statrl#
The Statistical Reinforcement Learning Toolkit — a research library for bandit and reinforcement-learning algorithms, organised as a taxonomy of settings, each with a matching environment, agent, and interaction loop.
statrl gives you a small, explicit protocol shared across settings: an
environment holding the problem, an agent that acts and learns, and an
interaction loop that runs the two against each other and returns a
cumulative-score time series. On top of that sit reference algorithms (IMED,
PSRL, IMED-RL, BIMED, BABA, an adversarial Lipschitz forecaster) and an
experiments harness for running many replicates in parallel and plotting
regret.
Install statrl and measure your first regret curve.
Narrative walkthrough of each setting and the algorithms it ships.
Auto-generated reference for every public module, class, and function.
Runnable scripts benchmarking agents and plotting regret.