Overview

Reinfors is a reinforcement-learning library for experiments involving games, planning, and neural networks. Game dynamics, search, episode orchestration, and batch assembly run in Rust. Model inference and training are caller-owned and connected through Python.

Execution model

In Engine workflows, policies request model evaluations while games and searches are running. The engine pools these requests into NumPy arrays, calls Python for inference, and routes the returned values to the requesting policies. Learners then convert completed decisions or trajectories into training records.

Rust engine and policies ── pooled NumPy observations ──▶  Python inference callback
Rust engine and policies ◀───── NumPy model outputs ───  Python inference callback
Rust learners            ──── typed training records ────▶  Python training loop

Engine.collect performs synchronous collection. Engine.collect_stream runs collection concurrently in Rust and passes batches to the Python training loop through a bounded queue that provides backpressure. Inference callbacks may use PyTorch, JAX, another local runtime, or a remote service; they may also differ by player.

Other execution surfaces cover workflows that do not use learner-shaped engine collection: Env exposes caller-controlled play, Arena runs paired evaluations, and standalone solvers own algorithm-specific traversals such as CFR.

Documentation map

• Getting started covers installation and a first collection call.
• Architecture describes the execution surfaces and component responsibilities.
• Training gives a complete Python training loop; the streaming guide covers concurrent collection.
• The game, algorithm, and compatibility catalogues describe the built-in components.
• Examples links runnable scripts for collection, training, evaluation, adapters, direct play, and solvers.
• Extending reinfors documents the Rust traits and Python registration path for new components.
• Reference specifies inference shapes, batch fields, telemetry, terminology, and current boundaries.

Scope

Reinfors uses native simulation and search within a reusable experimental interface. It is not intended to maximize throughput for one fixed workload; a fully fused implementation specialized for a particular model, game, and device topology may be faster.

The current scope includes:

• native game simulation, planning, episode collection, and inference batching;
• synchronous and concurrent collection with caller-owned models, optimization, and replay;
• sequential and simultaneous decisions, explicit chance, multiple players, and imperfect information where supported by the selected algorithm;
• resolved configurations, snapshots, telemetry, per-player inference, and evaluation surfaces;
• Gymnasium and PettingZoo adapters for compatible games.

Games currently use fixed discrete action and observation spaces, one decision phase, and native Rust component implementations. Reinfors is pre-1.0; the current boundaries page records the applicable API and modeling constraints.