New engine releases & news 2021

[voffka]

GreKo PSQ 1.1 - experimental version with piece/square only evaluation:
http://greko.su/greko-psq-1.1.zip

Congrats on the new GreKo release Vladimir!

[mvanthoor]

Megalodon 0.4.3 with hash tables!
https://github.com/megalodon-chess/mega ... tag/v0.4.3

But does it play legal chess already, and not crash anymore? Last time I tested it, it didn't play legally and wasn't stable.

Cosette 5.0
https://github.com/Tearth/Cosette/releases/tag/v5.0

Never heard of this engine (but there are many I never heard of, so that's not special...)

[mvanthoor]

After a lot of work and testing, Loki 3.0 is now released. Thanks to Marcel (who pointed out how underrated Loki was, its feature set taken into account), I have been working tirelessly to rework the search and evaluation function. This has paid off greatly!

I have tested the new version to have a strength of around 2510 (>600 elo points stronger than v1.2.0). This was done in three separate gauntlets against MadChess 2.2 and Raven 1.10.

The release: https://github.com/BimmerBass/Loki/releases/tag/v3.0.0

Congrats to be of help

I haven't had much time for chess programming lately. Next features for Rustic are going to be Killer, History, Aspiration Window and PVS; and a tuned and tapered evaluation after that. I still have to research the tuning. All of that stuff will require a lot of testing. After implementing the above features, I hope Rustic will have exceeded 2150 in my own testing, but it seems it always ends up 30-40 points lower on CCRL. (It could be due to opening book differences, or the fact that I test at 1m + 0.6s; maybe I should make that 2m+1s, like CCRL. But that will make the testing even longer...)

[Gabor Szots]

Megalodon 0.4.3 with hash tables!
https://github.com/megalodon-chess/mega ... tag/v0.4.3

But does it play legal chess already, and not crash anymore? Last time I tested it, it didn't play legally and wasn't stable.

I don't know but I am testing version 1.0.0 right now and it did not make an illegal move in the first 53 games.

[niel5946]

After a lot of work and testing, Loki 3.0 is now released. Thanks to Marcel (who pointed out how underrated Loki was, its feature set taken into account), I have been working tirelessly to rework the search and evaluation function. This has paid off greatly!

I have tested the new version to have a strength of around 2510 (>600 elo points stronger than v1.2.0). This was done in three separate gauntlets against MadChess 2.2 and Raven 1.10.

The release: https://github.com/BimmerBass/Loki/releases/tag/v3.0.0

Congrats to be of help

I haven't had much time for chess programming lately. Next features for Rustic are going to be Killer, History, Aspiration Window and PVS; and a tuned and tapered evaluation after that. I still have to research the tuning. All of that stuff will require a lot of testing. After implementing the above features, I hope Rustic will have exceeded 2150 in my own testing, but it seems it always ends up 30-40 points lower on CCRL. (It could be due to opening book differences, or the fact that I test at 1m + 0.6s; maybe I should make that 2m+1s, like CCRL. But that will make the testing even longer...)

Sounds like you have a lot of work ahead then! I hope it will pay off

Aspiration windows didn't work for me, even after trying numerous different window-widening strategies, but I guess it's lazy/hasty implementation is also partly responsible for the missing elo gain... If you need any inspiration for evaluation tuning, I have implemented a rather simple SPSA + Texel tuner for Loki, that you can check out

[Gabor Szots]

Sounds like you have a lot of work ahead then! I hope it will pay off

Aspiration windows didn't work for me, even after trying numerous different window-widening strategies, but I guess it's lazy/hasty implementation is also partly responsible for the missing elo gain... If you need any inspiration for evaluation tuning, I have implemented a rather simple SPSA + Texel tuner for Loki, that you can check out

IMO, Loki development news might be worth an independent, separate thread.

[mvanthoor]

Aspiration windows didn't work for me, even after trying numerous different window-widening strategies, but I guess it's lazy/hasty implementation is also partly responsible for the missing elo gain... If you need any inspiration for evaluation tuning, I have implemented a rather simple SPSA + Texel tuner for Loki, that you can check out

I'll take a look at the tuner. With regard to chess programming, and with most algorithms, I'm confident that I can either figure it out on my own, or learn how to do it. Something like the tuner though, I'm out of the water. Tuners like that are often based on math or stats, and I don't have a strong background there. I know high-school math and what you need to know to do a university level computer science course, but that's where it ends; I'm certainly not a mathematician or statistician.

Maybe I also have the concept of the tuner completely wrong, and I'd go like: "Oh, is THAT all? Easier than magic bitboards!" when I research it. I'll see. I have a plan to follow. It may take some time, but the engine will get there.

Sounds like you have a lot of work ahead then! I hope it will pay off

Aspiration windows didn't work for me, even after trying numerous different window-widening strategies, but I guess it's lazy/hasty implementation is also partly responsible for the missing elo gain... If you need any inspiration for evaluation tuning, I have implemented a rather simple SPSA + Texel tuner for Loki, that you can check out

IMO, Loki development news might be worth an independent, separate thread.

Create a "Progress on Loki" thread in the programming subforum There are more similar threads in there, including the one on Rustic.

[No4b]

Drofa 3.0.0 released.
https://github.com/justNo4b/Drofa/releases/tag/3.0.0

Changes include, but not limited to:

• Tuner implemented (AdaGrad, implementation closely follows Weiss logic and can be considered a c++ rewrite of it )

• Tuning of the evaluation features (using Andrew Grant positions dataset)

• Several new evaluation features added

• SEE implemented

• Aspiration window implemented

• Lazy SMP implemented

Drofa 3.0.0 is ~250 - 300 elo stronger than Drofa 2.2.0:

Code: Select all

STC	(10``+ 0.1``):
Score of Drofa_3.0.0 vs Drofa_2.2.0: 386 - 35 - 79  [0.851] 500
Elo difference: 302.70 +/- 35.74

LTC	(60`+ 0.6`):
Score of Drofa_3.0.0 vs Drofa_2.2.0: 205 - 19 - 76  [0.810] 300
Elo difference: 251.89 +/- 38.73

Special thanks to:
- Terje Kirstihagen (Weiss author)
- Andrew Grant (AdaGrad paper and Ethereal chess engine are great sources of knowledge; Ethereal tuning dataset published on talkchess)
- Kim Kahre, Finn Eggers and Eugenio Bruno (Koivisto team) for allowing Drofa on Koivisto OpenBench instance and motivating me to work on the engine
- OpenBench community for helping me in finding bugs, teaching me (even if unknowingly) good programming practices and interesting discussions

[Gabor Szots]

Thank you, Alexander. What is lunix?

[No4b]

Thank you, Alexander. What is lunix?

Misspelled `Linux`