Ouch.syzygy wrote:Current top engines (and in fact any regular chess engine developed over the last 70 years or so) have nothing to do with weakly solving the game.clumma wrote:Based on Regan's guestimate of perfect play at 3600 Elo, we are within 5 years of weakly solving game.
Houdini 6 has been released
Moderators: hgm, Rebel, chrisw
-
- Posts: 417
- Joined: Sat May 24, 2014 9:16 am
Re: Houdini 6 has been released
-
- Posts: 5566
- Joined: Tue Feb 28, 2012 11:56 pm
Re: Houdini 6 has been released
It seems this came up earlier:syzygy wrote:Current top engines (and in fact any regular chess engine developed over the last 70 years or so) have nothing to do with weakly solving the game.clumma wrote:Based on Regan's guestimate of perfect play at 3600 Elo, we are within 5 years of weakly solving game.
http://talkchess.com/forum/viewtopic.ph ... 654#689654
http://talkchess.com/forum/viewtopic.ph ... 656#689656
Absolutely not.clumma wrote:Incorrect. Such an approach can be used to *weakly* solve a game.syzygy wrote:Cerebellum and current top engines have little to do with solving the game. Solving the game is about obtaining mathematical certainty about the outcome of the game.
Weakly solving the game means getting mathematical certainty about the outcome of chess in a reasonable time (plus finding the necessary moves). To be somewhat more precise, this is what Wikipedia says:
The way Stockfish and all other top engines are developing, they will not be one inch closer to a proof of optimal play one thousand years from now.Weak
Provide an algorithm that secures a win for one player, or a draw for either, against any possible moves by the opponent, from the beginning of the game. That is, produce at least one complete ideal game (all moves start to end) with proof that each move is optimal for the player making it.
(Of course this is the wrong thread for this discussion, so I apologise.)
-
- Posts: 186
- Joined: Fri Oct 10, 2014 10:05 pm
- Location: Berkeley, CA
Re: Houdini 6 has been released
That's a very strong incorrect statement. Alpha-beta search was extensively used in the checkers proof, along with tablebases (which you know about) and a "proof manager" (analogous to Cerebellum) to handle graph history interaction and find the most forcing continuation:syzygy wrote:Current top engines (and in fact any regular chess engine developed over the last 70 years or so) have nothing to do with weakly solving the game.
"The manager uses the new approach of iterating on the error in Chinook's heuristic scores. The manager uses a threshold, t, and temporarily assumes that all heuristic scores ≥ t are wins and all scores ≤ –t are losses. It then proves the result given this assumption. Once completed, t is increased to t+∆, and the process is repeated. Eventually t reaches the value of a win and the proof is complete. This iterative approach concentrates the effort on forming the outline of the proof with low values of t, and then fleshing out the details with the rest of the computation."
Of course one must still prove a lot of positions, which is vastly easier in checkers because of the shorter distance to the tablebase. Nevertheless, it's absolutely clear that both 1.e4 and 1.d4, and more than one reply to each, draw. And since we know it, proving it is not inconceivable.
-Carl
-
- Posts: 417
- Joined: Sat May 24, 2014 9:16 am
Re: Houdini 6 has been released
I smell a good old debate brewing here.clumma wrote:That's a very strong incorrect statement. Alpha-beta search was extensively used in the checkers proof, along with tablebases (which you know about) and a "proof manager" (analogous to Cerebellum) to handle graph history interaction and find the most forcing continuation:syzygy wrote:Current top engines (and in fact any regular chess engine developed over the last 70 years or so) have nothing to do with weakly solving the game.
"The manager uses the new approach of iterating on the error in Chinook's heuristic scores. The manager uses a threshold, t, and temporarily assumes that all heuristic scores ≥ t are wins and all scores ≤ –t are losses. It then proves the result given this assumption. Once completed, t is increased to t+∆, and the process is repeated. Eventually t reaches the value of a win and the proof is complete. This iterative approach concentrates the effort on forming the outline of the proof with low values of t, and then fleshing out the details with the rest of the computation."
Of course one must still prove a lot of positions, which is vastly easier in checkers because of the shorter distance to the tablebase. Nevertheless, it's absolutely clear that both 1.e4 and 1.d4, and more than one reply to each, draw. And since we know it, proving it is not inconceivable.
-Carl
-
- Posts: 5566
- Joined: Tue Feb 28, 2012 11:56 pm
Re: Houdini 6 has been released
Alpha-beta is not going to give any proof unless you run it until terminal nodes. That is not going to work in chess where the distance from opening to tablebases is way too big. And modern top engines are not at all designed for that (on the contrary, they perform a selective search about which essentially nothing can be formally proven... but it finds good moves).clumma wrote:That's a very strong incorrect statement. Alpha-beta search was extensively used in the checkers proofsyzygy wrote:Current top engines (and in fact any regular chess engine developed over the last 70 years or so) have nothing to do with weakly solving the game.
But more importantly, you were equating the hypothesised 3600 Elo ceiling with weakly solving. There is no connection. Chess will still be nowhere from being weakly solved five years from now, and an incrementally improved Stockfish will not have weakly solved chess 1,000 years from now. Perhaps other techniques and technologies for solving chess will be invented, but that is another story.
