Q1: What are the characteristics of Intelligent Systems?

A1: Intelligent systems possess characteristics such as adaptability, autonomy, learning

ability, and problem-solving capabilities.
Examples include self-driving cars that can learn from their environment and adapt to
different driving conditions, or AI-powered chatbots that can understand and respond to
human language.

Q2: What are the differences between AI, Machine Learning, and Deep Learning?
A2:
AI is the broader concept of creating intelligent machines.
Machine Learning is a subfield of AI that involves creating algorithms that can learn from
data.
Deep Learning is a subset of Machine Learning that utilizes artificial neural networks to
model complex patterns and representations.

Q3: What are the differences between Computational Intelligence and Machine Learning?
A3:
Computational Intelligence is a subfield of AI that emphasizes biologically-inspired methods,
such as neural networks, evolutionary algorithms, and fuzzy logic.
Machine Learning is a broader subfield that also includes statistical and mathematical
methods for learning from data.

Q4: What is data mining? What is knowledge engineering? What are the differences
between them?
A4:
Data mining is the process of discovering useful patterns and knowledge from large data
sets.
Knowledge engineering is the process of creating knowledge-based systems by encoding
expert knowledge in a structured form.
Data mining focuses on extracting knowledge from data, while knowledge engineering is
about encoding and representing knowledge.

Q5: What are directed and undirected data mining?

Give an example for each.
A5:
Directed data mining, also known as supervised learning, involves learning a relationship
between input and output variables, like predicting house prices based on features.
Undirected data mining, or unsupervised learning, involves finding patterns without a
specific target variable, like clustering customers based on their purchasing behavior.

Q6: List and explain the data mining process.

Illustrate them with an example application.
A6: The data mining process includes:

Problem definition: Identify the problem to be solved (e.g., customer segmentation).

Data collection: Gather relevant data (e.g., customer transaction history).
Data preprocessing: Clean and transform the data (e.g., handling missing values, normalizing
data).
Data exploration: Understand patterns and relationships in the data (e.g., using visualization
techniques).
Model building: Apply data mining algorithms (e.g., clustering algorithms for segmentation).
Evaluation: Assess the model's performance (e.g., interpret clusters and validate with
domain knowledge).
Deployment: Implement the solution in the business process (e.g., targeted marketing
campaigns).

Q7: What are the 6 basic phases in Knowledge Engineering?

Illustrate with an example on how you will perform them.
A7: The 6 phases are:

Identification: Determine the problem and domain (e.g., diagnosing a medical condition).
Conceptualization: Define the main concepts and relationships (e.g., symptoms, diseases,
and their relationships).
Formalization: Represent the knowledge using formal methods (e.g., rules, ontologies).
Implementation: Create a knowledge-based system (e.g., a rule-based expert system for
diagnosis).
Evaluation: Assess the system's performance and correctness (e.g., comparing diagnoses
with expert opinions).
Maintenance: Update and refine the system as needed (e.g., incorporating new medical
knowledge).

Q8: Describe the 7 typical problem types addressed by intelligent systems and provide
examples to illustrate the types.
A8: The 7 problem types are:

Classification: Assigning data points to predefined categories (e.g., spam detection).

Regression: Predicting continuous values (e.g., house price prediction).
Clustering: Grouping data points based on similarities (e.g., customer segmentation).
Association rule learning: Discovering relationships between variables (e.g., market basket
analysis).

Anomaly detection: Identifying unusual data points (e.g., fraud detection).

Optimization: Finding the best solution to a problem (e.g., scheduling, resource allocation).
Reinforcement learning: Learning to make decisions based on interactions with an
environment (e.g., robotics, game playing).
Q9: Describe the incompatible, inconsistent, missing, noisy, and imbalanced data problem.
Provide discussions on the strategy and methods to handle these problems.
A9:

Incompatible data: Data collected from different sources or formats. Solution: Data
integration or transformation.
Inconsistent data: Contradictory information in the data. Solution: Data cleaning, resolving
conflicts using domain knowledge or expert input.
Missing data: Absence of data for certain attributes. Solution: Data imputation, using
statistical methods or machine learning to estimate missing values.
Noisy data: Random errors or fluctuations in the data. Solution: Data smoothing, outlier
detection and removal, or noise-robust algorithms.
Imbalanced data: Unequal distribution of classes in the data. Solution: Resampling
techniques (oversampling or undersampling), cost-sensitive learning, or ensemble methods.

