Artificial Intelligence

CS

Kaustuv Nag

/
Fuzzy Logic in Artificial Intelligence

▶ Fuzzy Logic (FL) is a method of reasoning that resembles human reasoning.

▶ The approach of FL imitates the way of human decision-making that involves all
intermediate possibilities between digital values YES and NO.
▶ Unlike computers, human decision-making includes a range of possibilities between
YES and NO, such as
▶ CERTAINLY YES
▶ POSSIBLY YES
▶ CANNOT SAY
▶ POSSIBLY NO
▶ CERTAINLY NO

/
Use of Fuzzy Logic

▶ Fuzzy logic is useful for commercial and practical purposes.

▶ It can control machines and consumer products.
▶ It may not give accurate reasoning, but acceptable reasoning.
▶ Fuzzy logic helps to deal with the uncertainty in engineering.

/
Fuzzy Logic Systems Architecture

Rules

Crisp Crisp
Fuzzifier Defuzzifier
Input Output

Intelligence
Fuzzy Input Set Fuzzy Output Set

/
Classical Set (Crisp Set)
▶ A classical set is a collection of distinct objects. For example, a set of students, a set
of even numbers etc.
▶ Each individual entity in a set is called a member or an element of the set.
▶ The classical set is defined such that the universe of discourse is split into two
groups: members and non-members. Hence, no partial membership exists.
▶ For example:
▶ A = {a1 , a2 , a3 , a4 }
▶ A = {2, 4, 6, 8}
Expressing the second example using a formula:
A = {x | x is an even natural number}
▶ Using a membership function µA (x), the membership value of an element x in a
crisp set A is: (
1 if x ∈ A
µA (x) =
0 if x ∈
/A
/
Fuzzy Sets
▶ Fuzzy set theory introduces vagueness (ambiguity) using the concept of a
membership function.
▶ Membership value µÃ (y) permits any value between [0, 1].
▶ Formal representation of a fuzzy set Ã:
à = {(y, µÃ (y)) | y ∈ U }.

▶ U is the universe of discourse.

▶ µÃ (y) is the membership value of element y in the fuzzy set Ã.
▶ Common shapes for membership functions: µ
▶ Triangular (most common).
▶ Trapezoidal. Membership function of Ã
▶ Singleton.
▶ Gaussian. 1

0 y /
Example for Fuzzy Set

▶ Let the universe of discourse be U = {a1 , a2 , a3 , a4 , a5 }.

▶ Examples of fuzzy sets over U :
. A = {(a1 , 0.4), (a2 , 0.7), (a3 , 0), (a4 , 0.1), (a5 , 1)}
. B = {(a1 , 0.2), (a2 , 0.6), (a3 , 0.2), (a4 , 0.1), (a5 , 0.8)}
. C = {(a1 , 0), (a2 , 1), (a3 , 0), (a4 , 1), (a5 , 0)}
▶ What is the corresponding crisp set representation for fuzzy set C?
▶ A crisp set contains only elements with full membership (value = 1).
▶ For fuzzy set C, the elements a2 and a4 have membership value 1.
Ccrisp = {a2 , a4 }

/
An Example for Fuzzy Set: Age
Young Middle-Aged Old

Membership, µ
.

“Age”

▶ Human age groups illustrated as fuzzy sets (“Young,” “Middle-Aged,” “Old”) with
corresponding membership functions (µ).
▶ Example: The set “Young” might include ages 0 to 37. Membership µ could be 1 up
to age 23, then decrease linearly to 0 by age 37. An age of 30 might have µ = 0.5 in
“Young”.
▶ Solid lines: Fuzzy membership functions. Dashed lines: Example of a crisp
(non-fuzzy) definition for “Young” (e.g., ages 0 to 30).
▶ Any specific age may belong to multiple sets (e.g., “Young” and “Middle-Aged”) but
with different membership values.
/
Operations on Fuzzy Sets
▶ Let A and B be fuzzy sets in universe U (A, B ⊆ U ), and u ∈ U .
▶ Standard complement (¬A or Ā):

µ¬A (u) = 1 − µA (u).

▶ Standard intersection (A ∩ B):

µA∩B (u) = min{µA (u), µB (u)} (Fuzzy AND).