-
- Posts: 6052
- Joined: Tue Jun 12, 2012 12:41 pm
Re: Houdini 6 has been released
Regan is as far from the objective estimate of perfect play as one possibly could be.clumma wrote:As an outsider to computer chess this always struck me as a strange architectural decision. If GUI handles book, engine cannot use its evaluations when searching. As Brainfish proved, engine strength can be substantially increased with only ~ 50 MB stored evals. Yet engine authors happily support ~ 1 TB tablebases that improve playing strength far less.Houdini wrote:Why? Our philosophy has always been that book management (in particular the probabilities of choosing between the different moves available in the opening book) is done by the GUI that runs the engine.
Each GUI has its own, different book format; if you’re using Arena, your book will be in the Arena ABK format; if you're using Fritz/ChessBase you'll be using ctg format book, etc.
To use an engine you'll eventually need a GUI; the GUI might then as well make the book move selection.
Based on Regan's guestimate of perfect play at 3600 Elo, we are within 5 years of weakly solving game. Idea of opening "book" as something created with human input is an anachronism. Absolute truth of many openings is already known, and indeed already found in Cerebellum.
Many other strange ideas in computer chess. Such as running engines on one's local machine...
-Carl
it is not 3600, but at least 5600.
-
- Posts: 6052
- Joined: Tue Jun 12, 2012 12:41 pm
Re: Houdini 6 has been released
not clear at all.clumma wrote:it's absolutely clear that both 1.e4 and 1.d4, and more than one reply to each, draw.
-Carl
they might draw, but we don't know that.
-
- Posts: 10296
- Joined: Thu Mar 09, 2006 12:37 am
- Location: Tel-Aviv Israel
Re: Houdini 6 has been released
I do not think that mathematical certainty is important.syzygy wrote:It seems this came up earlier:syzygy wrote:Current top engines (and in fact any regular chess engine developed over the last 70 years or so) have nothing to do with weakly solving the game.clumma wrote:Based on Regan's guestimate of perfect play at 3600 Elo, we are within 5 years of weakly solving game.
http://talkchess.com/forum/viewtopic.ph ... 654#689654
http://talkchess.com/forum/viewtopic.ph ... 656#689656Absolutely not.clumma wrote:Incorrect. Such an approach can be used to *weakly* solve a game.syzygy wrote:Cerebellum and current top engines have little to do with solving the game. Solving the game is about obtaining mathematical certainty about the outcome of the game.
Weakly solving the game means getting mathematical certainty about the outcome of chess in a reasonable time (plus finding the necessary moves). To be somewhat more precise, this is what Wikipedia says:The way Stockfish and all other top engines are developing, they will not be one inch closer to a proof of optimal play one thousand years from now.Weak
Provide an algorithm that secures a win for one player, or a draw for either, against any possible moves by the opponent, from the beginning of the game. That is, produce at least one complete ideal game (all moves start to end) with proof that each move is optimal for the player making it.
(Of course this is the wrong thread for this discussion, so I apologise.)
More interesting is if we have in 5 years a machine that practically never lose in chess(when of course program can include also opening book).
I consider chess to be practically solved as a draw as soon as the ssdf leader never lose a game for 3 years.(note that CCRL or CEGT that test with no book are not relevant here because they do not test with books).
Edit:I can replace SSDF by a different more relevant list with books if there is a relevant list at long time control.
Unfortunately I do not see in the ssdf list number of losses of every program but I will be surprised if there is a program with at least 1000 games and 0 losses.
I hope to see in the future better programs in this list.
https://ssdf.bosjo.net/list.htm
Edit 2:I think that we need also at least 1000 games at tournament time control that the top engine does not lose against engines that are not more than 100 elo weaker to consider chess to be practically solved.
If we have 1000 games at tournament time control that the engine does not lose with white then I consider chess to be practically weakly solved as not win for black.
It may be better to replace 3 years by 1000 games at long time control.
Uri
Last edited by Uri Blass on Tue Sep 26, 2017 10:13 am, edited 2 times in total.
-
- Posts: 6052
- Joined: Tue Jun 12, 2012 12:41 pm
Re: Houdini 6 has been released
too few data points.JJJ wrote:http://talkchess.com/forum/viewtopic.ph ... 3&start=30Lyudmil Tsvetkov wrote:where do you get that?JJJ wrote:Yes, congrats Robert, you did it, you really have made the number 1 engine again, some years after the reign of Komodo and Stockfish.
according to all current tests, H6 is way weaker than SF in SMP mode, and SMP is how tournaments like TCEC play.
even on single core, larger head to head H-SF samples give SF some marginal lead.
Ingo's, CCRL, CEGT etc. rating lists only reflect contempt.
As a real chess player, I don't care about contempt.
This test is good enough for me.
Andreas gives just +41 elo to H5 at 10 min. + increment.
-
- Posts: 5566
- Joined: Tue Feb 28, 2012 11:56 pm
Re: Houdini 6 has been released
In that case you do not think that "weakly solving" chess is important. That is fine, since chess is not going to be weakly solved in our lifetime anyway.Uri Blass wrote:I do not think that mathematical certainty is important.