Havannah Human versus Bot Challenge

Here's the current state of affairs on Google+

It seems to say both end of august and october 15. Since october 15 is the most recent (and precise) one, it seems to be the right one. Is this correct?

Yes, but I've requested the first game of a day to be played at 11.00 am and the second in the afternoon.

Here's the press release.

uh, I had forgotten about this :) interesting.

Can't wait to watch the live-streamed games at http://www.dgt.nl/site/index.php/dgtchesslab/activities/339-havannah-challenge-live-stream

Havannah challenge live stream

Since I'm trying to show that computers cannot yet be suspected of 'intelligence', I might enhance visual entertainment while proving that humans aren't always all that intelligent either, and hang Kobus around my neck on the last day:

But only if I win the games of the first four days of course, and if Kobus feels like it :).

Make sure he is not hungry before hanging him around your neck!

It cuts both ways: while intending to show that 'artificial intelligence' is still a thing of the future, I can thus illustrate that human intelligence is sometimes non-existent too ;-) .

For LG fellows interested here is the adress for following Christian's efforts lIve (or rather the bot's efforts) : "http://www.dgt.nl/site/index.php/dgtchesslab/activities/339-havannah-challenge-live-stream"

Game start was reported to start in one hour: at 12 GMT+1.

I wish you good luck Christian !

Well, I'm quite sure he does not need luck. The key word is entertainment, so he does not get bored. I expect some kind of demonstration by Christian, i.e., how to win by a bridge, by a large ring, by ignoring the bot's every third move ...

In any case Christian deserves much respect for offering the bet, and sticking to it. Anyone can ask themselves: would they venture such a 10 year prediction, even in their very field of expertise?

Of course the bot authors deserve respect as well. I know (because I failed, so far) how hard it is to take some "more or less obvious" ideas (UCT/MC, global evaluation) and put them together as a working piece of software.

PS: The live stream shows Christian in front of a screen, but where's the actual (or virtual) board?

You're lucky, Ring, all I get is blank rectangles. Adobe don't make a flash plucing for PowerPC linux machines. And that hosting site doesn't host anything that's compatible with the Gnash plugin that I do have.

Good luck, Christian.

It seems Christian won the first game agains Lajkonik in 30 moves
(that's what I figure from the ustream chat window).

The game would have been visible on iggamecenter (using just Javascript, I guess) but (1) you need to register to watch and (2) the game was private, so even then you couldn't watch.

I guess this "privateness" was necessary because bots were mis-interpreting the "join game as visitor" messages.

perhaps bot authors can post the games later, in some accessible way.

Is there a transcription of the commentary somewhere accessible?

I notice that Christian's ratings graph looks quite healthy - http://www.littlegolem.net/jsp/info/player_chart.jsp?gtid=havannah&plid=18176 - which I hope means that there's been some useful training, and in particular anti-bot practice!

@Ring: I'm a bit out of the loop, too, but I believe the games were made private so that Christian could not take advantage of any comments that might happen to pass by.

@FatPhil: As Ring says, I don't think much anti-bot practice was necessary. He just has to be sure he doesn't accidentally stumble into some kind of tactical morass.

@FatPhil: I've been putting a bit more thought into my games lately. However, 'anti-bot practice' can lead one to anticipating on a presumed weakness. I prefer to presume that the bots are at least as smart as I am. If there's a 'weakness' it will eventually show - no need to 'seek' it.

From the website of the match:
> In the second game an error happened and the arbiter decision
> was replay, later this week.

Some information on the nature of the error?

Ingo.

After a connection error Lajkonik left the game and couldn't return. Handling that eventuality should be fixed by now.

I presumed that the different bots have their own style, and that you could actually play towards their weak points. Even if that means just having a preferred opening against it.

For example, consider RoRoRo at Dvonn - I know I specifically aim my play to exploit its biggest weakness; all 1900+ players do, I'm sure.