▶ Standard union (A ∪ B):

µA∪B (u) = max{µA (u), µB (u)} (Fuzzy OR).

▶ Example: Let A = {(x1 , 0.5), (x2 , 0.7)} and B = {(x1 , 0.8), (x2 , 0.2)}.
▶ Complement (¬A): ¬A = {(x1 , 1 − 0.5), (x2 , 1 − 0.7)} = {(x1 , 0.5), (x2 , 0.3)}.
▶ Intersection:
A ∩ B = {(x1 , min{0.5, 0.8}), (x2 , min{0.7, 0.2})} = {(x1 , 0.5), (x2 , 0.2)}.
▶ Union: A ∪ B = {(x1 , max{0.5, 0.8}), (x2 , max{0.7, 0.2})} = {(x1 , 0.8), (x2 , 0.7)}.
/
Triangular Fuzzy Membership Function
▶ Defined by three parameters: a, b, c.


 0 if x ≤ a



 x−a
b−a if a < x ≤ b
µtri (x; a, b, c) =

 c−x
if b < x < c

 c−b

0 if x ≥ c

▶ max(0, min( x−a c−x

b−a , c−b )) is another form.

1.0

µA (x)

a b c

/
Gaussian Fuzzy Membership Function
▶ Defined by center c, width s, and sometimes fuzzification factor m.
▶ A common form (with m = 2 implicitly):
 2 !
1 x−c
µgauss (x; c, s) = exp −
2 s
▶ Generalized form from slide (with factor m):
 m
1 x−c
µ(x; c, s, m) = exp −
2 s
1
0.8
0.6
0.4
0.2
0
0 1 2 3 4 5 6 7 8 9 10 /
Introducing Linguistic Variables

▶ We often want to use natural language terms like ‘young’, ‘not young’, ‘middle age’,
‘possibly yes’, ‘possibly no’ etc. in our knowledge-base or rule-base.
▶ Linguistic variables allow us to represent and use these terms, incorporating
vagueness/ambiguity into rules.

/
Linguistic Variable: Definition and Examples

▶ A linguistic variable is a variable whose values are words or sentences in a natural

language (called linguistic terms).
▶ Examples:
. Height can be a linguistic variable.
▶ Linguistic values (terms): tall, very tall, medium, short, etc.
▶ ‘Tall’ itself is a linguistic value or primary term. Modifiers like ‘very’ can be applied.
. Age can be a linguistic variable.
▶ The set of terms T(age) could be: {young, not young, middle aged, old, very old, etc.}
▶ Each term (e.g., ‘young’) corresponds to a fuzzy set.
▶ Every member of the term set is a linguistic term, covering some portion of the
overall range of values (e.g., the range of possible ages).

/
Fuzzy Rules

▶ Fuzzy rules help model human-like reasoning and perception.

▶ The common fuzzy if-then rule structure is:
“IF x is A THEN y is B.”
▶ x, y are variables (often linguistic variables).
▶ A, B are linguistic values (fuzzy sets) defined over the universes of discourse for x and
y.
▶ “x is A” is the antecedent (premise or condition).
▶ “y is B” is the consequent (conclusion).

/
Examples: Fuzzy Rules

▶ Consider the rule: “IF pressure is high THEN volume is small”

▶ Here, ‘pressure’ (x) and ‘volume’ (y) are variables.
▶ ‘high’ (A) and ‘small’ (B) are linguistic values represented by fuzzy sets.
▶ The rule can be written as: IF x is A THEN y is B.
▶ If we have a specific input value for pressure (x), we can calculate its membership value
µA (x) in the fuzzy set ‘high’ using its pre-defined membership function. This value
indicates the degree to which the antecedent is true.

/
Fuzzy Inference System (FIS)
▶ Fuzzy Inference Systems (FIS) take inputs, process them using fuzzy reasoning based
on fuzzy rules, and produce outputs.
▶ Inputs and outputs are typically crisp (real-valued) numbers, while the internal
processing uses fuzzy sets, fuzzy logic, and fuzzy rules.
▶ Basic structure includes three main components:
. Rule Base: Contains the collection of fuzzy IF-THEN rules.
. Database (Dictionary): Defines the membership functions (MFs) for the fuzzy sets used
in the rules.
. Reasoning Mechanism (Inference Engine): Performs the inference procedure based on
the rules and input facts to derive a conclusion or output.

