ModelPaperSolvedAI_4

Section A
1. Compare Weak AI with Strong AI.
 Weak AI (narrow AI) is designed to perform a specific task intelligently,
such as speech recognition, language translation, or playing chess,
without any claim of real understanding or consciousness.
 Strong AI is a hypothetical form of AI that would have true human-like
intelligence and self-awareness, capable of understanding, reasoning,
and learning across many different domains.
 Today’s practical systems like chatbots, recommendation systems, and
self-driving cars are examples of Weak AI, while Strong AI does not yet
exist.

2. What is Goal based agent?

 A goal-based agent is an intelligent agent that chooses its actions by
considering a goal state and selecting actions that are expected to
achieve that goal.
 It uses a model of the environment to evaluate different sequences of
actions, often through search or planning, instead of reacting only to
the current percept.
 Example: A route-planning agent in Google Maps that, given the goal
“reach destination B from A,” searches for the best path that
minimizes travel time or distance.

3. List different types of learning.

 Supervised learning: The system learns from labeled examples (input
with correct output), e.g., classifying emails as “spam” or “not spam”
using past labeled emails.
 Unsupervised learning: The system finds patterns in unlabeled data,
e.g., clustering customers into groups based on their purchase history
without predefined labels.
 Reinforcement learning: An agent learns by interacting with an
environment and receiving rewards or penalties, e.g., a game-playing
program that improves its strategy by winning or losing games.

4. Justify – “Probabilistic Reasoning in AI – A way to deal with

Uncertainty”.
 Real-world data is often incomplete, noisy, or ambiguous, so AI
systems must represent beliefs as probabilities rather than as simply
true or false.
ModelPaperSolvedAI_4

 Probabilistic models such as Bayesian networks combine prior

knowledge with observed evidence to compute updated probabilities
and make rational decisions under uncertainty.
 Example: A medical diagnosis system uses uncertain symptom
information to compute the probability of different diseases and
recommends the most probable disease instead of a rigid yes/no
answer.

5. What do you mean by chunking in NLP?

 Chunking (or shallow parsing) is an NLP technique that groups words
into higher-level phrases such as noun phrases (NP) and verb phrases
(VP) without building a full parse tree.
 It usually uses part-of-speech tags and pattern rules or
machine-learning models to identify the boundaries of these phrase
“chunks” in a sentence.
 Example: In the sentence “The quick brown fox jumps over the lazy
dog,” chunking may produce chunks like “[The quick brown fox]NP
[jumps]VP [over the lazy dog]PP”.

6. State the relationship between Biological networks and

artificial neural networks.
 Biological neural networks in the brain consist of neurons connected by
synapses, where each neuron sums incoming signals and fires if the
combined signal exceeds a threshold.
 Artificial neural networks are simplified mathematical models inspired
by this idea: artificial neurons take weighted inputs, compute a
weighted sum, pass it through an activation function, and send the
result to other neurons.
 Learning in both systems is based on changing connection strengths—
synaptic plasticity in biological neurons and weight adjustment (for
example, using backpropagation) in artificial neural networks.

Section B

7. With an example, Explain A* search.

A* search is an informed search algorithm used to find the least-cost path from a start node to
a goal node. It combines the cost incurred so far with a heuristic estimate of the remaining cost.
Each node n is evaluated using the function:
f(n) = g(n) + h(n)
ModelPaperSolvedAI_4

where:
 g(n) = actual cost from the start node to node n
 h(n) = heuristic estimate of the cost from n to the goal
 f(n) = estimated total cost of the solution through n
At each step, the node with the lowest f(n) value is selected for expansion.

