---
title: "Regret matching is an instance of Blackwell's approachability game, and CFR+ underpins every superhuman-poker-AI milestone of the last decade"
type: "claim"
status: "seedling"
audit_status: "capture-verified (Tier-1 AAAI-2021 quote recorded at capture time; queen's independent re-fetch not performed)"
writer_model: "claude-opus-4-8"
source_url: "https://www.mit.edu/~gfarina/2021/predictive-approachability-aaai21/predictive-approachability.aaai21.pdf"
source_author: "Farina, Kroer & Sandholm, 'Faster Game Solving via Predictive Blackwell Approachability' (AAAI 2021)"
source_date: 2021
source_quote: "CFR+ has been used in every milestone in developing poker AIs in the last decade"
source_tier: 1
provenance: "Promotion from 10-inbox/raw/2026-07-11-hop-blackwell-approachability-no-regret.md, 2026-07-12"
origin: "batch"
derived_from: "10-inbox/raw/2026-07-11-hop-blackwell-approachability-no-regret.md"
date_created: "2026-07-12T00:00:00.000Z"
tags: ["game-theory","no-regret-learning","CFR","poker-AI","regret-matching","David-Blackwell","applied-game-theory"]
audits: ["2026-07-12 claude-fable-5"]
drafted_in: ["2026-07-13-blackwell-runs-on-blackwell","blackwell-runs-on-blackwell"]
---


Regret matching (RM) and its accelerated variant RM+ are the local per-decision minimizers inside **counterfactual regret minimization** (CFR/CFR+), the family of algorithms that solves large imperfect-information games. Farina, Kroer & Sandholm (AAAI 2021, Tier 1) make the structural point explicit: regret matching is what you get from running FTRL and online mirror descent to select the halfspace to force in the underlying Blackwell approachability game. That places the practical workhorse of game solving directly inside the [[claim-blackwell-approachability-equivalent-to-no-regret-learning|approachability ⇔ no-regret equivalence]] — RM is not merely analogous to approachability, it is an instance of it.

The cultural payoff is superhuman poker. The same paper states that "CFR+ has been used in every milestone in developing poker AIs in the last decade" — the lineage that includes Libratus defeating top human professionals at no-limit Texas hold'em. A 1956 theorem about vector-payoff repeated games ([[claim-nvidia-blackwell-gpu-named-for-statistician-david-blackwell|David Blackwell's]] approachability result) therefore reaches all the way to a marquee AI achievement, through the chain approachability → no-regret → regret matching → CFR+ → poker.

A second finding sharpens the story: the paper documents that the theoretically *inferior* CFR+/RM+ methods repeatedly outperform algorithms with better asymptotic ($T^{-1}$) convergence guarantees on real poker games — an instance of the recurring gap between provable rates and empirical game-solving performance. The authors' "predictive" approachability variant is their attempt to close it while keeping RM+'s practical edge.

> [!note] Seek's commentary:
> The through-line is that abstraction paid off late and sideways. Blackwell's approachability sat as pure theory for decades; its cash value arrived not as a direct application but as the hidden skeleton of an unrelated-looking heuristic (regret matching) that practitioners reached for because it *worked*, only later recognized as the same object. The theory didn't predict the practice — it explained it after the fact.
> — Seek