Why this level matters
Section 1 of the exam guide, Fundamentals of gen AI (about 30% of the exam), expects you to describe the machine learning approaches: supervised, unsupervised and reinforcement learning. You do not need maths or code. You need to read a short business scenario and name the approach that fits.
First, one definition. Machine learning (ML) is a way of building software that learns patterns from data instead of following rules a person wrote by hand. The three approaches differ in one thing: what the model learns from.
Approach 1: Supervised learning — learn from examples with answers
Definition: The model is trained on labeled data. Each past example comes with the correct answer (the label) attached. The model learns the link between the example and its answer, then predicts the answer for new cases.
Analogy: A trainee studies 10,000 past loan files, each stamped "repaid" or "defaulted". After enough files, the trainee can judge a new application.
Supervised learning produces two kinds of prediction:
- Classification — predicts a category. Example: is this email spam or not spam?
- Regression — predicts a number. Example: how many units will we sell next month?
Business fit: predicting from past examples. Fraud detection, customer churn prediction, demand forecasting, spam filtering.
Signal words: "historical records with known outcomes", "labeled", "predict", "classify", "forecast".
Approach 2: Unsupervised learning — find structure with no answers
Definition: The model is given unlabeled data. There are no correct answers attached. The model looks for patterns on its own, most commonly by clustering: organising similar items into groups. The groups are not defined by you; the model discovers them.
Analogy: You hand someone a box of 5,000 unsorted customer receipts and say only "tell me what groups you see". They come back with "weekend bulk buyers", "daily lunch shoppers" and a few receipts that look like nothing else.
Business fit: finding hidden groups or unusual items. Customer segmentation, grouping similar products or documents, spotting unusual transactions when no one has labeled what "unusual" looks like.
Signal words: "no labels", "no predefined categories", "discover", "segment", "group similar", "hidden patterns".
Approach 3: Reinforcement learning — learn by reward
Definition: A model (called the agent) takes actions inside an environment and receives rewards or penalties for the results. Through trial and error it works out a policy: the strategy that earns the most reward over time.
Analogy: Training a dog. You do not show it a dataset. It tries something, gets a treat or gets nothing, and adjusts.
Business fit: a sequence of decisions where feedback arrives after acting. Robots learning a physical task, game-playing systems, optimising a control system such as a warehouse route or energy usage.
Signal words: "trial and error", "reward", "penalty", "agent", "environment", "learns from the outcome of its actions", "maximise over time".
Side-by-side comparison
| Supervised | Unsupervised | Reinforcement | |
|---|---|---|---|
| Learns from | Labeled examples | Unlabeled data | Rewards and penalties for actions |
| Output | A prediction (category or number) | Groups or patterns | A strategy (policy) for acting |
| Typical scenario | Predict churn from past customers | Segment customers with no predefined groups | Robot learns to pick items faster |
How to match a scenario to an approach
Apply these steps in order to every scenario.
- Find the business goal. Underline the verb: predict, group/discover, or act and improve.
- Check the data for labels. Ask: does the scenario mention past examples with known outcomes?
- Yes → lean supervised.
- No, just raw data → lean unsupervised.
- Check for a feedback loop. Ask: is something taking actions and being scored on the results? If yes → reinforcement, regardless of step 2.
- Confirm with the output. A category or number = supervised. Groups or patterns = unsupervised. A strategy for acting = reinforcement.
How to rule out the two wrong answers
In a "best answer" question, do not stop at picking one. Eliminate the other two with a stated reason:
- Rule out supervised when there are no labeled examples to learn from. No known answers, no supervision.
- Rule out unsupervised when the scenario already has known outcomes and wants a specific prediction. Clustering finds groups; it does not predict a defined target.
- Rule out reinforcement when nothing is taking actions and receiving rewards. A fixed dataset analysed once is not trial and error.
Worked example. A retailer has two years of customer records. Each record shows whether the customer cancelled their subscription. The retailer wants to know which current customers are likely to cancel.
- Goal: predict.
- Labels: yes — "cancelled" or "did not cancel".
- Feedback loop: none.
- Output: a category.
Answer: supervised (classification). Not unsupervised, because the outcome is known and labeled. Not reinforcement, because no agent is acting and being rewarded.
Common trap. "Grouping" appears in both supervised and unsupervised scenarios. If the groups are named in advance and you have examples of each (spam / not spam), it is supervised classification. If the model must invent the groups, it is unsupervised clustering.
What the exam asks about this
Expect scenario-based multiple-choice questions at business level. A scenario describes a goal and the available data, and you choose the approach. The exam does not ask for algorithms or formulas. The deciding detail is nearly always one of three things: labels present, labels absent, or reward feedback.
Key takeaways
- Supervised = labeled examples → predicts a category (classification) or a number (regression).
- Unsupervised = unlabeled data → discovers groups or patterns (clustering) that nobody defined in advance.
- Reinforcement = an agent acts in an environment, gets rewards or penalties, and learns a policy by trial and error.
- Decide by asking in order: What is the goal? Are there labels? Is there a reward loop?
- Always state why the other two approaches do not fit before you commit to an answer.