AI-11 · Advanced · ENGLISH · PLANNED
Begin with a baseline, trace a model, measure errors and improve it.
Study time is an initial activity estimate, excluding extra reading, revision and extended laboratory or project work. Videos and question bodies have not been produced by adding this outline.
Complete these courses or demonstrate equivalent knowledge. These are recommended prerequisites, not enforced enrolment locks.
YouTube playlist: ARK Artificial Intelligence | Reinforcement Learning and Sequential Decisions
Playlist title reserved in the plan; no YouTube playlist or upload has been created yet.
Sign in to save this learning pathPractical outcome: Specify a tiny grid task with explicit rewards and failure states.
Explain markov decision processes using a visual and one worked real-world example. Context: Specify a tiny grid task with explicit rewards and failure states. End with a prediction or check-for-understanding.
Explain states and actions using a visual and one worked real-world example. Context: Specify a tiny grid task with explicit rewards and failure states. End with a prediction or check-for-understanding.
Explain rewards using a visual and one worked real-world example. Context: Specify a tiny grid task with explicit rewards and failure states. End with a prediction or check-for-understanding.
Explain discount factor using a visual and one worked real-world example. Context: Specify a tiny grid task with explicit rewards and failure states. End with a prediction or check-for-understanding.
Use the first concepts from this section. Specify a tiny grid task with explicit rewards and failure states. Produce a labelled diagram, trace or calculation; explain each step and check one changed input.
3–6 questions on the preceding first-half concepts only: recall, visual interpretation, application and one plausible misconception. Explain each answer after submission.
Explain policies using a visual and one worked real-world example. Context: Specify a tiny grid task with explicit rewards and failure states. End with a prediction or check-for-understanding.
Explain value functions using a visual and one worked real-world example. Context: Specify a tiny grid task with explicit rewards and failure states. End with a prediction or check-for-understanding.
Explain bellman equations using a visual and one worked real-world example. Context: Specify a tiny grid task with explicit rewards and failure states. End with a prediction or check-for-understanding.
Explain exploration versus exploitation using a visual and one worked real-world example. Context: Specify a tiny grid task with explicit rewards and failure states. End with a prediction or check-for-understanding.
Specify a tiny grid task with explicit rewards and failure states. Work without the example, compare two cases, record assumptions and limitations, then self-review correctness, reasoning and clarity.
6–12 questions covering Markov decision processes, States and actions, Rewards and the remaining section concepts, with cumulative retrieval from earlier sections. Include a new transfer problem; 70% unlocks progression.
Practical outcome: Compare learning curves using repeated seeded toy runs.
Explain dynamic programming using a visual and one worked real-world example. Context: Compare learning curves using repeated seeded toy runs. End with a prediction or check-for-understanding.
Explain monte carlo returns using a visual and one worked real-world example. Context: Compare learning curves using repeated seeded toy runs. End with a prediction or check-for-understanding.
Explain temporal-difference learning using a visual and one worked real-world example. Context: Compare learning curves using repeated seeded toy runs. End with a prediction or check-for-understanding.
Explain q-learning using a visual and one worked real-world example. Context: Compare learning curves using repeated seeded toy runs. End with a prediction or check-for-understanding.
Use the first concepts from this section. Compare learning curves using repeated seeded toy runs. Produce a labelled diagram, trace or calculation; explain each step and check one changed input.
3–6 questions on the preceding first-half concepts only: recall, visual interpretation, application and one plausible misconception. Explain each answer after submission.
Explain sarsa using a visual and one worked real-world example. Context: Compare learning curves using repeated seeded toy runs. End with a prediction or check-for-understanding.
Explain deep q networks using a visual and one worked real-world example. Context: Compare learning curves using repeated seeded toy runs. End with a prediction or check-for-understanding.
Explain policy gradients using a visual and one worked real-world example. Context: Compare learning curves using repeated seeded toy runs. End with a prediction or check-for-understanding.
Explain actor-critic methods using a visual and one worked real-world example. Context: Compare learning curves using repeated seeded toy runs. End with a prediction or check-for-understanding.
Compare learning curves using repeated seeded toy runs. Work without the example, compare two cases, record assumptions and limitations, then self-review correctness, reasoning and clarity.
6–12 questions covering Dynamic programming, Monte Carlo returns, Temporal-difference learning and the remaining section concepts, with cumulative retrieval from earlier sections. Include a new transfer problem; 70% unlocks progression.
Practical outcome: Explain why a high reward can still represent a bad policy.
Explain reward misspecification using a visual and one worked real-world example. Context: Explain why a high reward can still represent a bad policy. End with a prediction or check-for-understanding.
Explain off-policy evaluation using a visual and one worked real-world example. Context: Explain why a high reward can still represent a bad policy. End with a prediction or check-for-understanding.
Explain model-based learning using a visual and one worked real-world example. Context: Explain why a high reward can still represent a bad policy. End with a prediction or check-for-understanding.
Explain offline reinforcement learning using a visual and one worked real-world example. Context: Explain why a high reward can still represent a bad policy. End with a prediction or check-for-understanding.
Use the first concepts from this section. Explain why a high reward can still represent a bad policy. Produce a labelled diagram, trace or calculation; explain each step and check one changed input.
3–6 questions on the preceding first-half concepts only: recall, visual interpretation, application and one plausible misconception. Explain each answer after submission.
Explain multi-agent learning using a visual and one worked real-world example. Context: Explain why a high reward can still represent a bad policy. End with a prediction or check-for-understanding.
Explain safe exploration concepts using a visual and one worked real-world example. Context: Explain why a high reward can still represent a bad policy. End with a prediction or check-for-understanding.
Explain sim-to-real limits using a visual and one worked real-world example. Context: Explain why a high reward can still represent a bad policy. End with a prediction or check-for-understanding.
Explain evaluation variance using a visual and one worked real-world example. Context: Explain why a high reward can still represent a bad policy. End with a prediction or check-for-understanding.
Explain why a high reward can still represent a bad policy. Work without the example, compare two cases, record assumptions and limitations, then self-review correctness, reasoning and clarity.
6–12 questions covering Reward misspecification, Off-policy evaluation, Model-based learning and the remaining section concepts, with cumulative retrieval from earlier sections. Include a new transfer problem; 70% unlocks progression.
Combine all three sections: Explain why a high reward can still represent a bad policy. Submit an original worked artifact, evidence or tests, assumptions and a reflection. Rubric: correctness, method, practical usefulness, reproducibility and communication. Self-reviewed, not certified.
6–12 balanced questions sampling every section of Reinforcement Learning and Sequential Decisions, including interpretation, calculation or reasoning, a misconception and a new application. Passing is 70%; review errors before retrying.
Data experiments, evaluations, model tracing and reproducible projects. Guided work leads to an independent attempt, with a diagram, explanation, trace, calculation or tested artifact. Question bodies are generated when opened in your private learning path. Programming submissions are stored as text; the platform does not execute submitted code.
Use original or appropriately licensed datasets. Practical science activities use models or supplied data unless a qualified supervisor has reviewed a physical experiment.
Open the interactive practice sampler →