The big idea: Reinforcement learning has no labelled examples and no correct answers. An agent acts in an environment, receives a reward or penalty, and adjusts.
Nobody ever tells it the right move — only how things turned out.
| Supervised | Unsupervised | Reinforcement | |
|---|---|---|---|
| Learns from | Labelled answers | Structure in the data | Rewards |
| Feedback | The right answer, every time | None | A score, sometimes much later |
| Data | Fixed dataset | Fixed dataset | Generated by acting |
| Example | Spam detection | Customer segments | Game playing, robot control |
Free preview
This is the free notes preview
You're reading the free notes. Aimnova Pro unlocks the full study experience — and you can try it with your first topic free to keep:
- FlashcardsLock in vocabulary and key terms with spaced repetition.
- Practice questionsAnswer exam-style questions and get instant AI marking.
- Mock exams & past-paper vaultSit full mocks and see exactly how examiners award marks.
- Personalised study planA daily plan built around your exam date and weak areas.
Five terms, and they are all examinable
- Agent — the thing learning and acting
- Environment — everything it acts on and observes
- State — the situation it is in right now
- Action — what it may do from that state
- Reward — the number telling it how that turned out, positive or negative
- Policy — its current rule for choosing an action given a state
The loop: Observe the state → choose an action from the policy → the environment returns a reward and a new state → update the policy → repeat.
The policy is what is actually being learned.
Memorize terms 3x faster
Smart flashcards show you cards right before you forget them. Perfect for definitions and key concepts.
Explore or exploit
- Exploit — take the best action known so far
- Explore — try something else, in case it is better
- Exploit only and you never find the better move
- Explore only and you never use what you learned
Credit assignment
- A game is won after 200 moves — which ones caused it?
- The reward arrives long after the decisions that earned it
- Spreading credit backwards correctly is the hard part
Reward design
- The agent optimises exactly what you measure
- Reward "points scored" and it may loop a scoring glitch forever
- It is not cheating — it is doing what it was told
Reward hacking is the ethical point: An agent rewarded for time spent on a platform learns to maximise time spent — which may mean recommending whatever is most outrageous.
The system is working perfectly. The objective was wrong, and the harm follows from that, not from a bug.
How this is tested — you must identify reinforcement learning from the absence of labels, and reason about the reward function. It comes up two ways:
Paper 1 Section A
- Describe how an agent learns from its environment, 3-4 marks
- Name the agent, state, action and reward in a scenario
- Define the exploration-exploitation trade-off
Paper 1 Section B — case study
- Judge whether a reward function is well designed
- Explain a risk of the approach
The classic trap: Calling it supervised because there is feedback. Supervised learning is told the correct answer; reinforcement learning is told only how well things went — and often long afterwards.
A company trains an agent to control heating in an office building to cut energy use. Describe how reinforcement learning applies here, naming its components, and identify one risk in the reward design.
Model answer plan
See the mark-by-mark plan — for / against / judgement, with marking guidance — in study mode.