৪৯তম বিসিএস ⎯ তথ্য ও যোগাযোগ প্রযুক্তি [২৮১]

Old AI techniques (1950s–1980s) such as rule-based systems, expert systems, symbolic AI relied on explicit knowledge representation (e.g., IF–THEN rules, semantic networks).
The system could only work if all knowledge was predefined by humans → very hard to scale, inflexible, and brittle.
Modern AI (machine learning, deep learning) learns automatically from data rather than requiring humans to encode every detail.

Example:
An expert system for medical diagnosis needs thousands of rules like:
  IF fever AND cough AND chest-pain → THEN pneumonia.
If a new disease appears, system fails unless rules are manually updated.

The term “Artificial Intelligence (AI)” was first introduced by John McCarthy in 1956 at the Dartmouth Conference (Hanover, USA).
This event is considered the official birth of AI as a research discipline.

Before 1956, there were important foundations:
•1943 → McCulloch & Pitts created a model of artificial neurons.
•1950 → Alan Turing introduced the “Turing Test” in his paper Computing Machinery and Intelligence.

But 1956 Dartmouth Workshop is recognized as the official launch year of AI.

Example:
• At the Dartmouth Conference (1956), researchers like John McCarthy, Marvin Minsky, Allen Newell, and Herbert Simon met to discuss how machines could simulate intelligence.
•This led to early AI programs like the Logic Theorist (1956) and General Problem Solver (1957).

Source: Artificial Intelligence (2nd Edition) by Elaine Rich & Kevin Knight

Old AI (Classical AI / Symbolic AI) is relied on explicit rules, symbolic reasoning, and logic programming (expert systems, knowledge-based systems).

New AI (Modern AI / Machine Learning era) is focuses on data-driven learning, where systems automatically learn patterns from large datasets instead of relying on manually encoded rules.
Techniques include machine learning, deep learning, neural networks, reinforcement learning.

Logic programming is a key technique of old AI (symbolic AI).
It is based on formal logic (first-order predicate logic, rules, facts, and inference engines).
The most common language for logic programming is Prolog (Programming in Logic).

Predicate Logic (First-Order Logic, FOL) has four main components like below :-

Semantic networks represent knowledge in a graph form (nodes = objects, edges = relationships). They are excellent for showing hierarchical IS-A and HAS-A relations.

Knowledge Representation (KR) in AI is the process of encoding information about the world so that a computer system can use it to solve complex tasks (like reasoning, planning, learning).

Propositional logic (Boolean logic) uses connectives (also called logical operators) to combine or modify propositions.

Common connectives:
 AND ( ∧ ) is true if both propositions are true.
OR ( ∨ ) is true if at least one proposition is true.
IMPLIES ( → ) is true if whenever the first proposition is true, the second is also true.

• FOL uses predicates (e.g., Human(x), Mortal(x)) and quantifiers (e.g., ∀ “for all”, ∃ “there exists”) to talk about objects in a domain and their properties/relations.
• ∀x (Human(x)→Mortal(x))\forall x\,(\text{Human}(x) \rightarrow \text{Mortal}(x))∀x(Human(x)→Mortal(x)) reads: “For every object xxx, if xxx is human, then xxx is mortal.”

This has both predicates and a quantifier, so it’s properly FOL.
∀x(Human(x)→Mortal(x)) is the correct FOL with predicate symbols and a universal quantifier.

Example:
Domain = all people.
If Alice is human, FOL lets us infer Mortal(Alice) from the universal rule

∀x(Human(x)→Mortal(x)).

Source: Artificial Intelligence (2nd Ed.) by Elaine Rich & Kevin Knight

In First-Order Logic (FOL), the existential quantifier ( ∃ ) is used to express that at least one object in the domain satisfies a given property. It is used to indicate the existence of an object that has a specific characteristic.

Formal Meaning of Existential Quantifier ( ∃ ):
•The symbol ∃x P(x) means "there exists at least one object x in the domain such that P(x) is true."
•It asserts that there is some element (not necessarily all elements) in the domain for which the property holds.

Example:
If we say ∃x (Human(x) ∧ Alive(x)), it means "there exists at least one person who is alive."

