Wasp 5.00 released

[jstanback]

Version 5.00 of Wasp has been released and can be downloaded from Frank Quisinsky's web site:

https://www.amateurschach.de/

The primary change since Wasp 4.5 is that a neural network is now used for position evaluation. The NN has 1830 inputs, 192 nodes in the 1st hidden layer, 16 nodes in the 2nd hidden layer, and 2 nodes in the 3rd hidden layer. It was trained using positions from games between Wasp and other engines of roughly similar strength. More details on the NN and the training are included in the release notes. I am still new to understanding NN's and how to train them. It kind of amazes me that they work.

Many thanks to Frank for hosting Wasp and for testing and including it in his tournaments!

John

[BrendanJNorman]

Version 5.00 of Wasp has been released and can be downloaded from Frank Quisinsky's web site:

https://www.amateurschach.de/

The primary change since Wasp 4.5 is that a neural network is now used for position evaluation. The NN has 1830 inputs, 192 nodes in the 1st hidden layer, 16 nodes in the 2nd hidden layer, and 2 nodes in the 3rd hidden layer. It was trained using positions from games between Wasp and other engines of roughly similar strength. More details on the NN and the training are included in the release notes. I am still new to understanding NN's and how to train them. It kind of amazes me that they work.

Many thanks to Frank for hosting Wasp and for testing and including it in his tournaments!

John

Wasp was already a very unique engine in terms of style. Look forward to testing the NN version.

Thanks John for your lifetime of dedicated work in the field.

[Frank Quisinsky]

Hi John,

thank you from myself again!!

You do really a great work and without Wasp never I have so many fun with computer chess.

Sitting with Marina at home and looking what your engine do.
You have a fanboy and a fangirl in Trier.

Fam. Quisinsky is happy!

Best
Frank

Information about your devolpment for TalkChess will be set in Wasp News-33 in the next seconds.

[Wolfgang]

CEGT 40/4 running
https://cegt.forumieren.com/t1614-testi ... -00nn#3109

[Frank Quisinsky]

Hello Wolfgang,

very interesting on the Wasp results vs. the opponents I can see you will used.
John is thinking ... +70 and I am thinking +90.

The different between v4.65 dev and v4.66 dev is in my opinion bigger as John is thinking.
So I am thinking +90.

End of the day it's +75



Best
Frank

Thank you for testing!

[jorose]

Congratulations on the new release! I have always been a bit of a Wasp fan, so this is an interesting development. Its also funny to me, as for me Wasp feels somewhat like the OG NN killer.

[Frank Quisinsky]

Hi Jonathan,

thank you, John will like your message!



I think engines, strong in the earlier mid-games with a good king safty have a lot of room for more interesting improvements.
John is to 100% on the way with Wasp.

But what I never can understand ...
Why such a strong engine like Fizbo (also a great attacker, but holes in king safty) will be not longer developed.

Sources are free on github.
So many programmers with a lot of knowlege we have!

Example:
The Dark Toga programmer can produced the NN.
Other programmers can work on it in teamwork.

The style of Fizbo is great and it must be a reason why.
The sources are open but since years no persons like to work on it???

Maybe all wait of a comeback from the programmer of Fizbo!??

Best
Frank

PS: Speaking from Fizbo because the stats to the playing-phases from ... Spark-Fizbo-Wasp ... are _a bit_ comparable.
All have problems in the earlier endgame. All are very strong with many pieces on board in king attacks and Spark and Wasp have a very strong king safty. In case from Wasp the NN file helps for earlier endgames. But the strength of Wasp ... the aggressiveness in the earlier mid-games is again improved. I like that a lot.

[jstanback]

Congratulations on the new release! I have always been a bit of a Wasp fan, so this is an interesting development. Its also funny to me, as for me Wasp feels somewhat like the OG NN killer.

Thanks Jonathan! I'm a poor chess player so I can't comment on how the style of Wasp NN compares to the hand-crafted eval versions. It does seem to me that it manages to get into fewer situations where it's pieces have really low mobility and I think it somehow understands material imbalances better.

John

[Wolfgang]

Hello Wolfgang,

very interesting on the Wasp results vs. the opponents I can see you will used.
John is thinking ... +70 and I am thinking +90.

The different between v4.65 dev and v4.66 dev is in my opinion bigger as John is thinking.
So I am thinking +90.

End of the day it's +75



Best
Frank

Thank you for testing!

So far it is ~~+120, but 400 games only.
Linux Version is running very stable too
I'll try to update the thread as often as possible, but no guarantee for that...

[Frank Quisinsky]

Hi Wolfgang,

after 500 games now +123.
That is a lot but I very sure not more as +90.

Let us wait of more opponents.

If I remember (John's mails) the performance on AMD is improved too.

Hm ...

A sensation, Wasp will be in the next days on place 1.



No, no ...
Frank is thinking only!

Best and have fun with Wasp and thanks for testing again!
Very strong results!!

Frank