Unit –3

Knowledge Representation & Reasoning

Notes on Propositional Logic

3.1.1 Introduction to Propositional Logic

Propositional logic, also known as sentential logic, deals with logical relationships
between propositions (statements that can either be true or false). In propositional logic,
we analyze how propositions combine to form compound propositions and how their truth
values are determined.

• Propositions: A statement that is either true or false (e.g., "It is raining").

• Logical connectives: Operators that connect propositions, forming compound
statements (e.g., AND, OR, NOT).

3.1.2 Syntax and Semantics

• Syntax: The formal rules governing how propositions and connectives are written. A
proposition is often represented by a letter (e.g., ppp, qqq, rrr).
o Well-formed formulas (WFF): Expressions that are syntactically correct.
o Connectives: ∧\land∧ (AND), ∨\lor∨ (OR), ¬\neg¬ (NOT), ⇒\Rightarrow⇒
(implies), ⇔\Leftrightarrow⇔ (if and only if).
• Semantics: The meaning or interpretation of a logical formula. This involves
assigning truth values to propositions and determining the truth value of compound
formulas.
o Truth Value: The truth or falsity of a proposition, either true (T) or false (F).

3.1.3 Logical Connectives

1. AND (Conjunction, ∧\land∧): The result is true only if both operands are true.
a. Example: p∧qp \land qp∧q is true if both ppp and qqq are true.
2. OR (Disjunction, ∨\lor∨): The result is true if at least one operand is true.
a. Example: p∨qp \lor qp∨q is true if either ppp or qqq is true.
3. NOT (Negation, ¬\neg¬): The result is true if the operand is false, and vice versa.
a. Example: ¬p\neg p¬p is true if ppp is false.
 4. IMPLIES (Implication, ⇒\Rightarrow⇒): The result is false only if the first operand
is true and the second is false.
a. Example: p⇒qp \Rightarrow qp⇒q is false only if ppp is true and qqq is
false.
5. IF AND ONLY IF (Biconditional, ⇔\Leftrightarrow⇔): The result is true if both
operands have the same truth value.
a. Example: p⇔qp \Leftrightarrow qp⇔q is true if both ppp and qqq are either
true or false.

3.1.4 Truth Tables and Logical Equivalence

A truth table is a table that shows all possible truth values of a set of propositions and the
resulting truth value of a compound proposition.

• Logical Equivalence: Two logical expressions are equivalent if they always have the
same truth values for all possible assignments of truth values to their variables. For
example:
o p⇒qp \Rightarrow qp⇒q is logically equivalent to ¬p∨q\neg p \lor q¬p∨q.
o Tautology: A compound statement that is always true (e.g., p∨¬pp \lor \neg
pp∨¬p).
o Contradiction: A compound statement that is always false (e.g., p∧¬pp
\land \neg pp∧¬p).

Example Truth Table for p⇒qp \Rightarrow qp⇒q:

p⇒qp
p q \Rightarro
w qp⇒q
T T T
T F F
F T T
F F T

3.1.5 Normal Forms: CNF and DNF

• Conjunctive Normal Form (CNF): A conjunction of disjunctions. A formula is in

CNF if it is a series of ORs grouped by ANDs.
o Example: (p∨q)∧(¬r∨s)(p \lor q) \land (\neg r \lor s)(p∨q)∧(¬r∨s)
 • Disjunctive Normal Form (DNF): A disjunction of conjunctions. A formula is in DNF
if it is a series of ANDs grouped by ORs.
o Example: (p∧q)∨(¬r∧s)(p \land q) \lor (\neg r \land s)(p∧q)∨(¬r∧s)

3.1.6 Applications of Propositional Logic

Propositional logic is used in various fields to model and reason about systems:

1. Computer Science:
a. Boolean Algebra: Used in circuit design, search algorithms, and software
development.
b. Formal Verification: Used to verify correctness of hardware and software
systems.
2. Artificial Intelligence:
a. Knowledge Representation: To model facts about the world in a logical
manner.
b. Automated Theorem Proving: Solving logical problems automatically.
3. Mathematics:
a. Mathematical Logic: Propositional logic forms the foundation of more
complex systems like predicate logic.
b. Set Theory: Logical operations are often used in the manipulation of sets
and relations.
4. Philosophy:
a. Used to model logical arguments and reasoning.

Notes on First-Order Logic (FOL)

3.2.1 Introduction to First-Order Logic (FOL)

First-Order Logic (FOL), also known as predicate logic or first-order predicate calculus,
extends propositional logic by introducing quantifiers, predicates, and variables. It allows
for reasoning about objects, their properties, and relationships between objects.

