Kuhn Poker against the computer

Play Kuhn poker against the computer: three cards, one bet. Optimal play mixes its moves at random and cannot be beaten on average. Free, no sign-up, in the browser.

What this game is

Kuhn poker is the smallest poker there can be: three cards, one per player, a one-chip ante and a single betting round. Harold Kuhn built the game in 1950 and solved it analytically, on paper — decades before computers took on real poker.

Yet everything big poker is made of lives inside it: the value bet on a strong card, the bluff on a weak one, the awkward decision on the middle card, and the fold that gives up the pot unseen. That is why it is called the atom of poker: nothing smaller exists, and nothing can be thrown out.

Why the machine flips coins

The one thing this game demonstrates: optimal poker has to be random.

In the Nash equilibrium the first player bets J — a pure bluff — exactly one time in three, the second player calls a bet with Q exactly one time in three, and bluffs J after a check one time in three as well. These fractions are not a whim but the solution of a system of equations: shift any of them and the opponent gains a profitable answer. On the equilibrium level the machine draws these probabilities with an honest random number generator, and knowing its strategy does not help: it is printed on this page in plain text, and there is still nothing to exploit.

Every single hand is winnable — with any card there is a line that takes the pot on a lucky day. But on average the seats are almost level: with both sides playing optimally the first player loses exactly 1/18 of a chip per hand. Which is why the match alternates the first move: the second seat’s edge is shared out evenly.

About the levels

Naive plays honestly and never bluffs: it bets and calls only with the king, and checks and folds the queen and the jack. Such an opponent is an open book — its bet always means K, its check always means weakness — and can be exploited without mercy.

Nash equilibrium is the exact mixed strategy with a bluffing rate of 1/3. It cannot be exploited: whatever you do, on average you get no more than the seat is worth. This level cannot be beaten in the long run — you can only avoid losing extra.

Exploiter reads you instead. For the first hands it plays the equilibrium while quietly counting your frequencies: how often your bet is a bluff, how often you fold to a bet. Once it has observations, it leaves the equilibrium exactly where you are skewed: never bluffing means it stops paying off your bets with its middle card; folding too much means it starts bluffing every time. The only defence is to be random yourself, in the right proportions.

From three cards to real poker

The principle scales unchanged. In 2015 the Cepheus program from the University of Alberta essentially solved heads-up limit Texas hold’em — the same way: not by reading the opponent but by approximating the Nash equilibrium so closely that a human lifetime of play could not tell its losses from zero.

The only difference from Kuhn poker is volume: instead of a dozen probabilities that fit in a paragraph, a strategy table of hundreds of gigabytes. The mathematics is the one on this page.

Where to go from here