#1 · Bill LeBoeuf
24 Sep 2019
Wow, Leela got beat by Moo in a game where Leela had the swap advantage (game 2122088). Very amazing.
Public forum discussion.
24 Sep 2019
Wow, Leela got beat by Moo in a game where Leela had the swap advantage (game 2122088). Very amazing.
25 Sep 2019
Will be nice for mooone/two owner to introduce himself and to know smth. about these probably AI's.
25 Sep 2019
Hooray, mootwo confirms that I'm correct claiming that this is black win ;)
25 Sep 2019
Arek, hahah :D
25 Sep 2019
Arek, but indeed the game seems very principled and it could even be considered a refutation of the double 4-4 opening!
If you want to try 5-5 opening, then maybe try again this: https://littlegolem.net/jsp/game/game.jsp?gid=2090087&nmove=7
25 Sep 2019
Excuse me for skipping introduction, the bot will be introduced later at some point.
In 2122088, the bot thought it was behind until move 22.g10, after that it considered itself significantly ahead. As white, it would play g7. Seems like a blunder by Leela?
25 Sep 2019
22.g7 seems a natural human choice, but leela might have been right to look for complications.
26 Sep 2019
mootwo, is the bot MCTS-based or is it based on some minimax variant?
26 Sep 2019
I let it play against itself for a few moves link , and it does change its mind after 27.e8 is played to black being ahead, and in the end black would win this game.
26 Sep 2019
Sorry, fixed link
26 Sep 2019
mootwo, take a look at this, maybe is it an improvement on white play?
26 Sep 2019
The bot seems to like that slightly less. Black still won when I let it play against itself from that spot: link
26 Sep 2019
Well, 5 L6 is crazy and 6 I3 can't be the right reply... somehow bots win vs humans without understanding anything for real... :D
28 Sep 2019
Oh, so indeed leela was correct. But it only found another way to lose! :)
@lazy, in fact both L6 and i3 are reaaonable, they occur in most games of this kind but later :) From experience we know that in games of similiar nature there are key moves and sometimes indeed we should play them earlier than we like.
On the other hand L6 is dependant on L9 so I dont see the point of playing L6 first :D
4 Oct 2019
LeelaBot loses again to MooTwo: http://www.littlegolem.net/jsp/game/game.jsp?gid=2123807
this game had an unorthodox opening with MooTwo's h8 response to the first stone at k7
4 Oct 2019
MooTwo, congrats, it seems the initial stones are really hard even for the bots.
4 Oct 2019
The irony of hex is that intial stones seems impossibly hard but then a few moves later it's game over... so there has to be a way to play right even at the beginning but it's just a little too hard for us... maybe these MCTS bots can already approach perfect play on 13x13 with enough time on the clock and enough RAM.
4 Oct 2019
lazyplayer, how do you know that mootwo is mcts, also i do not believe that its just a 'few' moves until game over, if you look at games on here among top players, hex13 has about 30 moves per game on average
5 Oct 2019
With this win: https://www.littlegolem.net/jsp/game/game.jsp?gid=2123808 Mootwo is now tied 4-4 in games against LeelaBot, it seems that the title of best bot is unclear. Also Mootwo seems to be playing some unorthodox opening moves in its games.
5 Oct 2019
Iguser, the games last more moves but the outcome is already quite clear early on... :)
For example the game you've cited, already here it's clear black is probably winning. Mootwo plays very good, I wouldn't say unconventional, indeed what's convention here? There are no conventions, humans have their own conventions but they've been refuted by the bots.
5 Oct 2019
these are very exciting times for hex, its a shame we arent discussing on a half decent chatting platform... i tried to do a move towards discord but it failed to gain traction
5 Oct 2019
Mootwo now 5-4 against leela, https://www.littlegolem.net/jsp/game/game.jsp?gid=2124434 this one had mootwo k12, leela j4, j8. d10
5 Oct 2019
Good to see them struggling against each other!
Gives me hope to try seriously to avoid mistakes and one day I finally get a win :)
6 Oct 2019
I'm enjoying reviewing these games. Thanks to all who are maintaining the bots here.
6 Oct 2019
yes, i can't wait to see bots master 19x19 too
7 Oct 2019
It looks like Daniel has secured a win and with the swap disadvantage !
I think this will be the first bot loss to a human...
7 Oct 2019
I also think that Daniel seems to be winning. This would not be a first, lazyplayer beat leela in this game: https://www.littlegolem.net/jsp/game/game.jsp?gid=1999024 almost a year ago and daniel beat gzero in this game: https://www.littlegolem.net/jsp/game/game.jsp?gid=2033239 10 months ago.
7 Oct 2019
interestingly, both of those wins were also from a swap disadvantage
7 Oct 2019
Wow - congrats Daniel. Very similar to #2122370 - that might be inevitable given the opening, and leela will be playing similar moves since same model.
7 Oct 2019
Do the bots not learn from their mistakes?
7 Oct 2019
I have also tried to copy this game #2122370 once, but I was arrogant and tried to improve.
Anyway this game is the most likely weakness of leela and Daniel used it brilliantly! Till very late I was not sure who was winning because there were potential tricks, but Daniel solved it all, big congrats!
