Extend GPU-CFR to multiplayer and Monte Carlo CFR

Extend the GPU-CFR compiler beyond two-player games to games with more than two players and to sampled Monte Carlo counterfactual regret minimization variants whose traversal is not static.

Background

GPU-CFR compiles a fixed two-player zero-sum perfect-recall game into static dataflow, relying on a known game structure and a repeated, deterministic traversal. The compiler’s current scope therefore excludes settings in which the number of players exceeds two or in which Monte Carlo sampling makes the traversal dynamic rather than fixed in advance.

The paper identifies extending the compiler to these broader settings as an unresolved design problem. Addressing it would broaden GPU-CFR from full-tree tabular CFR to multiplayer and sampled CFR workloads while preserving, if possible, the compilation and accelerator benefits of the static representation.