Example
Consider a route-finding problem:
 Start: City A
 Goal: City G
 g(n): distance travelled so far
 h(n): straight-line distance to G
The algorithm expands paths that appear cheapest based on g(n) + h(n), guiding the search
efficiently toward the goal.
If the heuristic never overestimates the true remaining cost (i.e., it is admissible), then:
 A* is complete
 A* is optimal

Advantages of A*
 Efficient compared to uninformed search
 Guarantees optimal solution with admissible heuristic
 Widely used in pathfinding and AI planning

8. With a neat diagram, Explain 8-queens problem.

The 8-queens problem is a classic problem in Artificial Intelligence that involves placing eight
queens on an 8×8 chessboard such that no two queens attack each other. This means no two
queens can be in the same row, column, or diagonal.

Problem Representation in AI
 State: A partial or complete placement of queens on the board
 Initial State: Empty chessboard
 Operators: Place a queen in a safe square (usually one queen per column or row)
 Goal State: Eight queens placed with no conflicts
 Constraints:
o No two queens in the same row
o No two queens in the same column
o No two queens on the same diagonal
ModelPaperSolvedAI_4

Diagram

Solution Techniques
 Backtracking (State-Space Search): Systematically places queens and backtracks when
a conflict occurs
 Hill Climbing: Uses heuristic such as the number of attacking pairs of queens
 Constraint Satisfaction Problem (CSP): Treats rows, columns, and diagonals as
constraints
Conclusion
The 8-queens problem demonstrates how AI techniques such as state-space search, heuristics,
and constraint satisfaction are used to solve combinatorial problems efficiently.

9. How Propositional Logic in Artificial Intelligence Represents Data

to Machine? Explain all logical connectives with truth table.
In Artificial Intelligence, Propositional Logic (PL) represents knowledge
using propositions, which are atomic statements that can have only two
truth values: True (T) or False (F). Machines store and reason with these
propositions by combining them using logical connectives, allowing
symbolic manipulation and logical inference.
For example, the statement “It is raining” can be represented as a
proposition P, and “The road is wet” as Q. Logical connectives are used to
form compound statements such as P → Q.

1. Conjunction (AND, P∧Q)

Logical Connectives with Truth Tables

True only when both P and Q are true.

P Q P∧Q
T T T
T F F
F T F
F F F
ModelPaperSolvedAI_4

2. Disjunction (OR, P∨Q)

True when at least one of P or Q is true.
P Q P∨Q
T T T
T F T
F T T
F F F
3. Negation (NOT, ¬P)
Reverses the truth value of P.
P ¬P
T F
F T
4. Implication (P→Q)
False only when P is true and Q is false.
P Q P→Q
T T T
T F F
F T T
F F T
5. Biconditional (P↔Q)
True when P and Q have the same truth value.
P Q P↔Q
T T T
T F F
F T F
F F T
Conclusion
Propositional Logic enables AI systems to represent knowledge symbolically
and reason using truth tables. Logical connectives allow machines to
evaluate complex statements and make decisions based on logical inference.

10. Write a note on Nonmonotonic logics with advantages &

disadvantages.

Non-monotonic logic is a type of logic in which the set of conclusions

drawn from a knowledge base can decrease when new information is
added. This is different from monotonic logic, where once a conclusion is
derived, it can never be withdrawn.
ModelPaperSolvedAI_4

Non-monotonic logic is useful for representing common-sense reasoning

and default assumptions in Artificial Intelligence, where information is
often incomplete or uncertain.

Example
In real life, we often make assumptions that may later turn out to be false.
Example:
 Rule: “Normally, birds can fly.”
 Fact: “Tweety is a bird.”
→ Conclusion: Tweety can fly.
If new information is added:
 Fact: “Tweety is a penguin.”
The previous conclusion is withdrawn, because penguins cannot fly.
This ability to retract conclusions makes the logic non-monotonic.

