Font size
S
M
L
XL
WorksheetsML Unit 4& 5 MCQ
Total questions: 100
Worksheet time: 53mins
Name
Class
Date
1.
Given a query point $Q(3, 4)$ and two training points $A(1, 1)$ and $B(5, 7)$, calculate the Euclidean distance to both. Which point is the nearest neighbor?
a)
Point A ($d \approx 3.6$)
b)
Point B ($d \approx 3.6$)
c)
Both are equidistant
d)
Point A ($d = 13$)
2.
In a 1-NN classifier, if we have a point $P$ that is an outlier with an incorrect label, how does it affect the decision boundary?
a)
No effect
b)
Creates a "pocket" of misclassification
c)
Smooths the boundary
d)
Increases bias
3.
A dataset has 1000 points. If you increase $K$ from 1 to 100 in KNN, what is the most likely analytical outcome?
a)
Increased variance
b)
Increased bias; smoother boundary
c)
Overfitting
d)
Lower computational cost
4.
Calculate the Manhattan distance between $X(10, 2)$ and $Y(3, 8)$.
a)
9.21
b)
13
c)
11
d)
15
5.
In Locally Weighted Regression (LWR), if the kernel width parameter $\tau$ is very small, how does the model behave?
a)
Fits a global linear line
b)
Overfits to local points
c)
Becomes a rigid constant
d)
Ignores all data
6.
A Radial Basis Function (RBF) uses a Gaussian kernel $K(x, c) = \exp(-\|x-c\|^2 / 2\sigma^$. If $ $\|x$ increases, what happens to the influence of the center $e$ on $o$?
a)
Increases exponentially
b)
Decreases exponentially
c)
Remains constant
d)
Becomes zero instantly
7.
In KNN, if the features are measured in different units (e.g., Age in years vs. Salary in dollars), what is the necessary step before applying the algorithm?
a)
Feature selection
b)
Min-Max Normalization
c)
Dimensionality reduction
d)
Increasing K
8.
Compare KNN and RBF networks. Which of the following is an analytical similarity?
a)
Both use global approximations
b)
Both use local information/distance
c)
Both have fixed parameters
d)
Both use backpropagation
9.
For a 3-NN classifier, the nearest neighbors to a query point have labels {Red, Blue, Blue}. What is the predicted label?
a)
Red
b)
Blue
c)
Undefined
d)
Both Red and Blue
10.
In Case-Based Reasoning (CBR), the process of modifying a past solution to fit a new problem is known as:
a)
Retention
b)
Retrieval
c)
Adaptation
d)
Review
11.
Calculate the Minkowski distance ($p=3$) between $(0,0)$ and $(3,4)$.
a)
5
b)
7
c)
4.49
d)
12
12.
Why is KNN often called a "Lazy Learner"?
a)
It learns very slowly
b)
It postpones computation until query time
c)
It uses simple logic
d)
It requires high CPU usage
13.
In a binary classification problem, why is it analytically preferred to choose an odd value for $K$?
a)
To speed up distance math
b)
To avoid tie votes
c)
To reduce memory
d)
To increase accuracy
14.
In LWR, the cost function $J(\theta) = \sum w^{(i)}(y^{(i)} - \theta^T x^{(i)})^2$ is used. If $w^{(i)} = 1$ for all $i$, what does the model become?
a)
KNN
b)
Ordinary Least Squares
c)
RBF
d)
Logistic Regression
15.
Given a query point $X=5$. Training points: $x_1=4$ (label 10), $x_2=6$ (label 20). Using Distance-Weighted 2-NN regression, what is the prediction?
a)
15
b)
10
c)
20
d)
18
16.
Which distance metric is most suitable for comparing two documents represented as word-count vectors?
a)
Euclidean
b)
Manhattan
c)
Cosine Similarity
d)
Minkowski
17.
An RBF network has $M$ hidden units. What does each hidden unit represent?
a)
A linear decision line
b)
A local center in the input space
c)
A global weight
d)
A bias term
18.
In Case-Based Learning, "The Curse of Dimensionality" primarily affects which stage?
a)
Retrieval
b)
Retention
c)
Reuse
d)
Revision
19.
If a dataset has very noisy labels, which value of $K$ is likely to perform better?
a)
$K=1$
b)
$K=3$
c)
$K=15$
d)
$K=0$
20.
In Locally Weighted Regression, if $x$ is far from $x^{(i)}$, the weight $w^{(i)}$ typically:
a)
Approaches 1
b)
Approaches 0
c)
Becomes negative
d)
Remains constant
21.
