Explain like I'm five · AI Theory

What is Reinforcement Learning?

Reinforcement learning, told as a story about a rookie operator who, with no answer key, finds the knack purely from the score each action earns.

no answer key tells me pass or fail ahead of timeno teacher-given examples — the operator has to judge alone

The rookie operator has no answer key.No teacher-given examples — the operator has to judge alone.

this bin?✓+pointsthat bin?×-pointsright guesses score points, wrong ones lose them — the operator builds the knack from those points

The operator tries tossing produce into a bin, and learns from the bonus or the penalty.Right guesses score points, wrong ones lose them — the operator builds the knack from those points.

PASSFAIL✓reinforcement learning = no answer key — just finding good habits from the reward or penalty each action bringsover time, the operator's behavior shifts toward whatever racks up more points

Reinforcement learning finds good behavior from reward and penalty alone, with no answer key.Over time, the operator's behavior shifts toward whatever racks up more points.

Trial and errortry it, then see what happens
Reward signalpoints for right, penalties for wrong
Balance explore and exploitsometimes try something new
Maximize cumulative rewardit's the total score that counts, not one try
SUPERVISED (THE TRAINER)gives labeled examples upfrontcorrects by how far off it is from the examplelearns directly, example by exampleREINFORCEMENT (THE OPERATOR ALONE)no labels — just tries actionsadjusts using only reward or penaltylots of trial and error, so it takes time

Supervised and reinforcement learning learn in different ways.Supervised learning learns from labeled examples; reinforcement learning, from reward and penalty.

told fast tosses score, the operator tosses without checkingreward redesigned around accuracy: the operator learns properlyreinforcement learning's outcome hinges heavily on what reward you actually give

A poorly designed reward teaches the operator the wrong knack.Reinforcement learning's outcome hinges heavily on what reward you actually give.

In one breath

Reinforcement learning has no labeled examples — an agent changes its own behavior using only the reward or penalty each action earns, aiming to maximize cumulative reward.

Reinforcement learning has an agent observe the environment's state, choose an action, and receive a reward plus the next state in return. Through repeated interaction, the agent learns a policy that maximizes cumulative reward. Unlike supervised learning, there are no labeled answers — balancing exploration and exploitation, and designing the reward itself (reward shaping), are the central challenges. RLHF, used to align LLMs, builds on this same framework: human preferences become the reward signal that reinforces the model.

When grown-ups say it

State
What's currently on the conveyor belt. What the operator looks at before choosing an action.
Action
Choosing which bin to toss into. The choice the operator makes after seeing the state.
Reward
Plus points for right, minus for wrong. The only signal telling the operator whether the action was good or bad.
Policy
The whole knack the operator has built up. The rule that decides what action to take in each state.
Exploration vs Exploitation
Trying a new bin, or sticking with the usual one. Lose the balance and learning stalls or wanders.
RLHF
A special case where humans decide the score. Covered as an alignment method in the AI field. → see it in full
Time-Series Analysis (AR/MA)
The inspector who guesses tomorrow from the last few days. Also learns from data — but to forecast, not to act. → see it in full
AI Trustworthiness
The certification board accuracy alone can't earn. A bad reward design is one thing a trust check catches. → see it in full
Next story
The Inspector Who Rotates Which Group Becomes the Test →
Want another term as a picture book? Request →