UCCI2WB adapter available (XQ)

[H.G.Muller]

It is now possible to run any UCCI engine under WinBoard 4.3, using the UCCI2WB adapter that is available for download at

http://home.hccnet.nl/h.g.muller/ucci2wb.zip

UCCI is the protocol used by some of the strongest XiangQi engines in the World. It seems it is only described in Chinese, though. But fortunately I found a GPLd UCCI2QH adapter, which converts UCCI engines to Qianhong plugins. Now QH protocol is described in English, and extremely simple (just some 5 commands, or so). So I was able to hack UCCI2QH to make it recognize the WB-protocol commands in stead of the corresponding QH commands. I only had to add some time accounting (QH protocol does not support anything like the WB 'time' command) to get a working UCCI2WB command, practically without having to know anything about UCCI protocol at all!

I tested it with Elephant Eye (which is the major engine of Huang Chen, who was also the author of the UCCI2QH adapter). It is a very strong open-source engine. (The Fruit equivalent of XiangQi?) Other UCCI engines can be found at http://www.elephantbase.net (a site which crashes my Intrnet explorer, but could be viewed without problems with FireFox under Ubuntu).

[Josh Pettus]

That is excellent news. And was fast! Wow did you get any sleep? I will certainly enjoy getting trounced by Eleeye from now on. Is there any special syntax for adding it into the winboard.ini or is it similar to polyglot?

[Josh Pettus]

ok I got eleye working just by putting this in my winboard.ini

"ElephantEye 3.1" /fcp='C:\Program Files\WinBoard-4.3.15\Xiangqi\Engine\eleeye\UCCI2WB'
which then looks in the the UCCI2QH.ini (in the same folder as the adapter) to find where the engine is.

Couple of little things:
If I were to add a second UCCI engine I would have to have a second UCCI2WB adapter (probably in the second UCCI engine folder) because there is no way to tell UCCI2WB to use a separate .ini file(As you would have with polyglot) because it is looking for an .ini file specifically named "UCCI2WB.ini"

How do you tell it to play at a specific level, or does it default to the highest?

[H.G.Muller]

I already had to make a bugfix: not all UCCI engines want the time in msec, like Elephant Eye. Some expect it in sec, and they really think long when you give it to them in msec! It seems there is no default for this in UCCI: for some engines the default is msec, for others it is sec. There are engines that only accept seconds. Others have an "option use millisec type check default true", but when you try to set it, do not react to it. (Well, may be my fault, as I might have guessed an invalid comamnd to set it.) Still other engines tell you they have the option in a non-compliant format (leaving out the "type check", or send the options after sending "ucciok" (when the adapter no longer looks for them).

What a stupid protocol!

As a bonus the bugfixed UCCI2WB now also passes on the PV info. (Download from same link as before.)

Anyway, I downloaded a lot of UCCI engines from http://www.elephantbase.net, and tried them out. Indeed, at this point, each engine needs a copy of UCCI2WB.EXE in its folder (all identical, so a waste of space), and an UCCI2WB.INI (tailored for them). In the future I might allow passing of a filename as argument to UCCI2WB, to be able to do with a single copy of it. When you put all UCCI engines in subdirectories of the WinBoard folder, you can conveniently install one of them (say ufx, because it has a short name) in the winboard.ini as

ucci2wb -variant=xiangqi -fd=ufx

and for the second engine

ucci2wb -variant=xiangqi -sd=ufx

both as first item in their list. When you want to run an UCCI engine you then only have to overwrite the final "ufx" of the line that appears by default in the startup dialog of WinBoard by the folder where that engine is, that's all!

My results with the engines I tested were:

Code: Select all

engine          result

3DChess         OK
Anita           OK
Binghewusi      OK
BitStronger     seems to ignore clock, is ~2x too slow at 40/1
Blcx            OK
Coony           OK
EyChesU         Does get stuck after some moves
Hice            OK
Jiaolong        OK
Jupiter         OK
King            first moves fast, then very slow, forfeits in second session without moving
Kou             OK
Lily            Seems to use no CPU time (swapping?)
Mars            completely dead at UCCI level
mrsd            is the Ki11egg Qianghong plugin with QH2UCCI adapter
NymphChess      OK
QStar           seems to ignore clock, is ~2x too slow at 40/1 (does not print PV)
Sixteener       Seems to move instantly (no PV)
Swallow         OK
THT             OK
TLXJ            OK
XQ Cyclone      OK
XQ Spirit       OK
Yssy            OK


I guess some of the engines are buggy.

