Comments/Ratings for a Single Item

OK, I started a new thread for this to put an end to the off-topic
posting.

OK, the Seirawan algorithm Fairy-Max uses obviously sucks. Let me tell you
the full story about this:

A colleague engine author of mine recently converted his very strong Chess
engine (2800 Elo, so about 800 Elo stronger than Fairy-Max) to an S-Chess
engine, as a commercial project. I learned this last spring, and we came to
the conclusion that it would be nice if we could use XBoard as a GUIfor it.
So we agreed that he would convert his engine to WinBoard protocol (it was
originally UCI), and that I would add it to the variants supported by
WinBoard. And we defined the necessary protocol extensions that would be
needed to encode the gating moves.

After that, nothing happened for some time. (I was busy writing a Shogi
engine.) But a few weeks ago we set a release point for WinBoard 4.5.0,
(early December), and I wanted Seirawan Chess to be in before that. So I
just plunged ahead without an engine, and implemented it in WB. But I had
nothing to test the engine functions, which is a bit risky. So I decided to
do amakeshift implementation of Seirawan Chess in Fairy-Max, which already
knew the pieces, so that I just had to add the gating. But I implemented it
through a kludge, as my only interest was really to have an engine that
would support the protocol, and I did not worry much about playing quality.
(As that would be exceeded by 800 Elo when the other engine became
available anyway.)

So what I did was implement gating at the levelof the game move, but not in
the search. Fairy-Max selects its moves without being aware that any gating
is possible, and if it happens to move a piece from the back-rank, it gates
in any yet un-gated piece (first Hawk, then Elephant). I later refined this
a bit by checking that the score of the opponent reply after gating was not
suddenly much higher than it was in the original search without gating,in
which case I would cancel the gating, as I was apparently gating in a
super-piece on a square that was under attack. The idea was that by gating
in at the earliest opportunity, before there were any tactical
complications, I could get away with doing it without thinking about it.

Now I see this reasoning was flawed: Fairy-Max does his own gating, all
right, but the opponent can of course still delay the gating for as long as
he wants. And then surprise Fairy-Max with it while engaged in tactics.
This is exatly what happens here. Fairy-Max plays 9.Bg5, because he thinks
he can refute 9... Qxg5 with 10.Hc7+ to grab a Rook. (That this locks in
its Hawk this way, so that it might be doomed in the long run, is beyond
its strategic abilities to recognize, and even in engines with mobility
evaluation the poor Hawk mobility would never outweigh a Rook. Otho-Chess
engines program this as a pattern (with N in stead of H) to discount such a
doomed Knight.) So the plan is good, but of course did not count on an
Elephant suddenly appearing on d8(9... Qxg5/E) to defend c7, and thwart
white's ill intensions...

I guess this proves that my makeshift approach is just too simple,and that
you can never play decent Seirawn Chess this way. The gating really has to
be fully in the search. I amnot sure it would be worth it todo this in a
general-purpose variant engine, though. Anyway, thank you for exposing this
weakness.

As for the possibility to drop the pieces: the strong engine I was talking
about actually supports two variants: official Seirawan (or S-Chess, as
they call it nowadays), and a variant where you can drop the pieces on the
back rank whenever you want. I wanted to make it possible to play both
variants as 'Seirawan' under WinBoard, and let the engine decide which
drops or gatings are legal according to its current settings. WinBoard also
does not check, for instance, if the piece that effects the gating has
already moved or not. 

So one could say that Seirawan is only a partially defined variant in
WinBoard. This is not uncommon: some of the other variants are even
supported to a lesser degree, and can only be played with legality testing
off to prevent WinBoard from interfering with the game because it thinks
some of the moves are forbidden. In that case the engine is in full
control, and when it decides an entered move is illegal, it tells so to
WinBoard, and WinBoard will then refuse the move from the user.

I probably should let WB forbid drops anywhere else than on the back-rank,
however. These would be illegal under any circumstances, and it is easy to
do. Note that I only started implementing Seirawan a week ago, or so, and
that it is still in development.

