Supervised learning learns from labelled examples, unsupervised learning discovers patterns in unlabelled data, and reinforcement learning learns actions through rewards and penalties from an environment. The key difference is the feedback available during learning.
How each type works
A machine learning model improves by adjusting internal parameters from training experience. The source of guidance differs among the approaches.
| Learning type | Learning mechanism | Typical task and example |
|---|---|---|
| Supervised learning | The algorithm receives labelled training data: inputs paired with known correct outputs. It learns to predict outputs for new inputs. | Classification, such as identifying spam email; regression, such as predicting house prices. |
| Unsupervised learning | The algorithm receives unlabelled data and identifies structures without predefined correct answers. | Clustering customers by purchasing behaviour; detecting unusual patterns. |
| Reinforcement learning | An agent selects actions in an environment. Rewards evaluate outcomes, and the agent develops a policy that maximizes cumulative reward. | Training a game-playing agent or a route-finding robot. |
In supervised learning, feedback is direct because each prediction can be compared with a known label. In unsupervised learning, there is no answer key; similarity or another measure is used to organize data.
Reinforcement learning instead evaluates actions, not correct outputs. Rewards can be delayed, so the agent balances exploration of new actions with exploitation of actions already associated with rewards.
A common misconception is that reinforcement learning is supervised learning with rewards as labels. This is incorrect: it involves sequential decisions, and actions can alter later states and future rewards.
IB exam technique
For A4.1 Machine learning fundamentals, compare methods using three criteria: whether data is labelled, what feedback is supplied, and what is learned. When identifying a method from a scenario, justify it with evidence such as known outputs, hidden groupings, or rewards from interaction.
