No labels, no examples — just an agent, actions, and a reward that arrives late. RL is absent from the exam but core to robotics, game AI, RLHF and agentic decision-making. The arc: nail the tabular machinery (MDPs → Bellman → TD → Q-learning) until nothing is mysterious, then let a network replace the table (DQN → policy gradients → PPO) and finish by trading — with an honest out-of-sample backtest.
Finish the DQN trading capstone: linear Q → deep DQN (target net + ε-decay) → evaluate out-of-sample with your BYO-15 backtester — compare to buy-and-hold and report honestly (most RL traders lose after costs; saying so IS the result). Read the DQN or PPO paper with your research-skills method. You've got RL when you can explain MDPs/Bellman/Q-learning, implement DQN, and reason about exploration, variance, and why a replay buffer + target network matter.
Guides: dqn-trading capstone ↗ · module-a guide ↗