Features of Non-Monotonic Logic

 Supports default reasoning
 Allows exceptions
 Conclusions depend on current knowledge
 Models human-like reasoning

Advantages
 Represents real-world and common-sense knowledge effectively
 Handles incomplete and uncertain information
 Supports reasoning with exceptions and defaults
 More flexible than classical logic

Disadvantages
 Reasoning is computationally complex
 Difficult to design consistent rule sets
 May lead to ambiguity or contradictions
 Less efficient than monotonic logic

Conclusion
Non-monotonic logic plays an important role in AI by enabling systems to
reason realistically in dynamic and uncertain environments, even though it
introduces additional complexity in reasoning.

11. Explain the challenges of Pragmatic analysis in NLP.

ModelPaperSolvedAI_4

Pragmatic analysis in Natural Language Processing deals with

understanding language based on context, speaker intention, and real-
world knowledge, rather than just literal meaning. It focuses on what the
speaker means rather than what is explicitly stated.

Major Challenges
1. Context Understanding
Meaning often depends on situational and conversational context,
which is difficult for machines to model accurately.
2. Reference Resolution
Identifying what pronouns or expressions like “it”, “this”, or “there”
refer to requires context and prior knowledge.
3. Implied Meaning (Inference)
Humans often imply information rather than stating it directly. NLP
systems struggle to infer unstated intentions.
4. Speech Act Recognition
Determining whether a sentence is a request, command, promise, or
suggestion can be difficult when it is not explicitly stated.
5. Sarcasm and Irony
Statements may convey meanings opposite to their literal
interpretation, which is hard for machines to detect.
6. Commonsense and World Knowledge
Pragmatic understanding requires extensive background knowledge
about the real world, which is difficult to encode formally.
7. Dialogue History Tracking
Correct interpretation often depends on previous turns in a
conversation, requiring long-term context modeling.
8. Cultural and Social Factors
Politeness, indirectness, and meaning vary across cultures and
situations, adding complexity to pragmatic analysis.

Conclusion
Pragmatic analysis is challenging in NLP because it requires deep contextual,
social, and commonsense understanding. Despite advances, building
systems that fully understand pragmatic meaning remains a difficult
research problem.

12. Briefly explain machine learning life cycle.

The machine learning life cycle describes the sequence of steps followed
to build, deploy, and maintain a machine learning model.
ModelPaperSolvedAI_4

1. Problem Definition
Clearly define the objective, scope, success criteria, and constraints of
the ML problem.
2. Data Collection
Gather relevant data from sources such as databases, sensors, logs,
or APIs.
3. Data Preprocessing
Clean the data by handling missing values, removing noise, encoding
categorical variables, normalizing features, and splitting data into
training, validation, and test sets.
4. Feature Engineering and Selection
Extract meaningful features and select the most relevant ones to
improve model performance.
5. Model Selection and Training
Choose suitable algorithms and train the model using the training
data.
6. Model Evaluation and Tuning
Evaluate model performance using appropriate metrics and tune
hyperparameters to improve accuracy.
7. Deployment
Integrate the trained model into a production environment to make
predictions on real-world data.
8. Monitoring and Maintenance
Continuously monitor performance, detect data drift, and retrain the
model when necessary.

Section C

13. Any eight top technologies used in AI

Artificial Intelligence today relies on a set of core technologies that make
systems intelligent and autonomous.

 Machine Learning (ML): Algorithms learn patterns from data and

improve performance over time without being explicitly programmed,
used for prediction, classification, and recommendation.
 Deep Learning: A subset of ML using multi-layer neural networks to
automatically learn complex features, powering image recognition,
speech recognition, and game-playing AIs.
ModelPaperSolvedAI_4

 Natural Language Processing (NLP): Enables computers to understand,

interpret, and generate human language for tasks like chatbots,
translation, sentiment analysis, and question answering.
 Computer Vision: Allows machines to interpret visual information from
images and videos for applications such as facial recognition, medical
imaging, and autonomous driving.
 Expert / Knowledge-Based Systems: Use stored domain knowledge and
inference rules to give advice or decisions like a human expert in areas
such as medical diagnosis or troubleshooting.
 Robotics and Autonomous Systems: Integrate AI with sensors and
actuators so robots can perceive their environment, plan, and perform
tasks like navigation, assembly, and service operations.
 Reinforcement Learning: An agent learns optimal actions by interacting
with an environment and receiving rewards or penalties, widely used in
control, games, and recommendation systems.
 Generative AI: Models that create new content (text, images, audio,
video) similar to training data, used in tools like chat assistants, code
generators, and image creators.

