Comments by HGMuller

Ai, you mentioned the unspeakable word! Now we will be sued for sure!

I was out for the weekend, but I let the match run on unsupervised. When I
returned I saw it was still going.

I noticed that after 213 games, the program that values Falcon below Rook
has taken the lead, 120-93 (56%). This starts to be significant: the
statistical error in 213 games is only ~3%. So this is a 2-sigma
deviation, meaning that it is highly unlikely (only ~2%) that it is better
to value Falcons above Rooks.

I put up the zip-file with Fairy-Max, a confiuration file including Falcon Chess, and the WinBoard_F GUI, all packed together as a ready-to play combination on my website. For those who want to try it out, the download link is:

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

(Beside Falcon Chess it also contains definitions for normal Chess, Capablanca Chess, the unspeakable variant, Knightmate, Shatranj, Courier Chess, Cylinder Chess.)

I decided to terminate the Falcon-Chess match (I had let it run so long only because I was away anyway) after 676 games, at a score 356-320 (53%). Although this seems only a small difference, this is always to be expected in symmetric playtesting: they start out with equal material, and not all games will have R for F trades. And even for those that have: when I ran games that were imbalanced from the outset (RR vs FF), I never saw more extreme scores than 57% (and for only a single R-F imbalance, that would be halved).

Nevertheless, over this many games, a result of 53% is significantly
different from equality. One side scored 36 more wins than the other. The
standard deviation of the number of wins in 676 games is about 24. So the
deviation translates to 1.49 standard deviations, which has only a 6.7%
probability of occurring if the two engines were equal.

So the conclusion must be that it is almost certain (more than 93%
confidence) that it is better to value Falcons slightly lower than Rooks
than to value them above Rooks (in a simple scheme that has fixed piece
values throughout the game, like Fairy-Max). Current Fairy-Max has R=475, so I set F=450. (B=350, P=100.)

I have not had time to watch many of the games. (I will study the PGN
later.) Usually Fairy-Max is very materialistic: it has very little
positional evaluation, and the diffence between having a Knight on a1 and
on e4 is only about a quarter Pawn. If it lets undefended Pawns live, it
is usually because they are heavily poisoned.

This shows the 10x10 side, the Chess board is on the other side.
It was made by a Chess-supplies company called 'Revanche', I don't know
if they still exist. It measures 47 cm on the side. (A CD is shown in
the photograph for size reference.) Height of the shown King is 81mm.

If I wanted to play on a subset of the squares, I would clip something
over the edges that covers 10, 8 or 6 squares.

George:

I am close to releasing a new version of the WinBoard Chess GUI, which I expect to become widely distributed amongst Windows and Linux users. Most of the people no doubt will only be interested in playing 'normal' Chess, but I planned to include Fairy-Max with the distribution, as it can both play normal and many variants. The distribution package would look a lot like the experimental version that can be downloaded from the link I gave in the post below.

In this new WinBoard I included some support for Falcon Chess in the menus: people can click 'Falcon' as one of the options in the variant menu, and will then be able to use WinBoard as Graphical User Interface for an engine that could actually play it, provided they use one of the wildcard pieces supplied by WinBoard to represent the Falcons. I suppose you have no objection against this, and as WinBoard itself does not play the game, but only acts as a display, it probably does not fall under the patent anyway. As it would only be useful to use WinBoard this way if you did have a Falcon-Chess-playing engine, and the user can communicate with such an engine only through WinBoard, I make WinBoard pop up a licensing message, mentioning your name and the patent number. (Similar to what I make WinBoard do in Gothic Chess.)

Fairy-Max, in the download below, does include Falcon Chess as a pre-defined variant, in its fmax.ini file, and when programmed by this preset, actually does play the game (as you could watch in the broadcast). Can I leave this in, or would you prefer me to take out the Falcon-Chess game definition?

Please let me know ASAP, I hope to be able to releas this weekend.

