Hacker News Reader

Comment by martinced | Hacker News Reader

Texas Hold'em, contrarily to chess or Go, is a game of imperfect information.

So psychology comes into play and that is basically impossible to "solve" unless we have real AI (not anytime soon and the implication would be way broader than "beating Hold'em").

Fixed-limit Hold'em is easier to play because the decision tree is trivial. It's still an imperfect information game but the best bots can already beat some of the best professional players.

But No-Limit Hold'em is another thing entirely: the decision tree is so huge that it's impossible to bruteforce. Impossible as in: unless a major scientific discovery we'll never ever be able to bruteforce it (there's not enough entropy in the know universe to do so).

So we're stuck and waiting for amazing discoveries in the AI field (and psychology field).

View on HN

"psychology comes into play"

Actually, it does not, just like it does not come into play for Chess, or for Go, or for Othello.

"Fixed-limit Hold'em is easier to play because the decision tree is trivial"

Trivial? Write a bot for limit Hold'em and you'll see that the decision tree is not trivial. This is not Tic-tac-toe; limit hold'em is a fairly challenging game for computers, and it was only recently that computer players could even compete against the best human players.

"No-Limit Hold'em is another thing entirely: the decision tree is so huge that it's impossible to bruteforce"

Yeah, and the Chess game tree is impractical to brute force, and no Chess playing program actually attempts to do such a thing. Likewise with Go, and likewise with limit Hold'em. So what?

In general, the challenge with Hold'em (limit or no-limit) is not just the size of the game tree, but the branching factor. There are only four betting rounds in Texas Hold'em, and so the game tree is fairly shallow compared to Othello, Chess, or Go. On the other hand, just dealing the community cards adds a large number of branches, and in no-limit games the possible bets a player can make adds even more branching. This limits the utility of approaches based on exploring random paths.

This is not a problem of psychology, it is a problem of algorithms and of game theory.

As there's lots of misunderstanding in this thread, just wanted to explicitly +1 this. It's being downvoted, however it is entirely correct.

I'd be interested to see the results of two bots playing against humans.

Bot One just plays the odds.

Bot Two just plays the odd, but has a small chance of bluffing - it has a poor hand but plays as if it has a great hand.

I'd be interested to see the results after several thousand hands. And if people could tell the bots apart from people.

This is, obviously, nothing like AI. But sometimes trivial tricks are powerful. Look at the early discussions about ELIZA for an example.

I don't play poker, but my friend tells me that it's fun to play in tournaments because it's not just odds. You get people who have terrible poker faces, who are easy to read. That same friend also plays (and makes a little bit of money) from online poker. I have no idea how the psychology works there.

With no disrespect, it's clear that you don't play before you mention it.

For instance, "playing the odds" dictates that you bluff. Often. And since this discussion is about HU play, you can change "often" to "constantly."

So here is poker theory in a very, very brief nutshell.

One way to play is to (attempt to) play game theoretically, or more commonly (and less rigerously defined) "balanced". Your entire actions are defined in such that your opponent, even if they knew your strategy, could not exploit your decisions. Bets have the "perfect mix" (defined by pot size, but also other factors) of value hands to bluffs, so that your opponent is indifferent to calling down/rebluffing. This is what you called "not playing the odds" but, contrary to popular belief, involves bluffing (lots of it!), lots of aggression, and in general may look kinda crazy.

The second way is exploitative poker. This is the more common "not playing the odds" type poker, where you play in a way that maximally extracts value, because you think your opponent is not bluffing enough/betting too big with his range, folding too much, etc. The "too much" is basically "deviating from the correct GTO play". However, to exploit this, you yourself now need to deviate from optimal play, thus opening yourself to exploitation.

There is still a debate between which is "better." Obviously exploitive will, at least in theory, lead to higher winrates, but runs the risk of yourself being exploited. If your "read" is wrong, you could be losing. By playing GTO, you assure you never lose (except your share of the rake), and money naturally flows to you through opponents just basically exploiting themselves.

Chess bots don't know the 'perfect' move for most of the game either (only start and end games), that doesn't mean they can't beat human players.

Can you be more specific as to what is so impossible about building a smart betting poker bot?

That was said about Chess vs. Dames. And along came Deep Blue.

Reply on news.ycombinator.com