> Mats wrote:
> WinBoard for Xiangqi doesn't allow analysis mode, and I cannot retract 
> moves and make another move, so this software is largely useless as
> it stands now.

I was shocked to hear this, as I remembered explicitly having implemented
analysis mode in UCCI2WB. You seem to be right, however: for none of the
included engines analysis seems to work. For MaxQi this is no surprise, of
course; it does not even print a PV, so I never bothered to implement
analyze mode in it. Fairy-Max also does not support analysis, as it does
not support pondering at all, as analysis is a form of pondering.

I am ashamed to say that HaQiKi D does not support analysis either. There
really is no good reason for this, as it does support pondering. The SMP
version I am curently developing actually does support analysis (as I
support SMP by starting several engine processes with a shared hash table,
where one of the processes is playing, and will eventually perform a move,
and all other processes are analyzing the same position to fill the hash
table with results the master process can use). I have not worked on HaQiKi
D since May, I had forgotten that the single-CPU version did not have this
yet.

But you are definitely also right about another thing: HaQiKi D does not
support 'undo'! This escaped my attention completely, as I only use it in
engine-engine games. The 'undo' command is recognized in the interface
code, which I simply stole from Fairy-Max, but it only resets the game to
the initial position, and the routine that is supposed to re-enter all
played moves is commented out. I will fix all this as soon as possible.
Isee I also forgot to include HaQiKi D in the list of engines in the WB
startup dialog...

Elephant Eye simply seems broken. I have not figured out what the problem
there is yet, but both pondering and analysis seem not to work. This is not
a problem with the UCCI2WB adapter or with WinBoard, that much I already
know. I do have other UCCI engines that seem to work for analysis without
problems. In particular Binghewusi. A lot of free UCCI engines can be
downloaded from www.xqbase.com (you will need a translator, as it is all in
Chinese, but it is worth it). You can download Binghewusi (sometimes
referred to in translations as 'Soldiers Creek 54') there. If that does
not work, or you want a newer version, I can send it to you by e-mail.

I think I fixed HaQiKi D (implementing both 'undo' and 'analyze'
commands). The fixed version is now included in the
http://hgm.nubati.net/WinBoard-4.5.beta.zip download. In a quick test it
seemed to work (with still a small cosmetic error that it reports the score
from the POV of the side not to move, as it usually does when pondering).

I tried to figure out what is wrong with Elephant Eye that it would neither
ponder nor analyze, but it merely seems that this was because it was still
in book. When I first do a crazy move to get it out of book, pondering and
analysis seem to work normally.

I also fixed a bug in HaQiKi D. Analysis and undo both worked, but undo did not work in analysis mode... HaQiKi D is a bit hard to test, because it does not do illegal-move checking. (I was kind of in ahurry when I wrote it.) But from the things I tried now, it seems completely OK.

I also improved UCCI2WB, so that it now prints time and node-count for Elephant Eye (who sends those in separate infomessages) as well.

I put it all in the http://hgm.nubati.net/WinBoard-4.5.beta.zip package.

I improved the implementation of Seirawan Chess in Fairy-Max. It now has
the gating moves fully integrated in its search process, so it should be no
longer possible to surprise it by gating. (With the exception of gating at
the Rook position in castling.) In theory it can still be stupid, because
it does not realize there is a last opportunity for gating. But since I
built in a large preference for gating moves, it always gates early, so it
is not likely to be ever left with un-gatable pieces in hand.

Still obtainable at http://hgm.nubati.net/WinBoard-4.5.beta.zip .

Now it plays a very decent game. I'm thinking of publishing my Zillions
Seirawan implementation, too, since this WinBoard implementation exists
anyway. In 2007 I asked Seirawan whether he accepts that I publish this,
but he didn't like it. It turns out that they have always wanted a program
for online play against other people. Zillions can be used for online
play.
/Mats