Skip to content

Overview

The idea in use

You have evaluated many designs, and each came back with a handful of numbers: mass, power, settling time, pointing error. You want the shape of the trade: which numbers pull against each other, at what elasticity, where the evaluated designs end, and where the table is empty.

The package answers with trade-offs, not variance. Principal component analysis (PCA) finds the directions in which the data spread; this method finds the positive combination of the metrics along which the data stop, the binding constraint among metrics that are all better when larger. The answer is one weight per metric, and the weights sum to one. The metrics with positive weight in the exact-maximum fit (tau = inf) are the ones that contend. A small weight at the fit temperature is weaker participation. The data reveal them; you need not know in advance which dilemma or trilemma, two or three contending metrics, your problem holds.

The weights define a score V, a weighted geometric mean of the metrics. After a one-time normalization V = 1 is the fitted reference surface, V below 1 is inside it and V above 1 is beyond what the table reached.

The pipeline

A study has five stages and the package covers the last three:

  1. Sample configurations, by any scheme.
  2. Evaluate each into attributes, with a simulator, a test or an assessment. The package never sees how: any table works.
  3. Condition the attributes into metrics: positive, larger-is-better, comparable spread.
  4. Fit HyperbolicTradeSpace to the metrics.
  5. Read: the trade-off report and the tensions, the chart, a prediction from configuration.

What you supply, what you get back

You supply a table of designs and the rule that turns attributes into metrics. You get back:

  • the weights and the metrics that participate;
  • the elasticities between contending metrics;
  • for each design, its level offset to the fitted reference surface and its position along the surface;
  • how much one metric alone would need to improve to reach the surface;
  • a novelty score and candidate white space, the regions no design occupies;
  • the surface written back in your original attribute units;
  • a stopping record, termination_ and conditioning_, saying how far to trust the fit.

Where the method misleads is on the limitations page. Read it before acting on a result.