PRACTICE ASSESSMENT - FOR STUDY PURPOSES
WGU D682 TASK 3 - ARTIFICIAL
INTELLIGENCE OPTIMIZATION FOR
COMPUTER SCIENTISTS COMPREHENSIVE
PRACTICE (2026/2027 EDITION)
D682 - Artificial Intelligence Optimization for Computer Scientists | Western Governors
University
Task 3 Practice Assessment - Data Assumptions & Validation - Model Interpretation & Explainability -
Predictive Analysis & Forecasting - Model Optimization & Performance - Integrated AI Applications
Total items 100 multiple-choice questions (one correct answer each)
Section structure 20 + 20 + 20 + 20 + 20 = 100 questions
Suggested time 135 minutes (average 1.35 minutes per question)
Answer format Four options per item (A-D), with answer and rationale shown
Cognitive weighting 30% recall / 50% application / 20% analysis
Content Blueprint
Section Focus Items Weight
Section 1: Data Assumptions & Statistical Assumptions, Validation
20 20.0%
Validation Techniques, Data Quality
Section 2: Model Interpretation & Feature Importance, SHAP, LIME, Partial
20 20.0%
Explainability Dependence
Section 3: Predictive Analysis & Predictive Modeling, Time Series,
20 20.0%
Forecasting Forecasting Metrics
Section 4: Model Optimization & Hyperparameter Tuning, Regularization,
20 20.0%
Performance Ensembles
Section 5: Integrated AI Applications Complex Multi-Domain AI Scenarios 20 20.0%
Total All content domains 100 100%
WGU D682 TASK 3 - ARTIFICIAL INTELLIGENCE OPTIMIZATION FOR COMPUTER SCIENTISTS Practice - 2026/2027 Edition Page 1
, Instructions
1. Read each stem and select the single best answer among the four options (A-D). Every item has
exactly one correct option; distractors use common computational slips, misapplied theorems, or
incorrect interpretations.
2. Pace yourself: 100 items in about 135 minutes. If an item is unclear, mark your best answer and
return to it later; no item is worth more than another.
3. The correct option for each item appears in bold, tagged [CORRECT], followed by a short worked
rationale. For self-testing, cover the bold options and rationales; for review, read the full item together.
4. Item distribution follows the blueprint for this assessment, covering data assumptions and
validation, model interpretation and explainability, predictive analysis and forecasting, model
optimization and performance, and integrated multi-domain AI applications.
This is a practice assessment prepared for study purposes. Verify all course-specific content against current lecture
materials and the posted syllabus before relying on it for grading or instruction.
Section 1: Data Assumptions & Validation / Statistical Assumptions,
Validation Techniques, Data Quality
20 questions
Q1: A linear regression predicting delivery time shows a U-shaped pattern when residuals are plotted
against fitted values. Which assumption is violated and what is the standard remedy?
A. The linearity assumption is violated; add polynomial terms or use a nonlinear or
spline-based model [CORRECT]
B. The independence assumption is violated; use cluster-robust standard errors
C. Normality is violated; bootstrap the p-values only
D. Homoscedasticity is violated; weight the observations by distance
Correct Answer: A
Rationale: Curvature in the residual-versus-fitted plot indicates that the mean response is not linear in the
predictors, so adding polynomial or spline terms is the direct fix. Clustered errors, non-normal residuals,
and unequal variance produce different diagnostic patterns such as grouped residuals, heavy tails in a QQ
plot, or a fan shape.
Q2: A study measures student outcomes in 50 classrooms and fits an ordinary least squares model
treating each student as independent. What is the main problem?
A. Nothing; least squares requires no independence assumption
B. Observations within a classroom are correlated, so standard errors are understated; a
mixed-effects model or cluster-robust errors are appropriate [CORRECT]
C. The sample size is too small for least squares
D. Classroom membership should be used as the target
Correct Answer: B
Rationale: Clustered observations violate the independence assumption, which biases standard errors and
inflates significance. Sample size is not the issue, classroom membership is a grouping factor rather than a
target, and independence is a core least squares assumption.
WGU D682 TASK 3 - ARTIFICIAL INTELLIGENCE OPTIMIZATION FOR COMPUTER SCIENTISTS Practice - 2026/2027 Edition Page 2
, Q3: A model predicting house prices has residuals that are clearly non-normal, but the dataset
contains 80,000 observations and only predictions are needed. How should this be handled?
A. Transform the target until residuals are perfectly normal before predicting
B. Increase the number of predictors to absorb the non-normality
C. Proceed with prediction-focused inference, since the central limit theorem makes
coefficient inference roughly valid at this sample size and normality mainly affects
interval inference rather than point predictions [CORRECT]
D. Discard the model and use a rank-based method only
Correct Answer: C
Rationale: With large samples the sampling distribution of coefficients approaches normality, and normality
of errors is not required for point prediction. Abandoning the model, forcing perfect residual normality, or
padding the feature set addresses inference aesthetics rather than the modeling goal.
