5 kyu Wbaduk Fighting Play Questions

[Uberdude]

There are certainly some situations where LZ needs 10s of thousands of playouts to find the better move, and in some of these (e.g. involving ladders) I can do so in a matter of seconds. Of course even with these inferior moves its overall strength is still very high, but on 1k playouts there will be rare times it suggests a move 4d me can quickly and confidently say is wrong. Presumably a stronger player than myself would find more such occurrences, though care is needed your judgement isn't based on wrong human knowledge. In these ladder cases you can escape from the ladder LZ mistakenly thinks is working and see its winrate plunge as it realises its delusion.

[Bill Spight]

Hi Gomoto,

Thanks.
( I assume 'good' means it's either LZ's top choice or one of its top 3 choices for the given board. )

Without many playouts, you can't count on Leela's third choice, or maybe even the second choice.

How many playouts 1k is enough generally? I analyze my games with LZ and after 1k generally moves do not change.. Only win percentages change 0.5%-1%.

If you are talking about Leela, not Leela Zero, with only 1,000 playouts I would have doubts about even the first choice.

Leela Zero is more accurate, but I would still want to run it with at least a 10k setting.

[Bill Spight]

There are certainly some situations where LZ needs 10s of thousands of playouts to find the better move, and in some of these (e.g. involving ladders) I can do so in a matter of seconds. Of course even with these inferior moves its overall strength is still very high, but on 1k playouts there will be rare times it suggests a move 4d me can quickly and confidently say is wrong. Presumably a stronger player than myself would find more such occurrences, though care is needed your judgement isn't based on wrong human knowledge. In these ladder cases you can escape from the ladder LZ mistakenly thinks is working and see its winrate plunge as it realises its delusion.

I only have some preliminary results, but Leela Zero run at a 200k setting can come up with a different first choice than Leela Zero run at a 100k setting more than 15% of the time.

[Uberdude]

I only have some preliminary results, but Leela Zero run at a 200k setting can come up with a different first choice than Leela Zero run at a 100k setting more than 15% of the time.

How often does Leela Zero run at 100k come up with a different 1st choice than Leela Zero run at 100k another time? i.e. how consistent is it under the same settings. Control groups

[Bill Spight]

I only have some preliminary results, but Leela Zero run at a 200k setting can come up with a different first choice than Leela Zero run at a 100k setting more than 15% of the time.

How often does Leela Zero run at 100k come up with a different 1st choice than Leela Zero run at 100k another time? i.e. how consistent is it under the same settings. Control groups

Dunno. I am waiting both for having more time and more powerful hardware to do some real research. {shrug}