Willow 4.0 (final release for now)

[Whiskers]

Hi all, here is Willow 4.0 after several more months of development, it's 170 elo stronger than 3.1 using balanced books and likely much more on unbalanced books like Pohl. It's currently scoring 5/8 at its debut at TCEC and I feel so proud of it, and I've loved every step of the journey.

I'm moving on to my next project after this; it'll probably draw on a lot of code from Willow, but my goal is to fulfill the dream that I abandoned in the chase for ELO; to make an engine that plays with the most exciting style possible while still remaining clearly superhuman. I'll probably post a progress thread in the vein of Leorik's devblog. See you then.

https://github.com/Adam-Kulju/Willow/releases/tag/4.0

[Graham Banks]

Thanks, Adam.

[pohl4711]

but my goal is to fulfill the dream that I abandoned in the chase for ELO; to make an engine that plays with the most exciting style possible while still remaining clearly superhuman.

I wrote this already to the booot-author, but IMO this is important for anybody, who wants to build an engine with more aggressive playing-style:

Ed Schroeder used my EAS-Tool to make his legendary Rebel play more sacrifices (he refactored the learning-data, considering the EAS-Tool). And it worked! Look at my EAS-Ratinglist:
https://www.sp-cc.de/eas-ratinglist.htm

Rebel EAS played the most high sacs and the most sacs overall (except the 2 Stockfishes) and Rebel EAS gained Elo compared to Rebel 16.2

Ed Schroder wrote this:
"REBEL-EAS is an in between version based on the same, but more, neural net data as REBEL-16.2 but heavily modified during the last 5 months in order to make REBEL to play more aggressive and runs as a module under Chess System Tal 2.0
. REBEL 16.2 gains an EAS score of : 80.258
. REBEL-EAS gets an EAS score of : 160.569
As a cherish on the cake the playing strength of the neural net has increased with 20-25 elo using balanced positions. Using the TCEC positions composed by Jeroen Noomen and GM Matthew Sadler adds another 26 elo points while the draw rate dropped with almost 20%.
The REBEL-EAS approach: King Safety and Mobility were heavily used, the epd scores of King Safety and Mobility even increased with 25% with this function to get the desired effect, to play more aggressive measured with the EAS tool."

Perhaps you should try to go the same way as Ed did?

Here the link to my EAS-Tool:
https://www.sp-cc.de/files/engines_aggr ... cs_tool.7z

And her the link to the Ed Schroeder Tool (link-text is wrong, but leads to the correct site):
https://rebel7775.wixsite.com/rebel/kop ... rl-blitz-1

(Mention, the EAS-Tool needs a huge amount of games for valid EAS-scores. I strongly recommend at least 3000 played games per engine...but the more games, the better)

[pohl4711]

Perhaps a useful "benchmark" for aggressive and strong play is Revenge 1.0.

The most aggressive playing engines (strongest version) are:
Revenge 1.0 (later versions gained Elo, but lost aggressiveness)
Velvet 4.1.0 (later versions gained Elo, but lost aggressiveness)
Komodo 14 (or all later KomodoDragon-releases) with playing-style set to aggressive (mention, this switches off the nnue-net, only HCE is used)

But Revenge 1.0 is clearly the strongest engine of this 3 (around 100 Elo stronger, Velvet 4.1.0 and Komodo 14 aggressive are relativly close in strength).
Revenge 1.0 is on an Elo-level with PowerFritz 18 avx2...

Revenge 1.0 is commercial. But Velvet 4.1.0 and Komodo 14 are free.

You can consider my old SPCC-ratinglist (EAS-list is below the ratinglist in this textfile, just scroll down)
https://www.sp-cc.de/files/spcc_full_list.txt

These 3 engines have a sac-quote of more than 35%: Means more than 35% of their won games were won by playing a sacrifice!!!

[Frank Quisinsky]

It was always interesting to me when programmers tried to give their engines a special touch. The optimum is always exactly what they themselves want.

A good example is Rahman Paidar (chess program Ktulu, IRA):
For Rahman the rook endgames were always the measure of all things, everything else was less important to him, so he always optimised his rook endgames.

