What has Koshling been up to

Koshling

Vorlon
Joined
Apr 11, 2011
Messages
9,254
For those of you that knew me here, I thought you might be interested in what I've been up to recently.

I’ve been involved in working on a GGP (General Game Playing) program in my spare time for the past 9 months or so (following taking an online course in the subject: https://www.coursera.org/course/ggp). This is a game player that can play any game described to it at runtime (within the bounds of a description language), rather than any particular fixed game, such as chess (though chess might indeed be what is described to it in one run).

Anyway, this week is the 2014 world championships (http://games.stanford.edu/index.php/ggp-competition), and I’m taking part along with a co-developer who joined me a few months ago.

Over the past few weeks qualification rounds have been going on, that reduced a field of about 50 down to 17 qualifiers for the 2 day finals (of which we were one). Those finals take place today and tomorrow (Tuesday/Wednesday).

Day 1 was an 8-hour round of various puzzles, and some 2 player games with everyone playing the same games, and total scores being ordered at the end of the day. The top 4 qualify to a seed bracket for the final day, the next 8 to a lower bracket (the difference is that it’s a double elimination format on day 2, and getting into the top 4 mans you have to lose twice to be eliminated), and the remainder are eliminated on day one.

We started slowly (had our worst games first by luck of the draw), but in the end finished fairly comfortably in first place: http://arrogant.stanford.edu/ggp/administrator/leaderboard2014.php (we're Sancho)

Wish us luck for tomorrow!

You can also check out our blog if you're interested: http://www.sanchoggp.blogspot.com/
 
Good luck Koshling. Miss seeing all your amazing improvements on the SVN thread. Win first place and come back to c2c soon.
 
Congratz Kosh. Keep it up!
 

Attachments

  • Bez-nazwy-1.jpg
    Bez-nazwy-1.jpg
    18.2 KB · Views: 42
Good Luck kosh! :hatsoff::thumbsup:

Knock'em dead as they used to say back in my day. :D

JosEPh :)
 
Very interesting. Sounds like fun.
I checked out the descriptions of GGP and GDL on the site you linked and I guess while you start quite general you tend to make optimizations for the kind of pattern that humans often add into such games.

Checked out the twitch channel now and seen the first game of the semi finals. Amazing performance. Far better understanding that the mid fields are important than the other player.

Good luck :)
 
Extremely intriguing subject matter... I'll bet that course was really something in and of itself. Very impressive results in your competition as well... y'all should be quite pleased so far. It bodes very well for things going forward from here. Are you (and the other competitors) able to make many adjustments between today's events and tomorrow's?

Also... it certainly immediately makes me wonder if such a generic 'player program' would be able to adapt to any decent degree to the dynamics of a game as complex as Civ? Particularly in a way that would be extremely optimized?

I was wondering too from reading the blogs, are you guys rated on your processing speed during this competition? Or merely on the efficacy of the player's game choices?
 
I was wondering too from reading the blogs, are you guys rated on your processing speed during this competition? Or merely on the efficacy of the player's game choices?
Only winning counts I assume. But the faster you are, the further you can search in the game tree in the limited time available for each move and that results in better choices.


And congratulation, Koshling, entertaining games in the finals.
 
Yup - just results count. For reference we run on a Windows PC (single CPU), whereas some of the competition was running on much more powerful hardware (Ary runs on a 40-machine cluster for example).

Brief summary of the latter stages of today's play:

After winning the top bracket of 4 players, and the semi-final we were drawn in, the final was against a French player called LeJouer.

First game was hex, with him as first player (pretty sure that's actually a strong win). He had it sewn up (I think), but made one error, after which there was no looking back, and we won. Game record (use cursor back/forward to advance through/back moves): http://arrogant.stanford.edu/ggp/php/gameviewer.php?id=x1406759945001

Second game was Breakthrough. By some margin this was the best game of Breakthrough I have yet seen. LeJouer came up with a frankly brilliant move which clinched the game, and recorded our first Breakthrough loss for several months: http://arrogant.stanford.edu/ggp/php/gameviewer.php?id=x1406759945002

Third game was TTCC4_short, which I had very low expectations of. We started in second (inferior) role, with an assessment around 45%, but inexorably crept that up, and eventually found the win! : http://arrogant.stanford.edu/ggp/php/gameviewer.php?id=x1406759945003

Fourth game was Joint Connect4, which we are extremely strong in (because we can factorize independent sub-games into separately searchable spaces), and fairly rapidly arrived at forced wins on both boards, eventually winning in 33 moves on the normal board: http://arrogant.stanford.edu/ggp/php/gameviewer.php?id=x1406759945004

Since the final was best of 5, that clinched the win for us, and we became the 2014 champions.

In the carbon vs silicon the inferior squishy biologicals in-advisedly chose to play us at joint connect4, a mistake they did NOT live to regret, though it again took us a full 33 moves to grind them underfoot: http://arrogant.stanford.edu/ggp/php/gameviewer.php?id=x1406764968001

Draw tree and results are here: http://challonge.com/iggpc2014


Our blog, which details a lot of how our player works (though it does require a little background in programming, and assumes some knowledge of GGP) is at http://sanchoggp.blogspot.com/
 
Congratulations Champ! Well Done!:king::clap::hatsoff:

JosEPh
 
Out-STANDing!

Now then... aside from a heart swelling with pride of accomplishment, is there a prize for winning at all?

The first 2 years (2005 and 2006) had a $10K prize, but not since then (that was basically to gt the field going).
 
The first 2 years (2005 and 2006) had a $10K prize, but not since then

So for work done to C2C you earn the same ;)

I would love to see your posts in svn thread :)

for example we discovered that trade route callculation code is very unneffective....
 
Top Bottom