Classification & regression

Most supervised models answer one of two question shapes.

Classification — "which one?"

Pick a category from a fixed set. The racing model that chooses left / straight / right is a classifier; so is a spam filter or a cat-vs-dog model.

• Output: a probability per class (softmax); pick the highest.
• Measure: accuracy — but beware imbalance. A track that is 90% straight makes a "always straight" model 90% accurate and 100% useless in curves.

Regression — "how much?"

Predict a continuous number: an exact steering angle, tomorrow's temperature, a lap time.

• Output: a real value (or several).
• Measure: average error (MAE/MSE) — how far off, not right-or-wrong.

Choosing between them

The same problem can be framed both ways: steering as 3 buttons (classification) or as an angle (regression). Buttons are easier to label and learn; the angle is smoother to drive. Framing the task is part of the engineering — you saw the button framing in the supervised course.