[Josh Pettus]

Excellent!That's great! PV info too! Very good work.
By setting it up this way the engines will the UCCI engines play on their best settings?

[H.G.Muller]

Excellent!That's great PV info too! Very good work.
By setting it up this way the engines will all play on their best settings?

Like WB protocol, UCCI controls the engines by setting the (average) time per move. the engine is told before each move how much tme is left on its clock, and how many moves it has to do in that time. It is supposed to make best use of that time. So the only thing you have to do is sspecify the time control in WinBoard. The adapter will relay the information to the engine.

[H.G.Muller]

Note that when adding the PV, I accidentally broke something, making the time command do what the post command should have done (by deleting the wrong line after cloning code). As a result this version of the adapter makes all engines forfeit on time, by not updating the time according to what WinBoard sends it.

I just uploaded a corrected version.

[Josh Pettus]

Most Excellent! There are certainly enough UCCI engines that work to meet my needs. Thank-you so much.

Ah yes, twice the engines ran out of time moving slowly in a four min game. Sometimes when playing a game with a UCCI2WB engine (like Eleeye against 3DChess or something) it will checkmate and then the mated UCCI2WB engine will try to move and rather then be mated it will say Forfeit due to illegal move or something of that nature.

I think the one you just uploaded is the old one or perhaps the link changed. Had the same exact date on it as the older one I had and same size.

[H.G.Muller]

You are correct. I forgot to redo the zippng (it was about 1 a.m. at that time), so that the uploaded archive still contained the od executable. The correct one should be there for download now (same link). It can be identified from the release notes (in README.TXT) saying 1.1b, rather than 1.1.

A 40/1 tourney with 18 engines (15 UCCI that seemed to work without problems in the earlier test, and 3 WinBoard). The WB engines are trashed:

Code: Select all

Rank Name            Elo    +    - games score oppo. draws
   1 Binghewusi      316  188  150    17   88%   -19   12%
   2 Tlxj            226  160  139    17   79%   -13   18%
   3 Kou             223  162  143    17   76%   -13   12%
   4 Cyclone         167  153  139    17   71%   -10   12%
   5 Swallow         163  155  142    17   71%   -10   12%
   6 NymphChess      155  154  141    17   71%    -9   12%
   7 Blcx             98  144  139    17   62%    -6   18%
   8 Anita 0.2        94  150  143    17   62%    -6    6%
   9 Yssy             64  136  135    17   56%    -4   29%
  10 Thought          54  127  128    17   53%    -3   47%
  11 XQ Spirit        -8  138  144    17   47%     0   12%
  12 Coony           -11  138  146    17   44%     1   18%
  13 Jupiter         -34  137  142    17   44%     2   18%
  14 HIce           -129  144  160    17   32%     8    6%
  15 HaQiKi D flat  -223  151  183    17   21%    13    6%
  16 TJxiangqi      -324  164  218    17   12%    19    0%
  17 XQWLight       -325  166  222    17   12%    19    0%
  18 Jiaolong       -506  200  332    17    0%    30    0%


Jiaolong seems to use only a small fraction of its time, and as a result searches only 4 ply, where the top engines search 9 ply.

[H.G.Muller]

I added some more UCCI engines to the list:
Jiaolong did not play well because it announces it wants to have the time in milliseconds in a non-compliant format. I patched UCCI2WB to also recognize this format, and now it uses all its time. It still remains ne of the weakest engines, though.

Ultimate Flight (UFX) i had not installed before because the zip file was damaged during download, and I could not unpack it. I redid the download, and it works without problems.

3DChess I had simply forgotten to include in the tourney.

King does only calclate the correct time per move when movestogo<7. For larger moves to go it searches to d=1 only, and moves instantly. The work-around was to play it at 6 moves per 9 seconds, rather than 40 moves / min. It than plays fine, and turns out to be one of the strongest engines.

EyChessU seems to ly about wanting the time in msec. It just deepens forever when you do (based on the PV output). If I give it the time in seconds it does play reasonably normally, except that it still seems to ignore the time, and plays about 2x too slow at 40/1. (Like QStar and BitStronger.)

