Must a classical planner use one particular search algorithm?

No. Several algorithms can solve the same modeled planning problem.

Is the first valid plan optimal?

Not necessarily. Optimality depends on the objective and the algorithm's guarantee.

Can classical planning support real applications?

Yes, when the abstraction is appropriate and execution handles the differences between the model and the environment.

AI Glossary / Graphs and symbolic AI / Intermediate

Classical planning

Definition

Classical planning is the task of finding actions that transform a known initial state into a state satisfying a goal, under a simplified model with deterministic actions and a fully known state. Its usual setting is finite and discrete, with no outside changes occurring independently of the modeled actions. It is a foundational form of symbolic AI planning.

What it means

A planner receives descriptions of what is true now, what must become true, and what actions are available. It must work out an order that makes each action possible and ultimately achieves the goal. The answer is a plan rather than merely a prediction or a natural-language suggestion.
Classical describes the assumptions of the model, not the age or usefulness of the software. Real tasks can sometimes be represented effectively at this level of abstraction. However, uncertain outcomes, hidden information, concurrent external changes, and detailed timing may require richer planning models or monitoring around the classical planner.

How it works

Represent the state using logical facts. Describe each action by its preconditions and effects. Preconditions say when the action can be applied; effects specify how the state changes. A planner searches for a valid progression, or reasons backward from the goal, using the structure of these descriptions to guide its search.
In a simple STRIPS-style representation, applying an action removes its delete effects and adds its positive effects; other facts remain unchanged. A candidate plan must be checked step by step from the initial state. A valid plan reaches the goal, while an optimal plan additionally minimizes a specified measure such as action count or cost. The two requirements should not be confused.

Three practical examples

A simulated warehouse problem plans how a cart can collect a parcel and leave it at a destination under known movement and loading rules.
A software demonstration plans a sequence of configuration changes when each change has explicit prerequisites and predictable modeled effects.
An educational blocks-world task finds legal moves that arrange blocks into a requested pattern without violating support constraints.

Think of it this way

Think of solving a board puzzle whose pieces, legal moves, and goal arrangement are all visible. You can mentally apply moves before performing them. The analogy fits classical planning's known-state assumptions, but most business environments are less tidy: people interrupt, actions fail, and information can be stale. A valid puzzle plan is therefore not a guarantee of real-world execution.

Worked example

Consider two locations, A and B, one cart, and one parcel. Initially the cart and parcel are at A, and the cart is empty. The goal is for the parcel to be at B. There are three action types: load a co-located parcel into an empty cart; move the cart between the locations; and unload its carried parcel at the cart's current location.
A valid three-action plan is Load at A, Move A to B, Unload at B. Loading removes ParcelAtA and EmptyCart and adds CarryingParcel. Moving removes CartAtA and adds CartAtB while preserving CarryingParcel. Unloading removes CarryingParcel and adds ParcelAtB and EmptyCart.
Moving first would leave the cart at B without the parcel, so loading there would fail its co-location precondition. The three-action plan is shortest in this tiny model because reaching the goal requires loading, transporting, and unloading, and no action combines those changes. This conclusion depends on the specified action set; adding a direct delivery action would change the problem.

Illustrative case study

At fictional Maple Automation, a team prototypes an internal setup assistant using a classical planning model. Each setup operation lists the resources and settings it requires. Reviewing the generated sequence reveals a missing precondition: one operation also needs a resource that the model had treated as irrelevant.
The team adds the condition and validates candidate plans against the revised model. It also checks actual state before execution because the workplace can change after planning. This fictional case demonstrates why model review and execution monitoring matter; it does not claim that the prototype achieved a measured efficiency improvement.

Common misconceptions and mistakes

A fluent paragraph containing steps is not automatically a classical plan. The steps need formal meaning, and their preconditions must hold when reached. Another mistake is to assume that achieving subgoals separately always combines into a solution: an action for one goal can undo another.
Deterministic does not mean cost-free or easy to solve. It means the modeled outcome of an applicable action is fixed. Fully observable also concerns the planning state, not whether every physical detail has been measured. Finally, a plan can be valid within an incomplete model and still fail in practice. Model validity and real-world adequacy are separate questions.

Business use

Classical planning can help make a workflow's dependencies explicit and generate alternative action sequences. Businesses should define success conditions and action costs before asking for the best plan. Review the assumptions about resource availability, outside changes, and action reliability. If those assumptions do not hold, use appropriate extensions or execution checks rather than presenting the classical plan as an unconditional promise. Keep the final human-readable explanation connected to the actual validated actions.

Check your understanding

Why is Move A to B followed by Load invalid in the worked example?

Answer: The parcel remains at A after moving, so the cart at B is not co-located with it. The load action's precondition is false.

Related terms

Precondition
State-space search
Automated planning
STRIPS
effect
planning heuristic
plan validation
contingent planning
action cost

Learn next

Learn preconditions and effects next, then write the cart example as explicit state transitions. Change one action to have an uncertain outcome and identify exactly which classical assumption no longer holds.

Sources

Primary sources: Russell and Norvig, original planning chapter, https://aima.cs.berkeley.edu/newchap11.pdf ; Garrett, Lozano-Perez, and Kaelbling, Combined Task and Motion Planning as Classical AI Planning, https://arxiv.org/abs/1706.06927 . The parcel model and fictional case are original teaching examples.

Previous: Classical decomposition method Next: Classification