• Propositional Logic is limited to simple true or false values, whereas First-Order

Logic can express more complex relationships involving individuals and their
attributes.
 • FOL is more expressive and can represent statements like "All humans are mortal"
or "There exists a person who is taller than John."

3.2.2 Syntax and Semantics of FOL

• Syntax: The formal structure that defines how FOL expressions are written.
o Terms: Represent objects in the domain (e.g., constants, variables).
o Predicates: Represent relationships or properties (e.g., P(x)P(x)P(x) might
mean "x is a person").
o Atomic Formula: A predicate applied to terms (e.g., P(a)P(a)P(a), meaning
"a is a person").
o Complex Formula: A combination of atomic formulas using logical
connectives (e.g., P(x)∧Q(x,y)P(x) \land Q(x, y)P(x)∧Q(x,y)).
• Semantics: The interpretation of terms and formulas.
o Domain: A set of objects that the variables can refer to.
o Interpretation: Assigns meanings to predicates, functions, and constants in
the context of a domain.
o Truth Value: The truth of a formula depends on the assignment of objects to
the variables and the interpretation of predicates and functions.

3.2.3 Quantifiers: Universal and Existential

Quantifiers are used in FOL to express statements about all or some elements of the
domain.

1. Universal Quantifier (∀\forall∀):

a. Represents a statement that is true for every element in the domain.
b. Syntax: ∀x P(x)\forall x \, P(x)∀xP(x) means "for all x, P(x) is true."
c. Example: ∀x (Human(x)⇒Mortal(x))\forall x \, (Human(x) \Rightarrow
Mortal(x))∀x(Human(x)⇒Mortal(x)) means "All humans are mortal."
2. Existential Quantifier (∃\exists∃):
a. Represents a statement that is true for at least one element in the domain.
b. Syntax: ∃x P(x)\exists x \, P(x)∃xP(x) means "there exists at least one x such
that P(x) is true."
c. Example: ∃x (Human(x)∧Tall(x))\exists x \, (Human(x) \land
Tall(x))∃x(Human(x)∧Tall(x)) means "There exists a human who is tall."
3.2.4 Predicates, Functions, and Constants

• Predicates: Functions that return a truth value. They represent properties or

relations between terms.
o Example: P(x)P(x)P(x) might denote "x is a person," or L(x,y)L(x, y)L(x,y)
might denote "x loves y."
• Functions: Functions map objects to other objects in the domain.
o Example: f(x)f(x)f(x) might represent "father of x."
• Constants: Specific, named objects in the domain.
o Example: aaa could be the constant representing "John."

3.2.5 Differences Between Propositional and First-Order Logic

1. Expressiveness:
a. Propositional Logic: Deals only with whole propositions (true/false), with no
internal structure.
b. First-Order Logic: Allows for a detailed representation of individual objects,
their properties, and relationships.
2. Variables and Quantifiers:
a. Propositional Logic: Has no concept of variables or quantifiers.
b. First-Order Logic: Introduces variables (which can stand for objects) and
quantifiers (which express statements about all or some objects).
3. Scope:
a. Propositional Logic: Limited to true/false values for entire statements.
b. First-Order Logic: Provides a richer framework for expressing properties and
relationships between elements in a domain.
4. Reasoning:
a. Propositional Logic: Primarily uses logical connectives to form complex
statements.
b. First-Order Logic: Uses quantifiers and predicates to form statements that
can describe properties and relationships of individuals.

3.2.6 Applications of First-Order Logic

1. Artificial Intelligence:
a. Knowledge Representation: FOL is widely used in AI to represent
knowledge in a form that machines can reason about, e.g., expert systems,
rule-based systems.
 b. Automated Theorem Proving: FOL is used to automatically prove
mathematical theorems by defining axioms and inference rules.
2. Database Querying:
a. SQL Queries: FOL forms the basis of structured query languages (SQL),
where queries can express relationships between data objects and their
properties.
b. Relational Databases: FOL is used to define constraints and rules in
database schemas.
3. Mathematics:
a. FOL is used to formalize mathematical reasoning and proofs. Mathematical
structures, like groups, rings, and sets, can be described using FOL.
4. Formal Verification:
a. FOL is used in the formal verification of hardware and software systems,
ensuring that certain properties hold true in all cases.
5. Natural Language Processing (NLP):
a. In NLP, FOL can be used to model the syntax and semantics of natural
languages, making it possible to analyze and generate sentences in a
structured way.

Notes on Inference in First-Order Logic (FOL)

3.3.1 Introduction to Inference Mechanisms

Inference mechanisms in First-Order Logic (FOL) are used to derive new information or
conclusions from known facts or premises. The primary goal of inference is to use logical
rules to prove or deduce new truths based on existing knowledge.

