Hex bot battle

Pretty remarkable that both gzero and leela, completely independently, decided that k12 (c2) is the most balanced opening move.

First 2 games have started:

Game 2089896, similar to game 2080188 which gzero won, but gzero deviates on move 6

Game 2089655, similar to game 2080079 which gzero won, but leela deviates on move 6

My guesses for second and third respectively are a13/m1 and a3/m11 ;)

In both games above I think leela is ahead, but as lazyplayer has mentioned, I should not dare suggest solutions in bot games anymore ;)

GG(s) to leela!  Congrats!   :)

I think Arek you were right... early on gzero was a little pessimistic on the outlook of both games, and slowly became more and more pessimistic.  I've reduced the resign percentage, so it prolongs the game a bit - so gzero might play really desperate moves at the end (that is MCTS for you).

I am running a long job to find openings other than k12.  It seems that gzero is currently a fan of an unusual(?) trio f11/g11/h11 (and rotations).

Amazing! leela also likes g11 as an opening move!

Probably there are losing openings on first column too, but maybe MCTS is biased and doesn't perceive that? That is,  a7,a8,a9,a10...

Regarding f11,g11,h11, we know they're more or less balanced. :)

Regarding your last 2 games, it seems leela knows what it's doing and gzero does not, but somehow, gzero also wins some games :D

This position is probably worth analyzing in depth: https://www.littlegolem.net/jsp/game/game.jsp?gid=2089655&nmove=8

lazyplayer, I get the same impression! Leela seems to play correctly while gzero seems to play random "reasonable" moves and sometimes it's just enough to reach an endgame where gzero is monstrously brilliant like leela ;)

How about some 19x19 bot games? It might popularize this variant. Maybe humans could maintain a stronghold there for a while.

David, on 19x19 we stand no chance at all. We humans are mostly clueless on the big picture. Some more (me) some less (maybe Daniel and Arek).

But anyway, it would be nice to observe some 19x19 games played by bots according to "objective" standards.

I think probably playing mostly correctly at the global scale requires calculation, you really have to sample a lot of possible continuations. And this is exactly what the bots are doing. So basically on that front we can't outplay them at all. The only chance of outplaying them is in local tactics that are large enough to have escaped their patter recognition algorithms.

Well, technically, we could outplay them by intuitively making a correct global plan... but good luck doing that on 19x19! :D

What do you thinks guys about my games against leela and gzero? Especialky against leela on white and against gzero on black i feel disappointed

Daniel, in game 9895, after 7 k6 it's over, right? so maybe 6 i6 as alternative?

Hm no i6 loses easily, I'm trying i7 now...

Daniel, I think you blundered, at 6 you seem to have two choices that don't lose immediately, i7 and j6. I doubt they're enough to win but I've no idea really.

This game here is the same questionable idea: https://littlegolem.net/jsp/game/game.jsp?gid=2089660&nmove=4

And even here: https://littlegolem.net/jsp/game/game.jsp?gid=2089656&nmove=4

This idea of being very strong on one side isn't supported by bots. They don't play it and we shouldn't play it either. :)

Daniel, I think you've been losing pretty early in the first two...

here and here

the third I like your position here, later I will analyse this game deeper

In the fourth I don't like both #6 and #8 and it collapses more and more with every next move...

leela thought 7.d7 was a perfectly OK move. In fact things weren't quite over yet until 15.i3. It thought 15.j7 would have been much better.

Leela, you've not commented on this game: https://littlegolem.net/jsp/game/game.jsp?gid=2033216&nmove=19

I think that Daniel has blundered near the end, but maybe I'm missing some tactic?

Maybe 9.F6 instead E7?

@lazy & @arek - gzero would appreciate some game time if you are up for it.  you can tell me what the random(ish) moves are so I can try to debug it.  Thanks.

Gzero, I've no timefor another game now. In hex practice there are many things that are considered wrong but that are FAR from disproved. For example:

https://littlegolem.net/jsp/game/game.jsp?gid=2033245&nmove=4

https://littlegolem.net/jsp/game/game.jsp?gid=2089896&nmove=6

And basically, we don't know if they're REALLY wrong or not. Maybe in a FEW rare cases they're indeed optimal.