Source: Artificial Intelligence (2nd Ed.) by Elaine Rich & Kevin Knight

In First-Order Logic (FOL), we use quantifiers and predicates to represent statements about objects in a domain.
Existential quantifier ( ∃ ): This asserts that "there exists at least one object" that satisfies a given condition.
Universal quantifier ( ∀ ): This asserts that "for all objects in the domain" the condition is true.

To represent the sentence "Some students are intelligent," we need to express that there is at least one student who is intelligent. The existential quantifier ( ∃ ) helps us express this.

∃x(Student(x) ∧ Intelligent(x)) means that "there exists at least one object x such that x is a student and x is intelligent." This is the correct representation of "Some students are intelligent."


Source: Artificial Intelligence (2nd Ed.) by Elaine Rich & Kevin Knight

The rule says:

• If P is true,
• and P → Q is true,
• then we can infer Q.

This is exactly the definition of Modus Ponens (MP).

Modus Ponens
Rule:  P,(P→Q)⇒Q.

Example:
• Premise1: It is raining. (P)
• Premise2: If it rains, the ground will be wet. (P → Q)
• Conclusion: The ground will be wet. (Q)

Source: Artificial Intelligence (2nd Ed.) by Elaine Rich & Kevin Knight

Skolemization is a technique in First-Order Logic (FOL) used to eliminate existential quantifiers (∃) by introducing Skolem functions or constants. It is typically used in converting logical expressions into prenex normal form or in automated theorem proving.
In Skolemization, we replace existentially quantified variables with either a Skolem constant (if no universal quantifier is preceding) or a Skolem function (if preceded by a universal quantifier).

∃x P(x) would be replaced by P(c), where c is a new Skolem constant representing an arbitrary object that satisfies P(x).

Example:
Statement: “Someone loves Mary” → ∃x Loves(x, Mary)
After Skolemization → Loves(John, Mary) (John is a Skolem constant).


Source: Artificial Intelligence (2nd Ed.) by Elaine Rich & Kevin Knight

The primary goal of search algorithms in AI is to find the most optimal solution to a problem within a given problem space. This involves exploring possible solutions (or states) and selecting the best one according to predefined criteria (such as the least cost or shortest path).

Example:
In the A search algorithm*, the goal is to find the shortest path between a start node and a goal node in a graph, considering both the actual cost and a heuristic estimate.


Source: "Artificial Intelligence: A Modern Approach" by Stuart Russell and Peter Norvig

Breadth-First Search (BFS) is a complete search technique, meaning it is guaranteed to find a solution if one exists, provided the search space is finite. It explores all nodes at the present depth level before moving on to nodes at the next depth level.

Example:
In a maze, BFS will systematically explore each possible path level by level, ensuring that it eventually finds the shortest path to the exit (if one exists).


Source: "Artificial Intelligence: A Modern Approach" by Stuart Russell and Peter Norvig

A* Search uses a priority queue to select the next node to explore based on both the cost to reach the node and a heuristic estimate of the cost to reach the goal. The priority queue helps to prioritize nodes that appear most promising, ensuring optimal performance.

Example:
In pathfinding, A* uses a priority queue to prioritize nodes with the least total cost, where the total cost is the sum of the actual cost from the start node and the estimated cost to the goal.

Source: "Artificial Intelligence: A Modern Approach" by Stuart Russell and Peter Norvig

A blind search, also known as an uninformed search, is a search algorithm that does not have any additional information about the goal other than the problem definition itself. It explores the search space without any heuristics to guide the search. Breadth-First Search (BFS) is an example of a blind search because it simply explores all possible nodes level by level, without any additional knowledge or cost evaluation about which path may lead to the goal more efficiently.
BFS does not use any heuristic or additional information about the goal. It explores all possible paths, starting from the initial node and gradually expanding to all reachable nodes at the current depth level, and then moves to the next depth level.
It guarantees finding the shortest path in an unweighted graph, but it does not have any knowledge of which path is likely to be the shortest or optimal beyond exploring all possible options equally.

Example:
Consider a simple graph:
 
Goal: Find node F starting from A.
BFS will first explore A's immediate neighbors (B and C), then explore B's neighbors (D and E), and finally explore C's neighbor (F).
This process is systematic and does not use any additional information or heuristics to guide its search.

