trial-and-error learning

At its heart, RL learns by doing and seeing what happens—the way a child learns to ride a bike. There is no teacher saying "the correct action here is X." Instead the agent tries an action, observes the outcome and reward, and nudges itself to repeat what worked and avoid what didn't. Good behavior is discovered, not handed over.

This is powerful because we often can't write down the right answer in advance, yet we can recognize a good outcome when we see one. The price is that the agent must actually experience failures to learn from them, which can be slow and, in the real world, costly or unsafe. It also forces a hard question, explored throughout RL: which of the many actions I took actually caused this result?