Hacker News Reader

Comment by eries | Hacker News Reader

Can you share some about what strategies the bot prefers and how these compare with common professional human strategies?

View on HN

We talk about this a bit in the paper. Based on the feedback from the pros, the bot seems to "donk bet" (call and then bet on the next round) much more than human pros do. It also randomizes between multiple bet sizes, including very large bet sizes, while humans stick to just one or two sizes depending on the situation.

When human players donk bet it's almost always a weak player employing an extremely exploitable strategy, whereas pros almost never do it because the metagame has evolved around the presumption that nobody ever donk bets. I'd love to see what the bot's balanced GTO donking strategy looks like.

Donk bets exist in the meta, ie when the turn is extremely good for your range but is horrible for your opponent. ( if you have a fd on the flop and it hits on the turn you can overbet the pot on the turn with your bluffs and foushes then just go all in on the river) if they have top pair its pretty hard to play against that

Oh, sure. I more meant flop donk bets; I guess it doesn't specify which street the donking was happening.

The same logic can apply to flop donk bets. Some flops favor the donking player's range more than their opponent.

Yea I'm not saying it's impossible to devise an unexploitable flop donking strategy. I think the reason thinking players generally don't is because of the complexity of adding significantly more branches early in the game tree - basically going from 3 (check-{fold,call,raise}) to 6 (those 3 plus donk-{fold,call,raise}).

The other issue is that increasing the number of branches also decreases the number of hands that go into each bucket, to the point where it might not be effective any more without being able to randomize the branch choice for specific threshold hands. Most pros I know just have a hard cutoff for each branch and don't worry too much if they're slightly out of balance, but smaller bucket sizes could magnify errors. If you have 31 combos for one action when you're supposed to have 30.5, then whatever, but if you have 6 when it should be 5.5, that could become a problem faster.

It's basically been true along every step of the the poker bot evolution (HU limit, HU NL, and 6-max NL) that the bots donk a lot more than the humans. 10 years ago you could find pros arguing that donking in any situation is always wrong. That's been shifting for years, but still not to the level that the bots do it.

My personal belief is that the "no-donk" strategy is an adaptation by fallible human minds to reduce the branching on the decision tree to something tractable.

Your personal belief is likely correct. Balancing a donking range is incredibly difficult for humans and doing so perfectly likely yields only a very small EV bonus over just always checking. For humans it makes a whole lot of sense to reduce the branching in a case like that whereas for computers it doesn't really matter.

Another good example is varying continuation betting sizes. A true GTO strategy would mix in a number of different sizings (and I'm sure the bots adapted to do this), but you only sacrifice a very tiny amount of EV by basically betting the same size every time. Doing the latter limits humans risk for making errors which is far more valuable than squeezing out .05bb/100 more by varying the sizes.

If true in cash games, it is funny since it is a not uncommon strategy in high-level tournament play to control pot size.

Is there a way to see the EV the bot is calculating when it's deciding between checking and donk betting? When you place these spots in solvers, they actually advocate for a significant amount of donk betting on certain boards, but pros don't do it because the EV is marginal and it's better for pros to simplify their strategy so they make less mistakes. If you have a flop donk bet strategy, you also have to develop a corresponding turn and river strategy, which makes it extremely difficult.

Reply on news.ycombinator.com