PulseGTO
Solver engines
Engines at a glance
PulseGTO has one engine for whole Hold'em games, two for Hold'em postflop spots, and the PLO engine, which does both for pot-limit Omaha. They answer different questions, so their accuracy numbers cannot be compared directly.
Looking for the desktop software? Explore the heads-up and three-way postflop solver for your own ranges and bet sizes, or the preflop solver for custom game trees to train a whole-game blueprint. For measured comparisons, see the PulseGTO vs PioSOLVER benchmark: accuracy, time and memory.
| Preflop engine | RangeCFR (HU) | RangeCFR (3-way) | PLO engine | |
|---|---|---|---|---|
| Purpose | Train a whole-game blueprint | Postflop spot without hand abstraction, 2 players | Postflop spot from your own ranges or a blueprint node, without hand abstraction, 3 players | Pot-limit Omaha: a whole-game 6-max blueprint and heads-up postflop solves |
| Players | 2–6 | 2 | 2–3 | 2–6 in the blueprint, 2 in postflop solves |
| Streets | Preflop to river | Flop, turn, river | Flop, turn, river | Preflop to river |
| Hands | Preflop hand classes; postflop abstraction | Every combo | Every combo | 16,432 preflop hand classes; every hand in postflop solves |
| Board cards | Covered across training | Remaining cards covered | Full coverage where RAM permits; reduced flop coverage when needed | Covered across training; every card in turn and river re-solves |
| Opponent actions | Sampled | All | All | Sampled in training; all in postflop solves |
| Accuracy reported | Regret bound (convergence in the abstraction) | Exploitability, full best response | Average best-response gain per player (on a sampled flop: over the sampled turns) | Training deals and samples per hand class; exploitability for postflop solves |
| In the app | Preflop Settings | Run, HU mode, Recalculate | HU mode card (3 players), Recalculate on a 3-player node | PLO mode: Preflop Settings, Blueprint, Run, Re-solve, Change node |
Terms used on these pages
- No hand abstraction: every hand combination has its own strategy. RangeCFR works this way; blueprint postflop play does not.
- Hand abstraction: a lower-detail representation that makes whole-game training practical, while limiting postflop precision.
- Exact card removal: every combo is tracked with its two cards, so blockers between the players' hands and the board are applied exactly, not approximated.
- Exact payoffs: the value of each showdown or fold is computed exactly (pot, side pots, rake, split pots), not estimated. By default, rake is taken from the matched pot: when a hand ends with a fold, the unmatched part of the last bet goes back to the bettor and is not raked. A blueprint configured with Rake uncalled bets uses the full pot on a fold, including the uncalled bet; its re-solves retain that rule.
- Full enumeration of future cards: every remaining turn and river card is covered in the configured solve. Three-way flop solves do this by default when the tree fits in RAM.
- Card sampling: only some chance cards are dealt, either per iteration or for the whole solve. The answer then has sampling error on top of convergence error.
- Action sampling: some training paths are sampled to reduce cost. Individual iterations can therefore be noisy.
- Convergence: CFR-type solvers approach an equilibrium over iterations; after a finite number of iterations the result is an approximate equilibrium of the game being solved. For a blueprint, that game is the abstraction.
- Exploitability: how much a best response could win against the strategy, in % of the pot or bb/100. It measures the distance from equilibrium of the game being solved. Zero would mean an equilibrium of that game; in practice it is small but not zero.
- EV error: how far a displayed EV is from the value in a more accurate solution. A strategy can have low measured convergence error while its displayed blueprint EVs remain approximate (see Reading EV).
- The action tree: every engine solves only the bet sizes and raise levels in its tree. Sizes that are not in the tree are never considered, however accurate the solve.
- Blueprint accuracy: convergence plus the error of the abstraction itself (see Accuracy & limitations).