AI Glossary / Causal inference and experimentation / Intermediate

Causal link

Definition

In partial-order AI planning, a causal link records that one action establishes a condition needed by another action. It connects a producer action, the supported condition and a consumer action, helping the planner preserve that condition until it is used.

What it means

The phrase causal link can mean a cause-and-effect relationship in everyday language. In the planning sense used here, it has a more specific role: documenting why one action must support another within a plan.
For example, charging a battery may establish the charged condition required by an inspection action. The link says that this instance of charging supplies that precondition for the inspection. It also gives the planner something to protect from intervening actions that would remove the condition.
This is a symbolic planning commitment under an action model. It is not an empirical estimate that charging causes some measured outcome in observational data.

How it works

A planner identifies an unsatisfied precondition of a consumer action and selects an existing or new producer whose effects establish it. It adds a causal link and an ordering constraint placing the producer before the consumer.
Other actions are checked for threats: an action that could occur between them and negate the supported condition may break the link. A planner can resolve a threat by ordering that action before the producer or after the consumer, or through other valid constraint choices in the planning formalism. The complete plan must also satisfy its other preconditions and constraints. A single protected link does not prove that the entire plan is executable.

Three practical examples

• A robot plan links picking up a component to an assembly action requiring that component to be held.
• A workflow planner links approving a document to a later action requiring approval.
• A symbolic logistics plan links loading a parcel into a vehicle to a delivery action that requires the parcel to be onboard.

Think of it this way

Think of a handover note saying, “This earlier step provides the condition this later step depends on.” The note also warns against undoing that condition in between. The analogy explains support and protection, but a causal link is a formal structure checked against action preconditions and effects. It does not automatically handle every real-world delay, uncertainty or resource conflict.

Worked example

Consider three actions. Charge establishes BatteryCharged. Inspect requires BatteryCharged. Drain removes BatteryCharged. The planner creates the causal link Charge - BatteryCharged→ Inspect and requires Charge before Inspect.
If Drain is unordered relative to those actions, the sequence Charge, Drain, Inspect is possible and would invalidate the intended support. Drain therefore threatens the link. One resolution is to order Drain before Charge; another is to order Drain after Inspect, if the rest of the plan permits it.
The ordering Charge before Inspect alone is insufficient because it does not prohibit an intervening deletion. This example assumes a simple deterministic Boolean action model. A real battery system may require numerical energy levels, duration and uncertain consumption, which need a richer representation than a single charged flag.

Illustrative case study

Fictional case: Linden Automation builds a planning prototype for a document workflow. Approval supports publication, but an edit action invalidates approval. The first plan permits editing after approval and before publication.
By representing ApprovalValid as a supported condition, the planner identifies the edit as a threat and requires reapproval or a different ordering. The team checks these rules with process owners, including whether minor edits really invalidate approval. The demonstration shows how causal links expose a dependency in the model. It does not establish that the workflow complies with every external rule or that the symbolic actions capture all operational exceptions.

Common misconceptions and mistakes

A causal link is not merely an arrow saying that one action occurs earlier. It identifies a particular condition supplied to a particular consumer. A threat is a potential conflict under an allowed ordering, not necessarily an action that is always forbidden.
Do not assume that satisfying every local link automatically resolves all resource, timing or global constraints. The action model may omit important effects. Also distinguish this technical planning sense from causal links in statistical diagrams, where arrows encode substantive cause-and-effect assumptions rather than producer - consumer support between plan steps.

Business use

Causal links can make automated workflow and planning systems easier to inspect: reviewers can see which step establishes each required condition and what might invalidate it. This is useful where dependencies need to remain explicit as a plan is revised.
The main implementation challenge is accurate action modeling. Teams should identify preconditions, effects and invalidation rules with domain experts, then test plans against realistic exceptions. If conditions are uncertain or numerical, choose a planning framework that can represent them. A plan that is valid under a simplified symbolic model still needs operational checks before it is relied on in the real process.

People also ask

Is this the same as a causal effect estimate?

No. This entry uses the partial-order planning meaning.

Does a link require adjacent actions?

No. Other actions may occur between them if they do not invalidate the supported condition.

Can one action support several later actions?

Yes, where its effects and the plan constraints allow the required conditions to persist.

Check your understanding

Action A establishes condition P for action B, but action C can delete P between them. What should a partial-order planner investigate?

Answer: C threatens the causal link. The planner must resolve the threat through valid ordering or other permitted constraints, or find different support for B's precondition.

Related terms

• Partial-order planning
• Precondition
• Ordering constraints
• Causal graphs
• action effect
• threat
• ordering constraint
• plan-space search
• symbolic planning
• causal graph

Learn next

Learn partial-order planning and action preconditions first. Draw a small producer - condition - consumer link, then add an action that could undo the condition. Explore which orderings preserve the link and which additional constraints are needed for the whole plan.

• Partial-order planning

More in Causal inference and experimentation

• Causal estimand - A causal estimand is the precisely defined causal quantity that an analysis aims to learn.
• Causal inference - Causal inference is the process of drawing conclusions about how outcomes would change under specified interventions.
• Causal language modeling - Causal language modeling is a training task in which a language model predicts the next token using only the preceding context.
• Causal network - A causal network is a model that represents variables and their causal relationships, commonly using a directed graph together with probability...
• Causal positivity - Causal positivity is the requirement that the treatment choices needed for a causal comparison have positive probability within the relevant covariate...
• Causality - Causality concerns relationships in which changes in one part of a system can produce changes in another.

Sources

Daniel S. Weld, An Introduction to Least Commitment Planning, discussion of plans, causal links and threats: https://www.cs.cmu.edu/afs/cs.cmu.edu/user/reids/www/planning.mmv/readings/weld.pdf AIMA glossary supplied in the worklist: https://github.com/aimacode/aima-glossary/blob/master/glossary.md Primary planning-paper excerpt consulted. Planning sense is explicit; source-glossary alignment and final specialist review remain pending.

Previous: Causal language modeling Next: Causal network