Q4: Residuals from a cost model fan out as predicted values increase, with much larger spread at
high fitted values. Which assumption is violated, and which remedy is appropriate?
A. Linearity is violated; remove the highest-cost observations
B. Independence is violated; add a time index
C. Normality is violated; center the predictors
D. Homoscedasticity is violated; consider a variance-stabilizing transformation, weighted
least squares, or heteroscedasticity-robust standard errors [CORRECT]
Correct Answer: D
Rationale: A widening funnel indicates variance that changes with the fitted value, so transformations,
weighting, or robust errors are the standard responses. Curvature signals non-linearity, grouping signals
dependence, and centering predictors does not change error variance.
Q5: A regression predicting salary uses both years of experience and age as predictors, and their
variance inflation factors exceed 8. What is the practical consequence and a reasonable remedy?
A. Coefficients become unstable and hard to interpret individually; drop or combine one
predictor, use regularization, or apply dimensionality reduction [CORRECT]
B. Predictions become impossible to compute
C. The model will always overfit the training data
D. The intercept becomes the most important term
Correct Answer: A
Rationale: High multicollinearity inflates coefficient variance and makes individual effects poorly identified
while leaving overall predictions often acceptable. Predictions are still computable, overfitting is a separate
concern, and the intercept does not gain importance from collinearity.
Q6: A clinical dataset has a biomarker missing in 30 percent of patients, and the missingness is
related to disease severity. Which approach is most defensible?
A. Remove the biomarker from the feature set permanently
B. Model the missingness explicitly (for example with multiple imputation plus a
missingness indicator) and validate on a held-out set, since missing-at-random
assumptions are doubtful [CORRECT]
C. Delete all rows with missing values and proceed
D. Fill missing values with the column mean and report no limitation
Correct Answer: B
Rationale: When missingness depends on unobserved or severity-related factors, simple deletion or mean
imputation biases results, so principled imputation with diagnostics is needed. Listwise deletion loses the
sickest patients, mean imputation distorts relationships, and dropping the variable discards clinical signal.
WGU D682 TASK 3 - ARTIFICIAL INTELLIGENCE OPTIMIZATION FOR COMPUTER SCIENTISTS Practice - 2026/2027 Edition Page 3
WGU D682 TASK 3 - ARTIFICIAL
INTELLIGENCE OPTIMIZATION FOR
COMPUTER SCIENTISTS COMPREHENSIVE
PRACTICE (2026/2027 EDITION)
D682 - Artificial Intelligence Optimization for Computer Scientists | Western Governors
University
Task 3 Practice Assessment - Data Assumptions & Validation - Model Interpretation & Explainability -
Predictive Analysis & Forecasting - Model Optimization & Performance - Integrated AI Applications
Total items 100 multiple-choice questions (one correct answer each)
Section structure 20 + 20 + 20 + 20 + 20 = 100 questions
Suggested time 135 minutes (average 1.35 minutes per question)
Answer format Four options per item (A-D), with answer and rationale shown
Cognitive weighting 30% recall / 50% application / 20% analysis
Content Blueprint
Section Focus Items Weight
Section 1: Data Assumptions & Statistical Assumptions, Validation
20 20.0%
Validation Techniques, Data Quality
Section 2: Model Interpretation & Feature Importance, SHAP, LIME, Partial
20 20.0%
Explainability Dependence
Section 3: Predictive Analysis & Predictive Modeling, Time Series,
20 20.0%
Forecasting Forecasting Metrics
Section 4: Model Optimization & Hyperparameter Tuning, Regularization,
20 20.0%
Performance Ensembles
Section 5: Integrated AI Applications Complex Multi-Domain AI Scenarios 20 20.0%
Total All content domains 100 100%
WGU D682 TASK 3 - ARTIFICIAL INTELLIGENCE OPTIMIZATION FOR COMPUTER SCIENTISTS Practice - 2026/2027 Edition Page 1
, Instructions
1. Read each stem and select the single best answer among the four options (A-D). Every item has
exactly one correct option; distractors use common computational slips, misapplied theorems, or
incorrect interpretations.
2. Pace yourself: 100 items in about 135 minutes. If an item is unclear, mark your best answer and
return to it later; no item is worth more than another.
3. The correct option for each item appears in bold, tagged [CORRECT], followed by a short worked
rationale. For self-testing, cover the bold options and rationales; for review, read the full item together.
4. Item distribution follows the blueprint for this assessment, covering data assumptions and
validation, model interpretation and explainability, predictive analysis and forecasting, model
optimization and performance, and integrated multi-domain AI applications.