A query point is surrounded by 5 points of Class A at distance 2, and 2 points of Class B at distance 0.1. In weighted KNN ($1/d^2$), which class wins?
a)
Class A
b)
Class B
c)
Tie
d)
Neither
22.
What is the primary analytical disadvantage of KNN in a real-time system with millions of records?
a)
High training time
b)
High prediction latency
c)
Low accuracy
d)
High bias
23.
In RBF networks, how are the centers $c_i$ usually determined?
a)
Backpropagation
b)
K-means clustering or random sampling
c)
Genetic algorithms
d)
Gradient descent only
24.
Analyzing CBR: Which stage involves storing the newly solved problem and its solution back into the knowledge base?
a)
Retrieve
b)
Reuse
c)
Retain
d)
Revise
25.
In 1-NN, what is the training error rate?
a)
0.5
b)
1
c)
Depends on data
26.
If we use $K=N$ (where $N$ is total points) in KNN classification, the output will always be:
a)
The nearest neighbor
b)
The majority class of the entire dataset
c)
Random
d)
Always 0
27.
Which of the following is a "Global" method unlike Instance-Based Learning?
a)
KNN
b)
LWR
c)
Linear Regression
d)
RBF
28.
In RBF networks, the output layer typically uses which type of activation?
a)
Sigmoid
b)
Linear
c)
ReLU
d)
Tanh
29.
What is the effect of "irrelevant features" on KNN?
a)
No effect
b)
Improves accuracy
c)
Degrades performance by adding noise to distance
d)
Speeds up calculation
30.
A system uses CBR to fix laptops. A new laptop has a "broken screen." The system finds a case for "cracked monitor." This finding is part of:
a)
Adaptation
b)
Retrieval
c)
Retention
d)
Learning
31.
Assertion: KNN is highly sensitive to the scale of data. Reason: Distance metrics like Euclidean give equal weight to all dimensions regardless of their units.
a)
Both A and R true; R is correct explanation
b)
Both A and R true; R is not correct
c)
A is true; R is false
d)
A is false; R is true
32.
Assertion: Locally Weighted Regression is a non-parametric method. Reason: It does not require a fixed set of parameters to be learned during a separate training phase.
a)
Both true; R is correct explanation
b)
Both true; R is not correct
c)
A is true; R is false
d)
Both false
33.
Assertion: As $K$ increases in KNN, the decision boundary becomes more complex. Reason: Larger $K$ considers more neighbors, which smooths out local fluctuations.
a)
Both true; R is correct explanation
b)
Both true; R is not correct
c)
A is false; R is true
d)
A is true; R is false
34.
Assertion: RBF networks are faster to train than standard Multi-Layer Perceptrons. Reason: The weights from the hidden layer to the output layer in RBF can be solved using linear equations.
a)
Both A and R true; R is correct explanation
b)
Both A and R true; R is not correct
c)
A is true; R is false
d)
A is false; R is true
35.
Assertion: Case-Based Reasoning can solve problems that are not well-defined by rules. Reason: It relies on specific past experiences rather than general domain models.
a)
Both true; R is correct explanation
b)
Both true; R is not correct
c)
A true; R false
d)
A false; R true
36.
Assertion: Distance-weighted KNN is generally more robust than standard KNN. Reason: It gives closer neighbors more influence on the final decision.
a)
Both true; R is correct explanation
b)
Both true; R is not correct
c)
A true; R false
d)
A false; R true
37.
Assertion: LWR is preferred for large datasets over global linear regression. Reason: LWR requires keeping all data in memory and performing a fit for every new query.
a)
Both true; R is correct explanation
b)
Both true; R is not correct
c)
A is false; R is true
d)
A is true; R is false
38.
Assertion: RBF networks can approximate any continuous function. Reason: They use a combination of local basis functions to cover the input space.
a)
Both true; R is correct explanation
b)
Both true; R is not correct
c)
A true; R false
d)
A false; R true
39.
Assertion: KNN works well even if there are thousands of irrelevant features. Reason: Distance is calculated across all dimensions, and irrelevant features add random noise to the distance.
a)
Both true; R is correct explanation
b)
Both true; R is not correct
c)
A is false; R is true
d)
A true; R false
40.
Assertion: CBR is an inductive learning method. Reason: It generalizes from specific cases to create broad rules before seeing a test case.
a)
Both true; R is correct explanation
b)
Both true; R is not correct
c)
A true; R false
d)
Both false
41.
