Benchmarks & research

Sampling accuracy

Sampling cards saves memory or time, but adds error. These are the measurements behind the sampling choices in PulseGTO.

Three-way flop: reduced turn coverage

When a full flop solve exceeds available RAM, the engine may use fewer turn cards. In one short test, the reduced-card result was close to the full-card result in aggregate, but individual hands differed substantially. Both runs were far from convergence, so the test does not isolate sampling error. For a specific turn outside the solved set, start a solve from that turn. The displayed accuracy for a reduced-card flop does not include the difference from omitted turn cards.

Sampling river cards per iteration

Configuration: the three-way turn from Three-way research, with 12 of 48 river cards sampled per iteration.

With alternating updates, dealing every card reaches a given accuracy faster, so per-iteration card sampling is off.

Sampling in the preflop engine

Preflop training uses sampling, so rarely reached postflop decisions can be noisy after a short run. The headline accuracy number does not capture all of that uncertainty; see Exploitability metrics.