• Inference in FOL involves applying logical rules, such as modification of formulas

or quantifier manipulations, to generate new theorems or conclusions.
• Soundness and completeness are key properties that ensure the reliability and
thoroughness of the inference process.

Inference in FOL is more complex than in propositional logic because it involves variables,
predicates, and quantifiers. The most common methods used for inference in FOL are
forward chaining, backward chaining, and resolution.
3.3.2 Forward and Backward Chaining

Forward Chaining and Backward Chaining are two essential inference strategies used in
rule-based reasoning systems (like expert systems).

1. Forward Chaining:
a. Data-driven approach.
b. Starts with known facts and applies inference rules to derive new facts until
the goal is reached.
c. Procedure:
i. Begin with a set of known facts.
ii. Apply inference rules (e.g., P⇒QP \Rightarrow QP⇒Q) to generate
new facts.
iii. Repeat the process until the desired conclusion is reached or no new
facts can be derived.
d. Example: Given the facts "All humans are mortal" and "Socrates is a
human," forward chaining would infer "Socrates is mortal."
2. Backward Chaining:
a. Goal-driven approach.
b. Starts with the goal and works backwards, looking for the facts that support
it.
c. Procedure:
i. Start with a query (goal).
ii. Identify rules that could support the goal.
iii. Check if the premises of the rule are already known, otherwise work
backwards.
iv. Repeat until the premises are known facts or no solution is found.
d. Example: To prove "Socrates is mortal," backward chaining would check if
"Socrates is a human" and "All humans are mortal" are true.

3.3.3 Unification and Substitution

• Unification: The process of finding a substitution of variables that makes two terms
or formulas identical.
o Example: Unifying P(x)P(x)P(x) with P(a)P(a)P(a) results in the substitution
x↦ax \mapsto ax↦a, meaning xxx is replaced with aaa.
• Substitution: The process of replacing a variable with a specific term in a formula.
 o Example: Substituting xxx with aaa in the formula P(x)P(x)P(x) results in
P(a)P(a)P(a).

Unification and substitution are essential for making logical formulas match during the
application of inference rules, such as in forward and backward chaining.

3.3.4 Resolution in First-Order Logic

• Resolution is a powerful inference rule used to prove the unsatisfiability of a set of

clauses in FOL.
o It operates on clauses, which are disjunctions of literals (i.e., a literal is a
predicate or its negation).
o Resolution Process:
▪ Convert FOL formulas into clausal form (a set of clauses).
▪ Apply the resolution rule to combine pairs of clauses, aiming to derive
the empty clause (indicating a contradiction).
▪ If the empty clause is derived, the set of clauses is unsatisfiable
(meaning the negation of the conclusion is false).
• Example:
o Given two clauses: P(x)∨Q(x)P(x) \lor Q(x)P(x)∨Q(x) and ¬Q(a)\neg
Q(a)¬Q(a), the resolution will derive P(a)P(a)P(a).
o The empty clause □\square□ would indicate that the two clauses are
contradictory, confirming that the original premises imply the conclusion.

3.3.5 Soundness and Completeness of Inference

• Soundness: An inference system is sound if every conclusion derived using the

system is logically correct or valid. In other words, if we can derive a formula, it
must be true in all interpretations.
o For FOL: If a formula can be derived using an inference rule, it is guaranteed
to be true in any model that satisfies the premises.
• Completeness: An inference system is complete if every logically valid conclusion
can be derived using the system. In other words, if a formula is true in all
interpretations, it should be possible to derive it using the inference rules.
o For FOL: If a formula is logically valid, it should be possible to prove it using
the rules of inference in FOL.
3.3.6 Applications of Inference in FOL

Inference mechanisms in FOL are widely used in various domains, particularly in AI and
computer science. Some key applications include:

1. Expert Systems:
a. Inference engines in expert systems use FOL to derive new knowledge from
existing facts, supporting decision-making processes.
2. Automated Theorem Proving:
a. FOL is used to automatically prove mathematical theorems and verify the
correctness of statements in formal logic.
3. Knowledge Representation:
a. In AI, FOL is used for representing and reasoning about knowledge in a
structured way. Logical inference can derive new information from the facts
represented in the knowledge base.
4. Natural Language Processing (NLP):
a. FOL is applied in NLP for tasks such as semantic parsing, where sentences
are interpreted in terms of logical formulas and inferences are drawn.
5. Robotics:
a. Robots use logical inference to reason about their environment, plan
actions, and make decisions based on sensor data and previous knowledge.
6. Database Querying:
a. Logical inference mechanisms are used in databases to process queries and
derive answers from a set of known facts or data entries.

