Week 12: Capstone: how much of good poker intuition can be formalized

Grading the semester's own four promises, solving the real Kuhn poker equilibrium and checking it against the published result, and naming what stayed outside every tool this course built

10 May 2027

The question this course opened with

Week 1 asked how much of good poker intuition can be turned from a slogan into a computation — “a stated assumption, a derivation or enumeration from that assumption, and a number at the end that could in principle be wrong.” Four folk claims were catalogued there specifically so this page could be honest about how each one actually landed.

A scorecard, not a highlight reel

A method too small to matter, and one too big to do by hand

Week 7 solved a toy game with two actions per player by writing down two indifference equations and solving them on paper. Real Kuhn poker — the smallest fully solved poker game, named in week 7 but deliberately not derived there — has twelve information sets, each a hero-or-villain decision with its own mix of check/bet or call/fold. Writing that system of indifference equations out by hand is possible but not something to do live in a lecture. What scales is the idea behind week 7’s method, not the by-hand algebra: regret minimization — at every decision point, track how much better each action would have done than what was actually played, and shift probability toward whatever has been regretted most. Run for enough iterations, this converges to the same kind of indifference week 7 solved directly.

Solving the real Kuhn poker equilibrium

Kuhn poker (Kuhn, 1950) deals one of three ranks to each of two players from a three-card deck; a single round of check/bet, call/fold decides the hand. Its published equilibrium (Kuhn poker, Wikipedia) is a one-parameter family indexed by α ∈ [0, 1/3] — the lowest card bets (bluffs) with probability α, the highest card bets with probability 3α and always calls, the middle card never opens but calls a bet with probability 1/3 + α — with a fixed game value: the first player loses exactly −1/18 ≈ −0.0556 per hand at equilibrium, regardless of which α is played.

Running a from-scratch regret-minimization solver — the same idea as week 7’s indifference equations, automated — for 300,000 iterations over Kuhn poker’s actual game tree gives a game value of −0.05555, matching the published −1/18 to within 0.0001, and reproduces every piece of the published strategy: the lowest card never calls after checking and facing a bet, the highest card always calls, the middle card never opens betting and calls roughly a third of the time facing one, exactly as the cited result describes. The gap between “solved by hand” and “solved by a general method, checked against what’s already known” is close to zero here — the method is trustworthy precisely because it was checked against a case where the real answer was already published.

A bigger toy game nobody has published

The same solver, unchanged, run on a four-rank extension of Kuhn poker — a game no textbook states the equilibrium for — converges to a stable game value of −0.04167 (curiously close to −1/24, though this course doesn’t attempt to prove that pattern holds for a general number of ranks — flagging the observation rather than claiming the theorem). The resulting strategy is polarized: both the lowest and the highest rank open betting more often than either of the two middle ranks, which mostly check — the same “bet the very strong and the very weak, check the middle” shape week 7’s toy game assumed as its starting setup, except here it emerged from the solver rather than being assumed.

What no tool this semester captured

Every model this course built assumed a fixed, known, or at worst a Bayesian-updatable opponent. Three things a real table has that none of that captures: table talk and physical tells (information no probability distribution over hole cards was ever going to represent), opponent non-stationarity (week 8’s whole distinction between GTO and exploitative play exists because a real opponent’s tendencies drift mid-session in a way a fixed frequency x never does), and the meta-game across an entire session or tournament (a player’s image, table history, and what opponents remember about earlier hands shape a decision this course only ever modeled one hand at a time).

Real players versus solvers

A solver returns the equilibrium of whatever game was handed to it — the answer is only as good as the abstraction. A real player is working with a worse model of the game (nobody carries twelve exact information sets’ worth of mixed frequencies to the table) but a better model of the specific opponent sitting across from them, updated continuously the way week 6 formalized. Solvers win at removing the possibility of being exploited; they cannot, by construction, exploit a specific opponent more than the equilibrium already does. That trade is precisely what week 8 made numeric.

Answering week 1’s question

How much of good poker intuition can be formalized: a lot of the probability side completely (weeks 1–4, 9–10), a fair amount of the strategic side once the game is simplified enough (weeks 5–8, and this week’s Kuhn poker solve), and close to none of the psychological side — not because it’s unimportant, but because it was never this course’s subject. The folk claims that survived best were the ones that were already, secretly, claims about numbers. The ones that survived worst were about people.

Teaching team