How Poker Solvers Work: The CFR+ Algorithm Behind GTO Decisions
Every time your GTO strategy software recommends a raise, a call, or a fold, there is a mathematical engine running behind the scenes. That engine is built on Counterfactual Regret Minimization — known as CFR+ — the same algorithm that powers virtually every modern poker solver. Understanding how CFR+ works gives you insight into why GTO software produces the decisions it does, and why those decisions converge on Nash Equilibrium.
What Is Counterfactual Regret?
In game theory, "regret" measures how much better you would have done by choosing a different action at a specific decision point. Counterfactual regret extends this concept — it asks how much you would have gained if you had chosen action X instead of action Y, assuming your opponent played the same way. CFR+ iterates through millions of hypothetical poker scenarios, tracking regret for every possible action at every decision node. Over thousands of iterations, it gradually shifts strategy toward actions with lower cumulative regret. The result is a strategy profile that no opponent can exploit — the Nash Equilibrium.
From Iterations to Equilibrium
Each CFR+ iteration simulates a full hand of poker from every possible perspective. The algorithm evaluates every combination of hole cards, board runouts, and opponent actions, then updates regret values accordingly. Early iterations produce wildly unbalanced strategies, but as the number of iterations grows — typically into the tens of billions — the strategy converges toward a mathematically optimal solution. Professional desktop solvers like PioSolver or GTO+ require minutes or hours to converge on a single spot. Our GTO strategy software pre-computes these solutions across a massive database and augments them with real-time neural network inference, delivering results in under 200 milliseconds during live play.
Why Neural Networks Accelerate GTO
Running full CFR+ calculations during a live poker hand is computationally impossible — the game tree for No-Limit Hold'em contains approximately 10 to the power of 160 decision nodes. Instead, GTO strategy software trains neural networks on the outputs of solved scenarios. The network learns to approximate solver-level decisions from raw game state inputs: your hand, the board, pot size, stack depths, and opponent behavior. This approach trades a tiny margin of precision (typically under 2% exploitability) for enormous speed gains — making real-time GTO play practical across multiple simultaneous tables.
The Practical Edge
For the player, the technical distinction between CFR+ solving and neural network approximation is irrelevant at the table — both produce actions that are functionally optimal. What matters is the outcome: consistent, emotionless decision-making at a level that would require thousands of hours of study to replicate manually. GTO strategy software condenses the output of a solver farm into a single application that runs alongside your poker client and tells you exactly what to do, every hand, every street.
Frequently Asked Questions
Nash Equilibrium is a state where no player can improve their expected value by unilaterally changing their strategy. In poker terms, a GTO strategy at Nash Equilibrium guarantees a non-negative winrate against any opponent — no counter-strategy can exploit it. CFR+ is the algorithm that computes this equilibrium point.
Neural network approximations used in modern GTO software typically achieve under 2% exploitability compared to the exact solver output. In practical terms, this means the recommended action matches the fully solved action in approximately 94-97% of scenarios — a margin that is invisible at the table and far exceeds human accuracy.
Desktop solvers require you to manually input the game scenario, wait for convergence (often several minutes per spot), and then interpret the output. This process is impossible during a live hand with a time bank of 15-30 seconds. GTO strategy software automates the entire pipeline — screen reading, calculation, and recommendation display — in real-time, with no manual input required.