Q10: What are test cases? Give examples on how you would do it.
A10: Test cases are specific scenarios or inputs used to evaluate the functionality,
correctness, or performance of a system. Examples include providing input data to an AI
system and checking whether the output matches the expected result, or simulating user
interactions with a chatbot to test its response quality and accuracy.

Q11:
What is a prototype?
How is this useful in developing an AI system?
A11:
A prototype is an early, working version of a system, designed to demonstrate its core
functionality and gather feedback.
In AI development, prototyping helps to identify potential issues, refine the system's design,
and validate the system's usefulness before investing more resources in building the final
product.

Q12: Describe what are antecedents and consequent in rules? Give examples to illustrate
them.
A12: In rule-based systems, antecedents are the conditions that must be met for a rule to be
applied, while the consequent is the action or conclusion that follows. For example, in a
medical expert system, the rule could be "IF (symptom1 AND symptom2) THEN (diagnosis)",
where "symptom1 AND symptom2" is the antecedent, and "diagnosis" is the consequent.

Q13: List and describe the five members of the expert system development team.
A13:
Domain expert: Provides expertise in the specific subject matter.
Knowledge engineer: Translates domain knowledge into a formal representation for the
system.
Programmer: Implements the system using programming languages and tools.
Project manager: Oversees the project, coordinates the team, and ensures timelines and
objectives are met.
End-user: The person who will use the system, providing feedback and requirements to
ensure usability and relevance.

Q14:
Provide the basic structure of a rule-based expert system.
Describe each of the components in the structure.

A14: The basic structure includes:

Knowledge base: Contains the rules and facts representing domain knowledge.
Inference engine: Applies the rules to given facts or input data to derive conclusions.
Working memory: Temporarily stores facts and intermediate results during the inference
process.
User interface: Allows users to interact with the system, providing inputs and receiving
outputs.
Explanation facility (optional): Explains the system's reasoning and justifications for its
conclusions.

Q15:Describe the two common ways for conflict resolution in rule-based systems.
A15: The two common ways are:

Refraction: Prevents a rule from firing repeatedly by not allowing the same rule to fire again
until the working memory has changed.
Priority-based: Assigns a priority level to each rule, and when conflicts arise, the rule with
the highest priority is chosen.

Q16: What is forward chaining and backward chaining in rule inference?

A16: Forward chaining starts with known facts and applies rules to infer new facts or
conclusions. Backward chaining starts with a goal and works backward by finding rules that
support the goal, then recursively attempting to satisfy the antecedents of those rules.

Q17: What is the difference between knowledge-driven and data-driven AI systems? Give
examples for each.
A17: Knowledge-driven systems rely on predefined rules and expert knowledge, such as
rule-based expert systems for medical diagnosis. Data-driven systems learn from data, such
as a machine learning model for predicting house prices based on historical data.

Q18: What is the difference between linear and non-linear problems?

A18: Linear problems can be modeled using linear relationships between variables, such as a
straight line in a 2D space. Non-linear problems involve more complex relationships
between variables that cannot be represented by a straight line, such as a curve or more
complex shapes.

Q19: Using a basic structure of Artificial Neural Network (ANN),

explain the purpose of neurons, weights, transfer functions, and hidden layers.
A19: Neurons: Basic processing units that receive inputs and produce outputs. Weights:
Connection strengths between neurons, determining the importance of input signals.
Transfer functions (activation functions): Functions applied to the neuron's input to produce
an output, introducing non-linearity. Hidden layers: Layers of neurons between input and
output layers, allowing the ANN to learn complex representations.

