The bot does bluff, and in fact it learns from self-play that bluffing is (sometimes) the optimal thing to do. At the end of the day, bluffing is simply betting when you have a weak hand. The bot learns from experience that when it bets with a weak hand, the opponent (another copy of itself) sometimes folds and it makes more money than if it hadn't bet. The bot doesn't view it as deceptive or dishonest. It just views it as the action that makes it the most money.

Of course, a key part of bluffing is getting the probabilities right. You can't always bluff and you can't never bluff, because that would make you too predictable. But our self-play and search algorithms are designed to get those probabilities right.

> when it bets with a weak hand, the opponent (another copy of itself) sometimes folds and it makes more money than if it hadn't bet.

This makes no sense. If I am betting for thin value with a weak hand, then I make less money when my opponent folds. Does the bot not know whether it is bluffing or value betting?

It makes complete sense. There’s a component of value and a component of bluff for a given hand in front of you. They’re related.

Value betting and bluffing aren’t defined by the outcome of a hand — action yet to be completed. Poker is a game of hidden information so betting with “thin value” implies that your component of bluffing is larger. You want your opponent to fold more often than not when you have thin value because more often than not you’re actually beat.

QQ can get KK to fold based upon board texture, street, and prior action. But you don’t know the other person is holding KK when you’re betting for “thin value” on the river.

> You want your opponent to fold more often than not when you have thin value because more often than not you’re actually beat.

No, that is simply not true. If I am betting for value, then I want my opponent to call no matter how weak I am or how thin it is.

> But you don’t know the other person is holding KK when you’re betting for “thin value” on the river.

Then it's a value bet. As you said, it's not defined by the outcome.

“Value betting” and “bluffing” are human heuristics to simplify complicated situations.

The bot doesn’t “know” whether it’s value betting or bluffing—it’s not a relevant question. The relevant question is whether to bet, and what amount, in order to maximize value of the particular hand it has, with reference to the board and opponent actions taken.

Right, we agree on that, but the above comment lumps all of what you describe (“betting with a weak hand”) under “bluffing” and says the bot learns that it makes more money when its opponent folds.

Where does your quote say that the bet is a value-bet? I read it as saying that the bot learned to bluff (not value bet) by betting when it has a weak hand (I.e. The bot has a weak hand, so it's getting better hands to fold by betting). The phrase "value bet"was not used.

(This, in addition to what the other comments have said about there being spots where a bet can get better hands to fold with some probability AND get worse hands to call with some probability - see the chapter "The grey area between value betting and bluffing" in Applications of No Limit Hold Em)

"At the end of the day, bluffing is simply betting when you have a weak hand."

I was the one who introduced the term "value betting" to the conversation, applied specifically to weak hands.

I mean, unless only those who interpret it wrong would respond, then I must be the one reading it wrong. Because these responses aren't lining up with how I read it or what I meant.

At the highest levels of play psychological factors are pretty minimal. Before a showdown which cards you actually hold aren't particularly material, as the only information you convey is through your bids. This means if you predict that you're more likely to win a hand by bidding (and inducing a fold) than by calling and going to a showdown it makes mathematical sense to "bluff". I'm sure AIs have no trouble learning that fact.

The issue is that you don't know exactly the probability of your opponent folding.

This is psychology.

The probability of the opponent folding doesn't matter. The goal of bluffing in modern games is so that optimal players are indifferent in their decision (no matter how they play, you can't lose money). And because this is a zero sum game, if you can't lose money then you win if the opponent makes mistake.

You only need to know the probability of the opponent folding so that you can deviate from the theoretical optimal strategy to win even more money if they are a biased player

I'll have to go back and watch Data playing poker on Star Trek NG -- what do sci fi writers think of this.