Not logged inRybka Chess Community Forum
Up Topic The Rybka Lounge / Computer Chess / To MikeB : Solution to Tac=2 bad time management
- - By MrKris (***) Date 2020-06-08 10:03
From TalkChess: http://talkchess.com/forum3/viewtopic.php?f=6&t=73731&start=10#p846475
-----------------------------
Re: Tournament with Honey, Corchess, Bluefish-XI and Crystal

Post by MikeB » Sun Jun 07, 2020 1:48 pm

    MMarco wrote: ↑
    Sun Jun 07, 2020 1:13 pm
    I ran games with Bluefish tactical=2 at 40/2min repeating, and I can confirm it had a very bad time management, keeping often a second or less for the 10 last moves before the time control with an obvious negative impact on the result. I too found it was almost 300 elo weaker than regular Stockfish which was surprisingly low.

That's about right - tactical 2 is for analysis only but it's very bad at time management and plays much weaker. Last night I also noticed it breaks "go mate x"

-----------------------------

Tactical=2 seems to be related to PV=2 (Talkchess, somewhere).

Stockfish & family time management is probably partly based on https://www.chessprogramming.org/Time_Management :
" Time Management
...
Considerations
    How often did the best move change during the (last N) previous iterations?
    The score function over iterations and best moves, increase or decrease and/or oscillation, score amplitude, etc.
"

Tactical=2 and PV > 1, if used in game-playing T/C's distorts the standard measures best move changes and score oscillation.
Stockfish-family will use up all of its base-time (or time for x moves) in the first few moves because the standard measures are changing at a multiple of their normal rate.

Evidentially there has never been any time management code compensating for Tac.=2 and PV > 1 because they where intended for analysis.
(Which allowed the Komodo team to falsely claim that Ko.MCTS PV=6 'beats' Stockfish PV=6, pretending they did not notice that Stockfish then had very much worse than normal time management.
Not many games but see http://rybkaforum.net/cgi-bin/rybkaforum/topic_show.pl?tid=33114 Sf-PV=6-SM=20 24.5 - Ko.MCTS 11.5.
Also see the screenshots of some of my posts on this below.)

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
The solution for Honey Tactical and Stockfish multi-PV is simply to reduce the 'Slow Mover' setting.
   
It can be easily tuned by watching Tac.2 vs. Tac.0, or PV=x vs. PV=1,
    blitz games in any GUI that shows the clock times.
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

This works very well, as seen in my posts about Ko.MCTS and the one below with Blue-Honey tactical 2.
(Probably at 'Slow Mover' = 0 the formula is too simple to be affected by Tactical or multi-PV.)

Here http://rybkaforum.net/cgi-bin/rybkaforum/topic_show.pl?pid=586852 :
" G/6'+2" , 8-threads x 2 concurrently 2GBh defaults unless noted
2700X | 2060 , no book & no Syzygy
Dbl. RR 2                  Score     Sh Ho Ko Bl Ho Xi Et Bl
-------------------------------------------------------------
1: ShashChess-11_Capa.   9.5 / 14   XX == == =1 =1 == =1 11  (+5 -0 =9)
2: Honey-052520          9.0 / 14   == XX =1 == == =1 1= =1  (+4 -0 =10)
3: Komodo-13.3           7.5 / 14   == =0 XX == == 1= 1= ==  (+2 -1 =11)
4: Blue-Honey-0525 *     7.5 / 14   =0 == == XX == 1= == =1  (+2 -1 =11)
5: Houdini 6 Std         7.0 / 14   =0 == == == XX == == =1  (+1 -1 =12)
6: Xiphos-0.6            5.5 / 14   == =0 0= 0= == XX == ==  (+0 -3 =11)
7: Ethereal-12.21        5.5 / 14   =0 0= 0= == == == XX ==  (+0 -3 =11)
8: Bluefish-FD-0525      4.5 / 14   00 =0 == =0 =0 == == XX  (+0 -5 =9)

* Slow Mover = 25 (default 100), defaut on: Tactical=2
-- *Note: Tac=2 related to PV=4 {TalkChess} SlowM'r must be reduced
-- otherwise the engines use up all their base time in 3 - 6 moves after book.
-------------------------------------------------------------
"
True it only tied with Komodo in the few games, but the first game shown in post, Blue-Honey 1-0 Xiphos, was excellent, I thought; and at least it was 1.5/2 vs. Houdini.
 
Parent - - By Master Om (Bronze) Date 2020-06-08 18:24

>Tactical=2 seems to be related to PV=2


Its PV = 4
Parent - By MrKris (***) Date 2020-06-08 19:49

>> Tactical=2 seems to be related to PV=2
> Its PV = 4


Thanks! (I knew that but I did notice I had mis-typed.)