14. Knowledge-Based System in AI (with neat diagram)

A Knowledge-Based System (KBS) is an AI program that uses a stored body
of knowledge and an inference mechanism to solve complex problems
requiring expert-level reasoning.
Diagram
ModelPaperSolvedAI_4

components

 Knowledge Base: Stores domain knowledge as facts (e.g., symptoms,

equipment states) and rules (IF–THEN productions, semantic networks,
frames, etc.).
 Inference Engine: Applies reasoning methods like forward chaining and
backward chaining to the KB to derive new facts or reach conclusions
from user input.
 Working Memory / Percept Buffer: Holds current facts about the
problem instance received from the environment or user.
 Explanation Facility: Explains to the user how a conclusion was
reached, increasing trust and usability.
 Knowledge Acquisition Module: Helps experts or knowledge engineers
add, modify, or delete rules and facts in the KB.

Working of a KBS

1. User provides a problem description or the agent perceives the

environment and stores these data in working memory.
2. Inference engine matches these facts with rules in the KB using
forward or backward chaining to infer new facts or reach a goal.
3. The chosen action or solution is returned to the user or executed via
actuators, and the KB may be updated with new knowledge for future
cases.

15. Winston’s Learning Program

Winston’s learning program is a classic AI system that learns structural
descriptions of scenes (like blocks-world pictures) from positive and negative
examples.

 The program represents each scene as a network of objects (brick,

wedge, etc.) and their properties and relationships such as on, under,
supports, and adjacent.
 It uses Guzman’s algorithm to segment the visual scene into bodies,
determine which edges belong to which objects, and infer partially
occluded parts.
ModelPaperSolvedAI_4

Learning process

 The system is given a concept definition (model) through one or more

training examples labeled “positive” (instances) and “negative”
(non-instances)..
 For a positive example, it compares the example’s description with the
current model, constructs a comparison network showing similarities
and differences, and then generalizes the model so it still covers
previous positive instances but remains as specific as possible.
 For a negative example, it uses the comparison network to specialize
the model by adding constraints (e.g., a particular support relation
must hold) so that the negative example is excluded.
 Over time, the concept description converges to a structural pattern
that correctly distinguishes instances from non-instances of the
concept (e.g., “arches” made from blocks).

Significance

 Winston’s program demonstrated learning from examples by structural

description, not just numeric features, and influenced research in
symbolic machine learning and concept formation.
 It showed how knowledge about objects and relations in a scene could
be incrementally refined using generalization and specialization
operators driven by examples.

16. Robotic Engineering – forward and inverse kinematics for 2-DOF

using geometrical approach

Robotic engineering and kinematics

Robotic engineering deals with the design, construction, control, and
programming of robots that can sense, plan, and act in the physical world.
Kinematics in robotics studies motion (position, velocity, acceleration) of
robot links and joints without considering forces.

Consider a planar 2-DOF robotic arm with two revolute joints: link-1 of length
L1 attached to the base, and link-2 of length L2 attached at the elbow,
forming angles θ1 and θ2 from the horizontal.
ModelPaperSolvedAI_4

Forward kinematics (geometrical)

Forward kinematics finds the end-effector position (x,y) when joint angles are
known.
Using basic trigonometry for the 2-link planar arm, the equations are:
x=L1cos⁡θ1+L2cos⁡(θ1+θ2) y=L1sin⁡θ1+L2sin⁡(θ1+θ2)
These are derived by projecting each link on the x and y axes and summing
the components.

Example: Let L1=L2=10 cm, θ1=45∘, θ2=45∘.