The expanded list:

Code: Select all

Rank Name               Elo    +    - games score oppo. draws
   1 3DChess            331  172  140    21   86%     9   10%
   2 Binghewusi         280  159  135    22   82%   -13    9%
   3 Kou                222  144  129    22   75%   -10   14%
   4 King 1.07          204  144  131    21   71%    16   10%
   5 Swallow            175  139  129    22   70%    -8   14%
   6 Cyclone            170  136  125    22   70%    -8   14%
   7 Tlxj               164  132  123    22   70%    -7   23%
   8 NymphChess         155  138  128    22   70%    -7   14%
   9 Blcx                96  131  127    22   61%    -4   14%
  10 Thought             70  118  117    22   57%    -3   41%
  11 Anita 0.2           66  134  129    22   59%    -3    9%
  12 Yssy                66  125  123    22   57%    -3   23%
  13 Elephant Eye 3.1    18  127  129    22   50%    -1    9%
  14 Jupiter            -11  126  128    22   48%     0   14%
  15 XQ Spirit          -57  127  136    22   41%     3    9%
  16 Coony              -59  127  136    22   39%     3   14%
  17 Ultimate Flight    -98  131  144    21   33%    30   10%
  18 HIce              -106  129  139    22   36%     5    9%
  19 HaQiKi D flat     -204  137  157    22   25%     9    5%
  20 XQWLight          -268  145  181    22   18%    12    0%
  21 Jiaolong fixed    -303  152  194    21   12%    40    5%
  22 TJxingqi          -383  158  216    22    9%    17    0%
  23 Jiaolong          -530  206  359    18    0%    22    0%


[H.G.Muller]

I made a lot of progress with HaQiKi D. It turned out to have a cripling bug in the null move, due to a missing '=' sign in the stement that was supposed to flip the sign of the material eval. i already wondered why null move hardly affected the tree size. Fixing that bug was enough to get it 'amongst the UCCI crowd':

Code: Select all

Rank Name               Elo    +    - games score oppo. draws
   1 Binghewusi         296  157  131    25   84%   -17    8%
   2 3DChess            276  141  123    25   82%    11   12%
   3 Kou                219  135  121    25   76%   -14   16%
   4 King 1.07          209  145  132    21   71%    17   10%
   5 NymphChess         182  135  122    25   74%   -13   12%
   6 Cyclone            177  129  117    25   72%   -13   16%
   7 Swallow            176  131  120    25   72%   -13   16%
   8 Tlxj               166  126  116    25   72%   -12   24%
   9 Sixteen Men        141  371  553     1    0%   276    0%
  10 Blcx               101  125  120    25   62%   -10   12%
  11 Thought            100  115  111    25   62%   -10   36%
  12 Anita 0.2           76  127  122    25   60%    -9    8%
  13 Yssy                70  120  118    25   58%    -8   20%
  14 HaQiKiD eval         3  123  125    23   48%     9   17%
  15 Elephant Eye 3.1    -5  121  124    25   48%    -5    8%
  16 Jupiter            -12  118  121    25   48%    -5   16%
  17 HaQiKi D lmr       -30  120  123    23   43%    10   26%
  18 HIce               -45  123  128    25   44%    -4    8%
  19 XQ Spirit          -55  119  126    25   42%    -3   12%
  20 Ultimate Fight     -69  123  132    24   38%    20    8%
  21 Coony              -75  122  130    25   38%    -3   12%
  22 HaQiKi D n2       -130  126  139    23   30%    14    9%
  23 HaQiKi D flat     -245  132  155    25   22%     4    4%
  24 Jiaolong fixed    -256  137  165    24   17%    28    8%
  25 XQWLight          -303  141  180    25   16%     6    0%
  26 TJxiangqi         -412  154  215    25    8%    11    0%
  27 Jiaolong          -553  215  377    18    0%    23    0%


The version HaQiKi D eval is one which recognizes some patterns related to king safety, next to the piece-square tables stolen from XQWLight, and some dependence of the piece base values on total material added by myself. There still is a lot I want to improve there, but I was already so happy with the result that I released this version as HaQiKi D 0.4.

[edwardyu]

I added some more UCCI engines to the list:
Jiaolong did not play well because it announces it wants to have the time in milliseconds in a non-compliant format. I patched UCCI2WB to also recognize this format, and now it uses all its time. It still remains ne of the weakest engines, though.

Ultimate Flight (UFX) i had not installed before because the zip file was damaged during download, and I could not unpack it. I redid the download, and it works without problems.

3DChess I had simply forgotten to include in the tourney.

King does only calclate the correct time per move when movestogo<7. For larger moves to go it searches to d=1 only, and moves instantly. The work-around was to play it at 6 moves per 9 seconds, rather than 40 moves / min. It than plays fine, and turns out to be one of the strongest engines.

EyChessU seems to ly about wanting the time in msec. It just deepens forever when you do (based on the PV output). If I give it the time in seconds it does play reasonably normally, except that it still seems to ignore the time, and plays about 2x too slow at 40/1. (Like QStar and BitStronger.)

