Research
CFR without Unbiasedness: Deterministic Guarantees for Persistent Public-Chance Schedules
Deterministic convergence guarantees and exploitability certificates for persistent partial public-chance evaluation.
Central Question
When can persistent, conditionally biased public-chance feedback still give valid Counterfactual Regret Minimization guarantees?
Main Contribution
The paper establishes a deterministic target-transfer theorem for uniform, nonnested additive public cuts. The result combines delivered-feedback regret with a public-debit term that tracks coverage discrepancy and strategy motion.
Consecutively balanced schedules converge for additive signed regret matching and RM+ under predetermined averaging weights. A component-resolved form converts an execution trace into a numerical exploitability certificate.
On two heads-up no-limit hold’em turn endgames, persistent order improves over fresh reshuffling. The depth study identifies a crossover between 32 and 64 full-cut outcome budgets.
Citation
[1] J. Guo and L. Ye, “CFR without Unbiasedness: Deterministic Guarantees for Persistent Public-Chance Schedules,” arXiv:2608.14761, 2026, doi: 10.48550/arXiv.2608.14761.