Tuesday, March 09, 2010

Lee Jones, heretic

In the March issue of Bluff magazine, Lee Jones has a thought-provoking article about how online poker rules have lagged behind real-world considerations. He provides several suggestions for how online poker sites should deal with rules problems.

At the very end of the piece comes what was for me an entirely surprising idea:

The following is going to sound heretical, but stick with me. One or more
online poker sites may choose to embrace bots. They could set up an Application
Programming Interface (API) which allows a bot-developer to have his program
talk directly to the poker server at that site. That site might even make its
hand history database available for review, download, etc. I promise you that
such a site would be fascinating to watch. Computer science departments in
universities the world around would fall over themselves trying to develop the
best poker bot. Hackers in tiny apartments would compete byte-for-byte with the
computer science departments. It's hard to overestimate the academic
breakthroughs that might come out of such competition. Create the same profit
motive available to automated stock and optoins traders, and we'll see poker
played in ways that we never thought possible.
I think this is a great idea. As long as people know that they are or may be playing against a bot, why not let them?

It seems to me that what would happen is approximately this: The bots would start off playing pretty badly, as at present. (I'm talking no-limit, full-ring games here.) Good players would spot the programs' weaknesses, exploit them, and win money. The programmers, in turn, would have to figure out what holes in their games had been their downfall, patch them, and play with an improved version. Presumably the improved version could beat some players to whom it had lost before, thus generating revenue for continued investment in research and development and/or allowing the bot to have enough bankroll to play at higher stakes. Again, though, the best players (attracted by the bots now playing at stakes that interest them) would pick apart what the programs are doing, find weaknesses, and pounce on them. This arms race would continue until the bots are good enough that they can't be beat by humans.

Of course, as Jones says, at the same time the bots would be competing directly against each other. It seems likely that at first it would be the best human players driving the development, but eventually, one has to believe, the best bots will be competing only against each other, because no human will be willing to challenge them--at least not for more than occasional recreation.

What could we learn about optimal poker from such a competition? I have no idea, but I would love to zip ahead in my time machine about 20 years and see what has come of it.


Neil said...

Poker Bots rule the world and use humans as their "fuel", bacause we blackened the sky........wait a minute, wasnt this already done?

The Poker Meister said...

Unless there are major advances in computer hardware, you're going in an artificial intelligence direction. My background is computer science - and I can tell you that computers are black and white "thinkers." They think in terms of if / then / else, with "random" inter-spliced. While I do think that software can be built to excel at poker, I don't believe that computers, as we know them today, could ever excel at NLHE. Bet sizing represents far too many variables.

However, a bot can be written to excel at LHE fairly easily. LHE, for the most part, is an algorithm game. There is no "implied odds," for the most part. There is no "all in shove" to deter a player from drawing. The computer has 4 simple options: bet / fold, check / call. The computer can quickly compute his equity in a particular situation, as well as his odds of winning as his hand stands, and determine a proper course of action.

As far as NLHE, a computer needs to constantly "learn" and "adapt" to "random" human plays - more random than LHE - where a human can vary bet sizing.

Sorry for the rambling reply, but I've done some critical thought in the past regarding this topic. Univeristy of Alberta (I think) in Canada has a poker bot that played HU against a few pros and I think lost. There are a few non-accessible-from-work articles on it.

Butch Howard said...

Counter to The Poker Meister, and based on my 30 years of computer programming (call it engineering or science if you like) experience, I expect it to be a certainty that the bots will play exceedingly well at NLHE. I expect it will be on the order of 3-5 years after sites are opened up that the bots will be unbeatable by humans. Bill Chen's "Mathematics...", the CPRG work (even if LHE, the extent of the work on hand analysis and other parts still appy), and few other works point easily in that direction. The key will be that a decent bot will not have to completely 'solve' NLHE and the human-like element to be strong and profitable, especially at the beginning of the arms race. Even a modest poker heuristic combined with the ability to process huge hand histories, apply accurate statistics and probabilities on a per-opponent basis, and do these things faster and more consistently than humans is why am confident that computers will meet this prediction. Mix all of that in with the cash incentives of making money off of the bot-on-bot and bot-on-human battles for even greater likelihood of rapid improvements of the bots. The bot-on-bot wars alone will probably keep the market alive, but there will certainly be humans that believe, rightly or wrongly, that they have an edge over a lowly machine and will throw money at the market trying to prove it.