HekaStockfish 0.9

[MikeB]

btw - both current version of stockish honey are now benching with a nearly 9M node bench signature (~60%+ increase in bench signature - so having a much larger bench signature is not indicative that it will play worse at longer time controls - ( however anecdotally it does appear to have an impact at super bullet ( 10 sec /game with 0.1 sec increment - as an aside , the normal SF signature bench is run at depth of 13. This function blows up the search at depth 13 and it will never pass under SF testing protocols. What is interesting is that at depth 26, running on 64 threads - total nodes searched are just modestly higher 10%).

[Ovyron]

I no longer have any idea of the kind of numbers a function like 2000 * pow(2, int(Options["Profound"])) returns, but I really appreciate that you're improving the formulas and taking a very... profound look at this

[MikeB]

I no longer have any idea of the kind of numbers a function like 2000 * pow(2, int(Options["Profound"])) returns, but I really appreciate that you're improving the formulas and taking a very... profound look at this

final results

Code: Select all

 Date: 03/21/20 : 17:43:13
1800 game(s) loaded
Rank Name              Rating   Δ     +    -     #     Σ    Σ%     W    L    D   W%    =%   OppR 
---------------------------------------------------------------------------------------------------------
   1 Stockfish-032120   3551   0.0    8    8  1200  604.0  50.3  122  114  964  10.2  80.3  3549 
   2 Honey-dev-032120   3550   1.2    8    8  1200  600.5  50.0  105  104  991   8.8  82.6  3550 
   3 Stockfish-031420   3548   2.0    8    8  1200  595.5  49.6  110  119  971   9.2  80.9  3551 
---------------------------------------------------------------------------------------------------------
1800 game(s) loaded

considering both *032120's have bench signature of nearly 9M - I find this very promising, so I will now use "032120" versions as best and will test newer versions against this...

[mesilikas]

Is there an android version for this?

[MikeB]

I no longer have any idea of the kind of numbers a function like 2000 * pow(2, int(Options["Profound"])) returns, but I really appreciate that you're improving the formulas and taking a very... profound look at this

final results

Code: Select all

 Date: 03/21/20 : 17:43:13
1800 game(s) loaded
Rank Name              Rating   Δ     +    -     #     Σ    Σ%     W    L    D   W%    =%   OppR 
---------------------------------------------------------------------------------------------------------
   1 Stockfish-032120   3551   0.0    8    8  1200  604.0  50.3  122  114  964  10.2  80.3  3549 
   2 Honey-dev-032120   3550   1.2    8    8  1200  600.5  50.0  105  104  991   8.8  82.6  3550 
   3 Stockfish-031420   3548   2.0    8    8  1200  595.5  49.6  110  119  971   9.2  80.9  3551 
---------------------------------------------------------------------------------------------------------
1800 game(s) loaded

considering both *032120's have bench signature of nearly 9M - I find this very promising, so I will now use "032120" versions as best and will test newer versions against this...

Tested 3200 and 6400 yesterday, they both failed
. The tested 2600 against current best and it passed .

Score of Stockfish-0321-P2600 vs Stockfish-032120: 107 - 98 - 795 [0.504]
Elo difference: 3.1 +/- 9.7, LOS: 73.5 %, DrawRatio: 79.5 %

1000 of 1000 games finished.
All games were played at 8 min with 4.8 sec increment


I will now update a Curr-dev-Honey with Official latest SF commits , and the run a test vs Honey XI , if it passses with 10 Elo or more , we will have a our first release update to Honey XI.

[Ovyron]

That was very quick! I had expected that testing would take a lot longer.

[Alayan]

Edit: the code is quite invasive to the search function, although not a lines of code, the search explodes the bench node count signature up to nearly 7 M nodes up from 5 million - one reason why this version sucks at bullet chess - this version is for games 10 minutes plus 6 seconds incrment or longer - will not perform to expectations with shorter time controls.. The true Heka version has a much higher bench node signature , but it also has many other changes which I haven't really study yet - I'm really just looking at presearch now , and even that is way more limited than what Heka already has.

I assume the presearch value would have to be adaptive based on the expected search time and nps, in order to work well across a range of time control.

I'm looking forward to more experiments there.

I'm assuming one big upside of the presearch method is that it improves on move ordering for further iterations and helps reducing blindness to some shallow tricks.

If it can be demonstrated to be able to beat SF in h2h in reasonable conditions and with reasonable sample size, I think there'll be interest in refining it at fishtest.

[MikeB]

I assume the presearch value would have to be adaptive based on the expected search time and nps, in order to work well across a range of time control.
.....

Seems to be more or less linear...currently , my Elo lining code captures optimal time for each move. It looks like I might be able to piggyback off that code here

I'm assuming one big upside of the presearch method is that it improves on move ordering for further iterations and helps reducing blindness to some shallow tricks.

It clearly helps...

If it can be demonstrated to be able to beat SF in h2h in reasonable conditions and with reasonable sample size, I think there'll be interest in refining it at fishtest.

That's not quite clear - seems that it can at least equal right now - a collaborative effort would do well here to push it above and beyond. I would welcome support from anyone or if Fishtest adopts it.

[Damir]

You can try and ask Noobptw to try and test on his machines….