You may be familiar with the no-limit clairvoyance game from Mathematics of Poker. There are two players, an initial pot size of 1, and stacks of n. Player 1 is dealt 50/50 a winning and losing hand, and player 2 always has a bluffcatcher. Player 1 is allowed to bet any size up to n (or to check), then player 2 can call or fold (but not raise). They show that the NE is for P1 to always bet n (all-in) with a winning hand, and bet n with probability n/(1+n) with a losing hand (and check otherwise). And vs. any bet size x in (0,n] P2 calls with probability 1/(1+x).
It turns out that the analysis is actually a bit more nuanced. For suboptimal bet sizes x < n, player 2 can actually call with a range of frequencies while still being in equilibrium. If P2 calls with any probability in the range
[1/(1+x), min{n/(x(1+n)), 1}] then it is a Nash equilibrium. (The lower bound makes P1 indifferent between checking and betting x with losing hands, and the upper bound makes player 1 indifferent between betting n and betting x with a winning hand.)
So the question arises, if you are player 2 and facing a suboptimal bet size of x from player 1, how often should you call? Since there are infinitely many values in the equilibrium range, standard Nash equilibrium analysis doesn't give a conclusive answer. Obviously the answer doesn't matter from the perspective of a worst-case EV guarantee, since all strategies are NE. But against opponents who actually make suboptimal bet sizes it could make a big difference.
I've been thinking about the question for a long time, and I haven't been very satisfied with the solutions of several standard Nash equilibrium "refinement" concepts, which are game-theory solutions that select certain Nash equilibria over others.
To make this concrete, suppose that n = 2 (both players have 2*pot behind), and let x = 1 (player 1 makes a suboptimal pot-sized bet). Then player 2 can call any probability in the range [1/2,2/3] and be in equilibrium. But in practice is any one of these strategies better than the others?
Obviously to properly answer the question you'd need to think about reads on the particular opponent, population tendencies, maybe do experiments and/or look at historical data, etc. But I am also interested in a theoretically-motivated approach that is not totally reliant on this information (so that it can be also used for other similar non-poker situations).
So my question for you is what probability do you choose to call with in the game for n = 2 facing a bet of x = 1 and why?
If you want to see a more detailed discussion and analysis, I have a recent paper here (Section 3 is about the poker example specifically)
https://arxiv.org/pdf/2210.16506.pdf
To summarize, the prior main equilibrium refinement concepts say to call with prob 2/3, while a new one I came up with that I think is better motivated theoretically says to call with prob. 5/9. Though my intuition as a poker player says to call with prob 1/2.
Last edited by Sam Ganzfried; 12-19-2022 at 02:21 AM.