Glossary¶
Three optimizations are nested in this project, and one word can mean something different in each. The decision layer is the decision-maker's choice of a design: its words are objective (a metric the decision-maker wants more of), constraint (a requirement's limit) and the Pareto notions. The design layer is the designer's optimization of a configuration, inside which the fitted model appears as a function of the metrics: generalized geometric program and log-convex belong here. The method layer is the fit, an optimization over the exponents: its word is loss, the function the solver minimizes, which is scikit-learn's word. The words loss, metric, attribute and objective mean different things in decision analysis, in optimization and in machine learning. This glossary gives each one meaning, names the layer and the register of every entry, and the surfaces hold to it: plain words on the intuitive surfaces, the caveats in the entries and the detailed sections.
A¶
-
active face: The part of the simplex where the exponents that are zero at the fit stay zero; the fit moves only along it.
-
assignment: The term that scores a design highest; reading it off the fitted model is analysis. Only the soft fit (finite term temperature) weighs every term against every design. At the default (infinite), the assignments drive the refit.
-
assignment margin: The smallest gap by which a design's term beats the runner-up; positive means each design has one maximizing term.
-
attributes: What the evaluation returns (seconds, kilograms); raw, in mixed units and directions.
B¶
-
balanced point: At the exact-maximum fit, the mean of the log-metric rows under the optimal dual weights p; its components are equal on the support of the exponents.
-
baseline regime: The second regime of the case: the designs drawn with the proportional gain Kp from 0.03 to 1 (the stock gains), dataset cubesat-stock.
-
binding term: The term whose piece the surface follows at a point (the supervisor: "binds" in the infinite-temperature sense); at a finite temperature the term carries the largest weight.
C¶
-
cell: The designs assigned to one term, those whose largest term it is; the cells partition the rows, one per term.
-
certificate: That the optimizer reached a local minimizer of the loss; for a soft fit, that the second-order test holds.
-
characteristic point: A term's characteristic point is the point of the surface whose shares of the metrics are the term's exponents, where the term is largest for a given total.
-
chart: The reversible transformation between metric space and the fitting space; nothing is lost either way.
-
check: Evaluating a proposal once and comparing the measured chart point with the predicted one. Three outcomes: confirmed, refuted (a row the table lacked, so the fit is rerun) or infeasible.
-
conditioning: The analyst's step that turns attributes into metrics; LogAffinePreconditioner does it.
-
conditioning of the exponents: The smallest eigenvalue of the restricted Hessian at each term; small means a nearly flat direction.
-
configuration: The design variables the engineer chooses (wheel class, gains); carried as covariates, not fitted.
-
constraint: A requirement's limit on a measure: a design meets it or does not. The fit never sees constraints; they are read after it.
-
contending set: The metrics with positive exponent in the exact-maximum fit; the members of the n-lemma.
D¶
- design tensions: The tension formula at one design's own metric ratios, on the term it belongs to; from the data. In attribute units they carry the conditioning.
E¶
-
effective sample size: The inverse sum of squares of the support weights; returned as n_eff_.
-
elasticity: Within a term, the percentage of one metric given up per percentage gained in another.
-
energy differential: The log of the invariant value: a design's level relative to the fitted reference surface, by analogy with a potential; negative inside, positive beyond; returned by score_samples.
-
exponent vector: The weights of term k on the simplex over the metrics; its positive entries name the metrics that contend.
F¶
-
fit temperature: How hard the maximum is in the one-sided fit; infinite is the exact maximum, finite gives a band of near-frontier designs.
-
fitted reference surface: The level V = 1 of the fitted invariant at finite temperature, a model of the data underlying the table; rows above it exist by construction.
-
frontier: Everyday word for one of three objects; the pages name them as the fitted reference surface, the supporting envelope or the Pareto boundary.
G¶
-
generalized geometric program: A program that minimizes a generalized posynomial under generalized-posynomial inequality constraints and monomial equality constraints (Boyd et al. 2007, section 5.4); convex after x = exp(y).
-
generalized posynomial: A function built from posynomials by sums, products, positive powers and maxima (Boyd et al. 2007); the invariant V_K belongs to this class.
-
geometric program: The minimization of a posynomial under posynomial inequality and monomial equality constraints; convex after x = exp(y).
H¶
- hyperbolic: The level curves of V, products of powers of the metrics held constant, the rectangular hyperbola in two metrics. Not negative-curvature geometry, and not the graph measure of Färe, Grosskopf and Lovell.
I¶
-
improvement factor: The factor on one metric alone that brings a design to V = 1.
-
invariant value: A weighted geometric mean of the metrics; doubling every metric doubles it, so it reads as a factor of improvement.
L¶
-
launch cost: The launch rate, an illustrative USD per kilogram, times the total mass; the launch-budget requirement reads it. It is not a traded metric and not a dataset column.
-
level offset: Another name for the energy differential, read as the offset from the level V = 1.
