tinyrl

A minimal reinforcement learning toolkit built from scratch. Provides simple RL environments and training utilities for learning and experimenting with RL algorithms.

Why tinyrl?

• Simple — small codebase, easy to read and modify
• From scratch — no dependency on OpenAI Gym or Gymnasium
• Extensible — add new environments by subclassing Environment, new policies by subclassing Policy
• Batteries included — comes with a Runner for rolling out episodes and a TrainingMonitor for plotting training curves

Quick example

from tinyrl import GridWorld, Runner, RandomPolicy

env = GridWorld()
runner = Runner(env)
policy = RandomPolicy(n_actions=env.n_actions)

# run 100 episodes
for _ in range(100):
    result = runner.run_episode(policy)

# plot reward and episode length curves
runner.plot()

Package structure

tinyrl/
  core/          # Environment, Policy, Runner, TrainingMonitor, types
  envs/          # GridWorld, ...
  algorithms/    # RandomPolicy, ...