The Truth about Pluribus Poker AI Donk Bets

Much has been said in the press about Pluribus Poker AI donk bets. In their press release, Facebook says “ Pluribus disagrees with the folk wisdom that donk betting (starting a round by betting when one ended the previous betting round with a call) is a mistake; Pluribus does this far more often than professional humans do. ” So is it the case, how much, and in what specific context ?

Carnegie Mellon University published 10K hands played by Pluribus against the pros. So we can have a deep look at donk betting by Pluribus, but also by the pros. I’ll focus only on single raised pots, ie when there is only an open and a call preflop and no reraise/3bets. 3bet pots have a very different dynamic and there are less of them to look at.

As a reminder a donk bet is a bet where the preflop defender (player who just called) is out of position and leads the betting on the flop. This has traditionally been considered as bad play, hence the name “donk”.

Pluribus was in a position to donk bet 568 times and did it 11 times, so 1.9%. This is twice as much as the pros who did 1% of the times they had the opportunity to, but not huge either. Let’s look at every single one:

Every Pluribus Donk Bet in SRP

  • Strikingly, Pluribus has only donk bet in multiway pots. Interestingly Thomas Pinnock does mention multiway pots as one spot where donk bets “could” make sense as there is less a need to protect the checking range. This is because every player will need to be less agressive on the flop with lower equity.
  • If we’re donk betting, we need to protect that range. Pluribus here does 5 bets with very strong hands and 5 “bluffs” so it looks pretty balanced.
  • Maybe we can use this in live games that have many multiway pots with a decent frequency of value to bluff. We can’t go crazy on frequency, but Pluribus does it almost at 10% frequency (donk on multiway pots).
  • As with c betting, Pluribus varies bet sizes including some overbets.

Pluribus Overbets on the Flop (Poker AI)

Pluribus overbets are probably the most talked about behavior of the Poker AI. Yet we don’t know that much about these and, boy, are they ugly ! In the poker games I play, there are quite a few overbets on the turn with a combination of nuts and draws, overbets on the river with nuts or air. However overbets on the flop are super rare, other than the fish wanting to go all in preflop and as it doesn’t happen, shoves on the flop with his JJ or AKo whatever the board… or overbets on limped boards which well, family pots, who cares? But overbets on a normal heads-up single raised pot, I really see few of them.

Pluribus bets 200% pot on almost 10% of her Cbets. Now 200% is not a number Pluribus gradually optimised based on machine learning training or anything like that. On single raised pot, the researchers only allowed for 3 bet sizes: 50%, 100% or 200%. So the actual size really means nothing.

So in total, on the flop, Pluribus overbets the flop 27 times. In the following chart, I detail the kind of flop (single raised pot, 3bet pot, or limped raised pot), whether Pluribus was the agressor preflop, the position with detail, Pluribus hole cards, the flop cards, comments I could come up with, and range vs range equity for the flop (based on standard upswing ranges, not my estimated Pluribus ranges):

Pluribus Poker AI Overbets
  • What can I add to this ?
  • The majority of Pluribus overbets are IP vs BB , hence with a significant range advantage on most flops giving the opportunity to bluff at a given frequency
  • The range advantage of these specific flops is not particularly better than for random flops at the same positions.
  • The out of position and limped/raise Cbets seem super logical
  • The 2 overbets while defending are just super weird to me, especially the one donking multi way, looks like what the few remaing donk fishes still do in 1-2 live games. Probably just low low percentage AI weirdness.
  • Interesting to note, only 1 of these 27 overbets got a call and none a raise
  • Looking at this, it looks super exploitable to me, like reraise each time Pluribus overbets as the preflop agressor in position in single raised pot.
  • Obviously, I just don’t get it… Not a behavior I’ll add to my game 🙂
  • Tell me what you think, I can’t wait to have someone more competent than me analyse this !

Pluribus Poker Flop Cbet Strategy

We start the study of Pluribus flop strategy by an introduction to its Flop Cbet strategy.

We won’t have enough data to look at strategy by flop texture. But we sure have enough to see in position play and out of position play and bet sizes. There has been a lot of talk about over bet, so we’ll find out.

Pluribus Flop Cbet Strategy Frequencies

  • Cbet percentage is confirming reduction in frequencies of recent years at 52%
  • Cbet in position is 58%
  • Pluribus takes position in consideration a lot
  • Cbet out of position is 41%
  • (Edited) Warning: all 25% bets are on 3bet pot, never used on single raised pot. This is not what Pluribus found to be optimal, it is a strategy of the researchers to limit CPU resource/time on single raised pots flops.

Cbet Strategy Bet sizings

  • So most cbets are 50% pot
  • No cbet at 33% as is current habit of many competent players.
    • Personnally I still love to cbet 33% and follow on turn with 66-150%…. gets money and a fold from marginal hands that beat me which happens so much in my small games…
  • Almost 30% of cbets at 100% or 200% including a very significant 10% cbets at 200% pot.

In the next posts we’ll try to understand when those sizes are used or if it’s just a mixed strategy.