A house is 2000 sq ft and 5 years old. The nearest neighbors are $(1900, 4)$ at $300k and $(2100, 6)$ at $320k. Using 2-NN average, what is the price?
a)
$300k
b)
$310k
c)
$320k
d)
$350k
42.
If Square Footage ranges from 500-5000 and Age ranges from 1-100, which feature will dominate the Euclidean distance if not normalized?
a)
Age
b)
Square Footage
c)
Both equally
d)
Neither
43.
The company finds that house prices vary wildly by neighborhood. Which algorithm would best allow them to create "local" price models for different areas?
a)
Linear Regression
b)
LWR
c)
Perceptron
d)
Logic Gates
44.
A new house has a feature "Has Swimming Pool" (Binary 0/1). How should this be analyzed in a KNN model?
a)
Ignore it
b)
Use it with a custom distance metric (e.g. Hamming)
c)
Multiply it by 1000
d)
Convert to continuous
45.
The system needs to explain why it priced a house at $400k. Which method provides the most "interpretable" explanation to the user?
a)
RBF Network
b)
CBR (shows similar past sales)
c)
Deep Neural Network
d)
LWR
46.
If the company wants to reduce the "memory footprint" of their KNN model, which technique is most appropriate?
a)
Increase K
b)
Use Data Editing/Pruning (keep only representative prototypes)
c)
Decrease K
d)
Use more features
47.
In a CBR system for real estate, "The house has a view of the park, which adds 10% value." This rule is part of which phase?
a)
Retrieval
b)
Adaptation
c)
Retention
d)
Retrieval
48.
During a market crash, the "Retrieval" phase of CBR provides old high prices. Which stage is responsible for correcting this based on current market feedback?
a)
Retrieval
b)
Reuse
c)
Revise
d)
Retain
49.
An RBF network is trained on the data. If a house is located in a region where no training data exists, the RBF output will likely be:
a)
Very high
b)
Near zero (or the bias value)
c)
Extremely accurate
d)
Random
50.
The company switches from 1-NN to 10-NN. They notice the model no longer predicts extreme high or low prices. This indicates a decrease in:
a)
Bias
b)
Variance
c)
Accuracy
d)
Speed
51.
In a Sequential Covering algorithm, if a learned rule covers 10 positive examples and 2 negative examples, what is its precision?
a)
0.83
b)
0.8
c)
0.2
d)
1
52.
Given a Q-learning agent with $\gamma = 0.9$. If the current $Q(s, a) = 10$, the reward $r = 5$, and the maximum $Q(s', a') = 20$, calculate the new $Q(s, a)$ with learning rate $\alpha = 0.5$.
a)
16.5
b)
14
c)
19
d)
11.5
53.
Which of the following differentiates Analytical Learning (EBL) from Inductive Learning?
a)
EBL requires many examples
b)
EBL uses prior domain knowledge to explain a single example
c)
EBL is strictly stochastic
d)
EBL cannot generalize
54.
In First-Order Rule learning, what is the effect of "Inverting Resolution"?
a)
It simplifies rules
b)
It derives general rules from specific examples and background knowledge
c)
It removes noise
d)
It converts rules to bitstrings
55.
A Reinforcement Learning task has 4 states in a line. If the agent receives a reward of +100 at the end and $\gamma = 0.5$, what is the discounted value 2 steps away from the goal?
a)
100
b)
50
c)
25
d)
12.5
56.
What is the primary analytical goal of the FOCL algorithm?
a)
Pure reinforcement
b)
Combining Inductive and Analytical learning
c)
Maximizing entropy
d)
Reducing population size
57.
In Temporal Difference (TD) learning, the update is based on the difference between:
a)
Target and Current Estimate
b)
Two successive states
c)
Prediction and Ground Truth
d)
Reward and Action
58.
Given a domain theory: $Ancestor(x, y) \leftarrow Parent(x, y)$. If we see $Parent(Bob, Alice)$, EBL would deduce:
a)
$Ancestor(Alice, Bob)$
b)
$Ancestor(Bob, Alice)$
c)
$Friend(Bob, Alice)$
d)
Nothing
59.
In a rule-based system, if we add a constraint to a rule (specialization), what happens to its coverage?
a)
Increases
b)
Decreases or stays same
c)
Becomes zero
d)
Doubles
60.
The "Exploration vs. Exploitation" trade-off in Q-Learning is most commonly managed by:
a)
$\epsilon$-greedy strategy
b)
Increasing $\gamma$
c)
Decreasing $\alpha$
d)
Sequential covering
61.
