master level bot created?

[Rudbaeck]

[ QUOTE ]
You'd have to know it was a bot, and if you suspected that, then the poker site probably would, too. At least, it'd be in your interest to advise them of the possibility.

[/ QUOTE ]

No, you wouldn't. You'd have to know it was a multitabling SSH style player. Neither human nor bot would run any sophisticated counter-counterstrategies against you.

[Rudbaeck]

[ QUOTE ]
First, I have to ask why is a University making a poker bot? Don't they have some real research to do.


[/ QUOTE ]

Because making a bot for a task with incomplete information in a domain where natural learning is counterproductive is an incredibly interesting field.

Poker is well known, so it's easier to get grants. Which domain the bot actually operates on probably doesn't interest the researchers all that much. But poker is fairly well understood and popular, hence an easy choice.

[Rudbaeck]

[ QUOTE ]
It is *easier* to write a poker program that beats world class opponents in a non-exploitable way than it is to write a bot that beats crappy opponents on a maximum or close to maximum level.

[/ QUOTE ]

No, it's easier to write a bot that beats anyone HU than it is to write a decent ring game bot, as the HU problem can actually be solved.

It's not easier for bot nor human to beat a table full of good players than it is to beat a table full of fish.

[2ndGoat]

poker-academy.com is the commercialization of UofA's poker research work. It includes Poki, Sparbot, Vexbot.

Poki is the old code- rule-based AI
Vexbot is for HU play only, and it supposed to be unbeatable after it gets enough hands (large amount) to calibrate its neural net


You can play against these bots within the poker academy software as well as program your own via their API. You can also go their message board to find bot programmers. The software developers make sure to say there is nothing in the software to allow you to employ your bot in another poker client (like partypoker). However, people are writing their own interfaces and proceeding.

From a brief look at the message boards, it seems using full-scale neural nets just won't cut it for 10-handed games with the current state of computational power. Posters have mentioned that moving to 5 or 6 players would make their programs take days to make a single decision.

So I'm not so worried about poker being "solved" for now. 2 caveats to this statement:
1) I think it is possible to build a winning 15/30 bot with a rule-based approach. I don't know if one has ever been employed.
2) If Moore's law keeps up, I think poker will be solved in my lifetime.

A friend and I are in the exploratory stages of building one for the sake of improving our games. However, I would never use the bot during a real game of poker.

2nd

[KJ o]

[ QUOTE ]
No, it's easier to write a bot that beats anyone HU than it is to write a decent ring game bot, as the HU problem can actually be solved.

[/ QUOTE ]
Duh. That wasn't what we were talking about. (But yes, you are right, HU is easier than full ring.)

[ QUOTE ]
It's not easier for bot nor human to beat a table full of good players than it is to beat a table full of fish.

[/ QUOTE ]

Duh, that wasn't what we were talking about either. There is a difference between beating fish (easy, by the definition of fish) and betaing them for the maximum possible amount (which would require huge amounts of HH and specific learning against particular opponents, as opposed to just good poker).

And why were we discussing this? Because a number of posters indicated that huge amounts of HH or "PT" information would be necessary to build a good bot. Not true.

[Rudbaeck]

[ QUOTE ]
Duh, that wasn't what we were talking about either. There is a difference between beating fish (easy, by the definition of fish) and betaing them for the maximum possible amount (which would require huge amounts of HH and specific learning against particular opponents, as opposed to just good poker).

And why were we discussing this? Because a number of posters indicated that huge amounts of HH or "PT" information would be necessary to build a good bot. Not true.

[/ QUOTE ]

I'd say it's easier to beat fish for near the maximum than to beat a table where Lederer and Harman are in the middle of the field skill wise for any amount at all. I mean, if it were easier to be a tiny favorite in the Bellagio 4k/8k game than a massive favorite in the Party 2/4 games the happy days would soon be here for many of us. [img]/images/graemlins/wink.gif[/img]

A good bot has to adjust to opponents. Does he fold too much on the river? Bluff a bit more. Does he pay off to much on the river. Value bet a bit more. Does he raise preflop with too many hands? Isolate. Does he always slowplay made hands, raising on the turn with those while raising the flop with draws? Adjust, adjust, adjust.

Neither man nor machine can play the same style against every opponent and expect to win any grand sums. Ok, both could probably do it if the rake wasn't so large in the smaller games. But as it is you have to beat the fish in these games for 2BB/100 or so just to break even...

It's obviously easier to beat people if you've already profiled them prior to playing. I added the datamining not with the intention of it being a 'teaching' aid for the bot, but for player profiling. (The very same reason I datamine. If I already got 50+ hands on half the people at the table I obviously have a huge advantage over going in blind.)

A bot capable of winning any significant sum has to profile players. To do so datamining is our best option. Off course this gets updated as the hands from live play pours in.

[The once and future king]

[ QUOTE ]
[ QUOTE ]
Head. Arse.

[/ QUOTE ]

Ah, I see you're in deep.

[/ QUOTE ]

So deap my fringe is tickling your top lip.