Can an action achieve its main goal and still have a missing effect?

Yes. It may also cause an unmodeled resource or status change.

Is every unexpected outcome a missing effect?

No. Execution failure, sensing error, or an external event can also explain a discrepancy.

Why does the label example not primarily describe a missing precondition?

The label requirement is already present. The omitted consequence is that shipping consumes the label.

AI Glossary / Graphs and symbolic AI / Intermediate

Missing effect

Definition

A missing effect is a real consequence of an action that is absent from the agent's action model. The agent may correctly predict the action's intended result while overlooking another change it causes. Plans built from that incomplete model can therefore disagree with what happens during execution.

What it means

An action model describes when an action can be used and what changes afterward. If the model of moving a container says where it arrives but omits that its previous location becomes empty, the model can preserve a fact that should no longer hold. Missing effects can involve changes that add facts, remove facts, or alter numerical quantities.
This entry concerns omitted consequences in planning and model-based reasoning. A missing effect is different from an action that failed to produce an effect already represented in the model. It is also different from an unrelated event occurring in the environment at the same time.

How it works

A planner predicts a next state by applying modeled effects to the current state. In a common STRIPS-style representation, it removes facts in an action's delete list and adds facts in its add list. Facts not changed by the model are generally carried forward.
If an actual consequence is omitted, that carry-forward assumption can be wrong. Monitoring may expose a discrepancy between predicted and observed state. Repair can involve updating the state estimate, revising the action model, and reconsidering later actions. A single discrepancy does not establish its cause; observations, execution faults, and external events also require consideration.

Three practical examples

A delivery model can omit that using a disposable shipping label consumes it, leading to plans that reuse one physical label.
A robot model can include arrival at a new location but omit departure from the previous one.
An editing workflow can model a successful content change while omitting that the change invalidates an earlier approval status.

Think of it this way

Think of following a checklist that records the main outcome of a task but leaves out a side effect affecting the next task. The checklist can look internally coherent while the real situation changes differently. The analogy is limited because formal action models specify state transitions precisely, and diagnosing an omission requires evidence rather than merely noticing an unexpected result.

Worked example

Consider an invented packaging model. State variables already include LabelAvailable, PackedA, PackedB, and ShippedA and ShippedB. Initially LabelAvailable, PackedA, and PackedB are true; neither package is shipped. There is exactly one physical disposable label.
The incomplete action ShipA requires LabelAvailable and PackedA and adds ShippedA. ShipB similarly requires LabelAvailable and PackedB and adds ShippedB. Both action models omit the effect that consumes the label, so neither removes LabelAvailable.
The planner predicts that ShipA followed by ShipB works. After ShipA, its predicted state still has LabelAvailable, so ShipB appears applicable. In the real process, shipping A uses the only label. The actual state after ShipA has ShippedA but not LabelAvailable, making ShipB unavailable until another label is obtained.
Repair the action models by including deletion of LabelAvailable in each shipping action. Now, after ShipA, the model itself rejects immediate ShipB. If an authorized ObtainLabel action is available, the repaired plan can be ShipA, ObtainLabel, ShipB. This three-action sequence is only valid if ObtainLabel's preconditions and effects are correctly represented.
The example is specifically a missing-effect problem because LabelAvailable is already a modeled variable and the actions already check it as a precondition. The missing information is the change caused by shipping. If label availability had no representation at all, the diagnosis would involve a missing state variable as well.
The arithmetic is simple: one label permits one shipment before replenishment, while two shipments require two labels in total. The faulty model implicitly treated a consumable as reusable.

Illustrative case study

In a fictional publishing project, Cedar Pages marks a document approved, then permits a substantive edit without changing that status. An automated release plan assumes the approval still applies to the new version.
The team identifies the intended process rule that relevant edits invalidate approval and adds that consequence to the workflow model. It also revisits plans built on the stale status. This invented case illustrates a missing effect in a digital process; it does not prescribe a universal approval policy for every organization.

Common misconceptions and mistakes

A missing effect is not necessarily a harmful side effect. An omitted beneficial change can also make a model incomplete. The issue is whether a relevant actual consequence is represented accurately.
Do not automatically add every observed change to an action's effect list. Another actor or an unrelated event may have caused it. Also, correcting the current state without correcting the action model can allow the same planning error to recur. Conversely, revising the model without checking existing plans can leave decisions based on outdated predictions in place.

Business use

Review action models against observed transitions and process requirements. For each action, ask what becomes true, what becomes false, and which resources or statuses change. Consumables, location exclusivity, and invalidated approvals are useful areas to inspect.
When execution differs from prediction, preserve enough information to distinguish model omission from failed execution or observation error. Update the model only with justified consequences and validate the revised behavior on representative cases. Communicate whether the repair concerns the current state, a general action definition, or both. These are different changes with different implications for future plans.

Check your understanding

Why does adding a delete effect for LabelAvailable prevent the invalid two-shipment plan?

Answer: After the first shipment, the predicted state no longer satisfies the second shipment's label precondition. The planner must obtain another label or choose a different valid plan.

Related terms

Delete list
Add list
Execution monitoring
Missing precondition
Missing state variable
Action model
effect
state transition
model revision
replanning

Learn next

Choose one ordinary workflow action and list its intended result, resource changes, and status changes. Compare those consequences with the current model. Then create a two-action test in which an omitted consequence would make the second action appear possible when it is not.

Sources

Russell and Norvig, Artificial Intelligence: A Modern Approach, planning and acting in the real world, indexed textbook excerpt on missing effects: https://aihouse.univ-eloued.dz/wp-content/uploads/2024/11/1-Arti-ficial-Intelligence-A-Modern-Approach.pdf Original AIMA planning chapter, action representations: https://aima.cs.berkeley.edu/newchap11.pdf

Previous: Missing data Next: Missing precondition