7 Oct 2019
Oh damn, game 2122370 is identical, haha, this seriously devalues Daniel's win... but anyway, it's a weakness of the bots, they're not randomized! :D
7 Oct 2019
I'm also tempted to copy these games in my own game vs leela, I'll see... everything is fair game here... :D
https://www.littlegolem.net/jsp/game/game.jsp?gid=2123807&nmove=6
https://www.littlegolem.net/jsp/game/game.jsp?gid=2124448&nmove=10
IgUser, my win was just luck unless it's repeated... I hope it'll be repeated but I've to be really careful, hoping/dreaming is not enough to win.
7 Oct 2019
Bill, the bots learn from self-play not from the actual games they play. And they've limited capacity to learn (due to their limited "brain" size), then you've to change the neural net (basically, change "brain" and pick a larger one) and start over with training. And training is costly from a computational point of view. In the case of Lc0, the chess engine based on neural nets. they're using donated computing power to do the training of the network, and despite this it took many months.
7 Oct 2019
Bill, to prevent issues like this, that is, repeated games, they need to be randomized somewhat, basically, to play slightly weaker moves for sake of making the bot more unpredictable.
7 Oct 2019
I wouldn't be surprised if a bot was in the works that used high level computing power, maybe will show itself in the next computer hex tournament.
7 Oct 2019
Well the cats out the bag now. :D gzero (and AFAICT leela) have been playing the best move from move 2 onwards. We've kept the illusion of randomness by continually improving the models.
There is some randomness with MCTS when using multiple threads & virtual loss, but it isn't significant.
Still I am actually very surprised it has taken this long to see very similar games (especially between the games between gzero and leela).
7 Oct 2019
gzero, don't be ashamed, actually that's the principled thing to do, no randomness, but always change opinion on what's best! :)
7 Oct 2019
it looks like the 2019 hex computer championship was won by a bot named DeepEZO, https://icga.org/?page_id=2908
i think that if this bot were to play on littlegolem, it would be the best, sadly, i failed to find it online
7 Oct 2019
Ouch, I thought Daniel copied another game!
Indeed this #2122370 one is entirely identical, eh.
7 Oct 2019
whats the point of that, Daniel shows he cares more about rating that improving at hex
7 Oct 2019
Well, no, It's bot's fault it played the losing game. Daniel did the right thing.
I just wished it was a real human triumph over bot :)
8 Oct 2019
Iguser,
I played the most games with bots and lost most points and you write that i care more about points? Just look at my rating chart. Please, don't be funny (unkind you are).
See my previous games against bots, maybe you'll notice something (especially when i played black).
8 Oct 2019
Meanwhile new hex bot game has started link to 2125477
Even for me who is far from even average hex level these bots games are interesting. Want to give some support to gzero because of sympathy it plays very strong not only hex but also other games. gzero is a monster - give me the rules, give me time and I will show you how shit you are :P (or how you should play in moderate words :)) But can't bet on gzero with dark horse mootwo which already gave some new and surprising ideas at openings. Will be interesting.
Also I wonder about ability of leela and mootwo to play other logic board games. Personally I am interested in connect6 but it also could be other games. As far as I know here at LG plays strong twixt bot.
8 Oct 2019
gzero seems to have gotten stronger lately, I remember gzero saying that he was going to retire from little golem, I am glad that wasn't the case
8 Oct 2019
Depending on how fast will the game be I will try myself in a live commentator role :)
https://www.littlegolem.net/jsp/game/game.jsp?gid=2125477&nmove=7
So far it seems gzero on white has quite an advantage. K4 should be played and mootwoo's stones are almost useless!
8 Oct 2019
As usual for the bots mootwo started with a 4-4 corner which has been proven strong despite being less versatile than 5-5.
gzero immediately took favour of the strong connection on the right edge and expanded with j2. Tactically correct, but it's debatable if that is the optimal move. People have been playing this for long time with mixed results, but maybe it just takes a tactical master like bot to use properly.
I don't know what is the logic behind the mootwo's top-right corner fight becaues I see absolutely no gain for black!
8 Oct 2019
K4 was played indeed and mootwo tries to take a free nozoki on j3.
https://www.littlegolem.net/jsp/game/game.jsp?gid=2125477&nmove=9
But I'm not sure if gzero will allow this j3 for free - there is time to play tenuki like f7 or just escape from the corner with i5.
8 Oct 2019
The bots seem to pause for a while so I paste some random lines:
Likely white win:
http://www.trmph.com/hex/board#13,k7k7d10j2l3k3k5k4j3i5j5i6j7h8i9j4d5d3c3d2c2e6
Expected continuation (I bet on white):
http://www.trmph.com/hex/board#13,k7k7d10j2l3k3k5k4j3k2d5d4e4c12b12c11c9
8 Oct 2019
Eventually gzero played passive 10.k2 which made black j3 a free stone. Looks suboptimal to me.
Interestingly mootwo secured bottom with h7 which gives white just one possible option to connect through bottom like: http://www.trmph.com/hex/board#13,k7k7d10j2l3k3k5k4j3k2h7h6g6g7f7e9f8f10
If white cannot make use of the latter than maybe it's even black's advantage.