The expanded list:

Code: Select all

Rank Name               Elo    +    - games score oppo. draws
   1 3DChess            331  172  140    21   86%     9   10%
   2 Binghewusi         280  159  135    22   82%   -13    9%
   3 Kou                222  144  129    22   75%   -10   14%
   4 King 1.07          204  144  131    21   71%    16   10%
   5 Swallow            175  139  129    22   70%    -8   14%
   6 Cyclone            170  136  125    22   70%    -8   14%
   7 Tlxj               164  132  123    22   70%    -7   23%
   8 NymphChess         155  138  128    22   70%    -7   14%
   9 Blcx                96  131  127    22   61%    -4   14%
  10 Thought             70  118  117    22   57%    -3   41%
  11 Anita 0.2           66  134  129    22   59%    -3    9%
  12 Yssy                66  125  123    22   57%    -3   23%
  13 Elephant Eye 3.1    18  127  129    22   50%    -1    9%
  14 Jupiter            -11  126  128    22   48%     0   14%
  15 XQ Spirit          -57  127  136    22   41%     3    9%
  16 Coony              -59  127  136    22   39%     3   14%
  17 Ultimate Flight    -98  131  144    21   33%    30   10%
  18 HIce              -106  129  139    22   36%     5    9%
  19 HaQiKi D flat     -204  137  157    22   25%     9    5%
  20 XQWLight          -268  145  181    22   18%    12    0%
  21 Jiaolong fixed    -303  152  194    21   12%    40    5%
  22 TJxingqi          -383  158  216    22    9%    17    0%
  23 Jiaolong          -530  206  359    18    0%    22    0%


Hi HGM,

I am Edward Yu, author of UCCI Chinese Chess engine - EychessU. Congrualation to your success in the 2009 Olympaid
My previous engine may have bugs but now I think I have fixed it. I have tried it with UCCI2WB and it seems OK.
However, I don't have a decent Traditional Chinese Chess Board.bmp, so that I can only use the "marble" board from this forum.

I cannot attach attachment in this forum. Would you send your xiangqi.bmp (the middling boardsize) to my email eykm@yahoo.com ? I can then send my latest EyChessU engine to you.

Thanks!

Edward Yu
From Hong Kong - China

[H.G.Muller]

Thanks!

It seems that there is a World Championship Computer Xiangqi next year September in Taiwan, and the Taiwanese participant involved in organizing it even wants to invite me there. But, as you can see from the tournament resut above, my engine is not all that good on an absolute scale. (I expect your EyChessU to beat it, once I get it running.) But it would be fun to go there, and there is still more than a year to work on improving my engine.

Only just yesterday I posted a zip file on my website that contains a WinBoard configured to act as Xiangqi GUI. It also includes the XIANGQI font and the bitmap. The winboard.exe in there has somewhat improved XQ support, compared to the version 4.3.15 that is hosted on WB forum: when the option 'detect mates' is on, it will forfeit an engine that is perpetually checking. No chase detection yet, though. I hope to add that too, soon. The file can be downloaded from http://home.hccnet.nl/h.g.muller/XQgui.zip , but I already e-mailed it to you as well.

The bitmap is designed for -boardSize middling. With the font-based rendering, in principle any board size will have good Xiangqi pieces, as the font is scalable. But the board is more tricky: I tried to make the algorithm for cutting out squares of the bitmap such that it is reasonably resistant to requesting another board size than the original bitmap (by cutting out overlapping squares, if necessary), but it is not perfect, and the grid scales with a slightly larger factor than the board. In addition the diagonal lines in the palaces might show little gaps or excess line segments in other sizes. If the size is not too different, it is bearable, though. I am still thinking how this could be avoided. So perhaps I will change the cutting algorithm, in which case I will probably also have to start from a different bitmap.

When I get back home from Pamplona, I could test EyChessU against all the other engines in the list above, to add it to the ranking.

[edwardyu]

I try to use PSWBTM tournament manager for the UCCI2WB XiangQi engines, but cannot get the XiangQi starting position to work. I use the following xiangqi.fen :
rheakaehr/9/1c5c1/p1p1p1p1p/9/9/P1P1P1P1P/1C5C1/9/RHEAKAEHR w

When I run the tournament, PSWBTM always use the Western Chess default position for XiangQi. Is there a standard xiangqi.fen or the PSWBTM does not support XiangQi yet?

Edward Yu

[H.G.Muller]

There are two ways to do this.