Notes on Forward & Backward Chaining

3.4.1 Introduction to Chaining Methods

Chaining methods are fundamental inference strategies used in rule-based expert

systems to derive conclusions or find solutions. These methods determine how to apply
rules to facts or goals to infer new knowledge. The two primary chaining methods are:

1. Forward Chaining: Starts from known facts and applies rules to reach the goal.
2. Backward Chaining: Starts with a goal and works backward to determine the facts
needed to prove it.

Both methods are crucial for reasoning within expert systems, allowing systems to
generate conclusions based on available knowledge.
3.4.2 Forward Chaining: Working from Facts to Goals

Forward Chaining is a data-driven reasoning approach that starts with the available facts
and applies rules to derive new facts until the goal is reached.

• Procedure:
o Start with known facts: Begin with a set of facts or known data points in the
system's knowledge base.
o Apply inference rules: Use the rules (typically in the form "If X, then Y") to
infer new facts.
o Repeat: Apply rules iteratively to the new facts generated, continually
expanding the knowledge base.
o Reach the goal: Continue the process until the desired conclusion or goal is
reached.
• Example:
o Given facts: "All humans are mortal" and "Socrates is a human."
o Rule: "If X is human, then X is mortal."
o Forward chaining would conclude: "Socrates is mortal."
• Key Features:
o Data-driven: It moves forward from known data.
o Exhaustive: Can generate all possible conclusions that follow from the
facts.
o May require extensive processing: If many rules are involved, forward
chaining may go through a large number of facts.

3.4.3 Backward Chaining: Working from Goals to Facts

Backward Chaining is a goal-driven reasoning approach that begins with a goal and works
backward to find the facts that support it.

• Procedure:
o Start with a goal or query: Identify the goal that you want to prove (e.g., "Is
Socrates mortal?").
o Identify applicable rules: Look for rules that could lead to the goal.
o Check if premises are true: For each applicable rule, check whether the
premises (conditions) of the rule are known facts or need to be proven.
o Repeat: If premises are not known, recursively apply backward chaining to
those premises until facts are found.
 o Goal satisfied: The process ends when all necessary facts are found to
prove the goal.
• Example:
o Goal: "Is Socrates mortal?"
o Rule: "If X is a human, then X is mortal."
o Premise to check: "Is Socrates a human?"
o If "Socrates is a human" is true, then "Socrates is mortal."
• Key Features:
o Goal-driven: It starts with the goal and looks for facts that support it.
o Efficient: Only the necessary facts and rules are explored.
o Recursive: The process works backward through the knowledge base until
the facts are found.

3.4.4 Comparison of Forward and Backward Chaining

Feature Forward Chaining Backward Chaining

Starting Starts with facts and moves towards Starts with the goal and works
Point the goal. backwards.
Approach Data-driven (facts to conclusions). Goal-driven (goal to facts).
May explore all facts, leading to More efficient, only explores
Efficiency
higher processing time. necessary facts.
Ideal for scenarios where all possible Best suited for proving specific
Use Case
outcomes need to be explored. goals or queries.
Rule
Applies rules as facts become Works backward from the goal,
Applicatio
available. finding rules that can prove it.
n
Memory
May require storing many facts during Requires less memory as it only
Consumpti
the process. tracks relevant facts.
on

• Strengths:
o Forward Chaining: Ideal when the system needs to deduce multiple
conclusions from a set of facts.
o Backward Chaining: Efficient when working with specific goals, as it avoids
unnecessary processing by focusing only on relevant facts.
• Weaknesses:
 o Forward Chaining: Can be computationally expensive, especially in large
knowledge bases.
o Backward Chaining: Might be less efficient if the goal is not specific enough
or if there are many potential paths to check.

3.4.5 Applications of Chaining in Expert Systems

Both Forward and Backward Chaining are widely used in expert systems and other AI-
based applications for decision-making and problem-solving.

1. Expert Systems:
a. In expert systems, forward chaining is typically used in diagnostic systems
or situations where the system needs to explore all possible consequences
of known facts (e.g., medical diagnosis systems).
b. Backward chaining is often used in rule-based systems where a goal needs
to be verified or specific conclusions are required (e.g., troubleshooting
systems).
2. Medical Diagnosis:
a. Forward chaining can be used to infer symptoms based on known diseases
and conditions.
b. Backward chaining can be used to trace symptoms back to possible
diseases or conditions.
3. Problem Solving in AI:
a. Expert systems use both methods to solve complex problems in fields like
law, engineering, and finance. In these systems, backward chaining is often
used to trace a specific solution path, while forward chaining is used for
exploring all possible solutions.
4. Planning and Scheduling:
a. Forward chaining can be used to plan actions and reach goals, while
backward chaining can be used to plan backward from a desired goal state
to determine the necessary steps.
5. Learning Systems:
a. Chaining methods can be used in learning environments to derive new
knowledge from existing data and conclusions, applying inference to learn
patterns or make predictions.
6. Search Algorithms:
 a. In AI search problems, backward chaining can be used to explore paths
backward from the goal, while forward chaining can explore all potential
solutions from the initial state.