However, you're right, you should never /rely/ on your opponent making a mistake.

Hello Christian,

> After a connection error Lajkonik left the game and couldn’t return.
> Handling that eventuality should be fixed by now.

Thanks for the information.

Good luck to you - and of course also to the
bot teams.

Ingo.

Hmm, did Lajkonik really win
the third game this morning, as the website tells?

Ingo.

Copied from the site
http://www.dgt.nl/site/index.php/dgtchesslab/activities/339-havannah-challenge-live-stream

> The rounds and pairing (all times are CET)
> 1.Monday 15 Oct. 11:00 Lajkonik-Christian Freeling 0 - 1
> 2.Monday 15 Oct. 14:00 - Lajkonik Will be played later
> this week
> 3.Tuesday 16 Oct. 11:00 Christian Freeling-Lajkonik 0 - 1

Congratulations to the Lajkonik team !!!!!!

Ingo.

Looks like it. The screenshot shows at least a one-move advantage for black (because of the forcing move at B1) and maybe that's even two moves (since the marked black move at D9 needs to be answered). So - congrats!

Congratulations to Marcin and Piotr!

And now some interview questions:

Where does the name "Lajkonik" come from?

How long has Lajkonik been under development?

What improvements have been made since Lajkonik placed 3rd at the 2010 Computer Olympiad in Kanazawa?

What future enhancements are in the works?

One more question: Can someone please give
the full move list for round 3 game?

Thanks in advance, Ingo.

Oh wow, i didn't expect that...

But (after looking at the moves), the game is a simple race to connect 2 corners (each player "does his own thing" connecting corners, except answering ring threats), shouldn't you avoid such a type of game playing against the computer? The computer has a chance to make clumps, optimize this ring threats, it's easy to overlook something or defend to early or too often. So in the third move i would have put for example a stone in a third corner.

(but it's just my opinion from the time when i was playing, i have a lower rating than Christian and didn't follow the developments in the last two years here)

Number of planned games:

Lajkonik 4
Castro 4
Wanderer 2
So this is an indication that Lajkonik and Castro were considered the stronger programs.

@Ingo Althofer:
The first three of four games can be reviewed at iggamecenter.com (if you have an account), the fourth needs making public by the admin (Arty).
Filter by Havannah, you can copy/paste the move list into a text file on your computer.

So this is how fame and money feel like...

@MarleysGhost:
Lajkonik is one of symbols of Kraków, Poland where both of us live: a man disguised as a Mongolian warrior riding a hobbyhorse who dances through the streets beating passers-by with his mace.

I recall writing the first lines of code in May 2010.

First of all, RAVE; then simple patterns in playouts; multithreading with a lock-free transposition table; Havannah-mate in playouts and in the tree (stolen from Wanderer through Jan Stankiewicz's thesis); Johnson's algorithm for finding all the elementary cycles in a graph (for ring threat detection; not to be confused with a better known graph algorithm of him for finding all-pairs shortest paths); progressive bias for preferring moves close to player's own moves (stolen if I'm not mistaken from Jan Stankiewicz through Castro).

Not many. I was planning to steal also Castro's way of detecting if a ring closed in a playout should be treated as a win but apparently my way is good enough. Implementing pondering would be nice, too. Eventually, I plan to open-source Lajkonik, after removing deadwood code remaining from various experiments.

Thanks for the answers. More halftime locker-room interview questions:

Were you surprised that Lajkonik was able to manage a victory? Will you be surprised if Lajkonik wins another one of its two remaining games?

How important do you think Lajkonik's ring-threat-awareness adjustment was in the victory? Was anything else adjusted? ["How important do you think..." is not a question I would ask, but I think it's something an actual journalist might ask.]

How much time have you two put into Lajkonik development? How do your wives (or other significant others) feel about this project?

Christian Freeling, like David Levy, is not the world champion. Would you be open to a match against the Little Golem champion, currently fritzd?

