Partial Observability
A situation where an agent or system cannot directly observe all relevant information needed to make optimal decisions, forcing it to infer hidden states from incomplete data.
A fundamental constraint in decision-making systems where the agent lacks complete access to the true state of its environment. Unlike full observability, where all relevant variables are directly visible, partial observability requires the system to maintain beliefs about hidden factors and update them based on indirect observations. This concept is central to robotics, game AI, and autonomous systems operating in complex real-world environments where sensors provide limited or noisy information.
What this means in real life
When you're driving in fog, you can't see cars several hundred feet ahead, so you must infer traffic conditions from brake lights, engine sounds, and the behavior of nearby vehicles—making decisions with incomplete information about the full road state.
What it isn’t
It's not simply having noisy or imperfect sensors. Even with perfect sensors, an environment has partial observability if critical information remains fundamentally hidden—like an opponent's cards in poker or a robot unable to see around corners.
Commonly misused online
Tech discussions often conflate partial observability with uncertainty or randomness, treating any unpredictable outcome as 'partially observable.' True partial observability specifically means hidden state variables exist that the agent cannot directly measure, not just that outcomes are probabilistic.