We have programs with a human style (Smarthink or Hiarcs are good examples) and some others are strong in attacking the king. Nevertheless, these programmes quickly find their way into the hearts of many fans. Not speaking from the attackers, speaking from engines where we can find out the strengths.

Basically, it is important that strengths become apparent. If a lot of different programs do the same thing, it's not that exciting. Unfortunately, we need more statistic tools. Stefan has made a good point with his EAS tool. In the past it was quite tedious to create statistics by hand. I am very sure that more of these tools will follow and it would be best to have a GUI for them. That is my wish. This was on Martin's and my to-do list before I joined Arena. Unfortunately we never got around to it.

Thanks for your new Willow 4.0 and good luck with your idea!
Quite clear ... I like to read that.

--

In special for endgames / the late midgame / the transposition into the endgame we need more statistic programs. A big problem for me, I can create such stats but that need to many time. If I give hints to fast-wins ... people are thinking only this is interesting. In most of cases engines have problems with fast loses have other strengths, made the points in other gaming phases. With Klaus Wlotzka (Excel expert) I try here some things in the past but all that are not good enough and complicated also.

Elo is boring since programs are stronger as strongest humans.
I am not alone with this opinion.
Since Kasparow lost the important match vs. Deep Thought we lost a lot of people with interest on computer chess. About computer chess very rarely you can find a bit in TV or newspaper. The magic was gone.

If people can see ... aha, Stormphrax must be strong in the late midgame, maybe stronger as many others with around the same Elo ... such things are important. To give people like computer chess much more information as Elo only. The community is strong, sure in the near future we have nice and better statistic programs.

[Frank Quisinsky]

A wish for the programmers too:

To read ...
When a new version is available ...
Elo gain = x
Info is ok, but when I can read ...

I try to make my program stronger in the endgame or this and that about what the programmer likes to create as a final result. In combination with what the programmers say, when I look on github, many things are clear and sometimes I can find the information I am searching at first.

Important for stats are engine databases with games without resign mode.
Or some stats are not possible.

[Dariusz]

Adam, thank you beautifully

I compiled Willow 4.0 for Linux, Mac and Windows - downloadable from my site (Files section).

[JohnW]

Hi all, here is Willow 4.0 after several more months of development, it's 170 elo stronger than 3.1 using balanced books and likely much more on unbalanced books like Pohl. It's currently scoring 5/8 at its debut at TCEC and I feel so proud of it, and I've loved every step of the journey.

I'm moving on to my next project after this; it'll probably draw on a lot of code from Willow, but my goal is to fulfill the dream that I abandoned in the chase for ELO; to make an engine that plays with the most exciting style possible while still remaining clearly superhuman. I'll probably post a progress thread in the vein of Leorik's devblog. See you then.

https://github.com/Adam-Kulju/Willow/releases/tag/4.0

Excuse my ignorance, but why is there a folder for 4.0 containing 3.1 exes instead of 4.0 exes?

[Whiskers]

Hi all, here is Willow 4.0 after several more months of development, it's 170 elo stronger than 3.1 using balanced books and likely much more on unbalanced books like Pohl. It's currently scoring 5/8 at its debut at TCEC and I feel so proud of it, and I've loved every step of the journey.

I'm moving on to my next project after this; it'll probably draw on a lot of code from Willow, but my goal is to fulfill the dream that I abandoned in the chase for ELO; to make an engine that plays with the most exciting style possible while still remaining clearly superhuman. I'll probably post a progress thread in the vein of Leorik's devblog. See you then.

https://github.com/Adam-Kulju/Willow/releases/tag/4.0

Excuse my ignorance, but why is there a folder for 4.0 containing 3.1 exes instead of 4.0 exes?

I forgot to update the version info printed when uci is sent.
Here is a bugfix version that fixes that problem, use it instead of 4.0: https://github.com/Adam-Kulju/Willow/re ... ag/4.0.0.1

[Whiskers]

Adam, thank you beautifully

I compiled Willow 4.0 for Linux, Mac and Windows - downloadable from my site (Files section).

Could you update them with the newest release please? The info string still says "Willow 3.1" because of an oversight that has since been fixed.