Deep Sjeng

From Chessprogramming wiki
Jump to: navigation, search

Home * Engines * Deep Sjeng

Deep Sjeng,
a private, and former commercial chess engine by Gian-Carlo Pascutto, which emerged in 2002 from the 12.7 closed source branch of the chess variant and chess playing open source engine Sjeng [1] . Opposed to other commercial engines with the surename "deep" to indicate the version is able to play on multiple processors and sold for almost the double price than their "none deep" counterparts, Deep Sjeng, albeit able to play on multiple cores as well, is the native engine name for single as well as multiple processors. Deep Sjeng was market since 2003 by Lex Loep's company Lokasoft [2] . It came with the ChessPartner graphical interface and supports UCI and the Chess Engine Communication Protocol. Version 2.X with the Mayura Chess Board [3] and its third incarnation Deep Sjeng 3.x were distributed via Gian-Carlo's own site, but Deep Sjeng is no longer for sale [4] .


Deep Sjeng played many computer chess tournaments. It participated (so far) at six World Computer Chess Championships [5] :

Edition Tournament Ranking Participants Score Games
11th WCCC 2003 Graz 11 16 4.5 11
12th WCCC 2004 Ramat Gan 10 14 5.5 11
13th WCCC 2005 Reykjavík 3 12 7.5 11
15th WCCC 2007 Amsterdam 6 11 6 11
16th WCCC 2008 [6] Beijing 8 10 3.5 9
17th WCCC 2009 [7] Pamplona 1 9 6.5 9

Deep Sjeng further played various Dutch Open Computer Chess Championships, International CSVN Tournaments, Livingston Chess960 Computer World Championships, dominated The Chess Programmers Tournament with three wins so far from four editions, and won the Italian IOCSC 2010. Online Deep Sjeng played multiple CCT Tournaments, where Deep Sjeng won the CCT12 in 2010 and CCT13 in 2011. Since 2008, Deep Sjeng participated the ACCA World Computer Rapid Chess Championship always with top rankings, winning the WCRCC 2012.


Deepsjeng2 1.png

Deep Sjeng 2.5 with Mayura Chess Board [8]

Parallel Search

Gian-Carlo Pascutto in a reply to Georg von Zimmermann on Deep Sjeng's parallel search [9] :

How is Deep Sjeng going? What did you use to understand the parallel algorithms you are using (which ones) ?

I started out with ABDADA (described in ICCA journal article and used in Amy), which got me a speedup of +- 1.2. I went on to try PVS (Crafty 15.0 and described in several articles about parallel search) which got me a speedup of 1.2-1.3.
1.3 wasn't enough, so I 'bit the bullent' and started looking at DTS (Cray Blitz). Unfortunately, DTS is both hideously complicated and requires a nonrecursive search and a p2p design. I spent some time working on a variant of DTS that can work with a recursive search function and a master-slave design and that is what I am using now. It still needs a lot of test work, but current results indicate a speedup of about 1.6. 

Automated Learning

In 2007, Gian-Carlo's experimental program Stoofvlees aka Deep Sjeng 2.7 [10] with a set of feature recognizers coupled to a neural network [11], had its evaluation function entirely automatically learned from "watching" Grandmaster games. The results were incorporated into Deep Sjeng 3.0 [12]. The engine has noticeably improved in strength, particularly in the areas where it was less optimal before.

Forum Posts

External Links


Up one Level