Notes on Resolution in First-Order Logic

3.5.1 Introduction to Resolution in Logic

Resolution is a rule of inference used in First-Order Logic (FOL) and propositional logic
for proving the validity of logical statements. It is the primary method employed by
automated theorem provers to derive conclusions or refute hypotheses. Resolution is
based on the idea of refutation: a conclusion is inferred by proving that the negation of the
desired result leads to a contradiction.

• Key Concept: Resolution involves combining two clauses that contain

complementary literals (a literal and its negation) to derive a new clause.
• Goal: The aim is to derive an empty clause (a contradiction) to prove that a set of
clauses is unsatisfiable, which in turn proves the original statement.

3.5.2 Steps in the Resolution Process

The resolution process in FOL can be broken down into the following steps:

1. Convert Statements to Conjunctive Normal Form (CNF):

a. First, all logical statements are converted into conjunctive normal form
(CNF). In CNF, a formula is represented as a conjunction of clauses, where
each clause is a disjunction of literals.
b. Example: (P∨Q)∧(¬Q∨R)(P \vee Q) \wedge (\neg Q \vee R)(P∨Q)∧(¬Q∨R).
2. Identify Complementary Literals:
a. Next, complementary literals are identified in two clauses. Complementary
literals are pairs of literals where one is the negation of the other. For
example, PPP and ¬P\neg P¬P.
3. Apply the Resolution Rule:
a. The resolution rule is applied by combining the two clauses that contain
complementary literals. The resulting clause is a disjunction of the remaining
literals from both clauses after the complementary literals have been
eliminated.
b. Example:
 i. Clause 1: P∨QP \vee QP∨Q
ii. Clause 2: ¬P∨R\neg P \vee R¬P∨R
iii. Resolution: Combine the remaining literals: Q∨RQ \vee RQ∨R.
4. Repeat the Process:
a. The process of resolving clauses continues until no new clauses can be
derived or until an empty clause (a contradiction) is produced, indicating
that the set of clauses is unsatisfiable.
5. Conclusion:
a. If the resolution process results in an empty clause, the original set of
statements is unsatisfiable, and the negation of the goal is false, thus
proving the goal.
b. If no empty clause is derived, the goal cannot be proven.

3.5.3 Unification in Resolution

Unification is a crucial step in the resolution process, as it ensures that complementary

literals can be matched correctly across clauses. Unification is the process of finding a
substitution that makes two literals identical.

• Key Concept: Unification involves substituting variables with terms so that two
literals match.
• Example:
o Given literals: P(x)P(x)P(x) and P(a)P(a)P(a), where xxx is a variable and aaa
is a constant.
o The unification process would substitute x=ax = ax=a to make the two
literals identical.

Unification plays a central role in resolving clauses, as it allows the combination of

clauses with complementary literals even when the literals involve different variables or
terms.

3.5.4 Applications of Resolution in Automated Theorem Proving

Resolution is widely used in automated theorem proving systems due to its simplicity and
effectiveness in proving logical statements. Some of the key applications of resolution
include:

1. Automated Theorem Provers:

 a. Resolution is the backbone of many automated theorem provers, which
attempt to prove or disprove logical statements by iteratively applying
resolution to sets of clauses. These provers are used in fields like
mathematics, logic, and computer science to automatically verify proofs.
2. Artificial Intelligence:
a. In AI systems, resolution is used for reasoning in expert systems and
knowledge-based systems. It is used to deduce new facts from a set of
known facts and rules.
b. It is also used in constraint satisfaction problems where the goal is to find
a solution that satisfies all constraints, and logic programming (e.g., Prolog)
uses resolution to answer queries by searching through a database of facts
and rules.
3. Model Checking:
a. Model checking involves verifying whether a system satisfies a given
specification, which can often be expressed as a logical formula. Resolution-
based methods can be used to check if a system’s behavior is consistent
with its specification by proving the logical entailments of system properties.
4. Knowledge Representation:
a. Resolution plays a role in knowledge representation systems, where
complex facts and rules are represented in formal logic. By resolving logical
statements, the system can deduce new knowledge or check for
consistency.
5. Artificial Neural Networks:
a. Some techniques in neural network design and optimization utilize logical
resolution to verify correctness or to optimize the learning algorithms.

