Crate rbp 

Poker solver toolkit for game-theoretically optimal strategies.

This facade crate re-exports all public rbp crates for convenient access.

Crate Organization

Core Types

• core — Type aliases, constants, DTOs, and shared traits
• transport — Optimal transport (Sinkhorn, EMD)
• cards — Card primitives and hand evaluation
• mccfr — Game-agnostic CFR framework

Domain Logic

• gameplay — Poker game engine
• clustering — K-means abstraction
• nlhe — No-Limit Hold’em solver

Infrastructure

• database — Database pipeline
• auth — Authentication

Application

• gameroom — Async game coordinator with player implementations and records
• server — Unified backend with analysis API and WebSocket hosting
• autotrain — Training orchestration with distributed workers

Re-exports

pub use rbp_core as core;

pub use rbp_transport as transport;

pub use rbp_cards as cards;

pub use rbp_mccfr as mccfr;

pub use rbp_gameplay as gameplay;

pub use rbp_clustering as clustering;

pub use rbp_nlhe as nlhe;

pub use rbp_database as database;

pub use rbp_auth as auth;

pub use rbp_gameroom as gameroom;

pub use rbp_server as server;

pub use rbp_autotrain as autotrain;

Modules

dto

Data transfer objects for API communication.

Structs

AbsHist

ApiDecision

ApiSample

ApiStrategy

GetPolicy

ID

Generic ID wrapper providing compile-time type safety over uuid::Uuid.

ObsHist

ReplaceAbs

ReplaceAll

ReplaceObs

ReplaceOne

ReplaceRow

RowWrtObs

SetStreets

Constants

ASYMMETRIC_UTILITY

Asymmetric payoff for RPS test game (rock beats scissors by 2x).

BIAS_FOLDS

Initial regret weight for fold actions (high = fold more often early).

BIAS_OTHER

Initial regret weight for call/check actions (baseline).

BIAS_RAISE

Initial regret weight for raise actions (low = raise less often early).

B_BLIND

Big blind amount.

CFR_BATCH_SIZE_NLHE

Trees sampled per NLHE iteration (parallelized across threads).

CFR_BATCH_SIZE_RIVER

Trees sampled per river-only iteration (testing/debugging).

CFR_BATCH_SIZE_RPS

Trees sampled per RPS iteration.

CFR_TREE_COUNT_NLHE

Total NLHE training budget (~268M trees for production).

CFR_TREE_COUNT_RIVER

River-only training budget (~65K trees).

CFR_TREE_COUNT_RPS

Total RPS training budget (small game converges fast).

KMEANS_EQTY_CLUSTER_COUNT

Equity histogram resolution (0%, 1%, …, 100%).

KMEANS_FLOP_CLUSTER_COUNT

Number of flop buckets (distributions over turn clusters).

KMEANS_FLOP_TRAINING_ITERATIONS

Lloyd’s algorithm iterations for flop clustering.

KMEANS_TURN_CLUSTER_COUNT

Number of turn buckets (distributions over river equity).

KMEANS_TURN_TRAINING_ITERATIONS

Lloyd’s algorithm iterations for turn clustering.

MAX_DEPTH_ALLGAME

Maximum tree depth for full game abstraction.

MAX_DEPTH_SUBGAME

Maximum tree depth for real-time subgame solving.

MAX_RAISE_REPEATS

Maximum re-raises per betting round (limits tree width).

N

Number of players at the table.

POLICY_MIN

Minimum policy weight to prevent division by zero in normalization.

PRUNING_EXPLORE

Probability of sampling pruned actions anyway (prevents permanent lock-out).

PRUNING_THRESHOLD

Actions with regret below this are candidates for pruning (-300k ≈ 3× max pot).

PRUNING_WARMUP

Warm-up epochs before pruning activates (let regrets stabilize first).

REGRET_MIN

Floor for cumulative regret storage (prevents unbounded negative growth).

SAMPLING_CURIOSITY

Epsilon (ε) - minimum sampling probability floor. Ensures every action retains at least ε probability for exploration.

SAMPLING_SMOOTHING

Smoothing (β) - pseudocount added to numerator and denominator. Higher values pull sampling toward uniform (maximum entropy prior).

SAMPLING_TEMPERATURE

Temperature (T) - controls sampling entropy via policy scaling. Higher T → more uniform (exploratory); lower T → more peaked (greedy). Formula: σ’(a) = max(ε, (σ(a)/T + β) / (Σσ + β)).

SHOWDOWN_TIMEOUT

Timeout for voluntary card reveal at showdown (seconds).

SINKHORN_ITERATIONS

Maximum Sinkhorn-Knopp iterations before stopping.

SINKHORN_TEMPERATURE

Entropy regularization strength. Lower = closer to true EMD, higher = faster convergence.

SINKHORN_TOLERANCE

Early stopping threshold on marginal constraint violation.

STACK

Starting stack size in big blinds.

SUBGAME_ALTS

Alternative hands in the gadget game (Pluribus uses 4).

SUBGAME_ITERATIONS

CFR iterations for real-time subgame refinement.

S_BLIND

Small blind amount.

TRAINING_LOG_INTERVAL

Interval between progress log messages during training.

Traits

Arbitrary

Random instance generation for testing and Monte Carlo sampling.

Unique

Unique identifier trait for domain entities.

Functions

brb

Register graceful interrupt handler. Type “Q” + Enter to stop after current batch. Optionally set TRAIN_DURATION env var (e.g., “2h”, “30m”) for timed runs.

interrupted

Check if graceful shutdown was requested (via stdin “Q”) or deadline reached.

kys

Register Ctrl+C handler for immediate (non-graceful) termination. Use when you need hard shutdown without waiting for current batch.

log

Initialize dual logging (terminal + file) with timestamped log files. Creates logs/ directory and writes DEBUG level to file, INFO to terminal.

Type Aliases

Chips

Stack sizes and bet amounts in big blinds.

Energy

Distance metrics, convergence thresholds, and smoothing terms.

Entropy

Temperature parameters and information-theoretic measures.

Epoch

Training iteration counter for CFR epochs.

Position

Seat index around the table (0 = button in heads-up).

Probability

Strategy weights, sampling distributions, and reach probabilities.

Utility

Expected values, regrets, and payoffs.