The lesson
A nod and a grumble
I hired a new hand for the Friday rush. I did not hand her a manual. She learned the way I learned: try, and read the room. The uncle who nods and orders again. The aunty who grumbles about the bread. Empty plates, clean plates. Nobody wrote the rules down. The room taught her, bowl by bowl.
The room does the teaching
This is training by consequences, and machines do it too. Reinforcement learning needs no instructions, only feedback. The machine tries things, earns rewards or penalties, and keeps what paid off. The game machine that taught itself Go played millions of games against itself, with one signal at the end of each: win or lose. From that thin signal it found moves no human had named. The new hand runs the same loop. She tries the bread basket on table three. The table smiles, and the basket comes out faster next time. She tries it on table nine, and the table waves her away. The grumble is the penalty. Nothing was written down, and the lesson stuck.
The machine that taught itself Go
The plates do the teaching all over town. The machine that beat the Go champions learned from wins and losses alone. A warehouse robot learns to stack boxes in a simulation, dropping a thousand piles until the drops stop. The video app learns which videos you finish and shows you more of those. Every one of them: try, get a signal, keep what worked.
You set the rewards
Now you are the room. The new hand has to learn one rule: which table gets the bread first. You have no rulebook to give. You only have a nod and a grumble. Walk a week in your head. Every time she takes the bread to a table and the table smiles, that is a reward. Every time a table waves her away, that is a penalty. Watch her move from guessing to knowing without one word of instruction. That is reinforcement learning, and you were the training data.