Machine learning is not one way of learning. The type of feedback available to a system shapes the method you choose. In supervised learning, examples include answers. In unsupervised learning, the system looks for structure without supplied answers. In reinforcement learning, an agent improves a sequence of actions using rewards from an environment.
Supervised learning: learn from known outcomes
Suppose a library has past loans with their loan length and whether each book was returned late. The late/on-time value is a label. A supervised model uses many such examples to learn a relationship between inputs and outcomes. For a new loan, it predicts the label before the return happens.
Two common supervised tasks are classification and regression. Classification chooses a category, such as late or on time. Regression predicts a number, such as estimated days late. The same business question can lead to either task, but the target you define changes how you train and evaluate the model.
Example: A shop records order weight and delivery time. Predicting “late or on time” is classification; predicting the delivery time in days is regression.
Unsupervised learning: find structure without answers
Now imagine the shop has product measurements but no predefined category for each product. An unsupervised method can group products with similar measurements. This is clustering. The group numbers do not arrive with a built-in meaning: people inspect the groups and decide whether they help with storage, search, or analysis.
Unsupervised learning can also reduce many input measurements to a smaller representation that preserves useful structure. Neither grouping nor dimensionality reduction proves a cause. A cluster might reflect a real difference, an unhelpful measurement, or a flaw in data collection. You still need to check whether the result serves the original goal.
Tip: “Unsupervised” does not mean “no human review.” People choose the data, the method, and the way to judge whether the discovered pattern is useful.
Reinforcement learning: learn from actions and rewards
Consider a simple game in which a character moves through a maze. At each step, an agent observes its current situation, chooses an action, and receives a reward or penalty from the environment. It aims to learn a policy—a way to choose actions—that earns more reward over time. A useful move now may be valuable because it opens a better route later.
This differs from receiving a table of correct moves for every maze position. The agent must explore possibilities and learn from their consequences. Reward design matters: if you reward only speed, the agent may find a fast but unwanted shortcut. Training is often done in a simulation or another controlled setting before any real-world use.
Example: In a warehouse simulator, a robot receives a reward for completing a route and a penalty for collisions. It tries routes and adjusts its actions. A real robot would still need separate safety testing.
Compare the feedback each method uses
| Approach | Feedback during learning | Typical result | Example question |
|---|---|---|---|
| Supervised | Known labels or numeric outcomes | Prediction for a new case | Will this order be late? |
| Unsupervised | No supplied target label | Groups or compact patterns | Which products resemble each other? |
| Reinforcement | Reward from a sequence of actions | Policy for choosing actions | Which route earns the best long-term result? |
The approaches are not mutually exclusive. A project may use clustering to explore data and then train a classifier on human-reviewed labels. A system can also combine supervised training with later feedback from preferences or rewards. The categories describe learning signals, not fixed product labels.
Choose an approach step by step
- State the output. Do you need a prediction, a useful grouping, or a sequence of decisions?
- Inspect feedback. Do you have trustworthy answers for past examples, or can you define a meaningful reward?
- Check feasibility. Can you collect enough representative examples or safely simulate actions?
- Define success. Choose a measure that reflects the cost of mistakes, not just an impressive score.
- Compare with a simple baseline. A clear rule may solve the task without a complex model.
For the library reminder, labeled past returns point toward supervised learning. If the library only wants to explore groups of borrowing patterns, clustering might help. Reinforcement learning would be a poor first choice for predicting a single late/on-time outcome because no sequence of actions and rewards is required.
Key points
Choose supervised learning for known outcomes, unsupervised learning for patterns without labels, and reinforcement learning for sequences of actions with rewards. Start with the task and available feedback.