Ok, so you don’t know the mean error. It is relevant to the concept you were stressing, namely accuracy, don’t you think?Robert Pope wrote: ↑Thu Jun 06, 2019 3:55 pmThat would be dependent on the exact net you choose to examine, as well as the specific data you are comparing against. Also, not really relevant to his question or my response.
Is AlphaZero-LC0 (Leela0) A chess Engine?
Moderators: hgm, Rebel, chrisw
-
- Posts: 4315
- Joined: Tue Apr 03, 2012 4:28 pm
Re: Is AlphaZero-LC0 (Leela0) A chess Engine?
-
- Posts: 558
- Joined: Sat Mar 25, 2006 8:27 pm
Re: Is AlphaZero-LC0 (Leela0) A chess Engine?
Not at all, since I wasn't addressing a particular net, or even a particular topology. Clearly, a well-trained 20x256 net is going to have a greater capability to absorb the different nuances of chess than a 5x64 net, so it will have a lower mean error. Equally clearly, even the relatively young network structures that Lc0 has trained on so far do capture subtleties of chess understanding that are lacking in traditional engines. (And on the flip side, I will happily concede that there are specific cases of chess understanding that traditional engines have been able to explicitly code that Lc0 doesn't handle as well.)chrisw wrote: ↑Thu Jun 06, 2019 4:26 pmOk, so you don’t know the mean error. It is relevant to the concept you were stressing, namely accuracy, don’t you think?Robert Pope wrote: ↑Thu Jun 06, 2019 3:55 pmThat would be dependent on the exact net you choose to examine, as well as the specific data you are comparing against. Also, not really relevant to his question or my response.
However, knowing a specific value of mean error is irrelevant when we are discussing the "whys" as opposed to the "how wells". I can explain to someone that a sniper's rifle is more accurate at 100 meters than throwing a rock, without needing to measure the mean error.