Minor AI Bug: Asymmetric win-rate changes for symmetric board setup

Hello, this is minor but i thought i’d point it out to those in charge of tuning the AI network.

RE: game https://online-go.com/game/18367573, move 4 of full AI review (Dan-Supporter Strength).

Game setup is symmetric 4-4/4-4 openings for W+B. For blacks next move; it rates F17 -0.1 and C14 as -0.4. That by itself seems normal; what seems incorrect is it rates F3 as -0.3 and C6 as -0.3 as well.

By symmetry, the change in win rate for playing C6 should be exactly the same numerically as C14, and F17 the same as F3. Perhaps this is a minor, weird numerics issue? Or does it possibly represent a bigger issue, in that the AI networks seem to have a preferential direction, i.e. a bias?

Cheers and thanks for the wonderful website + great features as usual.

The Zero architecture explicity does not deal with symmetry.

1 Like

It doesn’t explicitly deal with it, but in principle it is incorrect. In the situation i’ve shown it’s minor, but if it happens here in a small way perhaps it happens in other positions in a large way?

Perhaps what should be looked into is rotating/reflecting more complicated positions, calculating the win rate changes again, and then seeing how significantly they change. If difference is almost always on the order of sub-1% then it’s mostly a non-issue.

There’s a difference between “does not explicitly” and “explicitly does not”. In this case, it is the latter.

To the algorithm, the input layers ‘look different’. Because they are. {4,3,2,1} is just {1,2,3,4} in reverse, but it’s not the same input.

You could write your own fork of LZ which takes symmetry into consideration (or find one that perhaps already does) though. Either way it’s got nothing to do with the website and everything with the design of LZ.

Try posting at https://github.com/leela-zero/leela-zero/issues. As far as I can tell, OGS is reporting LeelaZero’s output correctly, and it’s the output itself that’s imperfect.

1 Like

Leela Zero purpose is to learn game with zero ideas from humans. Look at symmetry is one of ideas it just didn’t invent. % of winrate in same positions has random nature like 2 coins drop in 2 identical places. Also it don’t care about estimation itself, all it needs is best move. Systematic taste preferences are not impossible.

This topic was automatically closed 91 days after the last reply. New replies are no longer allowed.