How does progression differ from regression?

Progression starts with the initial state and moves forward. Regression reasons backward from goal conditions toward requirements that the initial state can satisfy.

Does progression require a fully prebuilt state graph?

No. Successors can be generated from the action representation during search.

Can progression find an optimal plan?

Yes, with a suitable search method and assumptions, but the word “progression” alone does not guarantee it.

AI Glossary / Graphs and symbolic AI / Intermediate

Progression planning

Definition

Progression planning searches forward from an initial state by applying actions whose preconditions are satisfied and calculating their effects. It continues until it reaches a state satisfying the goal. It is also called forward state-space planning. The term describes the direction and representation of planning, rather than one particular search algorithm or a promise of optimality.

What it means

A progression planner asks: “Given what is true now, what could I do, and what would then be true?” It builds candidate action sequences from the beginning. Each search node represents a modeled state after some sequence of actions.
In classical planning, the initial state and action effects are known, and an action has a deterministic outcome. A goal description need not specify every fact in the final state; it states the conditions that must hold. The planner can generate reachable states as needed rather than constructing the complete graph in advance. Searching forward through a model is not the same as executing every considered action in the real world.

How it works

Place the initial state on the search frontier. Select a state using a chosen search strategy, test whether it satisfies the goal, and identify applicable actions. Compute each successor by applying the corresponding effects. In a basic STRIPS representation, remove the action's delete effects and add its add effects, retaining unaffected facts.
Record enough information to recover a plan when a goal is reached. Repeated-state handling can prevent cycling, while heuristics can prioritize promising states. Completeness and cost optimality depend on the search strategy and its assumptions. Forward search may consider many actions that are applicable but unhelpful, so applicability alone is not a measure of usefulness.

Three practical examples

A simulated warehouse planner can search through pick, move, and place actions until all requested items reach their destinations.
A document-production planner can apply drafting, checking, and release actions while tracking which prerequisites have been met.
A puzzle solver can expand legal moves from its starting arrangement and stop when the desired arrangement appears.

Think of it this way

Think of writing a sequence of possible next steps on copies of a checklist, updating each copy after its proposed step. You explore those copies before choosing a complete sequence. The analogy is limited because a real planner may store states compactly and evaluate enormous numbers of alternatives; it also depends on an accurate model of what each action changes.

Worked example

Consider an invented packing task with initial facts BoxOpen, ItemLoose, and LabelReady. All unlisted facts are false in this simple closed-world model. The goal is BoxSealed and LabelAttached. There are three unit-cost actions.
Pack requires BoxOpen and ItemLoose; it adds ItemInside and deletes ItemLoose. Seal requires BoxOpen and ItemInside; it adds BoxSealed and deletes BoxOpen. AttachLabel requires BoxSealed and LabelReady; it adds LabelAttached and deletes LabelReady.
At the initial state, only Pack is applicable. Its successor contains BoxOpen, ItemInside, and LabelReady. Seal then produces BoxSealed, ItemInside, and LabelReady. AttachLabel produces BoxSealed, ItemInside, and LabelAttached, satisfying both goal conditions. The plan has cost 3.
This plan is shortest in the stated model because each of its three actions is necessary: only AttachLabel establishes LabelAttached, it needs BoxSealed, and Seal requires ItemInside. Applying Seal at the beginning is invalid, even if it sounds useful. Applying Pack must also remove ItemLoose; forgetting that delete effect would misrepresent the state.

Illustrative case study

Fictional case: Wren Packaging builds a planner for a training simulator. An early action model lets sealing add BoxSealed without deleting BoxOpen. A later rule consequently allows the simulated system to add another item to an already sealed box.
The team repairs the action effect and replays the proposed plans state by state. It also distinguishes a plan generated in the simulator from instructions approved for physical equipment. The fictional case illustrates how progression makes state updates inspectable, while still depending on the quality of the action model. No measured operational improvement is claimed.

Common misconceptions and mistakes

Do not confuse progression with greedily choosing whatever seems best now. It can be combined with breadth-first, uniform-cost, or heuristic search, among other strategies. Do not assume the first plan found is the cheapest unless the chosen algorithm provides that guarantee under the relevant conditions.
Ignoring delete effects may be useful inside a relaxed heuristic, but it does not produce valid actual successor states for the original task. Keeping only the current location can also lose necessary information about resources and completed actions. Finally, a deterministic action sequence may be insufficient when observations are incomplete or action outcomes are uncertain.

Business use

Progression planning is attractive when a team can describe the current situation and action consequences clearly. Each intermediate state provides a concrete basis for discussing why a step is possible and what it accomplishes. The returned sequence can be checked independently against the model.
In a business prototype, start with a limited workflow and explicit prerequisites. Review both successful plans and states where no action is available. Measure search effort as the number of possible actions and state variables grows. A useful heuristic can reduce effort, but it should not conceal missing constraints or turn an unverified model into an apparently authoritative operational plan.

Check your understanding

Why can AttachLabel not be applied immediately after Pack in the example?

Answer: Pack establishes ItemInside but does not establish BoxSealed. AttachLabel requires BoxSealed, so Seal must occur first. The planner checks the facts in the current successor state, not merely whether the desired effect would be helpful.

Related terms

Regression planning
Precondition
Heuristic search
Classical planning
Forward search
state-space planning
STRIPS
add effect
delete effect
plan validation

Learn next

Learn regression planning and delete relaxation next. Add an action that prepares a missing label, then revise the initial state so LabelReady is false. Find a valid forward plan and explain which action orderings are required and which can vary without changing success.

Sources

David Poole and Alan Mackworth, Artificial Intelligence: Foundations of Computational Agents, Forward Planning: https://artint.info/3e/html/ArtInt3e.Ch6.S2.html Stuart Russell and Peter Norvig, AIMA planning chapter: https://aima.cs.berkeley.edu/2nd-ed/newchap11.pdf

Previous: Programming by example Next: Projected gradient descent