Reinforcement learning is a way to learn a decision-making strategy through interaction. A system observes an environment, chooses an action and receives feedback, often expressed as a numerical reward. Training aims to improve the outcomes of a sequence of decisions rather than simply reproduce a supplied correct answer.
The feedback must represent the intended goal, and the environment must provide useful information about the effects of actions. In research, an experiment's measurements could inform which experiment to try next. That possibility does not by itself establish that a laboratory system uses reinforcement learning or that its results are clinically validated.
ELI5
Reinforcement learning is a way to learn through consequences. Helpful choices earn a stronger signal, while unhelpful choices earn a weaker signal, and the model gradually favors actions that have worked before.
For example, a game-playing model can earn points for reaching the goal and lose points for moves that end the game. It may become very good at collecting points, but people still need to check that it has not found an unintended trick in the scoring rules.
Does reinforcement learning always need a human to give feedback?
No. Feedback can come from a simulation, a programmed scoring rule, measurements or human judgments. What matters is whether the feedback gives the system a useful signal about its decisions.
Can laboratory measurements be used as reinforcement-learning feedback?
They can contribute to feedback for choosing future experiments when the research problem is set up that way. Measurements alone do not prove that a system uses reinforcement learning, and learning to choose experiments is not the same as proving a treatment works in people.
