UCI Engine Update: BikJump v1.8

[abik]

version 1.8 of the UCI engine BikJump is now available at http://www.aartbik.com/MISC/chess.html for Windows 32-bit, Windows 64-bit, and Linux 64-bit.

Modifications:

(1) revised panic time
(2) added more chess rules to evaluation function
(3) revised various search methods
(4) added Nalimov endgame tablebases probing for 6-piece as well
(5) bug fix on missing \n in engine output when tablebases cannot be found

I installed a large number of 6-piece Nalimov endgame tablebases, and after some preliminary testing I am more confident to remove the artifical restriction up to 5-piece that was present in v1.7.

Sincerely,
Aart Bik

[Olivier Deville]

version 1.8 of the UCI engine BikJump is now available at http://www.aartbik.com/MISC/chess.html for Windows 32-bit, Windows 64-bit, and Linux 64-bit.

Modifications:

(1) revised panic time
(2) added more chess rules to evaluation function
(3) revised various search methods
(4) added Nalimov endgame tablebases probing for 6-piece as well
(5) bug fix on missing \n in engine output when tablebases cannot be found

I installed a large number of 6-piece Nalimov endgame tablebases, and after some preliminary testing I am more confident to remove the artifical restriction up to 5-piece that was present in v1.7.

Sincerely,
Aart Bik

Hi Aart

Welcome to the Winboard Forum, and thanks for the update

BikJump logo added to your profile.

Olivier

[abik]

Nice. Thanks for adding the logo to my profile.

[abik]

I am testing the "publish" feature of Chessbase with a tournament between all major BikJump versions, time 3'+1'', on the Nunn test set and 3-4-5-6 endgame tablebases (I realize that this is not a particular interesting tournament, but if testing goes well, I may ocasssionally run some minor tournaments with other engines in alternative time settings).

Games are posted live on http://www.aartbik.com/MISC/TOUR/

[abik]

Some results, somewhere past mid-way the round robin tournament (games are not completely even after suspending the tournament briefly). Results seem to imply that, with the exception of v1.7, strength does increase with version number at least.

Code: Select all

1    BikJump v1.8 (32-bit)  63.5 / 84
2    BikJump v1.6.1         47.5 / 80
3    BikJump v1.7 (32-bit)  46.5 / 82
4    BikJump v1.5           45.5 / 84
5    BikJump v1.4.1         44.5 / 82
6    BikJump v1.3           34.0 / 83
7    BikJump v1.2.1         27.5 / 80
8    BikJump v1.1.1         20.0 / 83


[Dr.Wael Deeb]

Some results, somewhere past mid-way the round robin tournament (games are not completely even after suspending the tournament briefly). Results seem to imply that, with the exception of v1.7, strength does increase with version number at least.

Code: Select all

1    BikJump v1.8 (32-bit)  63.5 / 84
2    BikJump v1.6.1         47.5 / 80
3    BikJump v1.7 (32-bit)  46.5 / 82
4    BikJump v1.5           45.5 / 84
5    BikJump v1.4.1         44.5 / 82
6    BikJump v1.3           34.0 / 83
7    BikJump v1.2.1         27.5 / 80
8    BikJump v1.1.1         20.0 / 83


This is definitely an improvement,thanks for your efforts
Do you plan a MP version of your engine

[abik]

I actually started BikJump because I wanted to explore some ideas on using SSE instructions and multi-threading in a chess engine. I planned to write a few versions v1.x using a simple mailbox array "to get all aspects of the UCI framework, search and opening/tablebase querying right". Versions v2.x were planned to switch to bitboards, and versions v3.x were planned for multi-threading extensions. As you can see, I am still stuck in v1.x getting all basic aspects right For one, it has taught me modesty and the highest respect for anyone writing a chess engine rated over 2000, let alone grandmaster level. But I enjoyed the whole development and progress enormously, and also continue on the original plan towards bitboards and threads. So weak or not, you have not seen the last of BikJump yet

[abik]

The live round-robin tournament is over and the table below compares the last reported RUEL scores (on exit) to the ELO performance of each version for this tournament (normalized to v1.6.1.'s RUEL). Performance trend matches release numbers now. Based on this I am looking forward to the first results of v1.8, although same-engine tournaments are known to be poor predictors (as probably already indicated by the differences in ELO perf and RUEL scores).

Code: Select all

                          NORM.        LAST
                          ELO PERF.    RUEL
1 BikJump v1.8 (32-bit)   2185         ????
2 BikJump v1.7 (32-bit)   2052         1965
3 BikJump v1.6.1          2025         2025 <==
4 BikJump v1.5            2016         1885
5 BikJump v1.4.1          2004         1875
6 BikJump v1.3            1935         1905
7 BikJump v1.2.1          1859         1785
8 BikJump v1.1.1          1800         1785


[abik]

Patrick just reported an initial RUEL of 2006 (which dropped to 2000 since) for v1.8. Too bad, this time I was really hoping for a little more

The good news is I just bought a Quad Q6700 with 6GB RAM and 64-bit Vista, so I can start more serious 64-bit and multi-threading work (and host tournaments with multithreaded engines and/or ponder on).

[H.G.Muller]

I would not worry too much about it if I were you: Patrick's ratings are wildly inaccurate for engines at the tail of the list. Most opponents in the standard gauntlet are simply too strong to provide meaningful information. A rating based on 14 points out of 128 games is less accurate than a rating based on 14 games agains equal opponents. The probability that the Bikjump with the highest initial rating determination is indeed the strongest is perhaps only 53% or so.

[Patrick Buchmann]

I would not worry too much about it if I were you: Patrick's ratings are wildly inaccurate for engines at the tail of the list. Most opponents in the standard gauntlet are simply too strong to provide meaningful information. A rating based on 14 points out of 128 games is less accurate than a rating based on 14 games agains equal opponents. The probability that the Bikjump with the highest initial rating determination is indeed the strongest is perhaps only 53% or so.

Hi,

You're absolutly right. But the first problem is to know who are the equal opponents.

If you look at the "Tournoi d'entrée" (the tournament with the 128 games for the first rating), you can find an other tournament just below wich BikJump 1.8 won. But with this tournament his rating falls to 2000.

But the most important is not the number, 2000 or 2300, it's very easy to have a rating of 3000, just change the offset. A rating is only valid in its frame of reference. The rating difference and the classification are important.
BikJump 1.7 was 79th above Ece 0.3 and below BigLion 2.23x. BikJump 1.8 is 75th above DelphiMax 2.9.

I know the problem in the UEL are the weak engines. They don't play enough games. The rule imposes 3 tournaments per division for an update of the divisions. So all the engines in Division Elite to Division Poussin are are about to play at least 30 games each month. But the "Accession & Réserve" is not affected by this rule.
It would take time and participants.

Regards,
Patrick

[abik]

I know the problem in the UEL are the weak engines. They don't play enough games.

On behalf of the weak engines, thanks for the new tournament at http://pagesperso-orange.fr/lefouduroi/ ... eserve.htm then!
This brings v1.8 rating to 2038, which is more consistent with what I was hoping for after the inter-BikJump tournament in this thread....