@lazy 38.k4 was 70% on policy from neural network - so it latched on to it because it gets more exploration due to policy percent and the winning rate was much higher than other candidates. The policy for L3 was 0.02%. The node is in optimistic mode due to latching on to k4, and with that and the low policy - l3 never gets much exploration.
I am running gzero now with more exploration at root (including some code changes), and a little longer search to compensate for search width. Replaying k12 opening game, which gzero lost to leela in first round of games.
gzero, actually I've just checked, maybe in 398 it's plainly winning. Leela plays a beautiful style and I like its move visually, but the evil is in the details.
Here is leela's current "swap file". The percentages are the winning percentage leela thinks white has if black plays it's first stone in the indicated hex. How does this compare with gzero's thoughts, gzero?
I think if you analyze a8 and a10 deeper, maybe they'll be closer to 50 than they're there.
Anyway, the really surprising result is H3 stronger than G3 and then even more surprising we see that J3 is almost as bad as G3. This latter fact (J3 almost as bad as G3) never crossed anyone mind as far as I know except maybe for Arek: https://littlegolem.net/jsp/game/game.jsp?gid=1597836&nmove=2. Probably it's because it allows white to play L2 at some point.
I don't have any swap file, as swap is baked into the rules/neural network. To get something similar I could generate, from blacks point of view, the win probability after N simulations for each board position with a single white stone from your swap file/heatmap. I'll try that at the weekend.
What I have done is has ran MCTS at depth=0 for a long time with high exploration settings. I'll paste a most recent result below. Also because it is MCTS, it is not very consistent results (the previous time I ran it, a13 was lower than f11). Nor is there any symmetry (so i took the top result for each symmetry).
(#simulations, policy probability, win probability for black after #simulations)
This is very useful, thank you Leela bot. Of course it will take a lot more time, and perhaps a different sort of algorithm, to nail down every cell as either 100% or 0. During that process, some of the relative values between specific cells might flip around.
Did you agree on a thinking time or better a number of playouts for this bot fight?
Could you explain the numbers for the openings? If swap is allowed, every number should be greater than 50% and else I think there should be lot more squares for first player that should be?
That is right. Put another way: if black plays it's first move on an indicated hex and that hex has a number smaller than 50% leela will swap. If bigger than 50% leela will play on with white. leela will also swap any empty hex (if black plays it's first move there).
Also, if leela plays first it would prefer to play k12 (or c2) because that is the cell with the number closest to 50%. In practice, for the sake of variety, it currently randomly plays anything within 6% of 50%.
Interesting results with the swap map. I might try more F3/G3 openings
For the first move selection, have you considered a weighted probability of all moves? Apparently this is what the TwixtBot does:
"For the first move, rather than pick the one point closest to 50%, I decided to make a weighted random selection: the closer the win rate is to 50%, the higher the chance I pick that move."
Ultimately that might reduce the bot's winning percentage, but it could potentially make for some more interesting games. And maybe give the lowly humans a chance ;)
Happy, if it plays less variation in first move, we can study its game and learn something. So actually, lowly humans have more to gain if it plays objectively better... ;)
"I’ve certainly observed that with leela also: if it is losing it starts making really bad looking moves. They are objectively probably still the best moves, but I think human players seem to know how to complicate the situation by playing tricky stuff when they are losing, but the bots are too “honest” for that :)"
Poor bot, these evil humans are tricking him, haha :D
Interesting game 2090408, it seemed leela had won but then played a couple of moves b10 and a12 - that gave gzero a chance to claw back into the game.
Score so far: leela 5, gzero 3. Think gzero is quite fortunate with that score line. :) Congrats leela!
I agree with lazy and arek, it always seems as if games with leela it is fighting a losing battle from move 2 onwards - with the win probability being slowly sapped away. If lucky, gzero seems to have jumps in win probability in the late game, such as the b10/a12 above, which lets it back in the game. Maybe leela just feeling sorry for her wee brother and throwing the game! ;)
One last game being played now, for reference: 2090522, with h11 opening, since g11 was interesting.
It would take longer analysis to prove but to me it seems in 2090408 at move 33 leela could even win in multiple ways... At least the move I would like is 33.h9 and I dont see many opportunities for white to turn the game...
Gzero, don't worry, leela is not making late game blunders. In the game 2090408 the "error" is probably 13 g8 or 15 i8. I say "error" in quotes because it's not obvious at all that there was a winning choice there for Leela at all. It's possible that all choices there were actually losing and Leela at best could have picked another that was less obvious.
Gzero, you can try a match at much higher time controls, like 6 hours per move instead of 10 mins. This should allow you to use your endgame skill to compute the right initial stones.
By the way, I think something similar is true between me and Arek. I'm slighly stronger later moves and/or much longer time controls and/or smaller boards.
Gzero, I also propose a funny experiment for you to do, if you've time. I see that you can easily change board size (your code is game-agnostic). Why don't you try to set board size = 100, or board size = 500, and then do some training on that, and then you show us the "heatmap" of openings? I'm wondering if center is a winning opening on arbitrarily large boards like 100x100.
@lazy on your game 2090087. Yeah, 18.k7 was wrong (leela jumped to above 70%). I wouldn't call 18.l2 winning: leela says 51%. It was actually expecting 18.l3 (44%). 14.l2 is also 51% (I didn't run leela super long on these so take these numbers with a grain of salt)
Well, no, 18 H7 indeed worse than 18 L2, maybe I've caused a mess at 16, and 16 D9 would be a clean white win? Or maybe there are no clean wins at all and it's genuinely complicated.