Calculate the Gain for a rule if it originally covered $p_0=4, n_0=4$ and after adding a literal covers $p_1=3, n_1=0$.
a)
1.58
b)
2.11
c)
3
d)
0.5
62.
In First Order Logic, what is a "Horn Clause"?
a)
A clause with no literals
b)
A clause with at most one positive literal
c)
A clause with multiple heads
d)
A recursive function
63.
Why is "Inverted Deduction" considered a "Bottom-Up" approach?
a)
Starts from goals
b)
Starts from data to build hypotheses
c)
It uses only biases
d)
It ignores background knowledge
64.
In EBL, "Operationality Criteria" ensures that the learned rule is:
a)
Mathematically complex
b)
Easy to evaluate in new instances
c)
Recursive
d)
Large in size
65.
An agent uses TD(0). If the reward $r_t=2$, $V(s_t)=10$, and $V(s_{t+1})=12$ with $\gamma=1$, calculate the TD error.
a)
4
b)
2
c)
-2
66.
Which of the following describes the "Credit Assignment Problem" in RL?
a)
Paying the agent
b)
Determining which action led to a distant reward
c)
Calculating $\alpha$
d)
Data entry errors
67.
In Sequential Covering, what happens to the training set after a rule is learned?
a)
It is deleted
b)
Covered positive examples are removed
c)
All examples are kept
d)
Negative examples are removed
68.
Comparing Q-Learning and SARSA: SARSA is "On-policy" because it:
a)
Updates based on the actual next action taken
b)
Always picks the best action
c)
Ignores rewards
d)
Uses a neural network
69.
In EBL, if the "Domain Theory" is "Perfect," what is the role of the training example?
a)
To provide new info
b)
To guide the search for a general rule already implied by theory
c)
To prove the theory is wrong
d)
To increase noise
70.
A rule $R1$ covers 100% of training data but has 20 literals. A rule $R2$ covers 95% with 2 literals. Which is analytically better?
a)
$R1$ (Higher accuracy)
b)
$R2$ (Prevents overfitting/Minimum Description Length)
c)
Neither
d)
$R1$ (More complex)
71.
If $\gamma = 0$ in an RL task, the agent is:
a)
Perfectly far-sighted
b)
Completely short-sighted (only immediate reward)
c)
Random
d)
Stuck
72.
In FOL induction, the "Literals" added to a rule can involve:
a)
Constants only
b)
Variables and predicates
c)
Only integers
d)
No variables
73.
What is the analytical function of the "Weak Theory" problem in EBL?
a)
Theory is too large
b)
Theory has gaps and cannot explain all examples
c)
Theory is recursive
d)
Theory is too simple
74.
In a Q-table for 5 states and 2 actions, how many total Q-values are stored?
a)
5
b)
2
c)
10
d)
25
75.
In the FOIL algorithm, the search for literals is:
a)
Depth-first
b)
Greedy (Hill Climbing)
c)
Breadth-first
d)
Random
76.
If a Q-Learning agent finds a path in 100 steps and then 80 steps, what should happen to the Q-values of the 80-step path?
a)
They should decrease
b)
They should increase
c)
No change
d)
Become zero
77.
Which model is used to describe the environment in Reinforcement Learning?
a)
Linear Regression
b)
Markov Decision Process (MDP)
c)
Decision Tree
d)
Logic Gate
78.
"Inverting resolution" is based on the principle that if $C = A \cup B$, then we can derive $A$ given $C$ and $B$. This is used in:
a)
Progol / Cigol
b)
Q-Learning
c)
EBL
d)
KNN
79.
In EBL, the process of "Regressing" the goal through the domain theory is called:
a)
Deduction
b)
Back-chaining / Regression
c)
Induction
d)
Resolution
80.
A "Policy" $\pi$ in RL is a mapping from:
a)
Action to Reward
b)
State to Action
c)
Reward to State
d)
Time to Action
81.
Assertion: EBL can learn from a single training example. Reason: It uses prior domain knowledge to justify why the example is a member of the target concept.
a)
Both A and R true; R explains A
b)
Both true; R doesn't explain A
c)
A is true; R false
d)
Both false
82.
Assertion: Q-learning is an "Off-policy" algorithm. Reason: It updates the Q-value based on the maximum possible future reward, regardless of the policy followed.
a)
Both A and R true; R explains A
b)
Both true; R doesn't explain A
c)
A is false; R true
d)
Both false
83.