Source: Introduction to Artificial Intelligence by Wolfgang Ertel

BFS explores each vertex once and checks all edges connected to each vertex. Therefore, the time complexity is O(V + E), where VVV is the number of vertices and EEE is the number of edges in the graph.

Example:
If you have a graph with 100 vertices and 200 edges, BFS will visit each vertex once and examine each edge once, giving a total complexity of O(100 + 200) = O(300).

Source: "Introduction to Algorithms" by Cormen, Leiserson, Rivest, and Stein.

In probabilistic reasoning, Bayes' theorem is the fundamental tool used to update beliefs based on new evidence. It provides a mathematical framework for revising the probability of a hypothesis in light of new data or evidence. This process of updating beliefs is called Bayesian updating.

Bayes' Theorem:
Bayes' theorem relates current knowledge (the prior belief) with new evidence to form an updated belief (posterior probability). The formula is:


Example:
If 1% of people have a disease and a test correctly identifies 99% of cases, Bayes’ theorem can calculate the probability a person has the disease given a positive test result.

Source: "Artificial Intelligence: A Modern Approach" by Stuart Russell and Peter Norvig.

In probability theory, a joint probability distribution refers to the probability of two or more events happening simultaneously. It is denoted by P(A, B), which represents the probability that both events A and B occur at the same time.

Joint Probability Distribution:
• Joint probability is the probability of two events A and B happening together.
• It is denoted as P(A, B) or P(A ∩ B), which indicates the probability that both events A and B occur simultaneously.
• It is crucial for understanding relationships between variables in scenarios where more than one outcome is possible.

Source: "Artificial Intelligence: A Modern Approach" by Stuart Russell and Peter Norvig.

Conditional independence allows variables to be independent given some other variable. This reduces the number of probabilities we need to specify and makes probabilistic reasoning computationally feasible.

Example:
Suppose “Sprinkler” and “Rain” both influence “Wet Grass.”
Given the status of “Wet Grass,” Sprinkler and Rain may become conditionally independent.

Source: "Artificial Intelligence: A Modern Approach" by Stuart Russell and Peter Norvig.

In a Hidden Markov Model (HMM), the states of the system are hidden, meaning we cannot directly observe them. However, we can observe emissions or outputs that are probabilistically related to the hidden states.

Example:
In speech recognition, the hidden states might represent the phonemes, while the observed output is the audio signal that is generated by these phonemes.


Source: "Pattern Recognition and Machine Learning" by Christopher Bishop

Bayesian Networks use directed acyclic graphs (DAGs) to represent the probabilistic dependencies between variables. Each node represents a random variable, and each directed edge represents a probabilistic dependency between two variables.

Example:
In a medical diagnosis system, a Bayesian Network might model the relationship between symptoms (e.g., cough, fever) and diseases (e.g., flu, pneumonia), allowing probabilistic inference based on observed symptoms.


Source: "Artificial Intelligence: A Modern Approach" by Stuart Russell and Peter Norvig

Part-of-speech (POS) tagging is the process of identifying the grammatical category (such as noun, verb, adjective) of each word in a sentence. It is an important step in understanding the structure of language.

Example:
In the sentence "The dog barked loudly," POS tagging would label "dog" as a noun and "barked" as a verb.

Source: "Speech and Language Processing" by Daniel Jurafsky and James H. Martin.

Word2Vec is a word embedding technique developed by Tomas Mikolov and his team at Google.
It was introduced in 2013 through two key research papers.

Word2Vec became revolutionary because it used a neural network–based shallow model to learn vector representations of words, capturing semantic and syntactic meanings efficiently.

Example:
Word2Vec can capture relationships like:
vector("king") – vector("man") + vector("woman") ≈ vector("queen")

This was a major breakthrough in NLP.

Source: Artificial Intelligence: A Modern Approach by Russell & Norvig.

FastText is a word embedding and text classification technique developed by Facebook AI Research (FAIR) in 2016.
It extends Word2Vec by representing each word as a bag of character n-grams, which helps it handle like below,
• Rare words
• Morphologically rich languages
• Out-of-vocabulary (OOV) words

Unlike Word2Vec, FastText can generate embeddings even for words that were not seen during training, by breaking them into subwords.

