HAL9000

HAL9000
"It just isn't conceivable that you can design a program strong enough to beat players like me."

January 9, 2016

Stockfish 6 vs all Development versions: Beta 2 wins as expected

This post is made for sharing the result of a closed tourney i've been conducting before the release of Stockfish 7, a kind of tourney i was not really willing to attempt for. Anyhow i could not resist anymore against my curiosity and thanks to one of the tablets standing free of load, i started such badly scheduled no sense.

* 8 different builds of 7 different versions, including Stockfish 6 which is supposed to be the weakest.
* Used the first 10 of the last 50 openings of TCEC-7 and each build played 20 games against all others, a total of 140 games per build.
* Ran on quad core RK3188 downclocked from 1.4 to 1.0 GHz to avoid overheating
* Time controls 180+2, way shorter than usual 900+2 of Rapidroid.

I was planning to go until all 50 openings are played, for a total of 700 games per build but the tourney became obsolete after Stockfish 7. Even though i stopped at 20% of the progress, below results are quite speaking, showing the progress achieved step by step.

The ratings are calculated according to latest Rapidroid rating of 121215 build by JA, anchored at 3325.

We will see soon how high Stockfish 7 (released with Droidfish 1.60) can climb in Rapidroid conditions. In any case, i don't expect 3367 ELO given that playing with other engines at longer time controls usually stretches the gaps.
click image to enlarge

15 comments:

Anonymous said...

strange JA compile sf 7 is better than PO? are u sure?

Gurcan Uckardes said...

Beta 2 performed better with 10 openings. It's not a serious rating list but only an interrupted tournament.
Regarding SF7 PO vs SF7 JA, PO is better. More info soon...

Anonymous said...

I'm new in this business, but regarding PO compile - I ran two interrupted and unorganized mini blitz tournaments 40/4 with 6 games from Silver and Tcec-7 first openings, and PO crushed Komodo 9.2 by a mile. From 12 games Komodo didn't win any, PO won 7/12.
128MB hash each, no tablebases.
Quite interesting :)



Gurcan Uckardes said...

Yes. PO compile of SF7 looks the best. My statement from last Rapidroid ranking is based on the fact that Komodo 9.2 used to overcome Stockfish 6 by 40 elo. Now Stockfish probably reaches 3350 but Komodo goes back from around 3330 to 3290. That's what surprised me most.

Gurcan Uckardes said...

Above tourney did not include latest SF7 compiles but only the betas. So take it for fun. Latest compiles are a different story. In Rapidroid i will use SF7 PO bundled with Droidfish 1.60.

Anonymous said...

After 16 games of continued mini tournament (40/4 repeated, 4 cores, 128MB hash, no TB), with first 8 openings from TCEC 7 opening book, Stockfish 7 DF 1.60 beats Komodo 9.2 with +5 =10 -1.

I know 16 games is nothing, but I love watching SF7 DF 1.60 in action ;)

I'm waiting for SF7 Rapidroid score :)

Anonymous said...

Now after 20 games it's +8 -1 =11 for SF7 (it's like 127 ELO difference!).
Something weird going on here, but I don't know what yet ;)

Алексей Чернаков said...

20 партий мало,надо как минимум 100.Ещё надо stockfish 7 (P.O.)на длинном контроле против komodo испытать.New stockfish 7 J.A. играет сильно и у меня есть основание полагать что на нескольких ядрах он будет сильнее stockfish 7 P.O.

Алексей Чернаков said...

Зачем хэш 128 mb с контролем 4/40?
Формула оптимального хэша:(T×S/100).
T-время на один ход.
S-количество kn/s.
Например:с контролем 4/40,среднее время на ход~6s.Скорость~500kn/s.Хэш=(6×500/100)=30.
Оптимальный хэш~32,64mb.

Anonymous said...

Thanks Alexei, I will try your advice with lowering the cache and running another mini tour.
There's clearly something wrong, Stockfish won the last 5 games in a row and winning +10 -1 =11.
I've noticed that Komodo 9.2 has considerably lower NPS most of the time on my phone, sometimes even 2x less than Stockfish 7 - is it normal?
For example opening position - Stockfish ~550 nps, Komodo ~350 nps (?).

Gurcan Uckardes said...

Exynos 830k vs 480k here using Komodo 9.3.
9.2 was not faster than that either.
Komodo has never been fast on Android.

Regarding the hash i keep 256mb in Rapidroid because some engines exceed 1200knps and in 900+2 time controls, earlier phase on the game can easily use 30-50 seconds per move. I believe more hash in no problem as long as we have enough ram.

Алексей Чернаков said...

Это нормально.У komodo другой алгоритм расчёта вариантов,делается больше акцент на оценку позиции чем на количество рассмотренных позиций.Komodo силен в партиях с длинным контролем времени.

Anonymous said...

So there's nothing unusual with lower nps in Komodo on Android.
Which version(s) of Android are you running? I'm on CM Lollipop 5.1.1.
I wonder if it could be the reason for Komodo losing so significally with SF7 DF Edition?
Maybe it's not well optimized for this certain system?

Anonymous said...

Haha, thanks Aleksei :)
I've heard about Komodo being slower in nps but better in position evaluation, but not that much slower. It's like 30-50% difference in speed on Android, while on PC this difference is much smaller.

Алексей Чернаков said...

Я думаю что проблемы у разработчиков Komodo с компилятором.Та же проблема наблюдается у stockfish 7 J.A,скорость которого меньше stockfish 7 P.O.