Assertion: Sequential covering algorithms are "Greedy". Reason: They learn one rule at a time and remove the examples covered by it.
a)
Both A and R true; R explains A
b)
Both true; R doesn't explain A
c)
A is true; R false
d)
Both false
84.
Assertion: Reinforcement learning requires a labeled dataset like Supervised learning. Reason: The agent learns from a scalar reward signal provided by the environment.
a)
Both true; R explains A
b)
Both true; R doesn't explain A
c)
A is false; R true
d)
Both false
85.
Assertion: FOCL is more robust than pure EBL. Reason: FOCL can use inductive methods when the domain theory is incomplete.
a)
Both A and R true; R explains A
b)
Both true; R doesn't explain A
c)
A is true; R false
d)
Both false
86.
If the agent is penalized (-10) for "Out of Stock" and penalized (-1) for "Holding Cost," what will it learn to do over time?
a)
Keep zero stock
b)
Balance stock to minimize both costs
c)
Keep infinite stock
d)
Ignore the penalties
87.
The agent discovers a strategy that works well in December but fails in July. Which RL parameter controls how much it values "Future" seasonal rewards?
a)
$\alpha$ (Learning rate)
b)
$\gamma$ (Discount factor)
c)
$\epsilon$ (Exploration)
d)
Reward $r$
88.
The company wants to encode a rule: "If Stock < 10 AND Day = Monday, THEN Order 50." This is a:
a)
First-Order Rule
b)
Zero-Order Rule
c)
Temporal Rule
d)
Random Rule
89.
In a simulator, the agent learns a "Perfect" model of the warehouse. When moved to the real warehouse, it fails. This is a problem of:
a)
High $\gamma$
b)
Model Mismatch / Domain Gap
c)
EBL error
d)
Sequential covering
90.
If the agent is "exploring" 10% of the time, what is the value of $\epsilon$?
a)
0.9
b)
0.1
c)
1
91.
The warehouse manager provides a rule set for the agent to start with. The agent then refines these rules using data. This is an application of:
a)
FOCL / Theory Revision
b)
Pure Q-learning
c)
Pure Induction
d)
Random walk
92.
To predict the stock level for the next 24 hours without an environment model, the agent should use:
a)
EBL
b)
Model-Free RL (e.g., Q-learning)
c)
Resolution
d)
Inverted Deduction
93.
During training, the agent's $Q$-values are not changing at all. Which parameter is likely set to zero?
a)
$\alpha$
b)
$\gamma$
c)
Reward
d)
State
94.
The agent learns a rule: $Order(x) \leftarrow LowStock(x)$. To make this "First-Order," we should add:
a)
A constant
b)
A relation like $Supplier(x, y)$
c)
A bias
d)
A weight
95.
The inventory system is modeled as an MDP. What does the "Markov Property" imply here?
a)
Future stock depends only on current stock and action
b)
Stock depends on 10 years of history
c)
Stock is random
d)
Actions don't matter
96.
Assertion: Distance-weighted KNN is generally more robust than standard KNN. Reason: It reduces the influence of distant, potentially noisy neighbors by assigning weights inversely proportional to distance.
a)
Both A and R true; R is correct explanation
b)
Both A and R true; R is not correct
c)
A is true; R is false
d)
A is false; R is true
97.
Assertion: Locally Weighted Regression (LWR) is computationally expensive during the prediction phase. Reason: LWR is a lazy learner that must perform a new regression fit for every specific query point using the entire dataset.
a)
Both true; R is correct explanation
b)
Both true; R is not correct
c)
A is false; R is true
d)
A is true; R is false
98.
Assertion: Radial Basis Function (RBF) networks are considered a bridge between local and global learning. Reason: They use local receptive fields (hidden units) but combine them linearly to create a global approximation.
a)
Both true; R is correct explanation
b)
Both true; R is not correct
c)
A true; R false
d)
A false; R true
99.
Assertion: The "Curse of Dimensionality" makes KNN perform poorly in high-dimensional spaces. Reason: In high dimensions, the volume of the space increases so fast that available data becomes sparse, making "nearest" neighbors very far away.
a)
Both true; R is correct explanation
b)
Both true; R is not correct
c)
A is false; R is true
d)
A true; R false
100.
Assertion: Case-Based Reasoning (CBR) is an "Eager" learning method. Reason: It generalizes training data into a set of explicit rules or mathematical functions before a test query is received.
a)
Both true; R is correct explanation
b)
Both true; R is not correct
c)
A true; R false
d)
Both A and R are false
Reset
