6.912 - MIT Pokerbot Competition | Hacker News Reader

6.912 - MIT Pokerbot Competition

pokerbots.mit.edu

I'm one of the guys behind this. I realize it's MIT students only right now (we have a special 1-month term in January with no classes for competitions like this) but if there's enough interest we would love to hold an open tournament in the next few months.

We're also looking at opening our scrimmage server up to the public this month if there are enough people who want the intellectual fulfillment of competing :)

Edit: Created a small mailing list. Feel free to sign up if you want to be updated when we decide to launch an open tournament (towards the end of the month, most likely).

http://mitpokerbots.com/interest

i'm interested.

Thanks for setting this up - it sounds really awesome. Could you set up a mailing list or something where we can subscribe to updates?

Yep, on it right now. Will post the link in a few minutes.

Edit: It's up, see my above comment.

Will we get a chance to compete against MIT students?

That's mostly what I'm interested in; I'll either get schooled by MIT, or school MIT. Yes this is a counterchallenge. ;)

This looks really interesting to me. I'm curious to see the rules-- does it differ much from the AAAI poker competition [1]?

[1] http://www.computerpokercompetition.org/

Sorry if this is a silly question, but where does the name come from? Googling 6.912 turns up references to "MIT 6.912 Introduction to Copyright Law".

All of our classes are identified using numbers. The 6 corresponds to the sponsoring department (EECS) and 912 happened to be available within the department, so we snagged it.

will the open tournament be limited to python as well? I've got a fair code base invested in F# already laying about.

Indeed, I am definitely interested in participating in the open tournament, but would love to be able to use something other than Python. I don't have any code except for some really old stuff I did for fun, but I would love to keep working on it using Ruby.

If I create an open platform where you can run your own server/bot, would you want to participate? Does this not already exist?

Really interested to hear more about the details, especially the format - will it be a single tournament or a cash game with thousands of hands? Since it's all bots playing, I assume they'll play through enough hands/tables to diminish/negate the role of luck.

Will the bots adopt to patterns in the other bots' playing styles? Most interestingly, will a bot come out of this that can go head-to-head with the likes of Tom Dwan, Phil Ivey and other poker pros? Most bots I've heard of in online poker just follow a set of rules for what to do with certain hands in different positions and scenarios. A bot that can adapt/pick up on betting patterns from the other player (over the course of thousands or tens of thousands of hands) would be really cool.

I would expect most competitors will attempt to design a simplified GTO/balanced (un-exploitable) strategy, with some base assumptions what's considered as "good poker". The only way I can see teams approaching a more exploitive strategy is if the tournament was filled with lots of very weak 'house bots' with highly exploitive strategies (like the previous rock-sissors-paper bot competion that was filled with lots of non-GTO house bots that could be predicted and exploited).

No known bots have come close to playing competitively heads up or ring in big-bet games (such as no-limit or PLO) above the micro/lowstakes. However bots already exist that are very strong at limit heads up.

With enough number of hands played and space for storage of history it should be possible to exploit each other, or at least the opportunity is there.

Certainly possible, and teams can utilise off-the-shelf analytics tools to aid in the analysis (if that's allowed); however using this information beyond very simple adjustments (such as preflop) is very difficult. Also any deviation from a balanced strategy will open up the bot to counter-exploitation.

As I say if dummy players are in there (playing easily exploitable passive or maniacal strategy) then I think there's good reason to implement an exploitive strat; otherwise the effort probably won't have the payoff amongst other static-strategy players.

Game theoretically optimal strategies for poker can only be reasonably approximated so there remains room for even these strategies to be exploited.

I doubt it. Not for another 7-10 years at least to beat the likes of pros.

Sources? I've always assumed that the bots are as good as the pros now even at high stakes. Haven't there been prized competitions where the bots beat the pros some years?

I can't give you a direct source; but there is very little public progress to bots in big bet (NL/PL) games past very simplistic, rule-based bots that may fair okay in nano-stakes games. Of course there an outside chance that powerful bots do exist and are being employed in live games, but there's been no academic work that claims much access.

However in limit games, where the game tree is finite and can be simplified, there's been great strides in poker AI. In heads up the level is expert-to-world class (see my other posts in this thread) and likely very strong in 6-max too. Academically the polaris team (uni of alberta) have published a great deal, and there are commercial/private bots that are considered very strong.

Perhaps I should I have clarified with No Limit. Yes heads up limit poker bots are pretty high standard. But No Limit (and 6+ players) is the real interesting game to solve.

In heads up limit, bots are already at a high class level, maybe world-class. If you want to test this, there's a HU limit bot machine in the Bellagio that's rake free and plays a fair game (fair deal, no knowledge of the cards to come or holecards). It currently spreads up to $10/20. The designers are confident enough to allow anyone to take it on. The bot employs a non-learning strategy.

Sure, but heads up limit has very little strategy compared with tournament no limit, which is what the pros tend to play.

Likewise, I understand Go has been completely solved on a 5x5 board, but Go bots rarely beat professionals even with several stone handicaps.

I beg to differ that HU limit has 'very little strategy'. The game tree is vast and well beyond computational power to solve by brute force.

TV pros may appear to 'tend to play' tournament NL as this is where they get the most exposure, but cash games are certainly where most make their money. HU limit may not get the action of NL/PLO these days, but it's still a popular game.

Nosebleeds games of 1k/2k and 2k/4k has been running a bunch recently, including action from PI, durrr and Antonius.

How about playing duplicate poker to eliminate the luck factor, too? You can even make all bots play all positions.

See http://en.wikipedia.org/wiki/Duplicate_poker

Another, potentially more relevant to HN MIT IAP this year is the web app security Capture the Flag: http://mitctf2011.wikispaces.com/IAP+Announcement

Reply on news.ycombinator.com