User tournaments

I see, that in the past there have been many user tournaments. Are there any people, wanting to revive the tradition? :)

Sure!

Why not?!

So we need a member user to create one :)

:D

I can create one :) What kind of parameters do we all want?

Nice! :) What kind of paremeters can you set? I would play anyway, asking just out of curiosity :)

Board size, date to start, rating limitations or free for all, public or private (by setting a password) are the options I remember.

For me: 13 X 13 the size of the board, no rating limitations and, no problems about the date :) What about you?

Same for me!

Same :)

So we are talking about Hex? Good.

(@Force majeure, I wasn't sure if you initially wanted a Hex or a Havannah tournament!)

I prefer hex, but would play havannah as well

Okay, I created the tournament, with a start date of November 1st, so just a little over a week from now.

The password is: gardner, in honor of one of the great popularizers of the game :) I decided to add a password just to prevent a lot of drive-by registrations that never show up to play. I put a link to this thread in the description so the password is not hard to find.

See you all over the board!

(Oops, I suppose I should include a link!)

https://littlegolem.net/ng/a/Ctrn.action?view=&dto.id=1112

Thanks!

https://www.littlegolem.net/ng/a/Ctrn.action?view=&dto.id=1112

Well done, thanks!

Already 6 players! Can we make it to double digit? ;)

9 players now

10 players now - and I'm the lowest rated by over 200 points! This should be interesting :)

Can I play? I realize some people don't like playing a bot, so I won't join if there are objections.

None from me, since you asked so nicely ;)

I don't mind personally either - although since you're here, I'd like to re-ask you a question I PMed you a few months ago: are the network weights and code you use published anywhere? I'm an avid leela-zero fan for Go and would love to train and play Hex with it too :)

leela, you should probably give handicap to players below 2000, just give them a strong start... :)

I don't want any handicap. We could have another tourney, with handicap, on the other hand.

But I think should be on another thread. How will the handicap be decided, for the various differences in rating?

Regarding handicap - are users tournament rated?

Force yes this is rated but leela is so strong and consistent, it'll win even with the handicap!

Ok nów I get it :) Looking forward to next chpionship

Force, in fact I've watched some of its games, maybe it's consistent but not strong, hehehe.

But consistency is very important in hex, one bad stone and you're dead!

apetresc: no the code isn't currently public. I haven't really decided where to go with it and am currently still working on the network training and don't want to spend time getting the code ready for "publication". It's a bit of a mess at the moment.

Row handicapping would not work for Hex, by the way. The only kind of granular handicapping system I can think of would be for strong players to agree on sets of  initial non-swappable moves that would fall into the same approximate category for a set ratings difference. One might even consider specific double initial moves for stronger handicaps. But that sounds like a total mess to deal with. I agree with hypercube to wait for later to implement any handicap. And by the way, how did this tournament turn into an AI allowed tournament? Who let THEM in here?

@David do you object on leela_bot playing in the tourney?

I suppose if there is at least one objection, it won't play. But I haven't seen any explicit objection yet.

I don't mind playing against bot as long as its rating reflects its strenght, but perhaps creating a leela challenge tournament is a better idea? Anyway, it would be another possibility to play some more games. Taking into account how popular this user tournament has already become, it seems that finding participants will not be a problem :)

@Force_majeure, since you initiated this thread and idea I appoint you as admin of the tournament (along with @apetresc who was kind enough to create the tourney).

I will gladly accept your decision on this.

I also notice that you mentioned "people" in the very first line, so perhaps it is better we avoid bots in this one (and as you suggested, create another one as "leela challenge) and/or some other, for all bots and humans to participate.

We have 13 people registered so far and 3 more days to go.

ypercube, this is very nice of you, however I don't feel like owner of the tournament :) Moreover, when typing "people" my intention was not to prevent bots from playing. However, if there are any concerns regarding bots, I think it would be safer just to create new tournament and enhance traffic around hex a little bit :)

It's fine with me if a bot participates.

leela is rated significantly above me, so I would not object to playing. If any bot rated lower than me participates, I will resign on the first move.

David, it's underrated anyway, even if it is above you in rating...

I would say we need to relax, it's only a silly game, there are no prizes or anything.

Btw, I've decided to challenge leela, see my game, commentary is welcomed even during the game... :)

https://littlegolem.net/jsp/game/game.jsp?gid=1999024&nmove=15

Hopefully I'm still winning. I think the initial stone was way too strong.

@lazyplayer: gzero thinks white has a 76% chance of winning after 16.i7. although gzero has a large bias for white.

ps: i am obviously supporting my compatriot leela in this game. ;)

gzero, in hex, 99% win can immediately turn into 100% loss, heheh... this is the funny part of this game...

in your bot, the eval function is very unstable, isn't it? from stone to stone... or stone to next stone of next player

@lazyplayer: the eval isn't unstable between moves... For example playing out your leela game it pretty much is at 70% ish all the way (to move 50 in your game) - and in self play continues to 72% at move 60, 74% at move 70 and then the game is pretty much done at move 74... [see here, hypothetical spoilers - ]....

The issue with instability with hex models is that certain board positions the network will think are good when they are clearer not.  And it seems that with extra training the instability doesnt improve (where in other game types, it does)... So in other words it is super optimistic about the outcome of the game, but as the search gets deeper it will end up flip flopping its opinion suddenly.  :)  I am still running an old network from early July when we had the bot competition, whereas (I assume) Leela has trained and improved much since then.

Also I reckon gzero is a lot stronger as white (so the bias might be accurate), but it is hard to prove.

PS: I won't be entering this competition (and sorry for gatecrashing this thread).

Stronger as white you mean stronger with an extra stone or stronger as 2nd player?

Thanks for explanation.

> And it seems that with extra training the instability doesnt improve (where in other game types, it does)...

Well this is exactly my point. Hex is kind of discrete/unstable. For example in my game, 32 d4 is black win I think, but it'll probably still be seen as 70% win for white by your bot? I think hex has low "branching factor" and low "depth", and yet it's so hard to approximate. It's kind of paradoxical.

> So in other words it is super optimistic about the outcome of the game, but as the search gets deeper it will end up flip flopping its opinion suddenly.  :)

Well this is the classic horizon effect.

> Stronger as white you mean stronger with an extra stone or stronger as 2nd player?

As 2nd player.

 Well this is the classic horizon effect.

I was going to write that.  :)

Basically my opinion is that, for hex,  either you play near perfectly, or you play near random. It's hard to find something in the middle.

I'm not sure what you mean by Hex having  low depth. The top player here is almost 2680, the other games on here which have a top player that has a rating at or above that are Go and Twixt. So Hex is roughly the third deepest game on little golem.

In the bots' play style I very often get the impression like the losing player is not even trying. It's very visible in the above example of gzero analysis. No wonder it's optimistic for white when it's playing hopelessly as black itself :D

Iguser, good point, I have a vague feeling on this but I don't know how to explain it. But I'm confident even top hex players play a ton of errors.

Wait, gzero, you say that the game is over at move 74 in your analysis... but it is BLACK who is winning in that position, not white!

http://www.trmph.com/hex/board#13,l1l1d5d10j4i10f9f8g8h6g7g6i6h7i8i7f7e6f6g4k6j7f5f4l7j11k10k11l10l11c7e3c4c3d3e1e2f1f2g1g2h1h2i1j2i2h4i3c2d2a4b4a5b5a6b7b6c6c5e5g5h3d6d8d7e7j9h9i5j3b9c10c9f10b12a11e10e11d11d9l2k3l3k4l4j6h8

Whereas if you rewind to move 71, white has a winning position. The mistake is in move 72 of the analysis.