What does Reinforcement Learning mean?

Reinforcement learning describes a learning method in which a system develops its own strategy for action through repeated trial and error. It observes a state, chooses an action, and receives a reward or penalty in return. Unlike supervised learning, there is no correct answer given in advance — only an evaluation of the consequences.

The method is described through four quantities: the states, the possible actions, the reward function, and the strategy. Q-learning estimates the expected future return for every combination of state and action; deep Q-networks take over this estimate using a neural network. A separate parameter controls how much the system explores new options versus exploiting known ones. Because many failed attempts are needed along the way, training usually starts in a simulation.

Reinforcement learning pays off where a sequence of decisions interacts and success can only be measured at the end. Examples include traffic light control, elevator dispatching, and warehouse reordering. For single, independent decisions, supervised methods are the better fit.

The advantage over a fixed schedule lies in the measured values. The strategy follows the state reported by the sensors instead of a predetermined plan. Because the evaluation exists as a number, success can be measured against a fixed metric rather than an impression.

The result depends entirely on the reward function, since the system optimizes exactly the quantity entered there. If a traffic light controller rewards vehicle throughput alone, pedestrian waiting times grow longer without the metric ever showing it. Which quantity gets rewarded is therefore a business decision, not a technical one.