decisionrl¶

A dependency-light, correctness-first reinforcement learning foundation.

Readable like CleanRL, composable like Stable-Baselines3, and batteries-included so it runs the moment you pip install it.

Highlights¶

20+ algorithms — tabular (Q-Learning, SARSA, Expected SARSA), model-based (Dyna-Q, MBPO), value-based (DQN, C51, QR-DQN, Rainbow), policy gradient / actor-critic (REINFORCE, A2C, PPO, IMPALA, Recurrent PPO, discrete SAC), continuous control (DDPG, TD3, SAC — with optional PER + n-step), offline (TD3+BC, IQL, CQL) — plus multi-agent PPO (self-play / IPPO).
Correctness-first — proper terminated/truncated bootstrapping, GAE, target-policy smoothing, automatic entropy tuning, orthogonal init.
Dependency-light — only NumPy + PyTorch in the core; Gymnasium optional.
Batteries included — built-in environments (classic control + applied), image observations (CNN), vectorized envs (sync & multiprocessing), a CLI and a tuned-hyperparameter registry.
One API for everything — every agent has predict / learn / save / load.

Agents learning applied tasks

Install¶

pip install git+https://github.com/DenisDrobyshev/decisionrl.git
# with Gymnasium environments:
pip install "decisionrl[gym] @ git+https://github.com/DenisDrobyshev/decisionrl.git"

See Getting started for a first training run.