Example:
• Word: “playing”
• FastText splits it into character n-grams: pla, lay, ayi, yin, ing.
• The embedding of "playing" is built from the embeddings of its subwords.

This helps the model understand new words like "playfulness" even if it wasn’t in the training data.

Source: Artificial Intelligence: A Modern Approach by Russell & Norvig.

The Porter Stemmer is one of the most widely used algorithms for stemming in NLP. It reduces words to their root form by applying a series of rules for removing suffixes.

Example:
“running” → “run”
“happiness” → “happi”

Source: Speech and Language Processing by Daniel Jurafsky and James H. Martin

Sentiment analysis is commonly used in NLP to determine the sentiment (positive, negative, or neutral) expressed in text. This is particularly useful for analyzing social media content like tweets, reviews, and posts.

Example:
In a tweet like “I love this new phone!” sentiment analysis would classify it as positive sentiment.

Source: Speech and Language Processing by Daniel Jurafsky and James H. Martin

Bag of Words (BoW) is a model that represents a text by the frequency of words, completely ignoring word order. In contrast, Word2Vec creates dense word vectors that capture the semantic relationships between words, meaning words that are similar in meaning are closer together in vector space.

Example:
• In BoW, the sentence "The cat sat on the mat" would be represented as a set of word counts, ignoring the sequence.
• In Word2Vec, “king” and “queen” will have similar embeddings because they are semantically related, while BoW treats them as distinct words.

Source: "Speech and Language Processing" by Daniel Jurafsky and James H. Martin.

Word2Vec is often used in machine translation tasks to represent words in a way that captures their semantic meanings. These word embeddings help the translation model to map words in one language to semantically similar words in another language.

Example:
In machine translation, Word2Vec can help map the word “dog” in English to “perro” in Spanish based on their semantic similarity.


Source: "Speech and Language Processing" by Daniel Jurafsky and James H. Martin.

The inference engine is the part of an expert system that processes the knowledge base and applies logical rules to draw conclusions or make decisions. It simulates the reasoning process of human experts.

Example:
In a legal expert system, the inference engine applies legal rules to facts and evidence to suggest legal actions.

Source: Expert Systems: Principles and Programming by Joseph C. Giarratano and Gary D. Riley

Expert systems typically use deductive reasoning, where they apply general rules to specific facts to derive conclusions. It is a top-down approach where the system applies known knowledge to arrive at new conclusions or solutions.

Example:
In a medical expert system, a rule like "If a patient has a fever and cough, then they may have the flu" would be applied to specific symptoms to deduce that the patient might have the flu.

Source: "Artificial Intelligence: A Modern Approach" by Stuart Russell and Peter Norvig,

The primary goal of machine learning is to create algorithms that can learn from data, recognize patterns, and make decisions or predictions without being explicitly programmed. The system improves its performance by analyzing and learning from data rather than following pre-programmed instructions.

Example:
In supervised learning, a model is trained using labeled data (inputs and outputs) to learn the relationship between them. For instance, a model might learn to predict the price of a house based on features like square footage, location, and age.

Source: Artificial Intelligence: A Modern Approach by Stuart Russell and Peter Norvig

In supervised learning, the model is trained on labeled data (data with known input-output pairs). The model learns to map inputs to the correct outputs, and its performance is evaluated based on how well it predicts the output for new, unseen data.
In machine learning, labeled data refers to data that has been annotated with correct answers or classifications. 

Example:
A model is trained to classify emails as spam or not spam using a labeled dataset of emails and their corresponding labels (spam or not spam).

Source: Pattern Recognition and Machine Learning by Christopher Bishop

In unsupervised learning, the algorithm is trained on unlabeled data, meaning the data does not contain specific output labels. The system must identify patterns or structures in the data, such as clusters or associations.

Example:
Clustering is an unsupervised learning task where the algorithm groups similar data points together. For example, a customer segmentation task might group customers based on purchasing behavior without predefined categories.


