How are draw's to be treated in Texel's tuning method? Consider the theoretical situation where all of the 64000 generated games ends in draws. The tuner would tune every evaluation parameter towards 0. This is evidently not what we want.
I assume the same holds for large number of draws for example 95% (at least in practice). But how many draws are too many draws? I assume also that simply removing all draws is a bad idea. I'm trying this right now to see the results. But should some draws be removed if my draw count is too high? Which draws should then be removed? What is the optimal percentage of draws in the training set?
Treating Draws in Texel Tuning
Moderator: Ras
-
spinosarus123
- Posts: 4
- Joined: Tue Jun 20, 2023 12:46 pm
- Full name: Isak Ellmer
-
spinosarus123
- Posts: 4
- Joined: Tue Jun 20, 2023 12:46 pm
- Full name: Isak Ellmer
Re: Treating Draws in Texel Tuning
Quick update:
When training regularly on my new training set around 17 million positions from 140000 games played at 1+0.04, I was consistently losing elo. Usually around -100 elo. Out of the 17 million positions around 45% were drawn games.
I ignored _all_ draws in the training set which gave me around 8 million positions. I have only trained my engine on this smaller set once but it was successful. It gained around 50 elo. As expected all the values got larger. Queen went from around 1000 to 2500. I use K=1 so sigmoid=1/(1+10^(-s/400)).
The question still stands, maybe I should still include some of the draws. The "useful" draws. I can see why I should definitely not include some draws (for example KNvK and all the material draws which the evaluation function does not know about).
When training regularly on my new training set around 17 million positions from 140000 games played at 1+0.04, I was consistently losing elo. Usually around -100 elo. Out of the 17 million positions around 45% were drawn games.
I ignored _all_ draws in the training set which gave me around 8 million positions. I have only trained my engine on this smaller set once but it was successful. It gained around 50 elo. As expected all the values got larger. Queen went from around 1000 to 2500. I use K=1 so sigmoid=1/(1+10^(-s/400)).
The question still stands, maybe I should still include some of the draws. The "useful" draws. I can see why I should definitely not include some draws (for example KNvK and all the material draws which the evaluation function does not know about).