3.5.5 Limitations and Challenges of Resolution

While resolution is a powerful tool for proving theorems, it comes with certain limitations
and challenges:

1. Computational Complexity:
a. The resolution process can be computationally expensive, especially in
cases involving a large number of clauses. The space and time complexity of
resolution can grow rapidly as the number of variables and clauses
increases.
2. Inefficiency with Large Knowledge Bases:
 a. In large or complex knowledge bases, the number of possible resolutions
can explode, leading to inefficiency. Many unnecessary resolutions may be
performed, leading to combinatorial explosion.
3. Need for Conversion to CNF:
a. Converting formulas to CNF can be difficult and time-consuming, especially
for formulas that are not naturally in CNF. The process of conversion may
also increase the size of the formula, further complicating the resolution
process.
4. Ambiguity in Unification:
a. In some cases, unification can be ambiguous, especially when there are
multiple ways to unify terms. This ambiguity can lead to errors or
inefficiencies in the resolution process.
5. Failure to Find a Proof:
a. While resolution is a refutation method (proving the negation of a statement),
it does not always find a proof for a statement unless it is explicitly
unsatisfiable. For some statements, it may not be able to derive a
contradiction, thus failing to prove the statement.
6. Limited Expressiveness:
a. Resolution works well in propositional logic and first-order logic but is less
effective in more complex or higher-order logics, where more advanced
techniques may be necessary.

3.6 Probabilistic Reasoning

3.6.1 Introduction to Probability Theory

Probability theory is a branch of mathematics that deals with the likelihood of events
occurring. In AI, probabilistic reasoning is used to make decisions and predictions in
situations involving uncertainty.

• Key Concept: Probability is a measure of the likelihood of an event. It is a value

between 0 and 1, where 0 means an event will not occur, and 1 means it will occur
with certainty.
• Basic Definitions:
o Random Variable: A variable whose value is subject to chance.
o Event: An outcome or a set of outcomes of a random process.
o Sample Space: The set of all possible outcomes of a random experiment.
 o Probability Mass Function (PMF): A function that gives the probability that a
discrete random variable is equal to a particular value.

3.6.2 Probability Distributions and Bayes’ Theorem

A probability distribution is a mathematical function that provides the probabilities of

occurrence of different possible outcomes in an experiment.

• Types of Probability Distributions:

o Discrete Distributions: Deal with discrete random variables (e.g., Bernoulli,
Binomial, Poisson).
o Continuous Distributions: Deal with continuous random variables (e.g.,
Normal, Exponential).
• Bayes' Theorem: Bayes' Theorem provides a way to update the probability of a
hypothesis based on new evidence.

P(H∣E)=P(E∣H)P(H)P(E)P(H|E) = \frac{P(E|H)P(H)}{P(E)}P(H∣E)=P(E)P(E∣H)P(H)

where:

o P(H∣E)P(H|E)P(H∣E) is the probability of the hypothesis HHH given the

evidence EEE.
o P(E∣H)P(E|H)P(E∣H) is the likelihood of the evidence given the hypothesis.
o P(H)P(H)P(H) is the prior probability of the hypothesis.
o P(E)P(E)P(E) is the marginal likelihood of the evidence.

3.6.3 Conditional Probability and Independence

• Conditional Probability: The probability of an event AAA occurring given that

another event BBB has occurred is called the conditional probability and is denoted
as P(A∣B)P(A|B)P(A∣B).

P(A∣B)=P(A∩B)P(B)P(A|B) = \frac{P(A \cap B)}{P(B)}P(A∣B)=P(B)P(A∩B)

where P(A∩B)P(A \cap B)P(A∩B) is the probability of both AAA and BBB occurring.

• Independence: Two events AAA and BBB are independent if the occurrence of one
does not affect the probability of the other.

P(A∩B)=P(A)P(B)P(A \cap B) = P(A)P(B)P(A∩B)=P(A)P(B)

If the equation holds, the events are independent.

3.6.4 Probabilistic Inference and Belief Networks

• Probabilistic Inference: The process of drawing conclusions from a probabilistic

model, given known evidence. This includes reasoning about uncertain variables.
• Belief Networks (Bayesian Networks): A graphical model representing a set of
variables and their probabilistic dependencies. Each node represents a variable,
and edges represent conditional dependencies.
o Inference in Belief Networks involves computing the probabilities of certain
events (or nodes) given other events (or nodes). This can be done using
algorithms such as variable elimination or Markov Chain Monte Carlo
(MCMC) methods.

