SF 11 vs T60 Lc0

[Laskos]

Question for Laskos: is 30m+30s considered fast slow or slow rapid?

I've seen few say that 30+30 is needed to be considered valid.

Probably a "rapid" in human terminology, but to me it's more of a slow, most of the properties of engines are there, including most of the scaling. Increment of 30s is too pretty large.

[Ozymandias]

Increment of 30s is too pretty large.

Absolutely, a lot of time would be wasted that way in easy endings.

[Ovyron]

Ooh, I wonder how would a "decreasing time control" do (it starts at 30 per second but decreases and becomes a 1 second increment by move 40 or so, so the engine is given the most time where it matters.)

[corres]

...
Interesting fact.-- Alpha Zero beat Stockfish 8 after 44 million games training( in a controversal setting).
T60 beat SF after 46 million games training in a seemingly fair setting
...

Especially if if you mean under "fair setting" no using opening book as it happened during A0 - Stockfish 8 and during the last match of CCC.

[Nay Lin Tun]

Stockfish vs Leelenstein (as bonus)
SF won by 103 vs 97

https://ibb.co/FqrtZ1y

Lco vs Leelenstein (final match )

Lc0 won by 108 vs 92

https://ibb.co/1RyDrtM

[jp]

Of course the people that “know”, know that a big grid search for parameters at different time controls is expensive, long winded and not fun. That’s why these “special settings” only come along every once in a while, probably aren’t best or even good at most time controls, and are passed around like an illicit drug or secret handshake.

I do not agree that tuning Lc0 engine parameters is not important. Without fine-tuning, Stockfish would be at least 200 Elo points weaker. The devs should be doing it more intensely and carefully.

The issue is that SF tunings are robust and will be good in many settings.
If Lc0 tunings depend on the exact time control and the exact network, etc., then there's an obvious problem.

[Laskos]

Of course the people that “know”, know that a big grid search for parameters at different time controls is expensive, long winded and not fun. That’s why these “special settings” only come along every once in a while, probably aren’t best or even good at most time controls, and are passed around like an illicit drug or secret handshake.

I do not agree that tuning Lc0 engine parameters is not important. Without fine-tuning, Stockfish would be at least 200 Elo points weaker. The devs should be doing it more intensely and carefully.

The issue is that SF tunings are robust and will be good in many settings.
If Lc0 tunings depend on the exact time control and the exact network, etc., then there's an obvious problem.

No, I don't agree. The three main runs T40, T59, and T60, and only 2 of them used on strong hardware, T40 and T60 as the strongest nets, all improve at a range from ultra-bullet to rapid TC significantly with Kiudee settings. I checked with T59 at 20k, 100k, 500k nodes per move, the improvement is very stable, only at 500k nodes per move seems slightly decreasing (buy still within error margins in matches of 500 games). When the longest, 500k nodes per move test finishes, I will write to Kiudee about the results (he was not sure what happens with more than 100k nodes per move). The settings depend slightly on time control and probably of training noise value.
The devs do have tot tune the parameters, and if the parameters are unstable under different conditions, they should change them to something more adequate to be tuned independently of conditions. Lc0 is a loosely tuned engine, and I guess additional improvement compared to Kiudee settings can be obtained. Also, compared to the sophisticated search of current top AB engines, Lc0 search is basic and probably can be drastically improved. Most of the computing effort seems to go to the training, overwhelmingly so, not on improving the engine, and that is a mistake. Yes, the devs can rely on the fact that GPU will probably improve faster than CPU, and thus still use rudimentary and untuned Lc0 engine with larger and larger nets on stronger and stronger GPUs to beat SF on big CPUs in competitions for years now, but really, is it improving the search of Lc0 a boring stuff for programmers? Or tuning Lc0 is so much worse than training nets?

[corres]

Of course the people that “know”, know that a big grid search for parameters at different time controls is expensive, long winded and not fun. That’s why these “special settings” only come along every once in a while, probably aren’t best or even good at most time controls, and are passed around like an illicit drug or secret handshake.

I do not agree that tuning Lc0 engine parameters is not important. Without fine-tuning, Stockfish would be at least 200 Elo points weaker. The devs should be doing it more intensely and carefully.

The issue is that SF tunings are robust and will be good in many settings.
If Lc0 tunings depend on the exact time control and the exact network, etc., then there's an obvious problem.

Really. For engine users it is more comfortable a Leela with fixed parameters than a Leela with such parameters what depend on TC and Net - even if the Default parameters would give weaker chess power with some Elo comparing to the optimal parameters. If the testers should use different Leela parameters for different TC and different Net it would be heavier the valuing of the test results also.