Playing the endgame like a boss !!

[Ovyron]

whether it plays moves that are shockingly out of this world.

The moves remain "shockingly out of this world" for the time control they're played in. You can't just download the PGN, analyze the positions for a few days, and conclude the moves weren't special. Because Leela found and played them insanely quick.

Coming from the correspondence chess world, there have been moves that have taken weeks to find. If anything found and managed to play them in a bullet game, that'd be right out of science fiction!

[jp]

Lc0 v0.21.2 was released, including

"Sticky endgames" feature (default on). This makes Lc0 play better when it sees a checkmate somewhere during the search.

This is sounding very non-zero, though we'd need to know the details to know.

[Peter Berger]

Lc0 v0.21.2 was released, including

"Sticky endgames" feature (default on). This makes Lc0 play better when it sees a checkmate somewhere during the search.

This is sounding very non-zero, though we'd need to know the details to know.

Why would this be non-zero? Similar things have been discussed about AlphaGo. Humans prefer a short and clean finish to a longer one in general, and it makes sense in general, too. If you add some random 50 moves to a game, it clearly adds to randomness of the result, simple logic.

When it is about "sticky endgames": no idea what it does, but certainly not what you are looking for. I just watched LC0 trying to mate with KRRKR .. - a painful thing to watch

[jp]

Why would this be non-zero? Similar things have been discussed about AlphaGo.

We won't know either way until we know exactly what it is doing. But if a non-zero feature was "discussed about AlphaGo", that wouldn't make that feature "zero"; it would make AlphaGo non-zero.

(Lc is already non-zero because of TB rescoring, but there are different degrees of non-zeroness.)

I agree it'd be better if it did not play endgames like that.

[hgm]

Where is this misconception coming from that EGT would be a non-zero technique? It uses nothing but the rules and the winning condition. If retrograde search is non-zero, then MCTS / PUCT should also be considered non-zero. Leela / Alpha did not design that algorithm by themselves.

[jp]

Where is this misconception coming from that EGT would be a non-zero technique? It uses nothing but the rules and the winning condition.

If it worked out itself that it would be a good idea to do retrograde search, then worked out itself how to do it, and then went ahead and did the retrograde search for itself, it would be "zero". "It uses nothing but the rules" is extremely misleading. It did not work out the consequences of those rules itself. Tablebase results are not the rules or even obvious consequences of the rules.

There's room for argument, just because "zero" is nothing more than a marketing slogan so we can argue about what it means. But by any sensible definition of "zero", using TB information in the training is non-zero.

[hgm]

So Alpha & Leela Zero are NOT zero at all. They would only be if they had worked out themselves that it would be a good idea to do tree search, then worked out themselves how to do it, and then went ahead and did the PUCT search for themselves, rather than just playing the move the NN's policy head says is best for them in the root... But they did nothing of the kind. All that was programmed into them, by programmers that knew it was a good idea to do that.

[carldaman]

Why would this be non-zero? Similar things have been discussed about AlphaGo.

We won't know either way until we know exactly what it is doing. But if a non-zero feature was "discussed about AlphaGo", that wouldn't make that feature "zero"; it would make AlphaGo non-zero.

(Lc is already non-zero because of TB rescoring, but there are different degrees of non-zeroness.)

I agree it'd be better if it did not play endgames like that.

Of course, it would be better, especially with all the hype about (Leela being) AI. The endgame play is anything but intelligent.

[jp]

So Alpha & Leela Zero are NOT zero at all. They would only be if they had worked out themselves that it would be a good idea to do tree search, then worked out themselves how to do it, and then went ahead and did the PUCT search for themselves, rather than just playing the move the NN's policy head says is best for them in the root... But they did nothing of the kind. All that was programmed into them, by programmers that knew it was a good idea to do that.

Yes, I'm okay with people saying: "Alpha & Leela Zero are strictly speaking NOT zero at all".

It's true Alpha & Leela did not program themselves.

But, as I wrote before, there are different degrees of non-zeroness. Some features are much more non-zero than others. Taking tablebase info is a high degree of non-zeroness. The way tablebases are created has nothing to do with Alpha & Leela. The info they provide is as far away from the rules of the game as you can hope to get.

Of course, at some point some humans had to program something. But e.g. PUCT is a very specialized idea, and furthermore DM & the Leela guys chose special values of parameters in PUCT and everywhere else to make it work. So, yes, they are strictly non-zero. There was already discussion in the technical forum about another technical point that could be non-zero. If the humans programmed up code that was very general, and the program had to work out itself how to specialize to a particular problem (e.g. the game of chess), without human help, then I'd be okay with people saying that program is zero even by high standards. That is not what happened with Alpha & Leela.

[hgm]

But, as I wrote before, there are different degrees of non-zeroness. Some features are much more non-zero than others. Taking tablebase info is a high degree of non-zeroness. The way tablebases are created has nothing to do with Alpha & Leela. The info they provide is as far away from the rules of the game as you can hope to get.

This is just not true. EGT info is a direct result of the rules and nothing but the rules. No strategic guidance is used at all. It is as close to the rules as you can get. Both PUCT and EGT generation are just minimax applied to a sequence of moves, the only difference is whether you build the tree forward, starting from the current game position, or retrogradely, starting from possible terminal positions that the current game state can lead to. And unlike PUCT, EGT generation doesn't require guiding criteria like UCT to decide which branches are most deservant of more search effort.

So it is PUCT has a very high degree on non-zeroness, compared to EGT generation. Apart of application of minimax logic it needs many tricks to make the search work acceptably, and even more tricks during training of the NN.

That it has "Nothing to do with Alpha & Leela" is just a circular non-argument.