What RL Is

state vs observation

The state is the full description of the situation—everything that matters for deciding what happens next. The observation is what the agent actually gets to see, which is often only part of the state. A poker player knows their own cards and the bets (observation) but not the opponents' hidden cards (part of the true state). The map is not the territory.

When the observation captures everything relevant, we call the problem fully observable and can treat the observation as the state. When it doesn't—a robot with a narrow camera, a trader without tomorrow's news—the problem is partially observable, and a snapshot can mislead. The usual fix is to remember history: stack recent frames or keep an internal summary, so the agent acts on more than a single blurry glimpse.

"State" sometimes means the true world state and sometimes the agent's internal state built from observations; context tells you which.