Q20: List and describe the commonly used transfer functions (aka activation functions) used
in artificial neurons.
A20: Common activation functions include:

Sigmoid: Maps input to a value between 0 and 1, smooth and differentiable.

Hyperbolic tangent (tanh): Maps input to a value between -1 and 1, also smooth and
differentiable.
Rectified Linear Unit (ReLU): Maps input to the maximum of 0 and the input value,
introducing sparsity and reducing the likelihood of vanishing gradients.

Q21: What is the difference between supervised and unsupervised learning?

A21: Supervised learning uses labeled data to learn a relationship between input features
and output labels (e.g., classification, regression).
Unsupervised learning finds patterns in the data without using labeled outputs (e.g.,
clustering, dimensionality reduction).

Q22: What is the difference between clustering and classification?

A22: Clustering groups data points based on their similarities, without predefined categories
(unsupervised learning). Classification assigns data points to predefined categories based on
learned patterns from labeled data (supervised learning).

Q23: Explain or illustrate what is function approximation.

A23: Function approximation involves learning a function that maps input features to output
values based on a set of examples. For example, a regression model that learns to predict
house prices based on features like size, location, and age is approximating the underlying
function that relates these features to the price.

Q24: Explain how backpropagation network do the learning – you can use figures and steps
or algorithms to explain.

A24: Backpropagation is a supervised learning algorithm for neural networks. It involves the
following steps:

Initialize weights randomly.

Perform a forward pass to compute the network's output for a given input.
Compute the error between the output and the target value (e.g., using mean squared
error).
Backpropagate the error through the network by computing the gradients of the error with
respect to each weight using the chain rule.

Update the weights using the computed gradients and a learning rate.
Repeat steps 2-5 for multiple epochs or until convergence.
Q25: In ANN, what is the learning rate? What could happen if the learning rate is too fast?
A25: The learning rate is a hyperparameter that controls the step size of weight updates
during training. If the learning rate is too high, the training may become unstable, leading to
oscillations or divergence instead of convergence to the optimal solution.

Q26: What is overfitting?

A26: Overfitting occurs when a model learns to perform well on the training data but does
not generalize well to unseen data, often because it captures noise or random fluctuations
in the data rather than the underlying patterns.

Q27: What is cross-validation?

A27: Cross-validation is a technique for assessing the performance of a model by splitting
the data into multiple subsets (folds) and training and evaluating the model on each fold,
then averaging the performance metrics. This helps to estimate the model's performance on
unseen data and reduce overfitting.

Q28: Given a set of available data, you normally divide them into 3 sets. What are the
Training, Testing, and Validation sets? How can you use them?
A28: Training set: Used to train the model. Testing set: Used to evaluate the model's
performance on unseen data. Validation set: Used to tune hyperparameters and select the
best model during training. The model is trained on the training set, its performance is
assessed on the validation set, and after selecting the best model, it is tested on the testing
set to estimate its performance on new data.

Q29: Provide the steps of how a Self-Organizing Map (SOM) do the learning.
A29: The steps for SOM learning are:

Initialize weights randomly.

Present an input pattern to the network.
Find the Best Matching Unit (BMU), the neuron with the closest weight vector to the input
pattern.
Update the weights of the BMU and its neighbors, pulling them closer to the input pattern.
Repeat steps 2-4 for multiple epochs, gradually decreasing the learning rate and
neighborhood size.

Q30: In RBF network, what is the purpose of the hidden layer?

A30: The hidden layer in a Radial Basis Function (RBF) network consists of radial basis
functions (usually Gaussian functions) that act as local feature detectors, measuring the
similarity between the input and the centers of the basis functions.
The hidden layer transforms the input into a new representation that can be more easily
separated by the output layer.

Q31: What is the main difference between Deep Learning and the Multi-layer Neural
Network?
A31:
Deep learning refers to neural networks with many hidden layers, allowing them to learn
hierarchical and complex representations.
Multi-layer neural networks can have multiple hidden layers,
but the term does not necessarily imply the depth or complexity of deep learning models.

