Built on Discounted CFR.
A short look at the engine behind ART/GTO Solver — the algorithm, the convergence profile, and the export formats that let a multi-terabyte library fit on a single external drive.
Discounted CFR, with a vanilla warmup.
The engine implements Discounted CFR (Brown & Sandholm, 2019) in its alternating-updates variant — a refinement of Counterfactual Regret Minimization that applies decaying weights to accumulated regrets and strategies as iterations progress. Same Nash-equilibrium target as vanilla CFR; reached in a fraction of the iterations.
Two practical refinements made the difference on deep postflop trees: a short vanilla-CFR warmup before discounting engages, and a periodic γ-reset that clears noisy early contributions from the running strategy average.
Minutes per spot, not hours.
On a typical desktop CPU (8 cores / 16 threads), a single spot converges to production-grade exploitability in minutes. Batch mode parallelizes solves across a CPU worker pool — four licences means four spots in flight at once.
Same input, same output. Byte for byte.
Same engine version, same config, same seed = identical output. The engine is deterministic from end to end, so the only thing affecting output is the input.