Is temporal projection the same as planning?

No. Projection asks what follows from a given sequence; planning searches for a sequence that achieves an objective.

Must the result be one fully known state?

No. Incomplete information or nondeterministic effects can leave multiple possibilities or unanswered queries.

Can it reason backwards?

Yes. Query regression can work backwards through actions to conditions on the initial state while answering a forward-consequence question.

AI Glossary / Graphs and symbolic AI / Intermediate

Temporal-projection

Definition

Temporal-projection is the reasoning task of determining what will hold after specified actions or events, given an initial description and a model of change. In symbolic AI, it often means inferring the truth of a property at a later state or time. It is conditional reasoning about consequences, rather than an unrestricted forecast of what the real world will actually do.

What it means

Suppose a model states where a parcel is, which actions can move it, and what those actions change. Temporal projection asks where the parcel will be after a particular sequence. Planning asks a different question: which sequence should be chosen to achieve a goal.

Projection depends on both change and persistence. Knowing that loading puts a parcel in a truck is not enough if the model fails to describe which other facts remain true. Incomplete initial information or uncertain action outcomes can also produce several possible later states rather than one definite answer. The model should distinguish a proven property from one that remains unknown.

How it works

Represent the starting conditions and define actions with preconditions and effects. Specify what persists when an action does not affect a property, addressing the frame problem. In situation-calculus approaches, successor-state axioms can describe when each changing property holds after an action.

For a given executable sequence, reason forward through the transitions, updating the relevant state description. Another approach regresses the desired final query through the actions, producing a condition to test against the initial information.

Check executability separately: an action whose preconditions fail cannot simply be applied as if it succeeded. If the model includes durations, overlapping actions, exogenous events, or probabilistic effects, projection must account for those features. A result follows from the stated initial conditions, action theory, and sequence; omitted real-world events can invalidate its use as an actual prediction.

Three practical examples

1. Logistics simulation: determine the location of an item after loading, vehicle movement, and unloading under a specified action model.

2. Workflow checking: determine whether a record remains approved after a sequence of review, revocation, and unrelated editing actions.

3. Robot planning: assess whether a proposed action sequence leaves the robot holding an object and whether the prerequisites for its next action will hold.

Think of it this way

Think of following a board game's moves from a known starting position. The rules let you determine the resulting board without physically carrying out every move.

The analogy is limited when real systems contain unobserved events, uncertain outcomes, or incomplete rules. A symbolic projection is as reliable as its starting information and transition model. It should not be mistaken for a guarantee that every planned action will occur successfully.

Worked example

Use a simplified parcel model with locations Depot and Shop, one truck T, and parcel P. Initially, At(T,Depot) and At(P,Depot) are true, and In(P,T) is false. Assume the state is fully known and only the actions described below can occur.

Load(P,T,L) requires At(P,L) and At(T,L). It makes In(P,T) true and removes At(P,L). Drive(T,L,M) requires At(T,L); it replaces that fact with At(T,M). It does not change In(P,T). Unload(P,T,L) requires In(P,T) and At(T,L); it removes In(P,T) and adds At(P,L).

Project the sequence Load(P,T,Depot), Drive(T,Depot,Shop), Unload(P,T,Shop). After loading, P is in T and has no direct At fact. After driving, T is at Shop and P remains in T. After unloading, At(P,Shop) is true and In(P,T) is false.

During the journey, a derived location rule can say that a parcel in a truck is located wherever the truck is. That is different from storing an independent At(P,L) fact while loaded. The chosen representation prevents two conflicting location records.

If unloading were attempted at Shop before driving there, its precondition At(T,Shop) would fail. Projection must flag the non-executable sequence rather than silently place the parcel at Shop. All conclusions are conditional on this toy model's deterministic effects and persistence rules.

Illustrative case study

Fictional case: FlowProof evaluates a proposed approval workflow. The model says that a revocation action removes approval, while a note-editing action preserves it. An early implementation clears every unstated property after each action, making approvals disappear when notes are edited.

The team corrects the persistence representation and checks both approval-changing and unrelated actions. It then tests a sequence whose required approval is absent, verifying that the action is reported as inapplicable. The case illustrates modeling and executability checks, not a validated production workflow.

Common misconceptions and mistakes

Projection does not choose the best plan. It evaluates consequences of specified actions; a planner may use many such evaluations while searching.

An action's intended effect is not evidence that its preconditions hold. Check whether the sequence is executable.

Do not assume every unmentioned fact becomes false, or that every fact persists, without an appropriate representation convention. Both shortcuts can yield incorrect conclusions.

Unknown initial information need not become known merely because time passes. Also distinguish chronological time from action order: a state-transition sequence does not automatically specify durations or concurrent behavior.

Business use

Temporal projection helps teams inspect proposed processes before execution and identify missing dependencies or inconsistent action descriptions. It can support scenario analysis in logistics, workflow design, and autonomous systems.

Document what the model excludes, such as unexpected arrivals, failed actions, or concurrent edits. Compare projected outcomes with representative real or simulated traces when available. Preserve the starting-state version and action sequence used for each result. When execution diverges from the model, update the state estimate and project again rather than continuing to rely on an obsolete consequence calculation.

Check your understanding

Why does P remain in T after the drive action?

Answer: The drive changes the truck's location but does not change the parcel-in-truck relation. The model explicitly preserves that unaffected property.

Related terms

Learn next

Study the frame problem and successor-state axioms next. Add a failed-drive possibility to the example and list the resulting possible states. Then decide whether the final parcel location is guaranteed, merely possible, or unknown under that expanded action model.

Frame problem

Sources

Primary reference: Gruninger and Fox, “The Logic of Enterprise Modelling,” section 3, https://www.eil.utoronto.ca/wp-content/uploads/enterprise-modelling/papers/gruninger-mmei96.pdf. The authors' temporal-projection and situation-calculus discussion was read directly. The parcel model and workflow case are original teaching examples with explicit assumptions and hand-checked transitions.

Previous: Temporal normalization Next: Temporal redundancy