MoHex plays rather well

Here is a recent game I played against MoHex on 13x13, with a time control of 30 minutes per player per game.

Using HexGUI as the interface, in the shell window I typed in "mohex_param use_time_management 1"

MoHex does not make the swap move, and when moving first does not try to make a balanced initial move. So, I always move first as black and try to make a balanced move. Here I played 1.A9.

I didn't play 11x11 because it usually beats me. But on 13x13 I usually do better.

I think leela and gzero are both better than me on 13x13. I expect leela to win the next championship

There will probably be a bot only 1-st championship league soon :/

Beautiful game, thanks for sharing :) I'm surprised to hear that A9 is considered a balanced first move by strong players; the only "opening theory" for Hex I've ever been able to track down is Cameron Browne's book and he seemed pretty certain that anything more central than A3 or A4 should always be swapped. Glad to hear it's not as simple as that!

apetresc: you can always check out my website to get an idea on the various openings. It's not "opening theory" but it's a start.

David, looking over your game I think both you and mohex are not playing well in the opening. If you played 5. b10 you should win pretty easily there.

Oh!!! So the time came when we face bots in championship...

I hope we can still win to them, but it takes huge discipline. Each move towards the end they play closer to perfection so we need flawless openings.

I'm very tempted to try my flag opening vs Leela to check if it's objectively good or just fit to my play vs people.

Arek, I actually beat gzero on 11x11 one time because it messed up near the end: http://www.trmph.com/hex/game/lg-1978111

(it has a winning position there before move 32)

But I could not beat gzero on 13x13; the developer said the 13x13 version is better trained, which might be the reason.

The point is, don't assume the bots play the ending perfectly! Their searching ability might be worse than yours. Neural nets are weird.

shalev, I get your point, but I would call this a bug rather than the bot's feature

I guess they should use some kind of balance between neural net solutions and sheer brute force power so that endgames are played flawlessly.

Also in practice it's rather useful to expect the best of opponents :)

And actually in the game you mentioned I don't see any way for gzero to win before 32. Are you absolutely sure?

Arek, 32 h10, then  36 i6, white wins...

I suppose correct play was 9 h5, let's play that immediately, then 10 h10, 12 h9, 14 i6, then it should be possible to see who is winning...

I think the bot play reasonable initial stones, it's clear that it's using the right approach, i.e. it has played lot of game vs itself on the empty board.

wrong http://www.trmph.com/hex/board#11,a3d8g5g8h8g9i9g6c8d7c7c9a10c5b6b4c4b5c6e5h5g4d5e3e4g2f3f2i3h2j3h10i10h9i8i6h7h6e6c10f5a11f7

Arek, wrong: http://www.trmph.com/hex/board#11,a3d8g5g8h8g9i9g6c8d7c7c9a10c5b6b4c4b5c6e5h5g4d5e3e4g2f3f2i3h2j3h10i10h9i8i6h7h6e6e9

Humans are just lazy, we can't bother to think seriously, bot will have easy win i suppose :D

Yeah the e9 stone is the trick for white. I didn't see it until somewhat late in the game, then realized I lost... and then gzero apparently didn't see it either, giving me the win.

lazyplayer, I feel like I should just resign after move 8.g6. Not sure if black has real chances after that. I think one of my first 3 stones need to change.

Indeed too lazy...

I abandoned analysis in favour of pure principled intuition and it does NOT work at all... Damn it, I had high hopes

shalev, well, after 3 g5, the other two are basically forced. Then 9 h5 is forced. 11x11 is all forced. :)

3 h7 and 3 i8 should have good chances, in fact h7 should be winning...