Q32: What is Recurrent Neural Network (RNN) and how it performs training?
A32: RNN is a type of neural network designed for processing sequences of data. It contains
loops, allowing it to maintain a hidden state that can store information from previous time
steps. RNNs are trained using backpropagation through time (BPTT), an extension of
backpropagation that unfolds the network through time and computes gradients across
multiple time steps.

Q33: What are the decision steps in using Artificial Neural Networks for a problem?
A33: The decision steps include:

Defining the problem and identifying the input and output variables.
Selecting the appropriate network architecture (e.g., number of hidden layers, number of
neurons per layer).
Choosing the activation functions for neurons.
Determining the learning algorithm and its parameters (e.g., learning rate, batch size).
Splitting the data into training, validation, and testing sets.
Training the network using the learning algorithm.
Evaluating the network's performance on the validation and testing sets.
Tuning hyperparameters or adjusting the architecture to optimize performance.
Deploying the trained model for use in real-world applications.

Q34:
What is the difference between probability and fuzzy logic?
Give examples to show the difference.
A34: Probability deals with the likelihood of events, quantifying uncertainty in terms of
frequencies or degrees of belief. Fuzzy logic represents uncertainty using fuzzy sets, which
allow partial membership to multiple categories.
For example, in probability, an object might have a 70% chance of being red and a 30%
chance of being blue.
In fuzzy logic, the same object could have a 0.7 membership in the "red" fuzzy set and a 0.3
membership in the "blue" fuzzy set.

Q35: What is the difference between Boolean logic and fuzzy logic? Explain the difference
between them with an example.
A35: Boolean logic deals with binary values (true or false) and crisp categories, while fuzzy
logic allows for partial membership in categories, using continuous values between 0 and 1.
For example, in Boolean logic, an object is either tall or not tall.
In fuzzy logic, an object could have a membership of 0.8 in the "tall" fuzzy set, indicating it is
somewhat tall but not entirely.

Q36: What are the universe of discourse and membership functions?

A36: The universe of discourse is the range of values over which a fuzzy variable is defined.
Membership functions map input values from the universe of discourse to degrees of
membership in a fuzzy set, usually between 0 and 1.
Q37: For a given value, how can you represent the value in fuzzy sets? You can explain using
an example.
A37: To represent a value in fuzzy sets, first define the fuzzy sets and their corresponding
membership functions. For example, let's consider the temperature variable with "cold,"
"warm," and "hot" fuzzy sets. Given a temperature value, say 25°C, calculate its membership
in each fuzzy set using their membership functions. The result might be 0.2 membership in
"cold," 0.8 membership in "warm," and 0.1 membership in "hot."

Q38: Explain the three main components of the fuzzy system.

A38: The three main components are:

Fuzzification: Converts crisp input values into fuzzy values using membership functions.
Inference: Combines fuzzy input values using fuzzy rules to produce fuzzy output values.
Defuzzification: Converts the fuzzy output values back into crisp values for decision-making
or control.

Q39: What is the difference between Mamdani and Sugeno fuzzy inference methods?
A39: Mamdani inference uses fuzzy sets for both input and output variables and typically
employs a defuzzification step to produce crisp output values.
Sugeno inference uses crisp functions (linear or constant) for output variables, simplifying
the defuzzification process and often resulting in more efficient computations.

Q40: What is the min-max rule?

A40: The min-max rule is a method used in fuzzy inference systems for combining the
output membership functions from multiple rules.
The method takes the minimum of the antecedent membership values and the maximum of
the consequent membership values to determine the final output membership value.

Q41: What is centroid defuzzification?

A41: Centroid defuzzification is a method of converting the fuzzy output values of a fuzzy
system back into crisp values.
It calculates the weighted average (or center of mass) of the output fuzzy set, with the
membership values acting as weights.

Q42: Provide the steps in developing a fuzzy system.

A42: The steps are:

Define the problem and identify input and output variables.

