1.g4 opening is losing?

[jp]

Realistically white isn't going to win any game, so do we get better %s if we take out possible white wins from the computation?

I'm not convinced taking out possible white wins is a good idea. Why would it be better than, for example, counting those wins as draws? It's the whole process Leela uses to get those stats that's being questioned and that's probably not fixable after the fact.

[Zenmastur]

Yeah, suppose 1.g4 is actually a draw, and you play a self-play match and get that 73% of games are indeed won. What you'll find is that in those games white played losing moves that it didn't play in the other 27%, so you create an opening book and forbid the engine to play those losing moves. What statistics do you get now?

If they're <73% now then it means the initial 73% was wrong and it should have been this new, more accurate value.

If the initial self-play match gives some 95% black wins then the percentage is also wrong.

That's why I'm not a fan of using %s instead of scores, those percentages mean nothing or they're inaccurate after you block the losing side's blunders (suppose there's only 1 line that draws, then it'd show high percentage of losing, except that you can play this drawing move every time and draw all the games, so the % chance of playing a losing move is irrelevant.)

I think it would be more interesting to play a long match with a learning book. Let the book root out the losing lines of play. Then look at what's left.

Regards,

Zenmastur

[jdart]

Permanent hash would have a similar effect.

[Ovyron]

I'm not convinced taking out possible white wins is a good idea. Why would it be better than, for example, counting those wins as draws?

Because if we're looking for the chess truth (whether 1.g4 is a draw or not) then we have to ignore the games where black has blundered. If black has let white win then a blunder happened and it's an irrelevant game that should be ignored.

Ideally, you'd also ignore all the games where black had a mate in X but blundered and drew, then if 1.g4 was lost by force we'd get 0% and know it.

[jp]

I'm not convinced taking out possible white wins is a good idea. Why would it be better than, for example, counting those wins as draws?

Because if we're looking for the chess truth (whether 1.g4 is a draw or not) then we have to ignore the games where black has blundered. If black has let white win then a blunder happened and it's an irrelevant game that should be ignored.

But there will also be (possibly an equal number of) games where White has blundered. That's just part and parcel of Leela.

You are skewing the statistics if you just ignore one.

[Ovyron]

But there will also be (possibly an equal number of) games where White has blundered.

By blunder I mean to say a move that changes the theoretical result of the game, in this case it'd mean white is in a position that can be drawn with perfect play and plays a losing move. If 1.g4 actually loses then all losing moves by white aren't blunders because the theoretical outcome doesn't change. There's no scenario where we're on a drawn position and white plays a losing move because that'd mean black had blundered the win and we're ignoring those already.

If g4 draws and we ignore all white blunders then Leela will show 50% score, and rightly so.

[jp]

If 1.g4 actually loses then all losing moves by white aren't blunders because the theoretical outcome doesn't change.

That's what we do not know with certainty. 1.g4 could be a theoretical draw. Just because you failed to draw does not prove it isn't.

[Ovyron]

But we know with certainty that 1.g4 is a loss or a draw, 1.g4 does not win by force, so we can ignore all the games where black lost, as we know for certain black blundered, and can safely ignore those games to get a more accurate % (where 50% is draw and 0% is lost.)

[jp]

No, even if that is known (and it's not with 100% certainty) that just biases the results.

[Ovyron]

But it biases them towards the correct answer.

Imagine a triangle, there is a dot at the center. At the top it's 1/2-1/2 with a drawn result. At the left it's the 1-0, white wins. At the right 0-1, black wins.

As Leela analyzes nodes, the dot moves. For every time black wins, it moves the dot towards the bottom-right corner of the triangle. For every time there is a draw, Leela moves the dot towards the top corner of the triangle. For every white win, it moves the dot towards the bottom-left corner of the triangle.

If 1.g4 is a win for black, we want the dot at the right corner.
If 1.g4 is a draw, we want the dot at the top corner.

That's it.

So every time the dot moves towards the left corner it's not productive. Ideally we want the dot somewhere in the line that connects the top and right corner, which would be achieved if we ignore the losing moves of black.

Probably someone can put an end to these discussions by producing a specific Leela Network that only trains for 1.g4. At the end it'd be able to always draw as white, or to always win as black.