Rule Base Database

(Fuzzy Rules) (Membership Functions)

Input Output

Fuzzy Reasoning

/
Structure of Fuzzy Inference Systems (FISs)
Provided by Experts
or Extracted from
Numerical Data

Activates the Provides a Crisp

Rules
Linguistic Rules Output Value

Crisp Crisp
Fuzzification Defuzzification
Input Output

Inference
Fuzzy Set of Input Variables Fuzzy Output Set

. Maps fuzzysets into fuzzysets

. Determines how the rules are activated and combined
/
Structure of Fuzzy Inference Systems (FISs)

▶ Fuzzification: Maps crisp inputs to fuzzy sets (degrees of membership).

▶ Inference: Applies fuzzy rules based on fuzzified inputs. Determines how rules are
activated and combines their results.
▶ Defuzzification: Maps the resulting output fuzzy set back to a crisp output value.
▶ Rules can be provided by experts or extracted from data.

/
Popular Models of Fuzzy Inference Systems (FISs)

▶ Popular Fuzzy Inference System models include:

▶ Mamdani Fuzzy Models (and Assilian).
▶ Sugeno Fuzzy Models (or Takagi-Sugeno-Kang, TSK).
▶ Tsukamoto Fuzzy Models.
▶ The primary differences between these models lie in:
▶ The format of the consequents (THEN part) of their fuzzy rules.
▶ Consequently, their methods for rule aggregation and defuzzification differ.

/
Mamdani Inference System

▶ The Mamdani model is the most commonly used fuzzy inference technique.
▶ It typically performs inference in four main steps:
. Fuzzification of input variables.
. Rule evaluation (applying fuzzy operators for antecedents and implication).
. Aggregation of rule outputs (combining results from all relevant rules).
. Defuzzification (converting the aggregated fuzzy output to a crisp number).

Example Application
Design a Mamdani FIS to predict risk in a project funding system.
▶ Inputs: Project Funding (pf ), Project Staffing (ps ) (crisp values).
▶ Output: Risk Prediction (e.g., linguistic values like low/normal/high, or a crisp value
after defuzzification).

/
Knowledge Base for the FIS: Example
▶ Crisp Inputs: x (for pf ), y (for ps ). Crisp Output: z (for risk).
▶ Linguistic Variable: Project Funding (pf )
▶ Terms/Fuzzy Sets: Inadequate (A1 ), Marginal (A2 ), Adequate (A3 ).
▶ Linguistic Variable: Project Staffing (ps )
▶ Terms/Fuzzy Sets: Small (B1 ), Large (B2 ).
▶ Linguistic Variable: Risk (z)
▶ Terms/Fuzzy Sets: Low (C1 ), Normal (C2 ), High (C3 ).
▶ Pre-specified membership functions exist for each fuzzy set
(A1 , A2 , A3 , B1 , B2 , C1 , C2 , C3 ).
Rule Base Example
. IF pf is Adequate (A3 ) OR ps is Small (B1 ) THEN Risk is Low (C1 ).
(IF x is A3 OR y is B1 THEN Z is C1 )
. IF pf is Marginal (A2 ) AND ps is Large (B2 ) THEN Risk is Normal (C2 ).
(IF x is A2 AND y is B2 THEN Z is C2 )
. IF pf is Inadequate (A1 ) THEN Risk is High (C3 ).
(IF x is A1 THEN Z is C3 )
/
Fuzzy Sets Defined Over Input/Output Variables

µA1 (x) µA2 (x) µA3 (x) µB1 (y) µB2 (y) µC1 (z) µC2 (z) µC3 (z)

x y z
Figure: Membership functions for Input x (Funding: A1 , A2 , A3 ), Input y (Staffing: B1 , B2 ), and
Output z (Risk: C1 , C2 , C3 ).

