MikeKyle analyses Hoshi, low approach, low 1 space pincer

[sorin]

One thing that bothers me about the highly unbalanced comparison between plays is that, when the winrate estimates are close between two plays, shouldn't the number of visits for second place be roughly the same as the number of visits for the eventual winner of the comparison? That is, in general the more visits a play gets, the smaller the margin of error of its evaluation should be. So it seems that we have a case where the margins of error of two competing evaluations overlap, and instead of reducing the larger margin of error, much more effort is spent reducing the smaller margin of error. You could get it down to a point and there would still be overlap.

The relative difference in number of visits may be actually "noise" since the total number of visits is still quite small in my analysis earlier in this thread. Maybe someone with a stronger computer or more time can let LeelaZero do a deeper analysis and, with many more total visits, the pattern that you expect may be true, as in relatively close winrate moves would also have relatively close number of visits...

I volunteered to try what I suggested earlier, and I let the best 15-blocks LZ network run until it reached about 1.4M total visits.
I was surprised to see that around the 500K total visits point, the traditional human move moved up from #2 to #1, and it kept the place. Yay humans!!

Also, as the number of visits increases, more and more weird moves are being explored, so basically any possible move has some non-zero probability of eventually being explored by MCTS.

It does seem to follow Bill's intuition, namely moves that are close winrate-wise get more attention and eventually get more visits; on the other hand, since new candidate moves pop up on the radar all the time, there will always be gaps, until a really huge number of visits happened (which I cannot experiment with on my laptop).
For instance, the low approach in the lower right corner is in the same winrate ballparck as top 2 moves on the left; it has way more visits than all other lower winrate candidates, but still way fewer than top two; I expect that with more visits, this 3rd candidate would also get more attention from MCTS.

[Uberdude]

sorin, very interesting about the press overtaking counter-pincer at 500k, I have little experience of such high visits. Makes one wonder how many of these other characteristic bot moves are just the dumb low visit intuition of the neural networks and with more time they would find the more subtle and refined human move .

Bill, about close winrates with low visits, that's why for reviewing I recommend playing that other move and letting it accumulate lots of playouts to get a winrate of similar accuracy. When playing as LeelaZero and manually doing this promising blue circle investigation to accelerate devotion of playouts to possibly better other moves I think I make it stronger, though as dfan says when programmatic changes have been made to switch the balance of exploration they've tended to make it weaker overall. Probably my strategy in when to do this is not as simplistic as fiddling a few MCTS hyper-parameters and might use a little of my Go skill. Also worth pointing out that all the moves in the AlphaGo teaching tool have 10 million playouts.

As for when you've found a good move, look for a better one, here's a very interesting one from the AlphaGo teaching tool Sinan Djepov pointed out. The knight move of 4 is a bit unusual compared to attach but human have played it, and AG and LZ and Elf all quickly prefer it. But the nozomi/probe/!! of 5 is an amazing idea. I didn't wait for a million playouts to see if LZ finds it itself, but it like it when you show it. AG recommends tenuki (tohugh they don't tell us if answer is bad and if so how much) but exploring with LZ shows how it can be a nice reduction kikashi before taking the corner if white makes normal h5 answer, and k4 is another nice followup if g4.

$$Bc
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . , . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . 3 . . . . 5 . . . . . . . . . . . |
$$ | . . . . . 4 . . . . . . . . . . . . . |
$$ | . . . O . . . . 2 , . . . . . , X . . |
$$ | . . . . . 1 . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+

Click Here To Show Diagram Code

[go]$$Bc
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . , . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . 3 . . . . 5 . . . . . . . . . . . |
$$ | . . . . . 4 . . . . . . . . . . . . . |
$$ | . . . O . . . . 2 , . . . . . , X . . |
$$ | . . . . . 1 . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]