Well, if we woud assume that the true score prcentage would be that of the
total run, i.e. 356/676 = 52.66%, the expected scores for the first 213
and last 463 games would be 112 and 244 points, repectively. So the actual
result is off by 8 points. This is kind of normal for 213 games: the standard
deviation over N games is ~0.5*sqrt(N), or 7.3 for 213 games. So the
observed deviation is indeed approximately 'standard'.

It is clear already from the 52.66 result that Rook and Falcon are very close matches, (although the difference from 50% is significant!), and that the details of the strategy (for which number of Pieces or Pawns would you prefer Falcons over Rooks, and for which the opposite) could improve the results much more than fine-tuning of a single Falcon value used during the entire game in the range R-0.5 to R+0.5.

Such subtle strategic evaluations are a bit beyond the scope of a simple program like Fairy-Max, which also doesn't use such considerations for the other pieces (e.g. it does not know about Rooks on open files, the Bishop pair or doubled Pawns). To do better, I would have to convert Joker80 to hndle the Falcon. (I already started with that, but I was a bit too busy recently to finish the job.)

Joker80 is supposed to do (b). But of course the FEN first has to pass
through WinBoard before it is handed to Joker. This version of WinBoar
doesn't necessarily require Rooks in the corners to castle with, as I
could imagine there would exist variants that had some other piece there,
with which you could castle. (Just like Knightmate allows Rooks to castle
with the Royal Knight.) 

Allowing empty squares to castle was a very bad idea, though.

I am not sure what the problem was. The link did not work (website was
down?) when it was first posted, and when I look now, it points to a Smirf
game.

If you implement the moving board by simply covering the squares that are supposed to be holes with an obstacle, as I suggested below, you would not have that problem. You could use draughts chips for marking the holes, but it would perhaps be nicer to have some square tiles of a third color that completely cover the squares.

This is much more likely to work than actually trying to construct a board where all squares can move.

My first reaction is that something must have gone wrong installing the new
WinBoard alpha, and that you are somehow still using an old one. When I
start the alpha with this FEN with two Joker80 engines, I can see in the
debug file that the FEN really sent to the engine only has KQ, not KQkq
castling rights.

So first check: click the 'help -> about' menu item on the WinBoard that
you are using, and check if it says WinBoard 4.3.14k. If not, we have a
mixup.

If this is indeed WinBoard 4.3.14, it is not possible for me to make any
sensible guess as to what is wrong without having the winboard.debug file
of this game. (This file records all communications between engine and
WinBoard.) WinBoard should be started up with the option /debug on the
command line to instruct it to make this file. (And when an irregularity
happens, immediately rescue that file by renaming it, before you restart
WinBoard!)

If Joker80 would castle when receiving a FEN that says it has no castling
rights, it is inconceivable that H would not have noticed that before, as
I played many thousands of games with it, also from positions without
Rooks (but then I also deleted the castling rights for that side, as
WinBoard 4.3.14 does here in my tests too). As you can see, WinBoard does
not accept such a castling as a legal move when it is actually done, not
even the version that erroneously passed the non-exsting castling rights
to the engine. And even though I don't examine every game I play, I
always scan the PGN file of a large run with a text editor for the
occurrence of the string 'forfeit' and 'on time', to see if there were
problems.

Checkering the board is not really useful in the absence of Bishops.
So in stead the light / dark square colors are used to contrast the
river banks and palaces. The Mandarins are indicated by a mandarin cap,
which is then used to represent Ferz throughout all variants in WinBoard.
The Elephants and Cannons speak for themselves.

The 'Generals', which we of course tend to think of as Kings, are represented
by the WinBoard Wazir symbol. I am not sure if it would be better to use the
ordinary King symbol here to represent royalty, or use the Wazir to indicate
the moves of the piece. The ornament on the Wazir turban is of course ideal
for indicating royalty. A normal Wazir could use a turban without ornament.