Learning by Trial and Reward
Every kind of learning in this book so far started from data somebody already had. Chapter 4's models copied answers out of labeled examples; this chapter's models hunted for structure in unlabeled ones. But one famous kind of learning fits neither shelf, because nobody hands the learner any examples at all. It acts, gets rewarded or stung, and gradually keeps more of what earned the reward — the way machines learned to beat world champions at the board game Go, and (a fact we'll cash in properly in Chapter 9) part of how chatbots got their manners.
That is reinforcement learning — the third and final family on your map of machine learning. Fair warning about the size of this visit: reinforcement learning is scientifically spectacular and, in ordinary business, rare. So this page is a short, honest tour — what the idea is, why it's genuinely different in kind, where it really runs, and why Plateful's data team has probably never trained one.
An Agent, an Environment, and a Reward
Think of training a puppy to sit. No lecture explains "sit" — no lecture could. The puppy just tries things: it jumps, it barks, it wanders, and occasionally its rear end touches the floor right after you've said the word. That attempt gets a treat; the others get nothing. Hundreds of attempts later, the behavior has converged on whatever earns treats — and the puppy never once saw an example or a rulebook.
Each piece of that scene has a proper name. The learner — puppy or program — is the agent: the thing that acts and adjusts. The world it acts in — your living room, a game board, an app full of users — is the environment. And the treat is the reward: a number that scores each attempt, which the agent adjusts itself to earn more of over many, many tries. Notice what's missing: a dataset. In supervised and unsupervised learning, the examples exist before training starts. Here the examples are generated by the agent's own attempts — it produces its experience by acting, one loop at a time, usually millions of times over. (That "millions" is not a flourish; it's the number that will decide this page's ending.)
The Third Kind of Learning
It would be tidy to file this away as a variation on what you know, but it is different in kind, and the difference is worth one clean paragraph. Supervised learning copies answers: it needs someone who already knows the right output for each example. Unsupervised learning finds structure: groups and oddities that were sitting in the data all along. Reinforcement learning discovers behavior: sequences of actions that nobody demonstrated, because nobody had to know them in advance — the reward did the teaching.
That last clause has a startling consequence: an agent can end up doing things beyond what its makers knew how to do. The Go-playing system that beat one of the world's top players in 2016 played one move so alien that expert commentators first called it a mistake — it wasn't, and no human had taught it, because it surfaced from millions of games the system played against itself. A supervised model can only approach the skill of its labelers. A reinforcement learner has no such ceiling, which is exactly why researchers love it — and, as the last section explains, why almost nobody else gets to use it.
Where It Fits Real Life
The famous homes are games and robotics. Games because they hand the method everything it wants — a perfect simulated world, a crisp score, and the freedom to play forever at no cost. Robotics because "walk without falling over" is a behavior nobody can write down as labeled examples, but a stumbling robot (usually a simulated one, at first) can find it by trial and reward.
The quieter homes share one signature: actions with delayed consequences. A recommendation system choosing what to show you next isn't making one isolated prediction — today's suggestion changes what you watch, which changes what tomorrow's suggestion should be. Pricing and ad decisions have the same shape: a discount today may earn a loyal customer next month or teach them to wait for discounts. Wherever each action reshapes the situation the next action faces, the loop above fits — and some large recommendation and pricing systems do run reinforcement learning, at companies with the traffic and the engineering to afford the experiment.
Why Your Data Team Probably Doesn't Use It
Here is the honest scoping note. Reinforcement learning learns by failing — millions of times, on purpose, as the only way to find out what works. A Go program can lose a million games overnight in simulation and nobody is harmed. But where, exactly, does a food-delivery company let an agent fail a million times? Assigning couriers absurdly to see what happens? Quoting random prices to real customers? Real businesses rarely have a simulator — a fast, faithful, consequence-free copy of their world — and building one is often harder than the original problem. Without it, trial-and-reward means experimenting on your actual customers, which is slow, costly, and occasionally a news story.
So when Plateful's data team ships something, the overwhelming default is the supervised and unsupervised machinery of Chapters 4 and 5 — reinforcement learning stays a specialized tool for those rare settings with a safe place to practice. One important exception, though, is already on your calendar: in Chapter 9 you'll see that a reward supplied by human feedback — people scoring a model's answers instead of a game scoring moves — is part of how modern chatbots are polished into being helpful and polite. Keep this page's loop in your pocket until then.
- "Reinforcement learning is how most ML works." Almost the opposite: most business ML is supervised, with unsupervised second. RL is a specialized family with a demanding requirement — a safe place to fail millions of times — that ordinary companies don't have.
- "The agent understands its goal." It climbs a numeric reward signal, nothing more; "goal" is our word for the shape of that number. Shape the reward badly and the agent will climb that instead — agents have learned to rack up points by driving in circles rather than finishing the race.
- "RL needs a training dataset like the other families." There is no dataset. The agent generates its own experience by acting — which is precisely why it needs somewhere safe to act.
- Your map of machine learning is complete. Any "the AI learned to…" headline now sorts into one of three families — copied from labeled answers, found in unlabeled structure, or discovered by trial and reward — and the sorting tells you what the system needed and where it can fail.
- It pre-loads Chapter 9: when you read that a chatbot was "trained with human feedback", the phrase will decode itself — human judgments playing the role of the reward.
Knowledge Check
The three families of machine learning differ by where their learning material comes from. Which line gets all three right?
- Supervised invents its own examples; unsupervised copies labeled answers; RL finds the groups
- Supervised learns from labeled answers; unsupervised finds structure in unlabeled data; RL makes its experience by acting
- All three start from a labeled dataset; they differ only in how large that dataset has to be before training can work
- Supervised handles numbers, unsupervised handles text and images, and reinforcement learning covers whatever data remains
In the reinforcement learning loop, what exactly is the "reward"?
- A final prize the agent receives once, after training finishes
- The labeled answer column of the agent's training dataset
- A numeric score for each action, which the agent adjusts itself to earn more of
- A trainer's written explanation of what the agent did right and what it should try differently next time
Why does a typical data team — like Plateful's — almost never use reinforcement learning?
- Learning by trial needs a safe place to fail millions of times, and most businesses don't have a simulator
- Reinforcement learning only works on board games and can't be applied to anything else
- It's an outdated technique that deep learning has since replaced
- It requires far more labeled examples than supervised learning, and labels are expensive to collect
Chapter 9 will describe chatbots "trained with human feedback". Based on this page, what is that phrase describing?
- Users correcting the chatbot's spelling mistakes until they stop appearing
- Humans writing an explicit rulebook of allowed and forbidden replies for the chatbot to consult in every conversation
- A reward supplied by human judgments — trial and reward applied to polishing the model's answers
- The chatbot reading transcripts of human conversations and grouping them into clusters
You got correct