Is an unknown battery level a missing state variable?

Not if battery level is already represented and its value is simply uncertain or unobserved.

Should every physical detail be added to the model?

No. Include distinctions needed for relevant decisions and predictions.

Why can adding a Boolean flag be insufficient?

A yes-or-no distinction may still merge situations requiring different actions, as HasSomeEnergy does in the worked example.

AI Glossary / Graphs and symbolic AI / Intermediate

Missing state variable

Definition

A missing state variable is a relevant aspect of the world that an agent's state representation does not include. Because the agent cannot distinguish situations that differ in that aspect, it may predict the same action outcomes or choose the same plan where different behavior is required.

What it means

A state model is a selected description of a situation, not a complete copy of reality. It may record locations, statuses, resources, and other quantities needed for decisions. An omission matters when it changes action availability, outcomes, goals, or costs relevant to the task.
This is different from having a represented variable whose value is unknown. If BatteryLevel exists in the model but has not been measured, the agent has uncertainty about an included quantity. If the model has no battery concept at all, acquiring another location reading will not by itself repair that representational gap.

How it works

An incomplete representation maps distinct real situations to the same modeled state. If those situations behave differently under an action, the model may make incorrect predictions or appear nondeterministic for an unexplained reason.
Repair can require adding a variable, defining its possible values, obtaining information about it, and revising action preconditions, effects, and objectives where relevant. Merely adding a field to a data structure is insufficient if the planner never uses it. The right level of detail depends on the decision: an exact numeric quantity may be necessary in one task, while a smaller set of meaningful categories suffices in another.

Three practical examples

A mobile-robot planner can omit remaining energy and therefore propose a route that its physical battery cannot support.
A workflow model can omit document version, treating an approval for an earlier version as equivalent to approval for the current one.
A stock-handling model can track an item's location but omit quantity, allowing repeated consumption of a depleted supply.

Think of it this way

Think of a map that shows roads but omits bridge-height restrictions when planning a tall vehicle's route. The map may accurately depict the features it includes while still lacking a decisive distinction. The analogy is limited because state variables can change during actions, whereas a printed map often represents relatively static information.

Worked example

Consider an invented robot moving among locations A, B, and C. The initial model records only Location. Each move between adjacent locations consumes real energy, but energy is not represented. The route A to B to C has two moves, and the robot needs 4 energy units per move.
Two real robots are both at A. Robot One has 10 energy units; Robot Two has 6. The location-only model represents both as exactly the same state, Location = A. It therefore accepts the same two-move plan for each.
For Robot One, energy changes from 10 to 6 after the first move and from 6 to 2 after the second. The route is feasible under this simplified energy accounting. Robot Two changes from 6 to 2 after the first move and cannot perform the second move if each move requires at least 4 available units.
Add EnergyRemaining as a state variable. Define a move precondition EnergyRemaining >= 4 and an effect subtracting 4. Now the two starting situations are distinguishable, and the model rejects the second move for Robot Two unless a legitimate recharge action or another feasible option exists.
Suppose instead the model adds only a Boolean HasSomeEnergy. Both 10 and 6 satisfy it, and after the first move both 6 and 2 still satisfy it. That coarse representation remains insufficient for this particular four-unit requirement unless its action rules and state categories are redesigned appropriately.
The numbers are hypothetical and omit all real battery physics. Their purpose is to show how an omitted quantity can affect both applicability and state transitions. The example does not claim that a simple constant-cost energy model is adequate for a physical robot.

Illustrative case study

In a fictional maintenance assistant, Alder Service records whether a machine is inspected but not which configuration was inspected. After a configuration change, the assistant treats the old inspection status as sufficient for a new task.
The team adds configuration identity and connects inspection records to it. It also updates the rules that decide which inspection is relevant. This invented case illustrates a representation repair spanning state and rules, not merely the addition of a database column.

Common misconceptions and mistakes

Not every omitted detail is a modeling error. Abstraction is necessary, and irrelevant details can make reasoning more expensive without improving decisions. The question is whether the omission hides distinctions that matter for the task.
Do not confuse missing variables with missing measurements. Adding a sensor does not help unless its information can be represented and used. Also, a new variable does not automatically restore prediction quality: its transition rules, initial values, and observation uncertainty must be addressed. Sometimes a history-dependent model needs memory or a belief representation rather than one casually added flag.

Business use

Look for repeated failures that occur in situations the current model treats as identical. Compare those situations to identify potentially relevant quantities, statuses, identities, or history. Check whether the proposed addition actually explains the difference rather than merely correlating with it.
Define how the variable is initialized, observed, updated, and used. Revisit stored plans and past state records that lack the new information. Test both a case where the added distinction changes the decision and a case where it should not. Keep the representation as simple as the task permits, while avoiding omissions that make valid and invalid actions indistinguishable.

Check your understanding

Why does the location-only model accept the same route for both robots?

Answer: It maps both real situations to Location = A and has no representation of their different energy amounts. The distinction that makes the second route infeasible is invisible to the model.

Related terms

Abstraction
Belief state
Missing precondition
Missing effect
State representation
state variable
hidden state
partial observability
action model
model revision

Learn next

Choose two real situations that an existing model represents identically. Ask whether the same action has the same relevant outcome in both. If not, propose the smallest additional distinction needed, then specify how it changes observations, action requirements, and state updates.

Sources

Russell and Norvig, Artificial Intelligence: A Modern Approach, indexed textbook passage on missing state variables and incomplete action models: https://aihouse.univ-eloued.dz/wp-content/uploads/2024/11/1-Arti-ficial-Intelligence-A-Modern-Approach.pdf

Previous: Missing precondition Next: Missing values