New release of Allie v0.4

[gonzochess75]

This is version 0.4 of Allie.

Some of the new things in this release:

* Introducing Minimax Monte Carlo hybrid search
* Extensive changes to time management
* Many speed improvements
* Changes to the way nodes, nps, tbhits, and other info is reported
* Support for Chess960
* This release is substantially stronger than v0.3

https://github.com/manyoso/allie/releases/tag/v0.4

[Eduard]

Thanks, interesting engine! I think this version only works with Cuda? Is a CPU version possible?

[gonzochess75]

I am really hoping to have a CPU only option available in the next release... sorry.

[CositasBuenas]

Hello, Allie is doing really well against Lc0. What are the best settings though?

[pohl4711]

This is version 0.4 of Allie.

Some of the new things in this release:

* Introducing Minimax Monte Carlo hybrid search
* Extensive changes to time management
* Many speed improvements
* Changes to the way nodes, nps, tbhits, and other info is reported
* Support for Chess960
* This release is substantially stronger than v0.3

https://github.com/manyoso/allie/releases/tag/v0.4

I started a comparative testrun for my NN-engine testings:

https://www.sp-cc.de/lc0-testing.htm

Testrun of at least 700 games with Allie 0.4 Net 42392 (exactly same conditions and same Net, than the testrun before, except Allie 0.4.exe instead of Lc0 v0.21.1.exe)- result not before Friday... stay tuned!

[jp]

Apart from how strong it is, does Allie play much differently from Leela? (e.g. can you notice the difference watching it play with the same NN?)

[CositasBuenas]

I used two different networks with similar strength (lco: 256x20.T8-swa-610000; allie: 42392) and got about 80 % congruence.

[pohl4711]

I started a comparative testrun for my NN-engine testings:

https://www.sp-cc.de/lc0-testing.htm

Testrun of at least 700 games with Allie 0.4 Net 42392 (exactly same conditions and same Net, than the testrun before, except Allie 0.4.exe instead of Lc0 v0.21.1.exe)- result not before Friday... stay tuned!

There is a little problem, using Allie 0.4 in LittleBlitzerGUI, because of the 50move-draw bug in the LBG. The LBG gives 50move-draw after 101 plies without pawn- or capture-move, instead of 100 plies. And Allie makes an empty or illegal move, when 100 plies are played. So, the LBG shows a "illegal-move"-result. With Thomas Zipproth's new LBG-tools, it is no problem, to correct such games to a draw. But it would be better, if Allie would continue playing, even if a 50move-draw is reached and let the GUI decide, if it is a draw or not... Lc0 does so.

Stefan (SPCC)

[Dokterchen]

Hi Adam,

thanks for the release. In opposition to v0.3 I have no lossen on time with v0.4.

Torsten

[Dokterchen]

Hi Adam,

I have tested AllieStein 0.4-n7. Conditions: 4CPU and GTX2070, 300+2 seconds, round robin with 30 opening positons and reversed colors:

Code: Select all

Place	Engine			Points		%
1	Lc0 0.21.1-nT40.T8.610	663,5/840	78,9
2	Stockfish 19050918	655,5/840	78
3	AllieStein 0.4-n7	601,5/840	71,6
4	Houdini 6.03		554,0/840	65,9
5	Komodo 12.3		554,0/840	65,9
6	Fire 7			440,5/840	52,4
7	Ethereal 11.25		398,5/840	47,4
8	Xiphos 0.5		393,5/840	46,8
9	Laser 1.7		354,5/840	42,2
10	Deep Shredder 13	328,5/840	39,1
11	Booot 6.3.1		317,0/840	37,7
12	Andscacs 0.95		302,5/840	36
13	Fizbo 2.0		286,0/840	34
14	Schooner 2.0.34		253,0/840	30,1
15	Fritz 16		197,5/840	23,5

For such a young projects that is an incredible results. Congrats!

KR
Torsten