The Range Reader, Part 8: The Arena Deals the Data

Part 8 of The Range Reader, a series on teaching a computer to read your mind at poker.

Machine learning starts with a dataset

Part 7 left the solver half-blind. It plays the hand well, then deals the opponents their cards at random. The range is a guess, and every value built on it inherits the error.

We want to learn the range instead. A model learns from data, so the data comes first.

Karpathy teaches on tinyshakespeare, a megabyte of the plays in one file (char-rnn). The corpus stays dumb on purpose, so the pipeline around it can prove itself: wired end-to-end on data you cannot botch, watched until it overfits a batch of two, then scaled (the recipe). Don’t be a hero.

So the first dataset here is rough by design. It proves the plumbing. Nothing more.

What one example is

Each example is an input and a label. The input is the public action - every seat, every bet, in order. The label is the two hole cards it was hiding.

The model learns P(hand | betting): a probability for each of the 1,326 possible hands, given the betting.

The solver from Part 7 is unchanged. It still solves for a strategy no counter can beat. A sharper prior - the starting guess over the opponent’s hands - gives it a truer range than uniform-random. The input gets better, and the play stays unexploitable.

v0.1: let the arena deal

The arena is the dealer, and it already writes what it deals (rs-poker). Seat the agents, press go, read the file.

Every hand comes out as Open Hand History, the standard JSON format for a poker hand. A real hand history shows cards only at showdown. The arena sees every seat, so it writes every hole card - folded, mucked, or shown. The label is free, and it is on all of them.

The field is the one from Part 3. RandomPotControl is the recreational player, the fish, betting without a plan. Three CFR agents each wear a different preflop chart - which hands to play from each seat - lifted from online solves: Pekarstas-6max-RFI, GTO-Experiment, and 6Max-RFI-GTO. One more, CFR-Configurable, runs no chart at all, just a real-time budget. Tables run two to six seats in an even split, so position and player count both move.

The result is public: otter-crew/range-reader-v0.1 on Hugging Face. 613,399 hands, 1.42 GB, one JSON hand per line.

What v0.1 gets wrong

The play over-shoves. One hand in six goes all-in. More than half never leave preflop. The agents trusted a read they had not earned.

They carried the crude estimator from Part 7 - opponents dealt uniformly at random - and committed as if the guess were fact.

There is a way to hold a real range without a neural net. Libratus (Brown & Sandholm, 2018) and Pluribus (Brown & Sandholm, 2019) reason over a precomputed blueprint: an MCCFR strategy (Lanctot et al., 2009) - Monte Carlo CFR - solved from self-play over an abstraction of the game, a shrunk and clustered copy of it. The blueprint gives every hand a probability, so a solver reading it holds a real range at any node instead of a point guess.

v0.1 has none of that. So the data leans loose and shove-happy. A known bias, written down.

That buys a working pipeline. The ranges get better from here.

Next time

Part 9 is the model. Predicting hole cards from betting is a translation problem: read one sequence, emit another you can read.