Is partial information the same as partial observability?

Partial observability is a standard formal setting in which observations do not fully reveal the state. Partial information is a broader phrase that can also include other missing knowledge.

Can action reduce uncertainty without sensing?

Yes. An action can map several possible states to one known state, as SetOff does in the example.

Must every belief state use probabilities?

No. Some planning methods use sets of possible states; probabilistic methods attach weights according to their model.

AI Glossary / Graphs and symbolic AI / Intermediate

Partial information

Definition

Partial information means that an agent's available evidence does not determine every relevant aspect of the situation. In AI planning and decision-making, the agent may need to act while several world states remain possible or while observations reveal only part of the state.

What it means

A warehouse assistant might know which aisle it occupies but not whether a shelf behind a closed door is occupied. The uncertainty concerns what the agent knows, even if the physical world has a definite state.

Partial information is different from random action outcomes. A system can have incomplete knowledge in a deterministic environment, and it can know the present state fully while future outcomes remain stochastic. These two sources of uncertainty often coexist, but they should be modeled separately.

It also differs from a partially completed computation. Running a search longer can improve a decision without revealing a hidden fact that requires an observation. An agent must distinguish uncertainty that reasoning can resolve from uncertainty requiring new evidence.

How it works

A planner can represent available knowledge as a belief state: a set of world states compatible with the evidence, or a probability distribution when probabilities are modeled. This is information about possible states, not several physical realities occurring at once.

Actions update the possibilities through a transition model. Observations then remove incompatible states or adjust their probabilities through an observation model. With noisy sensors, an observation usually changes probabilities rather than identifying the state with certainty.

A sensorless plan must achieve its goal across the relevant possible states under its assumptions. A contingent plan can branch on future observations. A probabilistic policy can instead choose actions according to expected outcomes and preferences.

The representation and solution method should match the task. Treating every unknown fact as false silently converts missing evidence into a claim about the world.

Three practical examples

A robot maintains several possible locations until a distinctive landmark narrows them down.
A planning assistant chooses a common preparation step that works whether a required resource is in one storage room or another.
A game-playing agent reasons about hidden opponent cards without assuming that the cards are randomly changing during play.

Think of it this way

Think of reading a map while a marker covers part of it. Several routes may remain possible until the covered area is revealed, even though the roads themselves are fixed.

The analogy is limited because AI observations may be noisy and the environment may change while the agent acts. Removing the marker is not always available or free. The essential point is that incomplete evidence and an uncertain physical process are different concepts.

Worked example

Imagine an agent controlling a lamp whose initial state is either On or Off. Its initial belief set is {On,Off}. The available Toggle action deterministically changes On to Off and Off to On.

After one Toggle, the possible states are still {On,Off}. Repeating Toggle any fixed number of times cannot guarantee the lamp ends Off because it preserves uncertainty about the initial state.

Now add an action SetOff that makes the lamp Off from either initial state. Applying SetOff maps the belief set to {Off}, so it achieves the goal without first observing the initial condition.

Alternatively, an accurate Read action followed by “if On, Toggle; if Off, do nothing” provides a contingent solution. If Read is noisy, that guarantee no longer follows without further assumptions. This original example separates deterministic transitions, incomplete information and the value of sensing.

Illustrative case study

The fictional Willow Automation team assumes that toggling a device after a restart will put it into a known state. Testing shows that the result depends on whether the device was already active before the restart.

Willow changes its reasoning to track both possible initial states. Where available, it uses an explicitly defined state-setting operation; elsewhere it obtains an observation and handles its uncertainty before choosing an action.

The fictional case illustrates a general design lesson: an action that changes a state is not necessarily an action that establishes knowledge of that state. The correct choice depends on the transition and observation models.

Common misconceptions and mistakes

Unknown means false. Lack of evidence is not automatically evidence of absence.
Partial information requires random physical dynamics. A deterministic world can be only partly observed.
A rational agent must know everything. Rational decisions can be defined relative to available evidence and objectives.
More computation always reveals the hidden state. Some uncertainty requires sensing, communication or additional data.
A sensor reading is always certain. Observation reliability must be represented rather than assumed.

Business use

Partial-information reasoning is useful when records are incomplete, sensors cover only part of a process or actions must precede full confirmation. It encourages teams to specify what is known, what remains possible and which decisions depend on missing facts.

A workflow can sometimes choose an action that succeeds across possibilities, as in SetOff, rather than paying to identify the initial state. In other cases, acquiring information is valuable because it changes the preferred action.

Document the assumptions behind any guarantee. If the model excludes a possible failure state or treats an unreliable observation as exact, a formally correct plan can still fail in operation. Update beliefs when new evidence arrives and preserve the distinction between observations and inferences.

Check your understanding

Why does Toggle fail to guarantee Off from the belief set {On,Off}?

Answer: It sends On to Off but Off to On. Since the initial state is unknown, both final possibilities remain. A deterministic action need not remove informational uncertainty.

Related terms

Belief state
Uncertainty
Rational agent
partial observability
sensing action
contingent planning
sensorless planning
observation model

Learn next

Choose a two-state device and list its possible initial states. For each action, compute the set of possible next states. Add an observation model and compare a fixed action sequence with a plan that branches on the observed result.

Sources

The AIMA search chapter explains searching with partial information and belief states: https://aima.cs.berkeley.edu/4th-ed/pdfs/newchap03.pdf

Brafman and Shani study planning with partial information and sensing actions: https://arxiv.org/abs/1401.6048

The lamp calculation and Willow Automation case are original examples. The entry uses the AI planning and agent-knowledge sense of partial information.

Previous: Partial derivative Next: Partial least squares path modeling