/
Designing FIS - Step : Fuzzification
▶ Convert crisp input values into degrees of membership in the relevant fuzzy sets.
▶ Example: Take crisp inputs for Project Funding (x1 ) and Project Staff (y1 ).
▶ Use the defined membership functions to find the membership values.
Example results from the graph:
▶ For input x1 : µA1 (x1 ) = 0.5, µA2 (x1 ) = 0.2, µA3 (x1 ) = 0.0.
▶ For input y1 : µB1 (y1 ) = 0.1, µB2 (y1 ) = 0.7.

µA1 (x) µA2 (x) µA3 (x) µB1 (y) µB2 (y)

.
.

.
.
x1 x y1 y

Figure: Example Fuzzification Step: Finding membership values for crisp inputs x1 and y1 . /
Step : Rule evaluation - Antecedent
▶ Determine the degree to which the antecedent (IF part) of each rule is satisfied by
the current inputs.
▶ If a rule’s antecedent has multiple parts connected by fuzzy operators (AND, OR):
▶ Apply the corresponding fuzzy operation to the membership values obtained during
fuzzification.
▶ The result is a single number representing the rule’s antecedent truth value (firing
strength).
▶ Evaluating operators (using standard fuzzy logic):
▶ OR (Disjunction): Typically uses the maximum operator. For antecedent “X is A OR Y
is B”:
Truth Value = max[µA (X), µB (Y )]
▶ AND (Conjunction): Typically uses the minimum operator. For antecedent “X is A AND
Y is B”:
Truth Value = min[µA (X), µB (Y )]

/
Mamdani Style of Rule Evaluation for Project System I

Figure: Calculating the firing strength for each rule using the fuzzified inputs and fuzzy operators
(min for AND, max for OR).
/
Mamdani Style of Rule Evaluation for Project System II

Rule Firing Strengths: µA1 = 0.5, µA2 = 0.2, µA3 = 0.0, µB1 = 0.1, µB2 = 0.7
▶ Rule (A OR B ): max(0.0, 0.1) = 0.1.
▶ Rule (A AND B ): min(0.2, 0.7) = 0.2.
▶ Rule (A ): 0.5 (single antecedent)

/
Step : Rule evaluation - Fuzzy Implication
▶ Apply the calculated firing strength of each rule to its consequent (THEN part). This
reshapes the consequent’s fuzzy set.
▶ The result of this step is one output fuzzy set for each rule, shaped by its firing
strength.
▶ Mamdani Implication Methods:
Clipping (min): The consequent Scaling (prod): The consequent
membership function is truncated (cut off) membership function is scaled (multiplied)
at the level of the firing strength. (Most by the firing strength.
common). 1
1 µC2 (z)
µC2 (z)

0.2 µR2 (z)

0.2 µR2 (z)
0 z
z Scaling
0
Clipping
/
 µC1 (z)
Step : Aggregation of Rules 1

▶ Unify the output fuzzy sets from all contributing µR1 (z)

rules into a single aggregate fuzzy set 0 z

representing the overall conclusion. µC2 (z)

1
▶ Input: The clipped or scaled consequent
µR2 (z)
membership functions from each rule.
▶ Output: One combined fuzzy set for the output 0 z

variable. µC2 (z)

1
▶ Common Method (max): The membership value µR3 (z)

of the aggregate set at any point z is the

maximum of the membership values of the 0 z

individual rule output sets at that point z. µC1 (z) µC2 (z) µC2 (z)
1
µAGG (z)
µAGG (z) = max[µR1 (z), µR2 (z), µR3 (z), ...]

0 z

/
Step : Defuzzification
▶ Convert the final aggregate fuzzy set (which represents the fuzzy conclusion) into a
single crisp output number.
▶ Input: The aggregate output fuzzy set. Output: A single crisp value.
▶ Methods: Many exist (Mean of Maximum, Bisector, etc.).
▶ Centroid (Center of Gravity - COG): Most popular method. Finds the point on the
x-axis (z) that represents the “center of mass” or “balance point” of the aggregate
fuzzy set.
R
▶ For a continuous set, COG = Rz·µAGG (z)dz .
µAGG (z)dz
▶ For discrete points or sampled representation:
P
zi · µAGG (zi )
COG = P
µAGG (zi )