Theory
CFR (Counterfactual Regret Minimization)
Also called: Counterfactual Regret Minimization · MCCFR · regret minimization
The algorithm behind poker equilibrium strategies: it plays millions of hands and corrects what it “regrets” not having done.
CFR (Counterfactual Regret Minimization) learns by playing against itself. At each decision it measures how much it would have won with each action versus what it won with its current strategy; that difference is the “regret”. On the next pass it plays the actions with more positive regret more often, and the average strategy across all iterations converges to a Nash equilibrium in two-player games.
Walking the whole tree every iteration doesn't scale to NLHE, so there are sampling variants: MCCFR and, within it, external sampling (ES-MCCFR). Even then the game has to be simplified, grouping similar hands and limiting bet sizes to a short menu.
EVolve has its own CFR engine. It computed the heads-up preflop range at 100 big blinds used in the heads-up GTO drill, on a menu of limp, a 3× raise or all-in, and it solves the river in the river GTO drill on your device. The lab shows it learning to bluff in two toy poker games.
Example with cards
With 7♣7♦ in the heads-up small blind, the engine ended up splitting almost evenly between limp (~50%) and raise (~49%). That is CFR at work: if raising had earned more than limping across iterations, regret would have pushed it to raise more often, and vice versa. When both actions earn the same against the opponent, the push balances out and the hand stays mixed.
Practice it
Related terms
- Solver — A program that computes an approximate equilibrium strategy for a specific spot, given ranges, stacks and bet sizes.
- GTO — Game Theory Optimal: an equilibrium strategy your opponent can't exploit, no matter how they play against it.
- Nash equilibrium — A set of strategies in which no player gains anything by changing theirs while the others keep theirs unchanged.