@ lazy - ok thanks.  The game 2033245 is a really old model, I threw away all the training data and started over.

https://littlegolem.net/jsp/game/game.jsp?gid=2089896&nmove=6

I will look that into that one, it was divergent from the games I had earlier with leela.  And I am not sure why.

And posted from wrong account. :)

Im up for it but do you mean through the "Rated games" mechanism?

gzero, it's hard for me to describe the randomishness clearly, because it's a long term observation. Leela seems to follow up on the initial moves. Gzero seems to play all the good stuff but without order.

An example is this https://littlegolem.net/jsp/game/game.jsp?gid=2080616&nmove=16

All white stones support each other. Black d5, g4 and j3 make no sense in terms of putting them on board together,

"gzero, it’s hard for me to describe the randomishness clearly, because it’s a long term observation. Leela seems to follow up on the initial moves. Gzero seems to play all the good stuff but without order."

That's interesting, thanks.  I'll try rerunning that example.  One thing is that if gzero gets pessimistic, the top moves end up shuffling down in priority and all the top candidate moves are bunched up with the same probability, and usually a randomish move remains at the top at the end of the search (so maybe what it thinks will prolong the loss).  It is generally whats happens with long searches and MCTS if it goes into pessimistic mode.  I am not a strong enough player to see such bad moves, so I have no idea when such bad moves are made, so it hard to even correlate this theory.  Maybe one reason we see less of this with leela is that leela doesn't lose very often!  I might be clutching at straws here, it is very hard to figure why it does something.

Sure rated or unrated if you want.

gzero, so basically it's more random when the game is less balanced? actually this is understandable, because it's hard to evaluate unbalanced positions...

it is more random for the player that is losing when the game is less balanced.

another way to view it - from a single board state and the current player to play:

the less chance of winning, the more it will explore - and the more random the top move will be.

[the famous lee sedol versus alphago is a prime example of this]

the greater the chance of winning, the more it will latch onto the top move and barely explore other moves.

but again, it is just one example of something that can go wrong.  If the network evaluation is in lala land, it will behave very different.

gzero, maybe I've understood, unbalanced position create some kind of instability, but this instability is a problem only at the root node, because at the root you've to eventually make a final decision.

Ok, I take back what I've just said. I guess it's due to the MCTS algorithm. It's simply more random when you're losing.

lazy - yes, what you said was right.  non-root nodes that are pessimistic don't matter, they just back-propagate a truer score.  In fact, pessimistic opponent non-root nodes are great since it will be exploring more, and less chance of a surprise move.

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 :)

Next pair of games are underway 2089922 and 2090091

I think k7/c7 is too strong and both game will end as white win. I will be watching carefully :)

It seems Leela is going to win both. It plays so confidently, as if it's knowing what it's doing.

Regarding root nodes, I can understand that, but there also has to be some asymmetry between losing and winning. Maybe there is a logarithm somewhere in the formulas of MCTS?

Well, maybe gzero has decent chances in game #...91. Both are after the 16th move right now.

In my experiences, for stones like c7 to lose, black has often to play b6-b9 threat, and it's really fun when it happens. :)

Well, but I was playing that on 11x11. On 13x13 it's really hard to imagine a way to fit b6 and b9 into play without losing on the spot. :D

Leela even found a way to play B6 in its game. Both games look complex and beyond human understanding. Maybe it would take hours to understand what's going on there.

fwiw, @ move 32 - gzero seems rather confident with game 22 (90% rising with search) - but then stranger things have happened!  91 is still in the balance, with (60% decreasing with search).

gzero, you may end up winning both against all previsions :D heheheh

i think you still play very dubious initial stones but not even leela seems able to refute them.

leela agrees 22 is over, also pretty balanced still on 91. I don't think I have ever seen it this close this deep into a game.

Yeah 22 is over, but I think gzero has just blundered at 91 with 38 k4...

Well, it's over anyway, congrats to gzero. Who wrote the code? I see it's in GitHub.

leela expected 38.k4.

38 L3 is of course better, as you can see locally. But it loses anyway.

Probably Gzero isn't even bothering analyzing, it picks first random winning move and stays with it.

Congrats to gzero on winning 22! Leela is now under 20% on 91 also, so it looks like the score is going to be leela 2, gzero 2 so far.