is something like zugzwang, where passing is the best choice. I don't know Go enough to be sure, but it's very likely.

But then still, this shouldn't be a problem for a computer: the set of possible moves is still finite: putting a stone on one of the empty squares, or pass.

Originally posted by tvochess
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.

Well the fact AlphaGo defeated a sitting world champ says comps can in fact do just that.

Originally posted by sonhouse
Well the fact AlphaGo defeated a sitting world champ says comps can in fact do just that.

2-0 in favor of comp. Doesn't look like he can win now. I haven't seen the win margin yet, it can tell roughly what level the comp is. Is it 10 Dan or 11?