Poker Bots

In some games, the only equilibrium is a mixed strategy. However, people are not good at randomizing with a set probability without outside help, especially when the probability is not 50%-50%. Exploiting this and other facts about, and drawbacks of, the way people think and decide, it is possible to design some "ultimate" poker bots that will defeat human gamblers every single time, as is depicted in this article. "The handwriting is on the wall. High quality bots are an online gambler’s worst nightmare." Maybe this kind of machine players will drive human gamblers offline? But it is not in the best interest of owners of online casinos. So what is the best way to implement such machine players? That's another question for which we may turn to game theory for the answer.

It is true that a poker bot playing the optimal strategy for a given preflop hand for a given board can beat human gamblers. But this strategy only works when the poker bot goes unnoticed. If a human gambler thinks the bot is just another player, then it could be difficult to beat it under normal circumstances. If the human gambler can pick up on tells that the player is actually a bot, then it becomes so much simpler to beat.

A poker bot's major weakness is a human gambler's greatest strength: bluffing. How do you program a poker bot to pull of bluffs? You can't because that defeats of having the poker bot in the first place. The idea of a poker bot goes back to one of the first games we played in the class, where some people are going for either the blondes or the brunettes. If you know the other person's strategy, then you can optimize your own and come out ahead. So if you know that the player is indeed a poker bot, then you know every single move the bot is capable of making and when the bot plans to implement it.

The bot was programmed by someone who wants it to implement perfect strategy. But the cards never work out every time, and that is where the human player can chip away at the bot, over and over again by bluffing, because the bot won't always hit the flop.

That depends how you define a "perfect strategy". If you include a perfect strategy to include the occasional bluff, then the poker bot just became a lot harder to spot. Granted, its significantly harder to "teach" a bot to bluff, but its not impossible. Lots of instructional poker books give you basic rules for when you should and shouldn't bluff, and those can be programmed into a bot with enough effort. You would just have to include a bluff as one of the strategies it can play. As long as you're willing to have your poker bot lose the occasional hand with a failed bluff, it's probably worth it in the long run to make a bot that much more unspotable and hopefully lucrative.

no doubt you could make a bot pretty good - say train it to spot when a particular player bluffs, and if you play him enough you can spot his tendencies against you and give a higher probability to call the bluff.

one thing that could totally stop bots though is if one of these
http://en.wikipedia.org/wiki/Captcha
popped up on your screen every few wins or something. might annoy humans, but could be worth it to get rid of the bots.

Also, don't forget that the average player skill on online poker sites, at least from what I've seen, it pretty bad. So bad in fact that I think I'd be falsely identifying bots all the time, since some people are really really bad. Really bad.

In other words, if a poker bot is betting, raising, and calling randomly throughout the game, I'm positive that I would ultimately end up losing money to it because I'd assume that it was an idiot after playing a few hands, and then lose a few monster hands.

And in that case, I'm pretty sure that that is the perfect strategy online: pretend to be an idiot.

I completely agree with Andrew Verrenti's point about adapting strategies.

If you know the other person's strategy, then you can optimize your own and come out ahead.

Yes, if the human knows the other player is a bot, it would be easier to beat it. However, bots also have another advantage, data mining. I'm not sure to what extent people have implemented this, but I do know that some bots mine the hand histories of games and use them for future games. So basically, people would create accounts and join a table as a guest. Then that account would record the communal cards, betting pattern of each player and the winning hand of each game. Bots would then use this information to adapt its own strategies and bluffing style depending on its opponent's history.

I'm not sure how well it works in real life since you would need a ton of data mining for it to be statistically significant.

There is simply no such thing as a perfect strategy in poker. Even if a poker bot bluffs, there can be no perfect strategy because poker is a game of imperfect information. This is why the U of Alberta group focuses on trying to develop the optimum strategy for an imperfect information extensive game. You can find information on their research at http://poker.cs.ualberta.ca/.

That's not to say poker bots are ineffective even if a perfect strategy doesn't exist for the game. Poker bots essentially play on the probabilities, which humans can get incorrect or even simply ignore. In the long run and with a large enough bankroll, a well-programmed bot will most definitely make a profit. However, I would say the experiments of "Man vs Machine" done on the U of A Polaris bot is statistically insignificant. They played four matches against Phil Laak and Ali Eslami in the first contest and six matches against other guys in the second "Man vs Machine" contest. The combined record for the matches is 4-4-2, pretty much dead even. This is as expected because these bots are preying on such small percentages to give them an edge (maybe 52/48 W/L ratio). They are going to have to put the bot through hundreds of matches to even begin to see if their bot will remain profitable indefinitely. Even the best human poker players only make money 20% of the time in tournaments (see http://www.officialpokerrankings.com/).

Side note: James, that is a very good idea to collect data prior to a matchup since people are creatures of habit. I don't believe the U of A group does this, although the Polaris bot learns on the fly against new opponents and uses player's tendencies in their algorithms to solve extensive form games.

If the only thing that's keeping poker bots from making tremendous amounts of money is that people play stupidly and are hard to predict, then any bot programmer could make alot of money by building a bot to detect less advanced bots. If a person catches a bot, there isn't that much they can do, because the bot's betting patterns and strategies may still be too complex for a person to predict. But if you had another bot monitoring it, there's a strong possibility that it could make use of data mining and perfectly pin down the other bot's preprogrammed strategies. It would be perfect information in a game that isn't meant to have it... which would be bad for regular human players.