Reinforcement learning is one of those terms that sounds far more forbidding than the idea underneath it. Strip the phrase back and what is left is something everyone has already done: try a thing, notice how it went, and let that shape the next attempt. Nicole Junkerman finds it one of the more approachable corners of a field that often seems to prefer not being approached.
The closest everyday parallel is learning to cook a dish nobody has written down for you. The first attempt is guesswork. The second is a small correction, because something was flat or something caught in the pan. By the sixth you are not following a rule at all but carrying a rough sense of what a little more salt does. Nobody handed you the right amount. You arrived at it by tasting.
A machine learning in this way has the same shape. It is placed in a situation, it does something, and afterwards it receives a signal about whether that went well or badly. It has no understanding of the dish and no opinion about dinner. What it has is an extremely patient willingness to try again, and to try again far more often than any cook would tolerate.
Three things follow from that.
- It learns from consequences, not from worked examples. Nothing shows it the right answer in advance. It only ever finds out afterwards, and only in the form of a score.
- It needs a great many attempts. Where each attempt is slow, costly or risky, the whole approach becomes awkward, which is a practical limit rather than a small inconvenience.
- It is only ever as sensible as the signal it is given. The measure of success is chosen by people, and everything else follows from that choice.
That third point is the genuinely interesting one. If the measure of success is slightly wrong, the result will be slightly wrong too, in a way that is perfectly logical and entirely unhelpful. Anyone who has set a target at work and watched a team meet it in the letter rather than the spirit will recognise the problem. It is not new, and it is not really technical.
None of this makes the approach magical. It sits happiest where an attempt is cheap, can be repeated often, and where a good outcome is easy to score without argument. Where attempts are expensive, or where what counts as a good outcome is honestly contested, it is far less at home. That limit is worth saying out loud, because the name suggests something grander than the method.
What stays with her is how ordinary the underlying habit is. Trying, noticing, adjusting and trying again is how a child learns to throw, how anyone learns an instrument, and how most good writing gets to its fourth draft. The vocabulary is new. The habit is very old.
Plainer definitions of a few other terms sit in the glossary, and the writing method page is the same idea applied to sentences rather than machines. An earlier note, the quiet, helpful side of AI, looks at where all this actually shows up in a day. Everything else is on the blog.