
Already know some of this? A 2-minute test skips the lessons you know.
Take the placement testA free 16-lesson course built from Nicholas Renotte's 3h 02m of YouTube video. 2 days at 90 minutes a day, with study notes and a quiz in every lesson. A 2-minute placement test skips what you already know.
Taught by Nicholas Renotte on YouTube. Videos play in YouTube's own player, so the creator keeps every view.
Understand the course structure, practical goals, and the five-step workflow for applying reinforcement learning with Stable Baselines3.
Explain the core RL framework (agent, environment, action, reward, observation), the exploration-exploitation trade-off, and real-world applications like autonomous driving and robotics.
Install Stable Baselines3 with the 'extra' flag, import core dependencies (os, gym, PPO, DummyVecEnv, evaluate_policy), and understand the library's algorithm zoo.
Distinguish simulated vs. real environments, navigate the Gym registry (classic control, Box2D, MuJoCo, Atari), and define Box, Discrete, MultiDiscrete, MultiBinary, Tuple, and Dict spaces.
Load the CartPole-v1 environment, inspect its observation and action spaces, and run a random-agent baseline to visualize the balancing task.
Configure logging directories, instantiate a vectorized PPO model with MlpPolicy, train for 20,000 timesteps, and interpret TensorBoard metrics (explained variance, losses, learning rate).
Use evaluate_policy to compute mean reward and std over episodes, render the trained agent, and launch TensorBoard to monitor training curves.
Implement StopTrainingOnRewardThreshold and EvalCallback to halt training automatically when mean reward exceeds 200 and checkpoint the best model.
Install Atari dependencies, download ROMs from AtariMania, import VecFrameStack and make_atari_env, and configure GPU-accelerated PyTorch if available.
Wrap four Breakout environments with VecFrameStack, train a CNN-policy PPO agent for 100k steps, evaluate, save, and compare against a 2M-step model averaging 22 reward.
Install SWIG and Box2D, load the CarRacing-v0 environment, inspect its image-based observation space (96x96x3) and continuous action space (Box[-1,1]^3).
Configure a CNN-policy PPO model, train for 100k steps on the racetrack, save the model, and reload it to verify persistence.
Evaluate a 428k-step model achieving ~50 mean reward and a 2M-step model reaching 900+ reward, demonstrating the impact of extended training.
Review Discrete, Box, MultiDiscrete, MultiBinary, Tuple, and Dict spaces with code examples to prepare for custom environment design.
Implement a Gym.Env subclass (ShowerEnv) with __init__, step, render, and reset; define a 3-action Discrete space, a Box observation space, a temperature-based reward function, and a 60-second episode limit.
Test the ShowerEnv with random actions, train an MLP-policy PPO agent for 40k steps, save the model, reload it, and verify evaluation works on tabular data.