Is MAGIQ a machine-learning model?

No. It is a structured multi-criteria scoring technique; it need not learn from training data.

Can it use subcriteria?

Yes. A hierarchy can distribute importance across groups and their constituent attributes, with branch weights multiplied during aggregation.

Does ranking preserve all measurement information?

No. Converting measurements to ranks discards the numerical distances between alternatives, which can matter to the decision.

AI Glossary / Mathematics and optimization / Intermediate

Multi-attribute global inference of quality

Definition

Multi-attribute global inference of quality, usually abbreviated MAGIQ, is a decision-analysis method that combines rankings on several attributes into an overall comparative score. It converts attribute priorities and alternative rankings into rank-order-centroid weights, then aggregates them. The resulting score describes a particular comparison model; it is not an independently measured amount of quality.

What it means

Choosing an AI tool can involve competing concerns such as answer quality, response time, accessibility, and operating cost. People may find it easier to rank these concerns than to assign precise numerical trade-offs. MAGIQ supplies a specific numerical convention for turning those rankings into a calculation.
The word global refers to combining the selected attributes into an overall assessment. It does not mean global mathematical optimization or evidence from every possible user. Different stakeholders can legitimately obtain different results when they prioritize different needs. Recording whose preferences the model represents is therefore part of interpreting its answer.

How it works

Arrange the decision criteria in a hierarchy if necessary. Rank the criteria within each group, and rank the alternatives on each lowest-level criterion. For n strictly ordered items, the weight for position k is w_k=(1/n) times the sum of 1/j for j from k through n. These positive weights sum to one.
Multiply weights along each branch of the hierarchy and add an alternative's contributions across branches. With a single criterion level, this is simply a weighted sum of its criterion scores. For a tie, average the weights belonging to the occupied positions. Keep enough precision during calculation to avoid rounding-driven differences.

Three practical examples

A software team can compare candidate AI assistants using documented rankings for task accuracy, ease of review, and response speed.
A training provider can structure a discussion of several learning platforms when stakeholders can agree on ordering but cannot justify precise utility differences.
A product team can inspect how different department priorities affect a shortlist, while keeping mandatory requirements outside the compensating score.

Think of it this way

Think of an awards panel that agrees on the order of judging priorities and then uses a published points schedule. Everyone can inspect how the winner was selected. The analogy stops at the points: a score of twice another score does not establish twice the usefulness, and the schedule cannot correct unreliable judging.

Worked example

Consider an original comparison of three tools, A, B, and C. Accuracy is ranked above speed. Their criterion weights are therefore 3/4 and 1/4. For three alternatives, first, second, and third place receive 11/18, 5/18, and 2/18 respectively.
Suppose the accuracy order is A, B, C, while the speed order is C, B, A. A receives (3/4)(11/18)+(1/4)(2/18)=35/72, approximately 0.4861. B receives 5/18, or 20/72. C receives (3/4)(2/18)+(1/4)(11/18)=17/72, approximately 0.2361. The three totals sum to one, and A ranks first.
Now reverse the priority of the two criteria. The corresponding weights become 1/4 for accuracy and 3/4 for speed. C's score becomes 35/72 and A's becomes 17/72. The evidence about tool performance has not changed; the preferred trade-off has.
If B and C instead tie for second on accuracy, both receive the average of 5/18 and 2/18, namely 7/36. Their combined allocation still equals 7/18. Giving both the second-place weight would allocate extra weight and change the method. Finally, if accuracy measurements differ only slightly, the rank conversion still imposes its prescribed spacing; it does not preserve the size of the original measurement gaps.

Illustrative case study

In a fictional example, Cedar Desk compares three internal document assistants. The support team ranks easy review first, while the engineering team ranks response speed first. Their initial presentations appear to disagree about which assistant is best.
An analyst shows both sets of rankings and reruns the same arithmetic under each priority order. The disagreement becomes a concrete discussion about intended users. Cedar Desk then runs a small task-based evaluation before choosing a pilot. This invented case illustrates transparent preference handling; it does not claim that MAGIQ predicts business outcomes or replaces performance testing.

Common misconceptions and mistakes

Do not describe the overall score as a probability that a tool is good. Normalization to a total of one is an accounting property of the comparison, not probabilistic calibration.
Do not infer that a large score difference establishes statistical significance. Sampling uncertainty in performance measurements needs its own analysis. Similarly, a detailed decimal display does not make subjective rankings precise.
Avoid counting the same advantage repeatedly through overlapping attributes. Separate eligibility requirements from preferences: a tool that fails a mandatory requirement should not qualify merely because attractive features offset that failure in an additive score. Changes to the comparison set also require recalculation and renewed interpretation.

Business use

MAGIQ can support a documented shortlist when collecting exact preference-strength judgments would be difficult. A useful decision record includes the alternatives, criterion definitions, evidence behind performance ranks, stakeholder priorities, treatment of ties, and sensitivity checks.
For an AI purchase, keep measured accuracy, running cost, and operational restrictions available alongside the aggregate score. If two options switch places after a plausible ranking change, present the choice as sensitive rather than declaring a decisive winner. The practical benefit is a clearer discussion of assumptions, not an assurance that the top score is the best eventual purchase.

Check your understanding

Two alternatives tie across the second and third positions in a three-item ranking. Should each receive 5/18?

Answer: No. Under the tie-averaging convention, each receives (5/18+2/18)/2=7/36, preserving the total allocation for those positions.

Related terms

Sensitivity analysis
Multi-objective optimization
multi-criteria decision analysis
rank-order-centroid weighting
analytic hierarchy process
SMARTER
Pareto dominance

Learn next

Rebuild the worked comparison in a small table, then introduce a tie and verify that scores still sum to one. Compare the result with a model that retains actual measurement differences. Explain which information each approach preserves before selecting one for a real decision.

Sources

Primary source: James McCaffrey and Nasa Koski, Competitive Analysis Using MAGIQ, original author article in MSDN Magazine: https://learn.microsoft.com/en-us/archive/msdn-magazine/2006/october/test-run-competitive-analysis-using-magiq . The method description and tie convention were consulted; all numerical comparisons and the fictional case above are original teaching examples.

Previous: Multi-agent collaboration Next: Multi-class