agent–environment interaction loop

This is the heartbeat of RL, repeated over and over. The agent observes the current situation, chooses an action, and sends it to the environment. The environment responds with a new situation and a reward. The agent takes that in and chooses again. Perceive, act, get reward, repeat—one turn of this cycle is called a time step.

All of RL's machinery hangs on this loop. The stream of states, actions, and rewards it produces—the agent's experience—is the only data the agent ever gets; there is no separate training set. Learning means using this stream to improve the policy so that, loop after loop, the actions chosen tend to gather more reward.