This is a practice assessment prepared for study purposes. Verify all course-specific content against current lecture
materials and the posted syllabus before relying on it for grading or instruction.
Section 1: Data Assumptions & Validation / Statistical Assumptions,
Validation Techniques, Data Quality
20 questions
Q1: A linear regression predicting delivery time shows a U-shaped pattern when residuals are plotted
against fitted values. Which assumption is violated and what is the standard remedy?
A. The linearity assumption is violated; add polynomial terms or use a nonlinear or
spline-based model [CORRECT]
B. The independence assumption is violated; use cluster-robust standard errors
C. Normality is violated; bootstrap the p-values only
D. Homoscedasticity is violated; weight the observations by distance
Correct Answer: A
Rationale: Curvature in the residual-versus-fitted plot indicates that the mean response is not linear in the
predictors, so adding polynomial or spline terms is the direct fix. Clustered errors, non-normal residuals,
and unequal variance produce different diagnostic patterns such as grouped residuals, heavy tails in a QQ
plot, or a fan shape.
Q2: A study measures student outcomes in 50 classrooms and fits an ordinary least squares model
treating each student as independent. What is the main problem?
A. Nothing; least squares requires no independence assumption
B. Observations within a classroom are correlated, so standard errors are understated; a
mixed-effects model or cluster-robust errors are appropriate [CORRECT]
C. The sample size is too small for least squares
D. Classroom membership should be used as the target
Correct Answer: B
Rationale: Clustered observations violate the independence assumption, which biases standard errors and
inflates significance. Sample size is not the issue, classroom membership is a grouping factor rather than a
target, and independence is a core least squares assumption.
WGU D682 TASK 3 - ARTIFICIAL INTELLIGENCE OPTIMIZATION FOR COMPUTER SCIENTISTS Practice - 2026/2027 Edition Page 2
, Q3: A model predicting house prices has residuals that are clearly non-normal, but the dataset
contains 80,000 observations and only predictions are needed. How should this be handled?
A. Transform the target until residuals are perfectly normal before predicting
B. Increase the number of predictors to absorb the non-normality
C. Proceed with prediction-focused inference, since the central limit theorem makes
coefficient inference roughly valid at this sample size and normality mainly affects
interval inference rather than point predictions [CORRECT]
D. Discard the model and use a rank-based method only
Correct Answer: C
Rationale: With large samples the sampling distribution of coefficients approaches normality, and normality
of errors is not required for point prediction. Abandoning the model, forcing perfect residual normality, or
padding the feature set addresses inference aesthetics rather than the modeling goal.
Q4: Residuals from a cost model fan out as predicted values increase, with much larger spread at
high fitted values. Which assumption is violated, and which remedy is appropriate?
A. Linearity is violated; remove the highest-cost observations
B. Independence is violated; add a time index
C. Normality is violated; center the predictors
D. Homoscedasticity is violated; consider a variance-stabilizing transformation, weighted
least squares, or heteroscedasticity-robust standard errors [CORRECT]
Correct Answer: D
Rationale: A widening funnel indicates variance that changes with the fitted value, so transformations,
weighting, or robust errors are the standard responses. Curvature signals non-linearity, grouping signals
dependence, and centering predictors does not change error variance.
Q5: A regression predicting salary uses both years of experience and age as predictors, and their
variance inflation factors exceed 8. What is the practical consequence and a reasonable remedy?
A. Coefficients become unstable and hard to interpret individually; drop or combine one
predictor, use regularization, or apply dimensionality reduction [CORRECT]
B. Predictions become impossible to compute
C. The model will always overfit the training data
D. The intercept becomes the most important term
Correct Answer: A
Rationale: High multicollinearity inflates coefficient variance and makes individual effects poorly identified
while leaving overall predictions often acceptable. Predictions are still computable, overfitting is a separate
concern, and the intercept does not gain importance from collinearity.
Q6: A clinical dataset has a biomarker missing in 30 percent of patients, and the missingness is
related to disease severity. Which approach is most defensible?
A. Remove the biomarker from the feature set permanently
B. Model the missingness explicitly (for example with multiple imputation plus a
missingness indicator) and validate on a held-out set, since missing-at-random
assumptions are doubtful [CORRECT]
C. Delete all rows with missing values and proceed
D. Fill missing values with the column mean and report no limitation
Correct Answer: B
Rationale: When missingness depends on unobserved or severity-related factors, simple deletion or mean
imputation biases results, so principled imputation with diagnostics is needed. Listwise deletion loses the
sickest patients, mean imputation distorts relationships, and dropping the variable discards clinical signal.
WGU D682 TASK 3 - ARTIFICIAL INTELLIGENCE OPTIMIZATION FOR COMPUTER SCIENTISTS Practice - 2026/2027 Edition Page 3