Any inclination to take on another game, e.g. Go, Slither, Connect-6? [Another journalism question]

What question do you think a journalist should be asking? What would you like to tell the world?

Yes and yes. As I wrote on October 4 -- "in base-10 Havannah, notwithstanding my personal involvement, my bet is on humanity."

The winning settings were a mixture of a lucky guess and some knowledge gained here through games with Castro and Wanderer and tested on a couple of positions with ring threats. To elaborate: they are two numbers that describe the y coordinate of a straight line at the beginning and at the end of a playout. When a ring is played in a playout, a random number from 0 to 99 is chosen, and if it falls below the line, the ring is considered valid. The idea is to filter out spurious rings that often happen in playouts on larger boards. In the first game, the numbers were +100 and -500. Then I changed them to +150 and -350. I made no other adjustments.

I set up the code repository for Lajkonik in December 2010. Since then, there have been 452 commits; almost all of them by me. My girlfriend says she is proud of me but there were ups and downs when I was coding for hours instead of hanging out with her. Piotr's involvement was not that big: he coded the GTP interpreter just in time for the olympiad in Kanazawa and ran various experiments afterwards.

Absolutely. Lajkonik is here to play. Anyone can challenge him, though in correspondence play his chances are even slimmer than in real-time games.

Maybe I will join Lukasz Lew and Witold Jarnicki in their research on a divide-and-conquer approach to computer go. They have some very interesting ideas.

I would like to thank many people: Timo Ewalds for many fruitful discussions and willingness to share his ideas with me, Przemek Drochomirecki for telling me about Havannah, Lukasz Lew for his consultations on MCTS and Havannah, Remi Coulom for his CLOP program, Witold Jarnicki and Adam Byrtek for their ideas on ring detection, Dawid Kurzyniec for helping me to travel to Kanazawa, and others whom I passed over by absent-mindedness. Last but not least, I thank my girlfriend for bearing with me.

I think that the moral of the story is: you can excel in anything if you consistently work on it.

Any games played today?

Yes. Just one between Castro and Christian ended:
Christian achieved a nice win (one move ahead).

Ingo.

Christian just won a snakelike game 6 against Wanderer_bot.

Ingo.

Am I missing something, or could Christian have won his lost game as late as in move 39 by playing b3? Also, his move 33 seems unnecessary to me, costing him one move on a non-threat?

I WAS missing something. Move 33 was obviously necessary. And although 39.b3 would have given Christian one extra move, he was still one move behind due to the ring threat caused by d9. Next time I'll study the game BEFORE I comment, instead of right after :-)

@Marius, If you were a member you could delete untimely posts. Your days of student poverty are in the past. I'll bet you have 19 Euros in your pocket at this very moment.

I do. They are. And I AM a member. I just feel it's more "right" to correct my errors with a new post, than erasing them and pretend I never make them :-)

Ah, my membership must just have run out, I see. Guess it's time for a renewal. How much for a life-time membership, so I didn't have to remember this every year? :-)

Hi Marius,

> ... I just feel it’s more “right” to correct my errors with a new post,
> than erasing them and pretend I never make them :-)

Good philosophy; I do it the same way.

Cheers, ingo.

The match is over. At the end of five hard days Christian had to
pay tribute: Today he lost one game each to Lajkonik and Castro.
For me it seemed that he was no longer as fresh as before.
So the final score is 7-3 for Christian.

Thanks to all for making the competition such an intriguing event!
Ingo.

Fascinating games!

I've collected all the game records of the 2012 Human vs Computer Havannah Challenge and published them online at trmph.com. No client download or registration is required to view the games. (JavaScript must be enabled, though.)

