gvsolver

A real-time solver for No-Limit Hold'em.

Poker is a notoriously hard game to solve: a vast action space, imperfect information, and stochastic outcomes. The standard approach to large-scale imperfect-information games is Counterfactual Regret Minimization (CFR) and its variants.¹ The two landmark results — Libratus² for heads-up no-limit, and Pluribus³ for six-max — both made the cover of Science. Both came out of Carnegie Mellon, built by academics rather than poker professionals. Strong players, in turn, rarely have the formal game theory and engineering background to build solvers of their own.

gvsolver was born of the question: what would the result look like if elite players did have those skills?

gvsolver is a real-time solver for No-Limit Hold'em. It runs on commodity hardware, returns solutions in under five seconds, and produces strategies with sub-1% exploitability against carefully chosen abstractions.

Stack

• CFR engine Vector-form CFR+: regret-matching+ clipping, linearly-weighted averaging, alternating updates. Empirical convergence is roughly O(1/t) versus O(1/√t) for vanilla CFR, and the vector form processes the full hand range at each node in one SIMD pass.
• Blueprint strategy 100TB of pre-computed solutions with hand-tuned tree construction and abstractions.
• Real-time subgame solving At runtime we anchor on the blueprint and re-solve the live subgame with updated gamestate including out-of-distribution actions.
• Vision & OCR layer Custom vision and OCR models, trained per-site, detect the live game state directly from screen capture.

End-to-end, gvsolver detects your live game state, solves it, and returns a sub-1% exploitability strategy in under five seconds, autonomously.

Access

€6,000 per seat per month. Currently no seats open.

Closing

We no longer operate commercially. From the beginning, our motivation was solving poker in real time and reach superhuman performance in the actual most competitive games in the world. We accomplished this mission.

1. Zinkevich, M., Johanson, M., Bowling, M., Piccione, C. (2007). Regret Minimization in Games with Incomplete Information. NeurIPS.
2. Brown, N., Sandholm, T. (2018). Superhuman AI for heads-up no-limit poker: Libratus beats top professionals. Science 359 (6374), 418–424.
3. Brown, N., Sandholm, T. (2019). Superhuman AI for multiplayer poker. Science 365 (6456), 885–890.