-
log-convex: Convex as a function of the log-metrics u = log m; the fitted reference surface is log-convex, so no number of terms represents a dent or a hole.
-
log-metrics: The metrics in log space, where every fit happens and a monomial becomes a hyperplane.
-
loss: The function the solver minimizes: the soft maximum of the scores at one term, the mean slack at several; scikit-learn's word. We choose it for the method layer.
M¶
-
mean slack: The mean shortfall of the designs below the fitted reference surface, F(W) = -(1/N) sum_i Phi_K(u_i), returned as mean_slack_. The several-term fit minimizes it; it falls with K by construction and cannot choose K.
-
metrics: The conditioned attributes: strictly positive, larger is better; the input to the fit.
-
MoE: Measure of effectiveness: a measure of how well a system serves its purpose that a stakeholder states without knowing the design; here the pointing error and the launch cost.
-
monomial: A positive coefficient times the metrics raised to real powers; one term of a posynomial, a hyperplane in log-metrics.
-
monomial term: One monomial of the posynomial: the model object, one trade-off; in log space one flat piece of the fitted reference surface.
-
MoP: Measure of performance: a measure of a system's performance under the scenario the engineer states; here the settling time, the total mass, the control effort, the power margin, the energy balance, the peak slew power, the depth of discharge.
N¶
-
n-lemma: How many of the d metrics contend in a term; a dilemma is 2, a trilemma 3; found from the data, not assumed.
-
near-frontier band: The designs with log V strictly between -1/tau and 1/tau. The fit weighs every design in its soft maximum, and a design inside the band weighs within a factor e of one on the surface.
-
novelty: How far a design lies beyond the fitted reference surface; a score for white space and new capabilities.
-
number of designs: The rows of the table; each row is one evaluated configuration.
-
number of metrics: The length of the metric vector; the dimension of the trade space.
-
number of terms: How many monomials the frontier has; the analyst's choice, never selected by the fit.
-
numerically zero: An exponent at or below 1e-12; the code treats it as zero, which is not the same as mathematically zero.
O¶
- objective: A metric the decision-maker wants more of; the objectives are a subset of the metrics, and a metric need not be one.
P¶
-
Pareto boundary: The nondominated points of the attainable set, a property of the system the data sample.
-
participation threshold: A working-temperature exponent above 1e-3 counts as participating; the threshold is a parameter of trade_off_report.
-
position along the frontier: The other d - 1 chart coordinates, free of overall scale; where in the trade space a design sits.
-
posynomial: A sum of monomials; the fit's K-term sum is one, and the fitted reference surface is where that sum equals one.
-
proposal: A configuration chosen for a chart point that has no datum, with the chart point predicted for it and a tolerance. A conjecture until it is checked.
R¶
-
reach: How far the data extend in direction w_k: the soft maximum of their scores along it, the largest score at infinite temperature.
-
region: In the ternary plot, the set of positions where one term is the largest; an analysis of the fit, not a label on a design.
-
ridge: A penalty on the squared size of the exponents added to the loss. Noise on the metrics adds one, and the penalty pulls the fitted exponents toward equal weights.
-
row share: The fraction of all designs in a term's cell; it weights the term gaps in the termination record's bound.
S¶
-
selection rule: How one configuration is picked for a chart point when many are predicted near it, because the way back from the chart to the design space is not unique.
-
share: A position in the triangle: a metric's fraction of the total of the three.
-
stable radius: A perturbation of the terms smaller than this leaves every design's assignment unchanged.
-
support weights: The softmax weights of the designs in the fit; which designs hold up the fitted reference surface.
-
supporting envelope: The fit at infinite temperature, the exact maximum; a hyperplane that supports the observed rows from above.
T¶
-
tension: How much one metric changes when another rises by a unit along the surface, with the others held, at a point; negative where both are carried by the term.
-
tension matrix: The tensions of every ordered pair of metrics along one term, at the term's balanced point; from the fit.
-
term gap: The first-order gap of a term's refit; zero exactly when the optimality conditions hold.
-
term share: The fraction of the near-frontier designs in a term's cell (term_share_), showing which terms carry most of the frontier. It depends on the band, and it gives the order of the terms, largest first.
-
term temperature: The temperature of the soft weights over the terms in the several-term fit; infinity is the sharp limit of one label per design.
-
termination status: Which level of guarantee the returned run meets: certified, stationary without margin, uncertified, guard stop, iteration cap, not applicable, or stationary (soft fit only).
-
ternary plot: A triangle whose corners are three metrics: a design sits by its shares of their total, tinted by region, with a numbered marker for each term's exponent vector. For reading, not for measuring distance.
-
trade-off: Metrics that cannot all rise together: raising one costs another. The plain word; its local measure is the tension, its constant-per-term form the elasticity.
W¶
- white space: A region of the chart where the fit speaks and no evaluated design is; the one word for it.