Engine release: CrazyWa 1.0.1

[hgm]

CrazyWa is a WB engine for playing Chess variants with drops. This is the first release of a fully functional version: all implemented variants now work, and have been tested. The supported variants are:

Crazyhouse (8x8)
mini-Shogi (5x5)
Judkins' Shogi (6x6)
Tori Shogi (7x7)
regular Shogi (9x9)
Wa Shogi (11x11)

As for the playing stregth: this was only tested in bullet games. In Crazyhouse it is close to that of Imortal1.0 and Sunsetter 7g. In Crazyhouse, mini-Shogi and Tori-Shogi it is about 100 Elo stronger than Sjaak II, in Judkin's Shogi about equally strong. In regular Shogi it scored ~34% against TJshogi. It scores about 90% against Sjaak II in Wa Shogi with drops; Sjaak II is the only other engine I know that plays this (but does not seem very strong at it).

Both a Windows executable and a Linux amd64 binary are available for download.

For Wa Shogi and Tori Shogi it is recommended you use the latest XBoard from http://hgm.nubati.net , or the WinBoard-AA package as GUI, because the piece images CrazyWa requests from the GUI are not supported in WinBoard/XBoard 4.8 or older. (The WinBoard-AA package uses the anti-aliased piece images of XBoard 4.9.)

[Evert]

It scores about 90% against Sjaak II in Wa Shogi with drops; Sjaak II is the only other engine I know that plays this (but does not seem very strong at it).

It can probably be greatly improved by actually giving it sensible piece values.

Anyway, this should be an interesting sparring partner for SjaakII in drop games (I can even give it time odds for more efficient testing)!

[hgm]

The piece values are not the problem. Sometimes Sjaak is +20 ahead, and CrazyWa agrees that it is at -18, and then Sjaak still almost always loses. It offers virtually no resistance against its King being drawn / chased to the center, after which it is toast. The same happens in regular Shogi. I would expect this to also be a problem in Crazyhouse, but it isn't.

[MikeB]

CrazyWa is a WB engine for playing Chess variants with drops. This is the first release of a fully functional version: all implemented variants now work, and have been tested. The supported variants are:

Crazyhouse (8x8)
mini-Shogi (5x5)
Judkins' Shogi (6x6)
Tori Shogi (7x7)
regular Shogi (9x9)
Wa Shogi (11x11)

As for the playing stregth: this was only tested in bullet games. In Crazyhouse it is close to that of Imortal1.0 and Sunsetter 7g. In Crazyhouse, mini-Shogi and Tori-Shogi it is about 100 Elo stronger than Sjaak II, in Judkin's Shogi about equally strong. In regular Shogi it scored ~34% against TJshogi. It scores about 90% against Sjaak II in Wa Shogi with drops; Sjaak II is the only other engine I know that plays this (but does not seem very strong at it).

Both a Windows executable and a Linux amd64 binary are available for download.

For Wa Shogi and Tori Shogi it is recommended you use the latest XBoard from http://hgm.nubati.net , or the WinBoard-AA package as GUI, because the piece images CrazyWa requests from the GUI are not supported in WinBoard/XBoard 4.8 or older. (The WinBoard-AA package uses the anti-aliased piece images of XBoard 4.9.)

Thanks HGM - would the source be available to build a macOS version?

[hgm]

I asked Joshua Pettus to compile it for Mac. (But that was only 10 hours ago.) He could also put it in the dedicated XBoard-Shogi OSX App. We even have Wa-Shogi kanji pieces for XBoard, with the aid of the -inscriptions option. (I.e. the SVG image of a blank Shogi tile, and let XBoard render the kanji from unicode.) That looks pretty decent:



Or for people that insist on 2-kanji representations:

[Ferdy]

CrazyWa is a WB engine for playing Chess variants with drops. This is the first release of a fully functional version: all implemented variants now work, and have been tested. The supported variants are:

Crazyhouse (8x8)
mini-Shogi (5x5)
Judkins' Shogi (6x6)
Tori Shogi (7x7)
regular Shogi (9x9)
Wa Shogi (11x11)

As for the playing stregth: this was only tested in bullet games. In Crazyhouse it is close to that of Imortal1.0 and Sunsetter 7g. In Crazyhouse, mini-Shogi and Tori-Shogi it is about 100 Elo stronger than Sjaak II, in Judkin's Shogi about equally strong. In regular Shogi it scored ~34% against TJshogi. It scores about 90% against Sjaak II in Wa Shogi with drops; Sjaak II is the only other engine I know that plays this (but does not seem very strong at it).

Both a Windows executable and a Linux amd64 binary are available for download.

For Wa Shogi and Tori Shogi it is recommended you use the latest XBoard from http://hgm.nubati.net , or the WinBoard-AA package as GUI, because the piece images CrazyWa requests from the GUI are not supported in WinBoard/XBoard 4.8 or older. (The WinBoard-AA package uses the anti-aliased piece images of XBoard 4.9.)

I am testing this now vs TJchess 1.3.

At a later time I will test this vs Sunsetter 9 at blitz on 50 games. If it gets 45% or more I will match it at TC 75m + 10s inc/move best of 24 games the winner will qualify for stage 2.

[flok]

Is CrazyWa op z'n Utrechts? Crazy, Wa?

[hgm]

No, it is from Wa Shogi, which can be played with or without drops. Like Chess with drops is known as Crazyhouse, I refer to Wa Shogi with drops as Crazywa.

[Evert]

Does it also play Wa without drops? If it can do that, it could play orthochess as well (not that this is in any way an interesting feature).

[hgm]

I am testing this now vs TJchess 1.3.

At a later time I will test this vs Sunsetter 9 at blitz on 50 games. If it gets 45% or more I will match it at TC 75m + 10s inc/move best of 24 games the winner will qualify for stage 2.

Well, it most likely will not achieve that. In my 40-moves/min tests is does better than 45% against Sunsetter 7g, but Susetter 9 should be much stronger.

As far as Crazyhouse is concerned, the only difference is that in 0.1.4 I disabled storing QS results in the TT. Because I limit QS depth, there is a hidden depth parameter.The proper implementation would really reserve depth 0-10 for QS, rather than call it all d=0. (I tried to do that, but I might have done it wrong, as this caused a regression of ~100 Elo.) Not storing QS nodes is a course way of preventing the full-width search of a later iteration arrives in QS in a node that before was very deep in QS, close to the depth limit, and thinks it can use the d=0 hash hit to satisfy the search request.

Not storing QS nodes makesit score only 45% agaist the version that does store them. But it scored better against the other test opponents. And I figured that at long TC an additional advantage would be a lower load factor of the transposition table.