Supervised, Unsupervised, and Reinforcement Learning

The three families of machine learning differ by the teaching signal: a correct label, raw structure, or a reward after actions.

Explain like I'm new to AI

Machine learning is “learn from data instead of writing rules.” That sentence hides a fork: what kind of feedback does the data give?

1. Supervised — each example already has the answer. Email + “spam.” House + sold price. You are learning a mapping you can reuse on new inputs.
2. Unsupervised — examples have no answer column. A pile of shoppers, or tickets, or pixels. You are looking for structure: groups, compressions, oddballs.
3. Reinforcement learning (RL) — nobody labeled the right move. An agent acts in an environment and gets a reward (win the game, the robot didn’t drop the mug). It learns a policy: what to do in which situation.

Deep learning is not a fourth family. It is a style of model (stacked neural nets) you can use inside any of the three. A neural net that predicts house prices is still supervised. A clustering method can be classical or deep. RL can use a neural net as the policy.

Mental model

Three teachers:

• Supervised is a marked exam. Every practice question has the answer key in the margin.
• Unsupervised is a box of unlabeled photos. Sort them into piles that look related. Nobody told you the pile names.
• RL is learning to ride a bike. You are not shown a labeled dataset of “correct lean angles.” You wobble, fall, get a delayed “you stayed up,” and adjust.

If you remember only one distinction: did a human (or a system of record) already write the target for each row? Yes → supervised. No, and you want groups or compression → unsupervised. No, and the feedback is a score after a sequence of actions → RL.

How it works

Supervised

1. Collect pairs (x, y) — features and a target.
2. If y is a number, you are heading toward regression. If y is a category, classification.
3. Train: predict ŷ, measure error vs y, adjust the model.
4. Evaluate on rows the model never trained on. That split is what later lessons on overfitting and cross-validation are for.

Unsupervised

1. Collect x only.
2. Choose a notion of structure: distance (clusters), reconstruction (compression), density (anomalies).
3. There is no y to score against in the supervised sense. You judge clusters by usefulness, stability, or a downstream task.

Reinforcement

1. Define states, actions, and a reward.
2. The agent tries, observes the next state and reward, and updates a policy (or a value function that scores states).
3. Credit assignment is hard: the winning move may have been twenty steps ago. That is why RL is powerful and sample-hungry.

Real-world example

A support org has three different ML jobs on the same tickets:

• Supervised: tag “billing / bug / how-to” because agents already labeled last year’s queue. Train a classifier.
• Unsupervised: you have no tags yet. Cluster embeddings of tickets to discover that “SSO timeout” is a hidden theme. A human then names the cluster.
• RL (rare on this desk): an agent that chooses which knowledge article to show, scored by “did the user stop writing?” — delayed, noisy reward. Most teams should not start here.

A game studio trains an RL policy to play a title. A bank trains supervised models on historical defaults. Same word “training”; different signal.

Technical explanation

Formally:

• Supervised: estimate p(y | x) or E[y | x] from i.i.d. labeled samples.
• Unsupervised: estimate structure in p(x) — mixture components, a low-dimensional manifold, outliers in the tails.
• RL: maximize expected discounted reward E[Σ γ^t r_t] by interacting with a Markov decision process. Data is on-policy unless you deliberately reuse old trajectories.

Cousins you will meet later, still not extra families:

• Self-supervised (masked language modeling, contrastive vision) invents a pretext label from the data itself. It is unsupervised in spirit, supervised in the training loop. Foundation models use it, then you fine-tune with real labels.
• Semi-supervised mixes a few labels with a lot of unlabeled x.
• Imitation / offline RL blurs into supervised learning when you just clone expert actions.

RL sits beside that table, not on it: the unit of data is a trajectory, not a row with a static y.

Common mistakes

• Treating clustering output as if it were a certified class label. Clusters are hypotheses. Classification requires names you already trust.
• Reaching for RL because it sounds advanced. If you can write (x, y) pairs, use supervised learning. RL is for sequential decisions with delayed reward, not “we don’t feel like labeling.”
• Confusing deep learning with a learning paradigm. Depth is the model; the paradigm is the signal.

When to use it

• Supervised when history already recorded the thing you want to predict, and new cases will look like old ones.
• Unsupervised for exploration, segmentation, anomaly hints, or compressing x before a later supervised step.
• RL when the system must act over time and you can define a reward you are willing to optimize (games, robotics, some bidding and control).

When NOT to use it

• Do not pick RL as a first production tool for a spreadsheet problem. Label or extract y.
• Do not run unsupervised clustering and then ship the cluster id as “fraud” without a labeled holdout. That is a story, not a detector.
• Do not force supervised learning when there is no stable target (the business has not decided what “good” means). Fix the definition first.

Alternatives

• Rules and heuristics when the mapping is short and legal (“if amount > 10k and country in list → review”).
• Retrieval and tools (later AI stack) when the “answer” is a document, not a learned function of columns.
• Human labeling so an unsupervised pile becomes a supervised dataset.

Quick quiz

Last reviewed: 2026-09-04 · Written by ByHeart AI · Reviewed by ByHeart AI