Then x=10cos45∘+10cos90∘=7.07+0=7.07 cm, and y=10sin45∘+10sin90∘=7.07+10=17.07
cm (approximate values shown verbally in exam).

Inverse kinematics (geometrical)

Inverse kinematics finds joint angles θ1,θ2 for a desired (x,y).
For the 2-link planar case, using the cosine rule for the triangle formed by
the two links and the line from base to target:
cosθ2=2L1 L2 x2+y2−L12−L22

(sign of sin⁡θ2).
Then θ2=arc cos(cosθ2), giving two configurations: elbow-up and elbow-down

For θ1,
θ1=tan−1(xy)−tan−1(L1+L2cosθ2L2sinθ2)
which comes by splitting the triangle into two right-angled triangles and
using tangent relations.

Example: For L1=L2=10 cm and desired (x,y)=(7,17) cm, the above formulae
yield approximately θ1≈45∘ and θ2≈45∘, matching the previous forward
example (numerical working can be shown stepwise in the answer booklet).

17. Leftmost and rightmost derivation with parse tree

In context-free grammars, a derivation is a sequence of rule applications
producing a string from the start symbol.
In a leftmost derivation, the leftmost non-terminal is replaced at every step,
while in a rightmost derivation, the rightmost non-terminal is always
expanded.
ModelPaperSolvedAI_4

Example grammar
Consider grammar G with start symbol E:

1. E→E+T∣T
2. T→T∗F∣F
3. F→(E)∣id.

Leftmost derivation of id + id * id

1. E⇒E+T
2. E⇒T+T
3. T⇒F+T
4. F⇒id+T
5. T⇒T∗F
6. T⇒F∗F
7. F⇒id∗F
8. F⇒id∗id giving string id + id * id.

Rightmost derivation of id + id * id

1. E⇒E+T
2. E⇒T+T
3. T⇒T∗F (rightmost T)
4. First T on left: T⇒F∗F
5. Rightmost F: F⇒id giving F∗id
6. Left F: F⇒id∗id so right side becomes id * id
7. Leftmost T in T + T replaced:T⇒F
8. F⇒id to get id + id * id.

Parse tree construction

 Draw root node E.

 Split E into children following the structure implied by any derivation: E
at top, then left child E, plus sign, and right child T, etc., until all
non-terminals are expanded to terminals id, +, *.
 The yield (leaf nodes from left to right) must be id + id * id; the
same parse tree corresponds to both leftmost and rightmost
derivations for an unambiguous grammar like this one.
ModelPaperSolvedAI_4

18. CNN, RNN, and LSTM

(a) Convolutional Neural Network (CNN)

A CNN is a deep neural network designed primarily for grid-like data such as
images, using convolution and pooling layers to automatically learn spatial
features.

 Convolution layers slide learnable filters over the input to detect

patterns like edges and textures, producing feature maps.
 Pooling layers down-sample feature maps (e.g., max pooling) to reduce
dimensionality and make features more translation invariant, followed
by fully connected layers for classification, e.g., object recognition.

(b) Recurrent Neural Network (RNN)

An RNN is a neural network for sequential data where output at each time
step depends on current input and a hidden state carrying information from
previous steps.

 The same set of weights is reused across time, making RNNs suitable
for text, speech, and time-series modeling.
 Standard RNNs suffer from vanishing and exploding gradients during
training, which makes learning long-term dependencies difficult.

(c) Long Short-Term Memory (LSTM)

LSTM is a special type of RNN that introduces a memory cell and gates to
overcome vanishing gradients and remember information over long
sequences.
ModelPaperSolvedAI_4

 Each LSTM cell has input, forget, and output gates that control how
much of the past cell state is kept, how much new information is
added, and what part is exposed as the hidden state.
 This allows LSTMs to model long-range dependencies effectively and
they are widely used in language modeling, machine translation,
speech recognition, and sequence prediction.