@RwSF75 said in #20:
> - Stockfish has used its own games since the WDL output was introduced in the engine.
Where is it explained how they relate the output to their own games outcomes?
I thought that, in the first place, the WDL output was based on a conversion curve using other population of games, such as lichess top games.
Is there now a parameter search to make their internal scoring** behave as WDL according to their games? something would be floating otherwise, in my blurred mind (blurred or from afar), but sometimes appearance of pulling by own bootstrap is misleading.
I actually think that in all WDL approaches Lichess and previous SF, using populations of games is indeed improving on the centipawn limited point of view, but that some scrunity on the assumptions that allow the conversions or relations should be made explicit and explained.
How can a position score wihtin a game be completely determining the odds from then on (I guess having the constraint on input data set to be human ratings above threshold, but then the question of position space covering might come in, which LC0 would have somehow adressed through at least some wider exploration early in its training self play scheduling, in its compromise between exploratoin and "exploitation".
I also had that question with the Maia paper (Figure 11, then) that showed how lichess data was powerful when taken over all qualities of play. It was lacking as often in chess data display, notions of error bars, or I could not read it around. But I was also not clear on their definition of positoin complexity. I have a tendency to watch of outer information flow in data analysis, and other contrust, what seems floating, and hence likley having hidden or obvious to author assumptions, or even dissonance.
But I can be wrong in that. I welcome alternate points of view (even questions) so I am not myself floating for ever. As above. I may be asking too much and not well written.. But not asking won't give myself a chance.. thanks for above.
** scoring which is not based on game outcomes directly, but some loop of NNue (master?) training with SF itself as moderate depth search oracle --- possible old news as well, all of this post.
> - Stockfish has used its own games since the WDL output was introduced in the engine.
Where is it explained how they relate the output to their own games outcomes?
I thought that, in the first place, the WDL output was based on a conversion curve using other population of games, such as lichess top games.
Is there now a parameter search to make their internal scoring** behave as WDL according to their games? something would be floating otherwise, in my blurred mind (blurred or from afar), but sometimes appearance of pulling by own bootstrap is misleading.
I actually think that in all WDL approaches Lichess and previous SF, using populations of games is indeed improving on the centipawn limited point of view, but that some scrunity on the assumptions that allow the conversions or relations should be made explicit and explained.
How can a position score wihtin a game be completely determining the odds from then on (I guess having the constraint on input data set to be human ratings above threshold, but then the question of position space covering might come in, which LC0 would have somehow adressed through at least some wider exploration early in its training self play scheduling, in its compromise between exploratoin and "exploitation".
I also had that question with the Maia paper (Figure 11, then) that showed how lichess data was powerful when taken over all qualities of play. It was lacking as often in chess data display, notions of error bars, or I could not read it around. But I was also not clear on their definition of positoin complexity. I have a tendency to watch of outer information flow in data analysis, and other contrust, what seems floating, and hence likley having hidden or obvious to author assumptions, or even dissonance.
But I can be wrong in that. I welcome alternate points of view (even questions) so I am not myself floating for ever. As above. I may be asking too much and not well written.. But not asking won't give myself a chance.. thanks for above.
** scoring which is not based on game outcomes directly, but some loop of NNue (master?) training with SF itself as moderate depth search oracle --- possible old news as well, all of this post.
