Kiwi 0.6a beta

[Alessandro Scotti]

Hi folks,
after seeing Kiwi play so poorly I've finally followed Guenther's advice and gave him a completely new evaluation.
Because of this change, what previously was a clueless and timid bird is now a bloodthirsty attacking monster that will relentlessly go after the enemy king, leaving behind a trail of destruction and flaming pawns.
Well, not quite. Actually, it's the same old patzer but at least let me dream for a moment!
At any rate, evaluation now includes king safety, shield, mobility, passed pawns and some more. When I tested that against version 0.5a, I got more or less 50%, and the same thing in a further test against Yace (here, make it 20% of course). This amazes me to no end. I have to wonder what Kiwi does with his evaluation data, if a radically different code causes no change at all in the engine strength, either for bad or good!
Anyway, some search bugs have now been found and fixed. Also, null move and futility were throwing away a whole lot of good moves and that should have been somewhat improved now... these changes can have a great impact on the engine strength!
I could only run a few test games at super-blitz time, but the results are encouraging so far. At least, it's refreshing to see Kiwi trying to move his pieces and *do something*, I've seen some very interesting games lately!
If not for the strength, I hope you will enjoy the "original" playing style... and oh, I can always use the occasional test tournament!

P.S. Kiwi is available at the usual place:

http://www.ascotti.org/programming/chess/chess.htm

[Guenther Simon]

Thanks for the new Kiwi Alessandro. It complains already
about the missing bitbases

Regards,
Guenther

[Tord Romstad]

At any rate, evaluation now includes king safety, shield, mobility, passed pawns and some more. When I tested that against version 0.5a, I got more or less 50%, and the same thing in a further test against Yace (here, make it 20% of course). This amazes me to no end. I have to wonder what Kiwi does with his evaluation data, if a radically different code causes no change at all in the engine strength, either for bad or good!

I've seen the same thing many times. Even very radical changes to my evaluation function often do not change the playing strength measurably. It sometimes feels like I could replace my eval with a random number generator without any noticable effect.

I could only run a few test games at super-blitz time, but the results are encouraging so far. At least, it's refreshing to see Kiwi trying to move his pieces and *do something*, I've seen some very interesting games lately!

Sounds interesting. Is there any hope of Linux and Mac OS X versions of 0.6a?

Tord

[Alessandro Scotti]

I've seen the same thing many times. Even very radical changes to my evaluation function often do not change the playing strength measurably. It sometimes feels like I could replace my eval with a random number generator without any noticable effect.

Well it's comforting to know I'm not alone! Unfortunately in my case I've got a random search too, so the end result is... random play, or something similar!

Sounds interesting. Is there any hope of Linux and Mac OS X versions of 0.6a?

Yes of course! I do almost all developing under Linux and recompiling for Mac is a matter of minutes. Right now I'm running some tests to decide if keeping "history" reduction or not, and as expected it will have to go. I suspect to use that stuff successfully you must have a stable and bug-free search as well as a solid null-move.

Anyway I'm now a bit tired of bug hunting in Kiwi. Too little design, some years old code and way too much coding at late nights (*) don't make any good for the engine. So, Kiwi 0.6 will be the last version and I'm now cleaning up the source code to put on the site, which will remove from everyone the suspect that Kiwi is a clone of Chinook.
I will miss this bird but I don't want to leave it all alone, so there could be a new pet coming...

(*) That will only get worse in the future, unfortunately!

[Alessandro Scotti]

Thanks for the new Kiwi Alessandro. It complains already
about the missing bitbases

Regards,
Guenther

I know I know! I will put them on the site eventually, but I mainly wrote the code to understand something more about bitbases. There will be no difference in strength. (Nothing can improve Kiwi's strength, this is know as the Great Kiwi's Invariance Theorem.)

[Tord Romstad]

I've seen the same thing many times. Even very radical changes to my evaluation function often do not change the playing strength measurably. It sometimes feels like I could replace my eval with a random number generator without any noticable effect.

Well it's comforting to know I'm not alone! Unfortunately in my case I've got a random search too, so the end result is... random play, or something similar!

I don't think Glaurung is any better in this respect, and I am sure Gothmog (my old engine) is worse.

Sounds interesting. Is there any hope of Linux and Mac OS X versions of 0.6a?

Yes of course! I do almost all developing under Linux and recompiling for Mac is a matter of minutes.

Great!

Right now I'm running some tests to decide if keeping "history" reduction or not, and as expected it will have to go. I suspect to use that stuff successfully you must have a stable and bug-free search as well as a solid null-move.

I am not so sure about that. Gothmog uses something similar to history reductions, and it appears to work fairly well. The search is about as far from stable and bug-free as you can possibly come.

On the other hand, it is clear that the effectivity of history reductions and similar tricks varies enormously between different engines. For some reason which I don't understand at all, it seems to work far better in Glaurung than in any other engine. I haven't tested it recently, but I guesstimate the Elo difference with and without history reductions to be at least 100.

Anyway I'm now a bit tired of bug hunting in Kiwi. Too little design, some years old code and way too much coding at late nights (*) don't make any good for the engine.

I know how it feels. I've been there, too. The really scary thing is that Glaurung is alarmingly quickly approaching the same state. I think I'll have to do a major rewrite soon, too. The only good thing is that my program is so small that even a complete rewrite isn't such a big task.

So, Kiwi 0.6 will be the last version and I'm now cleaning up the source code to put on the site, which will remove from everyone the suspect that Kiwi is a clone of Chinook.
I will miss this bird but I don't want to leave it all alone, so there could be a new pet coming...

We'll all miss it. I hope your new pet will be ready soon!

Tord