Honey-branch (Honey, Bluefish, Black Diamond etc.) Tactical = 2 is related to PV=4. <<<<
 

What are the other Tactical #'s > 0 ?
- By MrKris (***) Date 2020-06-08 20:16
 
Honey Slow Mover = 25, Tactical =2
is crushing
Komodo 13.3 Contempt=0
Nooman's TCEC 17 Su.Fin. openings + 50 Nooman 2 moves

I'll make a new thread when its done.

Screenshot:
Cute Chess .config file verifying Tactical = 2 setting. (blue text at left),
2 x 8 threads R7 2700X (boards),
correct 8 threads each (100% th.x8 each), and 2 GB hash (middle white text),
current score (bottom white text).
 
- - By Clementin (*) Date 2020-06-13 03:27
Thanks you for this post Mr Kris! Indeed, lowering Slow Mover is a good idea!

I  ran a match on Android, Bluefish-XI (Defensive= off, Tactical=1, Slow Mover =66) vs Komodo 13.1. 4 threads, Snapdragon 429, 256 Hash, 5 mens syzygy, Noomen 2016 opening suite, 40/4min repeating. Time usage by Bluefish-XI was alike Komodo's from the few games I could watch. Results: Bluefish-XI 53  - Komodo 13.1 47.

I reran it with Defensive= true. Everything else the same. Results : Bluefish-XI 57.5 vs Komodo 13.1 42.5.

Given that my version of Bluefish is older than yours, and that Honey is stronger, the results make sense. I'd be curious what kind of results you get with tactical=1 (and what Slow Mover value you then consider best). Honey and Blue are already tactically strong with default settings, so from a pure playing strenght point of view, they should be stronger with tactical=1 instead of tactical=2 (but then of course tactical=0 would be even better, but we want blood!!).
Parent - By MrKris (***) Date 2020-06-13 17:09

> Given that my version of Bluefish is older than yours, and that Honey is stronger, the results make sense.


It looks a litte low to me, You might try SM=50 (Tac2 related to PV=4, TalkChess, I used SM 25 (guess 100/4 worked out very well). I assume Tac1 is related to PV=2 so maybe 100/2=50).
Or, watch 1 game Bf vs. BfTac1 at a time, quit about 10 - 15 moves after the book.
I did that in Arena, alternating colors, and in a few tries the times were very close, within seconds - much closer than the HnT2-Ko games later where they seemed to use time differently.

Or it could be Bluefish is weaker than Honey/Sf by more than I think.
I'll try to incorporated Tac1 & 2 in my next event.

- - - -
https://play.google.com/store/apps/details?id=com.komodochess.komodo13&hl=en_US
Komodo 13 Chess Engine
Reviews: 3.8 / 5 Stars, 589 total

Jonel Boljanac [latest]
May 8, 2020 (7 likes)
I had Komodo13 play in the Javiolo engine tournament app vs Stockfish 10, which it lost 1 game and drew 30+ games. Also vs other Stockfish9 derived engines Cfish, Corchess, Brainfish from Javiolo, of which it lost 1 vs Corchess and drew 40+ games. In Chess for Android GUI vs SF10, Komodo lost 12 games and won 4. Maybe there are optimal settings? Doesn't show strength or superiority over SF10 & for sure not SF11. It's still very strong, but free apps are equal or stronger. So what's the point?!?!
 

http://fastgm.de/h2h60.html G/60+0.6 secs. 1-core Xeon list, head-to-head:
1) Stockfish 11                3487 :  33000 (+29627,=3180,-193),  94.6 %

     vs.                               :  games (     +,    =,   -),   (%) :   Diff,   SD, CFS (%)
     asmFish 190518                    :    250 (    61,  166,  23),  57.6 :    +37,    3,  100.0
     SugaR-NN 140919                   :    250 (    35,  190,  25),  52.0 :    +55,    2,  100.0
     Houdini 6                         :    250 (   122,  117,  11),  72.2 :   +128,    3,  100.0
     Komodo 13.2                       :    250 (   112,  123,  15),  69.4 :   +132,    5,  100.0
     Komodo 13.3                       :    250 (   114,  118,  18),  69.2 :   +133,    5,  100.0
     Komodo 14                         :    250 (   108,  128,  14),  68.8 :   +135,    3,  100.0
...
Up Topic The Rybka Lounge / Computer Chess / To MikeB : Solution to Tac=2 bad time management

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill