Does ATMS stand for automated truth verification?

No. It means assumption-based truth maintenance system and maintains dependencies within a supplied reasoning model.

Can a conclusion have several supports?

Yes. Its label can contain multiple minimal consistent assumption sets that independently support it.

Does a nogood reject every assumption in the set individually?

No. It rejects their combination; individual assumptions may remain consistent in other environments.

AI Glossary / Graphs and symbolic AI / Intermediate

ATMS

Definition

ATMS stands for assumption-based truth maintenance system. It is a reasoning support mechanism that records which combinations of assumptions support each conclusion and which combinations lead to contradictions. It can track several alternative contexts without declaring that every assumption, or every supported conclusion, is simultaneously true.

What it means

Many reasoning tasks involve possibilities: perhaps a component failed, a delivery route is available, or a configuration option is selected. A conclusion may follow under one combination of assumptions but fail under another.

An ATMS keeps those dependencies explicit. Instead of storing only a single yes-or-no acceptance decision, it associates a conclusion with the assumption sets sufficient to support it. The main problem solver supplies facts, assumptions, and justifications; the ATMS maintains the resulting support information.

Here, truth maintenance concerns consistency and logical dependencies within the supplied model. It is not an independent fact-checking service. Incorrect premises or missing rules can still produce conclusions that are logically supported but unsuitable descriptions of reality.

How it works

Information is represented by nodes. Some nodes are designated assumptions, while justifications describe how combinations of nodes support other nodes. An environment is a set of assumptions considered together.

Each node receives a label containing minimal consistent supporting environments. Minimal means that no listed environment unnecessarily includes another listed supporting environment. It does not mean the most likely explanation or the explanation with the fewest assumptions among all alternatives.

When a justification combines several antecedents, their supporting environments are combined. The system removes redundant supersets and environments containing known contradictions. A contradictory assumption combination is called a nogood; any larger environment containing that combination is also inconsistent.

This representation lets the problem solver ask whether a conclusion follows in a chosen consistent environment. It avoids treating every alternative as a separate independent reasoning session, although the number of relevant support sets can still become large.

Three practical examples

1. A diagnostic reasoner tracks which combinations of possible component faults explain an observation, while excluding combinations contradicted by other measurements.

2. A product configurator records the assumptions behind compatible feature combinations and identifies combinations that violate supplied constraints.

3. A planning prototype preserves conclusions under alternative resource-availability assumptions so that a change in scenario does not erase the distinction between their supports.

Think of it this way

Think of keeping a conclusion on a card with several alternative bundles of supporting conditions attached. Crossing out an impossible bundle need not remove the card if another valid bundle remains.

The analogy captures dependency tracking. Real ATMS labels follow formal rules, and a surviving bundle is not automatically an established real-world explanation or a preferred decision.

Worked example

Let a, b, and c be assumptions. Suppose the justifications say that a supports X, b supports X, c supports Y, and X together with Y supports Z.

Initially, X has the two minimal supports {a} and {b}. Y has the support {c}. Combining the alternatives for X with Y's support gives Z two candidate supports: {a, c} and {b, c}.

Now suppose the problem solver supplies a contradiction showing that a and c cannot consistently hold together. The set {a, c} becomes a nogood, so it is removed from Z's label. Z retains {b, c}.

Neither {a} alone nor {c} alone is ruled out by that nogood. The contradiction concerns their combination. X therefore still has both original minimal supports, while Z remains supported in the consistent environment containing b and c. The example assumes no additional justifications or contradictions.

Illustrative case study

Fictional case: Elm Devices builds a troubleshooting assistant for a training simulator. One symptom can be explained under either of two fault assumptions. A second observation rules out one combination of assumptions but leaves another explanation possible.

The team uses ATMS labels to show which conclusions survive the new constraint. The interface avoids presenting the remaining explanation as a confirmed diagnosis: the simulator's observation model may be incomplete, and further tests may distinguish additional possibilities.

The fictional benefit is a clearer account of conditional reasoning. A reviewer can see which assumption combination supports a conclusion and why another combination was excluded, without confusing those supports with probabilities.

Common misconceptions and mistakes

An ATMS does not automatically choose the correct assumptions. Choosing what to investigate or adopt belongs to the surrounding reasoning process.

A label is not a confidence score. Two supporting environments do not imply twice the probability of a conclusion, and a shorter environment is not necessarily more likely.

Do not confuse an empty label with a label containing the empty environment. An empty label has no consistent support. A label containing the empty environment indicates support without requiring any assumptions, given the supplied premises and justifications.

Finally, compact dependency tracking does not eliminate worst-case combinatorial growth. Scope the modeled alternatives and monitor the size of labels and nogood collections.

Business use

For business applications, ATMS concepts are useful when explanations must preserve alternative assumptions, such as configuration and model-based troubleshooting. They can help make conditional conclusions inspectable.

A practical implementation needs clear ownership of the domain rules and assumptions. Reviewers should be able to distinguish observed information from hypothetical conditions and examine the justification behind a nogood. Test scenarios with overlapping supports and contradictions, because simply demonstrating one successful inference does not verify that alternative contexts are maintained correctly.

Check your understanding

Z has supports {a, c} and {b, c}, and {a, c} becomes a nogood. Must Z lose every support?

Answer: No. The support {b, c} remains if it contains no known nogood. Z can still follow in that consistent environment.

Related terms

Justification
Environment
Dependency-directed backtracking
truth maintenance system
assumption
nogood
abduction
model-based diagnosis

Learn next

Draw the worked example as a dependency graph. Add a second justification for Y and recompute Z's minimal supports. Then introduce a nogood and explain which environments disappear without falsely rejecting each assumption separately.

Justification

Sources

Primary sources consulted: Johan de Kleer, An Assumption-Based TMS, https://dekleer.org/Publications/An%20Assumption-Based%20TMS.pdf ; Raymond Reiter and Johan de Kleer, Foundations of Assumption-based Truth Maintenance Systems, https://cdn.aaai.org/AAAI/1987/AAAI87-033.pdf . The dependency example and fictional case are original.

Previous: Asynchronous policy iteration Next: ATN