1 · RL foundations
Agent, environment, reward, and the formal frame: Markov Decision Processes, returns, and policies.
●Track 11 · live
Reinforcement Learning
Learn by trial and error — agents, rewards, Q-learning, policy gradients.
The third pillar after supervised and unsupervised learning: an agent learns to act by maximising reward. Markov Decision Processes, dynamic programming, Monte Carlo and temporal-difference learning, Q-learning and SARSA, policy gradients (REINFORCE, actor-critic), and the ideas behind Deep RL (DQN). Every algorithm is implemented from scratch on a tiny environment you run in the browser (gridworld, bandits) — PyTorch DQN is read-along.
19 / 19 lessons7 chapters
2 · Bandits & exploration
The exploration-vs-exploitation dilemma in its purest form — multi-armed bandits, epsilon-greedy, and UCB.
3 · Dynamic programming
When you know the environment's rules: the Bellman equations, value iteration, and policy iteration, solved on a gridworld.
4 · Model-free prediction
Learning value from experience alone — Monte Carlo returns and temporal-difference (TD) bootstrapping.
5 · Model-free control
Learning an optimal policy from trial and error — Q-learning, SARSA, and a full gridworld training project.
6 · Policy gradients
Optimise the policy directly instead of values — the policy-gradient theorem, REINFORCE, and actor-critic.
7 · Deep RL & capstone
Scaling RL with neural networks — DQN and its tricks, the modern Deep RL landscape, and a capstone agent trained end to end.