Source: Artificial Intelligence: A Modern Approach by Stuart Russell and Peter Norvig

The main difference between supervised learning and unsupervised learning lies in the type of data used. Supervised learning uses labeled data (input-output pairs) to learn the relationship between inputs and outputs, while unsupervised learning involves discovering patterns in unlabeled data without predefined outputs.

Example:
• Supervised learning: Predicting the price of a house based on features (e.g., square footage, number of rooms).
• Unsupervised learning: Grouping customers into clusters based on purchasing behavior without predefined labels.


Source: Pattern Recognition and Machine Learning by Christopher Bishop

Precision is a common evaluation metric for classification tasks, especially in imbalanced datasets. It measures the proportion of true positive predictions out of all positive predictions made by the model.

Example:
If a spam email classifier identifies 100 emails as spam, and 80 of them are actually spam (True Positives), then the precision is 80%.

Source: Pattern Recognition and Machine Learning by Christopher Bishop.

Decision trees are an example of a supervised learning algorithm, where the model is trained on labeled data. It uses the feature values to create a tree-like structure for classifying data or predicting continuous outcomes.

Example:
A decision tree could be used to classify emails as spam or not spam, where the features are attributes like the email content, sender, and other relevant data.

Source: Pattern Recognition and Machine Learning by Christopher M. Bishop

Supervised learning is used for tasks where the input data (features) is associated with known output labels. In this case, the house price is a continuous variable, making this a regression problem in supervised learning.

Example:
A regression model can be trained on historical data of houses with their features and prices, and then the model can predict the price of new houses based on their features

Source: Machine Learning: A Probabilistic Perspective by Kevin P. Murphy,

k-Means clustering is an unsupervised learning algorithm that partitions data into clusters based on similarity, without using labeled data. The algorithm tries to group similar data points together.

Example:
In a customer segmentation task, k-Means clustering could group customers based on their purchasing behaviors without prior knowledge of the clusters (labels).

Source: Pattern Recognition and Machine Learning by Christopher Bishop.

Principal Component Analysis (PCA) is an unsupervised learning technique used for dimensionality reduction. It reduces the number of variables in high-dimensional data by transforming it into a set of orthogonal components, while retaining as much variance as possible.

Example:
PCA can be used to reduce the number of features in a dataset of customer preferences while maintaining the most important variations between customers.

Source: Pattern Recognition and Machine Learning by Christopher Bishop.

Linear regression is used for predicting a continuous output (like house price) based on multiple input features. It assumes a linear relationship between the independent variables (house features) and the dependent variable (price).

Example:
In predicting house prices, linear regression would learn a model that might look like this:



Source: Machine Learning: A Probabilistic Perspective by Kevin P. Murphy.

Collaborative filtering is a popular method used in recommendation systems. It works by recommending items to users based on the preferences of other users who have similar tastes. It can be user-based (recommending items liked by similar users) or item-based (recommending items that are similar to those a user has liked).

Example:
In a movie recommendation system like Netflix, collaborative filtering might recommend movies to a user based on the viewing habits of users with similar preferences.


Source: Machine Learning: A Probabilistic Perspective by Kevin P. Murphy

In linear regression, the model tries to minimize the sum of squared errors (SSE), which is the sum of the squared differences between the actual output values and the predicted values. The goal is to find the best-fitting line that minimizes this error.

Example:

Source: Pattern Recognition and Machine Learning by Christopher Bishop.

Mean Squared Error (MSE) is the most common metric used to evaluate the performance of regression models. It measures the average squared difference between the predicted values and the actual values.


Example:

Source: Machine Learning: A Probabilistic Perspective by Kevin P. Murphy

• K-means clustering requires the number of clusters (k) to be specified in advance.

• DBSCAN (Density-Based Spatial Clustering of Applications with Noise) automatically detects clusters based on density and does not require a predefined number of clusters.

• Agglomerative Hierarchical Clustering starts with each data point as its own cluster and merges them based on a distance measure, so the number of clusters is not predefined.

Example:
• DBSCAN can discover clusters of arbitrary shapes by analyzing density, and it can also identify outliers.
• Agglomerative Hierarchical Clustering can generate a dendrogram to help decide the number of clusters based on hierarchical merging.

Source: Pattern Recognition and Machine Learning by Christopher Bishop.