The first way is to edit the winboard.ini file just before you start the tournament, and add the line:

/variant=xiangqi

(does not matter where; I usually append it at the end). This will than remain preserved during the entire tournament, as PSWBTM suppresses the saving of settings on exit of WinBoard. (If you run WinBoard by hand from the command line, or from the PSWBTM engne manager, it usually does save its settings on exit, and this erases the /variant setting, as this is one of the options that is not saved from run to run, and starts then at the default next time you run WinBoard.)

The second method is to install all Xiangqi engines in PSWBTM with in the parameters field the options:

WBopt /variant=xiangqi

(At the end, after other options, when these are needed.) This then automatically passes the /variant=xiangqi option on the command line with which PSWBTM invokes WinBoard for every game in which this engine plays.

You can then play all games from the default position, no need to pass a FEN. Unfortunately WinBoard cannot identify the variant from a FEN that you paste. It can of course determine the board size, but there are too many 8x8 variants that use different pieces or use different letters for the same pieces to make this possible in general. (I could of course make it slightly less stupid, and have it guess that a 9x10 board means a Xiangqi position, and only use the FIDE-Chess position as default for all 8x8 FENs.)

Unfortunately PSWBTM has no input field in its tournament-setup dialog for passing extra options to WinBoard that are needed for the entire tournament. This would be a bit more user friendly than having to edit them in the winboard.ini file. The latter method works if you only use that WinBoard for the PSWBTM tournament. But if you want to use WinBoard during the tournament for something else, you now have to make a separate copy that uses its own winboard.ini file.

[H.G.Muller]

There definitely is still something very wrong in the time management of EyChessU. On my 2.4MHz Core 2 Duo is consistenty behaves as follows at 40 moves/min:

The first 3 sec it does about 32 moves. Then the remaining moves get progressively longer, and use up the remaining 30 sec of the first session. So closely, that it often forfeits on tme on move 39, while it had 4-5 sec for that single move.

If it survives the first time contol, it suddenly starts to play very quick, like the next 10-15 moves in only 1-2 sec total. Then it starts to progressively slow down agan, and very ofteh forfeits on time at move 79.

[Norbert Raimund Leisner]

Hello HGM,

what´s about the prgram Qianhong by Jeremy Craner? URL = http://www.jcraner.com/qianhong/index.html

Regards,
Norbert

[H.G.Muller]

Qianhong is all of the following three: a GUI, and engine and a protocol. In Qianhong terminology engines are called plugins.

QH protocol has no provisions for time management. It requests the engine to play at a certain 'level', and ech engine interprets a given level in its own way. (Some as fixed depth, other as fixed time.) In some modes the QH GUI can send a "move now" command to an engine, but not all engines support that. All these modes of time management are very wastefull.

There are not many native QH engines, and most of them are rather weak. The Qianhong engine itself is extremely weak: it is a non-Shannon engine comparable to N.E.G. 0.3d for normal Chess. As it does not search ahead, it uses near zero time.

UCCI engines can play under QH using the UCCI2QH adapter, which implements a level as a fixed time per move (like the WB st command). This kind of wrecks the advanced time management supported by UCCI.

I wrote a simple QH2WB adapter to play QH engines (called plugins) under WB; of the non-UCCI engines only Ki11egg is strong. VSCCP is already much weaker than MaxQi (the Xiangqi version of micro-Max). Now that UCCI2WB exists it is much better to play UCCI engines under WB with the aid of that. Unfortuatelu QH2WB is dependent on cygwin1.dll for runnning.

[Josh Pettus]

In light of your very recent commit on your branch where you can invoke a second adapter the way you would polyglot e.g. /uxiAdapter="UCI2WB -s %fcp %fd" with /fUSI e.g., what is the sytax for involking UCCI2WB? As far as I know, UCCI2WB isn't like your UCI2WB and still needs to be directed to a text file with an engine's name and directory. Or did you come up with another system?

[H.G.Muller]

At some point I also implemented a file-less mode in UCCI2WB. The commnd syntax is:

UCCI2WB -noini ENGINE.exe ENGINEDIR

(The latter is optional; if it is missing UCCI2WB derives it from the engine path.) This would be acheived by

/uxiAdapter="UCCI2WB -noini %fcp %fd"

Currently I have made it such that the /fUCCI and /fUSI options are synonyms; I am not 100% sure if this was a good idea. The underlying philosophy was that people would either use WB for Chess, Xiangqi or Shogi, nd hardly ever for all. If they wanted to use it for Xiangqi and Chess, (say), they are likely to configure it anyway through an ini file for one or the other (because they want different looks), and might as well redefine the /uxiAdapter in that file. There was a need for two adapter commands, though, because for Xiangqi alone, you would want to be able to specify both UCI and UCCI engines.