Game #1Lajkonik (White) vs Christian Freeling (Black) - Black won.Game #2Christian Freeling (White) vs Lajkonik (Black) - Black won.Game #3Christian Freeling (White) vs Castro (Black) - White won.Game #4Christian Freeling (White) vs Wanderer (Black) - White won.Game #5Castro (White) vs Christian Freeling (Black) - Black won.Game #6Wanderer (White) vs Christian Freeling (Black) - Black won.Game #7Christian Freeling (White) vs Lajkonik (Black) - White won.Game #8Christian Freeling (White) vs Castro (Black) - White won.Game #9Lajkonik (White) vs Christian Freeling (Black) - White won.Game #10Castro (White) vs Christian Freeling (Black) - White won.

"For me it seemed that he was no longer as fresh as before."

@Ingo
You saw that right ;-) Nevertheless it was great fun :) .

@All
The games: I hate analysis but here's the overall impression.http://www.trmph.com/havannah/game/hvc2012-1

The first game Lajkonik presented itself as dumb - that's perfectly legal. It's not reversable either: I can't lure a bot into thinking ...

There was a second game that went wrong due to a connection bug. I wasn't too happy with the position at the time it was aborted. I should have been warned.

The next second game I chose the wrong strategy, plain and simple. You can't outrace a bot when it indicates readyness to engage. I thought I had the decisive tempo. Mea culpa.

Then wins up to and including the 8st game. All based on a strategy of enclosure at two sides because the bots prefer cluster bombing with ring threats for making speed, and will usually keep expanding close to the main structure, and very effectively at that, given the object.

In the 9th game I was already fairly exhausted, but I [i]think[/i] I had the decisive tempo when I accidentally submitted a move instead of removing the stone I had been analyzing. I had a direct threat in two to answer and was well aware of that. Analysis will probably reveal I had a won position, but (sh)it happens :( .

In the final game against Castro, I saw a catastrophic move in the replay. At black_5 (move 10) I placed a 'defensive' stone under a delusion that I was unware of at the time, like all delusions, but that I immediately spotted on replay: the danger was an illusion. That not only cost a tempo (and I was playing second as it was), but once you stick your foot in it, there are suddenly additional moves necessary, and Castro expoited that very good. From that point on I was running behind the facts.

My congratulations go to Timo and Marcin, and my thanks to all involved, Richard Lorentz, Ton van der Valk, Albert Vasse, Jan Krabbenbos, Frans Faase and Ed van Zon.

P.S.
I burned my ships behind me in the 5th game when I went for a bridge, because I didn't see how it could be prevented. I could have kept the option while securing a forkframe by cutting above it, but because I couldn't see it, putting it all on that one card seemed more aesthetical (hear the bots laughing).
Fortunately the bot didn't see it either. Jos Dekker saw it, and Ton van der Valk confirmed it: the bridge could be prevented. So I got away with it :) .

@Z
Capital Z! :)

Thanks to Christian for his comments.
Thanks to z for making the games publicly available on his side.

Three questions.

(i) to all:
Can someone provide the "original" game 2, which was aborted?

(ii) to the programmers:
Did (some of) you improve your bot during the match?

(iii) to Christian:
How is the prize money split amongst the teams of Lajkonik and Castro?

Also my thanks to the Dutch team of organizers:
Ton van der Valk, Albert Vasse, Jan Krabbenbos, Frans Faase, Ed van Zon.

Ingo.

(iii):
Equally, including Wanderer, or equally excluding Wanderer, or 2/3-1/3 excluding Wanderer. Point is the interpretation of 'the winning bots' - are they referred to as one team or as three seperate opponents. I don't care, it's not my call, and most probably none or the competitors will really care. The fun was in the event, as far as I'm concerned :)

Another photo with the programmers, taken at the
Computer Olympiad in Tilburg(NL), November 2011.

Sitting from left to right:
Timo Ewalds (Castro), Marcin Ciura (Laykonik), Lukasz Lew,
Richard Lorentz (Wanderer).
Standing in the middle of the background Ton(y) van der Valk.

Ingo.

Hmm, I used the command
img scr="http://www.althofer.de/havannah-tilburg-2011.jpg"
with brackets around. But somehow it does not work.

Ingo.