Salas 1 (130)

What is the fundamental goal of DoE in AI?

To systematically determine the relationship between input factors and model performance.

Define "Factorial Design"

An experimental strategy where all combinations of factor levels are tested.

What does "2^3 design" signify?

An experiment with 3 factors, each tested at 2 levels.

What is a "Lurking Variable"?

An unmeasured variable that influences both the input and the response, potentially creating a false correlation.

Define "Experimental Error"

The variation in the response variable that cannot be explained by the factors in the design.

What is "Nuisance Variation"?

Variability that comes from sources we know exist but aren't the primary focus of the study.

What is the "Unit of Abstraction" in AI DoE?

It can be a single inference, a training epoch, or an entire dataset split.

Define "Homoscedasticity"

The assumption that the variance of the residual errors is constant across all levels of the factors.

What is "Independence of Errors"?

The requirement that the error in one experimental run does not influence the error in another.

What does the "Intercept" represent in a linear DoE model?

The average response value when all coded factors are at their mid-point (zero).

Define "Main Effect of Factor A"

The change in response produced by a change in the level of Factor A, averaged over the levels of other factors.

What is a "Positive Interaction"?

When the combined effect of two factors is greater than the sum of their individual effects.

What is a "Negative Interaction" (Antagonism)?

When one factor reduces the effectiveness of another factor on the response.

Why are "Coded Units" (-1, +1) used?

To compare factors with different physical scales (e.g., Temperature vs. Learning Rate) on a fair basis.

What is the "Red Dashed Line" on a Pareto Chart?

The t-limit or Bonferroni limit for statistical significance at a chosen alpha level.

If a bar in Pareto does not cross the limit, what should you do?

Consider removing that factor or interaction to simplify the model.

Define "Model Sparsity Principle"

The idea that a system is usually dominated by a few main effects and low-order interactions.

What is "Hierarchical Ordering"?

The principle that if an interaction (A*B) is significant, the main effects (A and B) should remain in the model regardless of their p-values.

Define "Response Surface Methodology" (RSM)

A collection of mathematical techniques used to find the optimal settings for a process.

What does a "Curved Contour" in a contour plot indicate?

The presence of interaction or quadratic (non-linear) effects.

What is a "Steepest Ascent" path?

The direction in which the response variable increases most rapidly.

Define "Robust Design"

A design that identifies factor settings where the response is least sensitive to noise.

What is "Degrees of Freedom for Error"?

The number of independent observations minus the number of parameters estimated.

Define "Standard Deviation of the Units"

A measure of the spread or "noise" within the experimental measurements.

What is "Blocking Factor"?

A factor used to group experimental units that are similar to reduce known noise.

Define "Randomized Complete Block Design" (RCBD)

A design where every treatment is present in every block, and the order within blocks is random.

What is the main risk of not randomizing?

Confounding the effect of a factor with a time-dependent trend.

Define "Replicate" vs "Repeat"

A replicate is a full reset of the experimental setup; a repeat is just a second measurement of the same setup.

Why is Replication better than Repeating?

It captures the true "setup-to-setup" variability.

What is "Confounding"?

When the effect of one factor is indistinguishable from the effect of another factor or interaction.

Define "Resolution III Design"

A design where main effects are confounded with 2-factor interactions.

Define "Resolution IV Design"

A design where main effects are clear of 2-factor interactions, but 2-factor interactions are confounded with each other.

Define "Resolution V Design"

A design where both main effects and 2-factor interactions are clear of each other.

What is a "Screening Experiment"?

An initial experiment with many factors used to "weed out" the unimportant ones.

What is "Power of the Test"?

The probability of correctly rejecting the null hypothesis when it is actually false.

How does increasing Sample Size affect Power?

It increases the power to detect smaller effects.

What is "Signal-to-Noise Ratio" in DoE?

The ratio of the magnitude of the factor effect to the experimental error.

Define "Alpha (Type I Error)"

The probability of saying a factor is significant when it actually is not.

Define "Beta (Type II Error)"

The probability of saying a factor is not significant when it actually is.

What is a "Box-Behnken Design"?

A type of response surface design that does not contain any points at the extremes (corners).

What is a "Central Composite Design" (CCD)?

A factorial design augmented with "star points" to estimate curvature.

Define "Stationary Point"

The point on a response surface where the slope is zero (can be a maximum, minimum, or saddle).

What is a "Saddle Point"?

A point on a surface that is a maximum in one direction and a minimum in another.

Define "Model Overfitting" in DoE

Including too many terms (like high-order interactions) that describe noise instead of the process.

What is "Residuals vs. Fits" plot used for?

To check for non-constant variance (heteroscedasticity).

What is a "Normal Probability Plot of Residuals"?

A tool to verify if the errors follow a normal distribution.

What does a "Run Order" column indicate?

The sequence in which the experiments were actually performed.

Define "Coded Variables"

Mapping real values (e.g., 0.001 to 0.1) to a scale of -1 to +1 for easier math.

What is the "Lack of Fit" test?

A statistical test that determines if the chosen model is adequate to describe the data.

What is "Parsimony" in modeling?

The preference for the simplest model that adequately explains the data.