Is one axiom written for each action?

The characteristic organization is one successor-state axiom per fluent, covering relevant actions and conditions.

Does an unrelated action make a false fluent true?

Not under a complete axiom with no positive-effect condition for that action; the fluent stays false.

Does persistence turn unknown into false?

No. If the prior value is unresolved and the action only preserves it, the later value can remain unresolved too.

AI Glossary / Graphs and symbolic AI / Intermediate

Successor-state axiom

Definition

A successor-state axiom is a logical rule that specifies exactly when a changing property is true after an action. It combines the circumstances that make the property true with the circumstances under which an already true property persists. In AI planning and action reasoning, it addresses the representational frame problem: describing what stays unchanged without writing a separate persistence statement for every action-property combination.

What it means

A property whose value can change between situations is called a fluent. Examples include a door being open, a package being loaded, or a task being approved. An action may change some fluents while leaving many others alone.

A successor-state axiom focuses on one fluent across possible actions. Its basic message is: the fluent holds afterwards if and only if the action establishes it, or it held before and the action does not remove it. The “if and only if” matters because the axiom rules out unexplained changes within the stated model, not merely asserts some known effects.

How it works

Identify a fluent F and all modeled conditions under which an action makes it true or false. Combine them into a rule of the form F_after ↔ CausesTrue OR (F_before AND NOT CausesFalse). In situation-calculus notation, the after-situation is often written do(a,s), meaning the situation reached by performing action a after situation s.

Use separate action-precondition axioms to specify whether an action is possible. Add facts about the initial situation and appropriate assumptions about action identities. Then reason forward about consequences or backward about what must have held before a desired outcome.

The rule depends on a complete account of relevant changes. A new action, an indirect effect, or an external event may require revising the axioms. Simultaneous or conflicting effects also need a coherent model; the simple formula should not be treated as an automatic conflict-resolution policy for arbitrary real-world processes.

Three practical examples

Planning lessons: track whether a robot holds an object after pickup, release, and unrelated movement actions in a deliberately simplified domain.
Workflow modeling: state when an approval flag becomes true, when it is revoked, and which actions preserve it, separately from who is authorized to act.
Model checking: identify unexpected changes in a transition system by comparing proposed next states with explicit fluent-update rules.

Think of it this way

Think of a status register with two instructions: explain every way it can be set, and otherwise retain its value unless something explicitly clears it. The analogy is limited because a logical axiom describes all models satisfying the rule; it is not necessarily executable assignment code, and uncertain initial information may leave several possible states.

Worked example

Consider an original toy document model with fluent Approved(s). Its actions are Approve, Revoke, and Read. Assume exactly one action occurs per step, the three action names are distinct, and no other event changes approval. For possible actions, use:

Approved(do(a,s)) ↔ [a=Approve OR (Approved(s) AND a≠Revoke)].

Initially Approved is false. After Read, the first branch is false and the persistence branch is false, so approval remains false. After Approve, the first branch is true, so approval becomes true. After another Read, the first branch is false but the persistence branch is true, so approval remains true. After Revoke, both branches are false and approval becomes false.

For the sequence Read, Approve, Read, Revoke, the successive values are therefore false, true, true, false. The rule avoids needing a separate “Read does not change approval” axiom while still expressing that fact.

Now introduce a new action ExpireApproval that is intended to remove approval. If the old axiom is left unchanged, it will incorrectly preserve an already true approval after that action. The negative-effect condition must be extended, for example by requiring a≠Revoke AND a≠ExpireApproval in the persistence branch.

An additional rule must still determine when Approve is authorized or possible. The update axiom alone does not grant permission to approve anything. This is a teaching model.

Illustrative case study

Fictional case: a team models a document workflow for an AI planning demonstration. It correctly records that an approval action sets a flag but initially says nothing about what happens to the flag after ordinary reading or tagging actions.

The team writes a successor-state axiom to prevent unsupported flag changes, listing revocation and expiry events separately from access-control preconditions. This fictional model is neither a deployed workflow nor a complete real policy.

Common misconceptions and mistakes

An effect axiom saying “Approve makes Approved true” is not by itself a complete successor-state axiom. It leaves open other ways the fluent might change. A successor-state axiom is also not the same as an action precondition: describing an outcome does not establish that the action is allowed.

Do not assume that solving the representational frame problem automatically solves every planning difficulty. Missing initial facts, indirect effects, uncertain outcomes, concurrency, and qualification of actions remain separate modeling concerns. A fluent can remain unknown when its prior value is unknown and an action merely preserves it.

Business use

The business value is precise change semantics. When an automated system reasons about operational status, reviewers should be able to identify every event that creates, preserves, or removes that status. This supports discussions about exceptions and prevents silent assumptions that all unmentioned properties either reset or remain fixed.

Review these rules whenever the event catalog changes. Keep real observations distinct from predicted states and define how discrepancies are handled. A logically consistent transition model can still be incomplete or wrong about the external process it represents.

Check your understanding

Approved is true and the toy model executes Read. What value follows, and why?

Answer: Approved remains true because Read is not Revoke, so the persistence branch of the successor-state axiom holds.

Related terms

Fluent
Situation calculus
Frame problem
effect axiom
action precondition
state transition
persistence
regression
logical planning

Learn next

Write an initial value and successor-state axiom for a simple on/off property. Trace three actions, including one unrelated action. Then add an expiry event and identify which part of the rule must change, keeping action permission separate from state effects.

Sources

Russell and Norvig, First-Order Logic slides, successor-state axioms and the frame problem: https://aima.cs.berkeley.edu/slides-pdf/chapter08.pdf UC Berkeley CS188, Logical Agents, fluent persistence and transition formula: https://inst.eecs.berkeley.edu/~cs188/textbook/logic/logical-agents.html The original examples assume a complete deterministic action description for the toy domain.

Previous: Subwords Next: Sufficient dimension reduction