3.6.5 Applications of Probabilistic Reasoning

Probabilistic reasoning is widely used in AI, including:

1. Medical Diagnosis:
a. Used to determine the likelihood of a disease given the observed symptoms
and patient history.
2. Robot Navigation:
a. In uncertain environments, robots use probabilistic models to infer their
position and make decisions about where to move next.
3. Speech Recognition:
a. Probabilistic models help identify the most likely words or phrases given
audio input.
4. Natural Language Processing:
a. Models like Hidden Markov Models (HMM) and Bayesian Networks are
used in text understanding, part-of-speech tagging, and language modeling.
5. Machine Learning:
a. Naive Bayes classifiers and Gaussian Mixture Models (GMMs) are
examples of machine learning techniques that use probabilistic reasoning.
3.7 Utility Theory

3.7.1 Introduction to Utility Theory

Utility theory is concerned with the modeling and quantification of preferences in decision-
making. In AI, utility theory helps to make rational decisions in uncertain environments by
assigning values (utilities) to different outcomes.

• Key Concept: Utility represents the satisfaction or value derived from a particular
outcome. The goal is to maximize the utility of a decision or action.

3.7.2 Rational Decision Making and Preferences

Rational decision-making involves choosing actions that maximize the expected utility,
based on a decision-maker's preferences. Preferences refer to the way a decision-maker
ranks different outcomes or alternatives.

• Assumptions:
o Completeness: A decision-maker can rank any two outcomes.
o Transitivity: If outcome A is preferred to B and B is preferred to C, then A is
preferred to C.
o Independence: If two options are equivalent in some respects, the decision
maker will make choices based on the remaining differences.

3.7.3 Expected Utility and Decision Theory

• Expected Utility: In decision theory, the expected utility is the weighted average of
the utilities of all possible outcomes, where each outcome is weighted by its
probability.

EU(A)=∑P(Oi)⋅U(Oi)EU(A) = \sum P(O_i) \cdot U(O_i)EU(A)=∑P(Oi )⋅U(Oi )

where P(Oi)P(O_i)P(Oi ) is the probability of outcome OiO_iOi , and U(Oi)U(O_i)U(Oi ) is the

utility of that outcome.

• Decision Theory uses expected utility to help make rational choices in the face of
uncertainty, balancing the probabilities of outcomes and their respective utilities.
3.7.4 Risk and Uncertainty in Decision Making

• Risk: A situation where the probabilities of different outcomes are known, and
decision-makers can calculate the expected utility.
• Uncertainty: A situation where the probabilities of outcomes are unknown, making
it difficult to calculate expected utilities.

In AI, decision-makers often must make decisions under both risk and uncertainty, where
decision trees or Markov decision processes (MDPs) are used to model and evaluate
possible decisions.

3.7.5 Applications of Utility Theory in AI

Utility theory plays a significant role in several areas of AI, including:

1. Automated Decision Systems:

a. AI systems use utility theory to make decisions that maximize long-term
benefits, like recommending actions in a game or selecting optimal business
strategies.
2. Robot Path Planning:
a. In path planning, robots use utility functions to choose paths that balance
cost, safety, and time, maximizing their expected utility.
3. Game Theory:
a. AI systems use utility theory to make decisions in competitive environments,
such as choosing strategies in adversarial games (e.g., chess or poker).
4. Reinforcement Learning:
a. In reinforcement learning, agents learn policies that maximize the
cumulative expected utility (reward) over time.
5. Risk Management:
a. Utility theory helps in balancing risk and reward in investment strategies,
healthcare treatment choices, and other decision-making processes.

3.8 Hidden Markov Models (HMM)

3.8.1 Introduction to Hidden Markov Models

A Hidden Markov Model (HMM) is a statistical model used to represent systems that are
modeled as a sequence of hidden states, with observable outputs that are probabilistically
dependent on the state. HMMs are widely used in time-series analysis and sequential data
problems, where the goal is to infer the hidden states based on observable data.

• Hidden States: These are the unobserved states of the system that generate the
observations.
• Observations: These are the observable variables that depend on the hidden
states.

3.8.2 Components of HMM: States, Observations, Transitions

An HMM is defined by the following components:

1. States: The system has a finite number of hidden states. The states are typically
discrete and not directly observable.
2. Observations: These are the observable outputs that are generated by the hidden
states. In most applications, each state corresponds to a probability distribution
over the possible observations.
3. Transition Probabilities: These define the probability of transitioning from one
state to another. This is represented by a matrix where each element aija_{ij}aij is
the probability of transitioning from state iii to state jjj.
4. Emission Probabilities: These define the probability of observing a particular
observation from a given state. Each state has a distribution over possible
observations.
5. Initial Probabilities: The probabilities of starting in each state at the beginning of
the sequence.

