What you should be able to do
Model states/actions/rewards
Implement value updates
Understand exploration
Evaluate policies offline
Every module is now a full lesson.
Bandits
Learn which action performs best while continuing controlled exploration.
MDPs
Model sequential decisions with states, actions, transitions, rewards and discounting.
Dynamic programming
Build a working mental model of Dynamic programming then connect it to implementation evaluation and production trade-offs.
Monte Carlo
Build a working mental model of Monte Carlo then connect it to implementation evaluation and production trade-offs.
TD learning
Build a working mental model of TD learning then connect it to implementation evaluation and production trade-offs.
Q-learning
Learn action values off-policy from temporal-difference targets.
Policy gradients
Optimize a parameterized policy directly using sampled returns or advantages.
Projects that prove the skill
Response policy lab
Use a baseline first then instrument the result and document failure cases.
design with cortex →Adaptive difficulty controller
Use a baseline first then instrument the result and document failure cases.
design with cortex →SEO bandit
Use a baseline first then instrument the result and document failure cases.
design with cortex →How progress is judged
Reward design
Evidence should come from code, results, explanation quality and the ability to identify when an approach should not be used.
Policy stability
Evidence should come from code, results, explanation quality and the ability to identify when an approach should not be used.
Regret
Evidence should come from code, results, explanation quality and the ability to identify when an approach should not be used.
Off-policy reasoning
Evidence should come from code, results, explanation quality and the ability to identify when an approach should not be used.
Turn the track into a personal plan.
Cortex can break this path into daily sessions and adjust based on completed lessons, interview results and your saved learning goals.
Ask Cortex to plan it →