Creating NNUE Nets from NNUE evaluation?

bastiball · Post by **bastiball** » Sun Jan 03, 2021 11:50 pm

Did anyone try this or is it not possible?

tmokonen · Post by **tmokonen** » Mon Jan 04, 2021 12:25 am

How about a texelish style tuning of hand crafted evaluation parameters using NNUE scores?

AndrewGrant · Post by **AndrewGrant** » Mon Jan 04, 2021 1:14 am

I do believe it is common practice to train the NNUE by using an earlier NNUE.

As for "Texel Tuning", typically in my experience it has been very bad trying to fit your eval to match a particular score, and has been much better to have it try to match a WDL result. Although, you can map CP scores onto [0.0, 1.0] and try that.

tmokonen · Post by **tmokonen** » Mon Jan 04, 2021 2:17 am

AndrewGrant wrote: ↑Mon Jan 04, 2021 1:14 am I do believe it is common practice to train the NNUE by using an earlier NNUE.

As for "Texel Tuning", typically in my experience it has been very bad trying to fit your eval to match a particular score, and has been much better to have it try to match a WDL result. Although, you can map CP scores onto [0.0, 1.0] and try that.

OK, thanks. I was looking for a cheap shortcut, but I guess I will forego that idea.

AndrewGrant · Post by **AndrewGrant** » Mon Jan 04, 2021 2:20 am

tmokonen wrote: ↑Mon Jan 04, 2021 2:17 am
AndrewGrant wrote: ↑Mon Jan 04, 2021 1:14 am I do believe it is common practice to train the NNUE by using an earlier NNUE.

As for "Texel Tuning", typically in my experience it has been very bad trying to fit your eval to match a particular score, and has been much better to have it try to match a WDL result. Although, you can map CP scores onto [0.0, 1.0] and try that.
OK, thanks. I was looking for a cheap shortcut, but I guess I will forego that idea.

The cheap shortcut being to take SF NNUE evals and train your HCE to match it? I'de advise against it.

dkappe · Post by **dkappe** » Mon Jan 04, 2021 3:15 am

bastiball wrote: ↑Sun Jan 03, 2021 11:50 pm Did anyone try this or is it not possible?

That is the RL that was done on the SF nets.

bastiball · Post by **bastiball** » Mon Jan 04, 2021 3:27 am

dkappe wrote: ↑Mon Jan 04, 2021 3:15 am
bastiball wrote: ↑Sun Jan 03, 2021 11:50 pm Did anyone try this or is it not possible?
That is the RL that was done on the SF nets.

Oh, I thought they still use HCE

Thanks for the info !

voffka · Post by **voffka** » Mon Jan 04, 2021 4:01 pm

bastiball wrote: ↑Mon Jan 04, 2021 3:27 am Oh, I thought they still use HCE Thanks for the info !

In Igel I have not yet gained a lot of elo due to RL (perhaps 1-5 elo but I spent big amount of time/money on data generation), so I keep training on HCE until I reach it's limits before switching to NNUE data generation.

bastiball · Post by **bastiball** » Tue Jan 05, 2021 12:28 am

voffka wrote: ↑Mon Jan 04, 2021 4:01 pm
bastiball wrote: ↑Mon Jan 04, 2021 3:27 am Oh, I thought they still use HCE Thanks for the info !
In Igel I have not yet gained a lot of elo due to RL (perhaps 1-5 elo but I spent big amount of time/money on data generation), so I keep training on HCE until I reach it's limits before switching to NNUE data generation.

Training Igel HCE could help for more future NNUE improvement/development and it is impressive with the patience you have

Creating NNUE Nets from NNUE evaluation?

Creating NNUE Nets from NNUE evaluation?

Re: Creating NNUE Nets from NNUE evaluation?

Re: Creating NNUE Nets from NNUE evaluation?

Re: Creating NNUE Nets from NNUE evaluation?

Re: Creating NNUE Nets from NNUE evaluation?

Re: Creating NNUE Nets from NNUE evaluation?

Re: Creating NNUE Nets from NNUE evaluation?

Re: Creating NNUE Nets from NNUE evaluation?

Re: Creating NNUE Nets from NNUE evaluation?