3.8.3 Forward and Backward Algorithms

The Forward Algorithm and Backward Algorithm are used for calculating the likelihood of
a sequence of observations given an HMM.

• Forward Algorithm: Used to compute the probability of observing a sequence of

observations up to a certain time, given the model parameters. It involves recursive
calculation over time, integrating over all possible state sequences.
• Backward Algorithm: Used for computing the probability of the future
observations, given the model parameters, starting from the end of the sequence
and moving backwards.

Both algorithms are essential for the efficient computation of probabilities in HMMs.
3.8.4 Training HMM: The Baum-Welch Algorithm

The Baum-Welch Algorithm is a type of Expectation-Maximization (EM) algorithm used

to train HMMs. It is used to estimate the parameters of the model (transition and emission
probabilities) given a set of observation sequences.

• Expectation Step (E-step): Calculate the expected value of the hidden states given
the current model parameters.
• Maximization Step (M-step): Update the model parameters to maximize the
likelihood of the observed data.

The algorithm iteratively refines the model parameters to better fit the observed data.

3.8.5 Applications of HMM in Speech and Pattern Recognition

HMMs are widely used in applications where the goal is to model sequential or time-
dependent data, such as:

1. Speech Recognition: HMMs are used to model the sequence of phonemes (basic
sound units) in speech, and the task is to recognize words from the spoken audio.
2. Pattern Recognition: HMMs are used in various pattern recognition tasks such as
handwriting recognition, gesture recognition, and bioinformatics (e.g., gene
sequence analysis).
3. Part-of-Speech Tagging: HMMs are employed in natural language processing
tasks, where the hidden states represent parts of speech, and the observations are
the words.

3.9 Bayesian Networks

3.9.1 Introduction to Bayesian Networks

A Bayesian Network (also called a Belief Network or Bayes Net) is a graphical model that
represents probabilistic relationships among a set of variables. Each node in the network
represents a random variable, and the edges represent conditional dependencies between
these variables. Bayesian networks are particularly useful for representing complex
systems involving uncertainty and for performing probabilistic inference.
3.9.2 Structure and Representation of Bayesian Networks

• Structure: The structure of a Bayesian Network is a directed acyclic graph (DAG)

where:
o Each node represents a random variable.
o Each edge represents a probabilistic dependency between the variables.
o An edge from node AAA to node BBB means that BBB is conditionally
dependent on AAA.
• Conditional Probability Tables (CPTs): Each node in the network has an
associated CPT that quantifies the effects of the parents on the node. This table
contains the probabilities of each possible value of the node given the values of its
parents.

3.9.3 Conditional Independence and d-Separation

• Conditional Independence: Two variables XXX and YYY are conditionally

independent given a third variable ZZZ if the knowledge of ZZZ makes XXX and YYY
independent. In a Bayesian network, this is represented by the absence of a direct
path between XXX and YYY once ZZZ is known.
• d-Separation: A criterion for determining whether a set of variables is independent
of another set, given a third set. It is used to identify conditional independencies in
the network.

3.9.4 Inference in Bayesian Networks

Inference in Bayesian Networks involves computing the probability distribution of one or

more variables, given evidence about other variables. This is typically done using methods
such as:

• Variable Elimination: A method of summing out variables in a systematic way.

• Belief Propagation: A message-passing algorithm used for inference in tree-
structured Bayesian networks or loopy networks.

3.9.5 Learning Bayesian Networks

Learning a Bayesian network involves:

1. Parameter Learning: Estimating the conditional probability distributions (CPTs)

from the data.
 2. Structure Learning: Discovering the structure of the network from the data. This
can be done using algorithms such as score-based methods or constraint-based
methods.

Learning Bayesian networks from data is computationally challenging, especially for large
networks.

3.9.6 Applications of Bayesian Networks

Bayesian networks are used in a wide range of applications:

1. Medical Diagnosis: Bayesian networks are used to model the relationships

between diseases, symptoms, and test results to diagnose illnesses.
2. Expert Systems: Used in decision support systems where expert knowledge is
encoded in the network structure.
3. Risk Analysis: In fields such as finance and engineering, Bayesian networks are
used to model risk factors and predict potential outcomes.
4. Machine Learning: Bayesian networks serve as a foundation for various machine
learning algorithms, including classification and regression tasks.
5. Natural Language Processing: Used for tasks like part-of-speech tagging and
parsing where the structure of language is probabilistic.