Houdini 6 has been released

[Nordlandia]

Critter 1.6a 64-bit assesses the position as equal.

4k3/p1p1p1p1/PpPpPpPp/1P1P1P1P/8/8/4R3/4K3 w - - 0 1

Analysis by Critter 1.6a 64-bit:

1.Kd2 Kd8 2.Re3 Kc8 3.Rg3 Kb8 4.Kc2 Ka8 5.Kc1 Kb8 6.Kb2 Ka8 7.Ka1 Kb8 8.Rg2 Ka8 9.Rg4 Kb8 10.Kb2 Kc8 11.Rg2 Kb8 12.Kc3 Ka8 13.Kd3 Kb8 14.Kd2 Ka8 15.Kc3 Kb8
= (0.05) Depth: 53/79 00:00:02 36639kN
(28.09.2017)

[Uri Blass]

From a mathematical point of view, chess is a finite game and can be trivially solved by application of the minimax algorithm, but I don't believe you can produce a proof that chess is a draw which can be verified within our lifetime. Zermelo's theorem will in any event not be of help in tackling the computational effort. So chess may be "solved" according to some definitions, but it is not "weakly solved" within the normal meaning of that term.

again a very definite statement, but eg. Prof Jaap vd Herik states that chess probably will be solved in a few decades, with the developments going faster than expected; that is, weakly solved, i presume.
Minimax ? i thought that alfa-beta is universal validity ? maybe not for some math purist, but for a weak solution that doesn't matter.

Then i don't think it's a pure computational problem, i once had this discussion with Bob Hyatt, showing him some games which have been solved without brute force computation. So i do think one of the theorems of Zermelo can be applied, at least in a common sense matter.
If you cannot find a winning advantage (from the opening) than it's a draw, simple as that. Look at some results H6 vs Kom11, only one win out of 33 games. We are approaching the dreaded draw zone...

If the two strongest engines have a billion draws in a row and zero wins between them at very long time control on a giant array of CPU cores, it won't prove anything as far as weakly or strongly proven result for chess.

That is because strong chess engines do unsound pruning because it wins more games. But for the same reason, these same engines can miss important tactical shots. Sometimes, to solve a problem, you have to turn off or at least turn down null move pruning, for instance.

One can argue that the game may be "practically proven" and in a common sense sort of way this is true. But to actually prove a game it has to be exhaustive. So, for instance, alpha-beta pruning can be used because it is sound, but null move pruning cannot be used in a proof.

The fabulous depths and results that you see with advanced chess engines is due to un-sound pruning. Mathematically, it does not produce the same result as a mini-max search {as alpha-beta does} which is to say it does not always produce the right answer. What it does do is to help find a really good, really deep answer really fast.

If the 2 strongest engines have a billion draws in a row and zero wins between them at very long time control then it will not be a mathematical proof that chess is a draw(or even a statistical proof that say that if chess is not a draw the probability that it happen is close to 0) but it is going to be a convincing evidence for me assuming the games are different and they do not repeat the same game again and again.

In theory they can miss the win because of pruning but in that case I expect to see a better engine than them because in this case it seems that pruning is a weakness of the engine and you can have a better engine by not pruning.

[shrapnel]

Took the Plunge ! Got a Discount as I'd bought Houdini 5 earlier.
Figured that whatever my doubts about the strength of Houdini 6, private enterprise should be encouraged.
Also, a computer chess world with only Stockfish and its derivatives would be a very boring one !
Cheers !

[Leo]

Took the Plunge ! Got a Discount as I'd bought Houdini 5 earlier.
Figured that whatever my doubts about the strength of Houdini 6, private enterprise should be encouraged.
Also, a computer chess world with only Stockfish and its derivatives would be a very boring one !
Cheers !

Please let us know what you think of it.

[david]

So i played a longer series of matches between houdini 6 and stockfish 8, using a 15 minute time control, giving 8 threads on a ryzen 1800x.

Houdini +45 =340 -40
So after 425 games, houdini has a net score of +5.
i can't do the Elo calculations, but it wouldn't surprise me if the two engines are not statistically significantly different in terms of the error for 400 games.

cheers
david

[JJJ]

50,59%
So something around 4 elo.

[tpoppins]

Right, +4 Elo with +/-15 error margins. So it could be anything from -11 to +19 Elo for H6.

Why SF8, though? It's almost one year old. Get the current dev version from
abrok.eu - slowest compiles but synchronized with the official repo automatically
ultimaiq.net - some 8% faster than abrok, not updated as often
Brainfish - identical to SF dev but with large pages and NUMA, thus faster than the ultimaiq compiles when using hash over 1GB and/or on multiple processors; don't know how often it's updated
ASMfish - some 25% faster than the ultimaiq builds; same bench as the corresponding SF dev build; there may be weeks between updates, though.

[david]

i ran 1 minute blitz matches, again between houdini 6 and stockfish 8, 8 threads on a ryzen 1800x.

Houdini +136 =547 -117

if anyone can convert into Elo and the corresponding uncertainty limits, i would be grateful.

so that is a net +19 from 800 games, or 2.375% net wins at 1 minute time control.
this compares with 1.1765% net wins at a 15 minute time control.

So it looks like Houdini is comparatively stronger at shorter time limits.
As to why SF8, it is just a convenient standard.

[Nordlandia]

david bell: SF8 is almost a year old by now. Use latest developmental build.

http://chess.ultimaiq.net/stockfish.html

[JJJ]

i ran 1 minute blitz matches, again between houdini 6 and stockfish 8, 8 threads on a ryzen 1800x.

Houdini +136 =547 -117

if anyone can convert into Elo and the corresponding uncertainty limits, i would be grateful.

Something around + 8 elo.