Determine the fuzzy sets and their corresponding membership functions for each variable.
Develop the fuzzy rules that relate input fuzzy sets to output fuzzy sets.
Implement the fuzzy inference system, including fuzzification, inference, and defuzzification
processes.
Test and validate the system using real-world data or simulated scenarios.
Tune the membership functions or rules to optimize system performance.
Q43: What problem is Genetic Algorithm (GA) good at solving? Give an application example
that GA can be used.
A43: GAs are good at solving optimization problems, especially in complex search spaces
where traditional methods may struggle.
An example application is optimizing the design of an aircraft wing to minimize drag while
maintaining structural integrity.

Q44: Explain the steps in the basic Genetic Algorithm.

A44: The steps are:

Initialize a population of candidate solutions randomly.

Evaluate the fitness of each individual in the population.
Select individuals for reproduction based on their fitness (e.g., using roulette wheel
selection or tournament selection).
Perform crossover (recombination) and mutation on the selected individuals to create
offspring.
Replace the old population with the new offspring (possibly using elitism to retain the best
solutions).
Repeat steps 2-5 until a stopping criterion is met (e.g., a maximum number of generations
or a satisfactory fitness level).

Q45: What are the advantages of using GA systems?

A45: Advantages of GAs include:

They are well-suited for complex search spaces and global optimization problems.
They can explore multiple solutions simultaneously, reducing the likelihood of getting
trapped in local optima.
They are robust to noisy or incomplete data.
They can be applied to a wide range of problem domains with relatively little problem-
specific knowledge.

Q46: What is the difference between multi-objectives and single objective problems?
A46: Single objective problems have one optimization goal, such as minimizing cost or
maximizing profit. Multi-objective problems involve optimizing multiple objectives
simultaneously, often requiring trade-offs among conflicting objectives, such as minimizing
cost and maximizing performance.

Q47: What is Bayesian Networks and based on what theory does BN perform its reasoning
with?
A47:
Bayesian networks are graphical models that represent the probabilistic relationships
among a set of variables. They perform reasoning based on Bayesian probability theory,
allowing for the efficient computation of conditional probabilities and updating beliefs given
new evidence.

Q48: Describe the two main approaches of how Bayesian networks can be created.
A48: The two main approaches are:
Expert-driven: Domain experts provide the structure and parameters of the network based
on their knowledge and experience.
Data-driven: The structure and parameters of the network are learned from data using
statistical learning algorithms, such as the Expectation-Maximization (EM) algorithm.

Q49: In SVM, how does the algorithm find the separation between linear and non-linear
problems?
A49: Support Vector Machines (SVMs) can find separation between linear and non-linear
problems by using kernel functions. Kernel functions implicitly map the input data into a
higher-dimensional space, where a linear hyperplane can be used to separate the data.
Common kernel functions include the linear kernel, polynomial kernel, and radial basis
function (RBF) kernel.

Q50: What is the 'maximum margin' in SVM and why does it need to be maximized?
A50: The 'maximum margin' in SVM refers to the distance between the separating
hyperplane and the closest data points (support vectors) from each class. Maximizing the
margin helps to improve the generalization of the model, as it seeks to find the hyperplane
that best separates the data while keeping the largest possible distance from the support
vectors.
This approach tends to produce models that are less sensitive to noise and more robust to
new, unseen data.

Q51: In GAN, there are two deep networks, explain what they are and their purpose.
A51: In Generative Adversarial Networks (GANs), there are two deep networks, the
Generator and the Discriminator:

Generator: This network aims to generate realistic data samples by learning the underlying
data distribution. It takes random noise as input and outputs synthetic data samples.
Discriminator: This network is trained to distinguish between real data samples and
synthetic data samples generated by the Generator. Its goal is to correctly classify real
samples as real and synthetic samples as fake.
During training, the Generator and Discriminator play a two-player minimax game, where
the Generator tries to fool the Discriminator by producing increasingly realistic samples,
while the Discriminator tries to improve its ability to differentiate between real and fake
samples. As a result, the Generator learns to create highly realistic data samples.