Multi-Objective Optimization: Introduction

Debasis Samanta

Indian Institute of Technology Kharagpur

[email protected]

01.04.2024

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 1 / 56

 Solving Multiobjective Optimization Problems

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 2 / 56

Multiobjective optimization problem: MOOP
There are three components in any optimization problem:
F: Objectives
minimize (maximize) fi (x1 , x2 , · · · , xn ), i = 1, 2, · · · , m

S: Constraints
Subject to
gj (x1 , x2 , · · · , xn ), ROPj Cj , j = 1, 2, · · · , l

V: Design variables

xk ROPk dk , k = 1, 2, · · · , n
Note :
1 For a multi-objective optimization problem (MOOP), m ≥ 2
2 Objective functions can be either minimization, maximization or
both.
Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 3 / 56
A formal specification of MOOP

Let us consider, without loss of generality, a multi-objective

optimization problem with n decision variables and m objective
functions

Minimize y = f (x) = [y1 ∈ f1 (x), y2 ∈ f2 (x), · · · , yk ∈ fm (x)]

where

x = [x1 , x2 , · · · , xn ] ∈ X
y = [y1 , y2 , · · · , yn ] ∈ Y

Here :
x is called decision vector
y is called an objective vector
X is called a decision space
Y is called an objective space

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 4 / 56

Illustration: Decision space and objective space

3 2

1
1

Thus, solving a MOOP implies to search for x in the decision space

(X ) for an optimum vector (y) in the objective space (Y ).

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 5 / 56

A formal specification of MOOP (contd...)

In other words,
1 We wish to determine X̄ ∈ X (called feasible region in X ) and any
point x̄ ∈ X̄ (which satisfy all the constraints in MOOP) is called
feasible solution.
2 Also, we wish to determine from among the set X̄ , a particular
solution x̄ ∗ that yield the optimum values of the objective functions.
Mathematically,

∀x̄ ∈ X̄ and ∃x̄ ∗ ∈ X̄ | fi (x̄ ∗ ) ≤ fi (x̄),

where ∀i ∈ [1, 2, · · · , m]
3 If this is the case, then we say that x¯∗ is a desirable solution.

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 6 / 56

Why solving a MOOP is an issue?

In a single-objective optimization problem, task is to find typically

one solution which optimizes the sole objective function

In contrast to single–objective optimization problem, in MOOP:

Cardinality of the optimal set is more than one, that is, there are
m ≥ 2 goals of optimization instead of one
There are m ≥ 2 different search points (possibly in different
decision spaces) corresponding to m objectives

Optimizing each objective individually not necessarily gives the

optimum solution.
Possible, only if objective functions are independent to their solution
spaces.

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 7 / 56

Illustration: Single vs. multiple objectives

f1 f2 f3 f4
Objectives

Objectives
Search space Search space

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 8 / 56

Why solving an MOOP is an issue?

In fact, majority of the real-world MOOPs are with a set of trade-off

optimal solutions. A set of trade-off optimal solutions is also
popularly termed as Pareto optimal solutions
In a particular search point, one may be the best whereas other
may be the worst

Also, sometime MOOPs are with conflicting objectives

Thus, optimizing an objective means compromising other(s) and
vice-versa.

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 9 / 56

MOOP: Trade-off and conflicts in solutions

5 f2
4
i ze
Maximize f2

3 im

Objectives
x
Ma
2

1
Minim
ize f
1

Minimize f1 Search space

Trade-off optimal solution Conflicting objectives

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 10 / 56

Illustration: ideal solution vs. real solution
It is observed that in many real-life problems, we hardly have a
situation in which all the fi (x̄) have a minimum in X̄ at a common point
x̄ ∗ .
This is particularly true when objective functions are conflicting in their
interests.

F2 (minimize)
F2 (minimize)

F1 (minimize) F1 (minimize)

Ideal Situation Real Situation

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 11 / 56

 Multiobjective Evolutionary Algorithms

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 12 / 56

GA-based approach to solve MOOPs
MOEA : Multi Objective Evolutionary Algorithm

Solution
Found ?

MOEA follows the same reproduction operation as in GA but follow

different selection procedure and fitness assignment strategies.
There are also a number of stochastic approaches such as
Simulated Annealing (SA), Ant Colony Optimization (ACO),
Particle Swam Optimization (PSO), Tabu Search (TS), etc. could
be used to solve MOOPs.
Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 13 / 56
MOEA: GA-based approach to solve MOOP

There are two board approaches to solve MOOPs with MOEA

A priori approach (also called preference-based approach)
A posteriori approach (does not require any prior knowledge)

Two major problems must be addressed when a GA is applied to

multi-objective optimization problems.
1 How to accomplish fitness assignment and selection in order to
guide the search toward the optimal solution set?
2 How to maintain a diverse population in order to prevent
premature convergence and achieve a well distributed trade-off
front?

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 14 / 56

Schematic of a priori MOEA approach

A MOOP problem Estimate a Single-objective

Minimize f1 relative optimization problem
Higher level m
Minimize f2
information
importance F   wi  fi
......... vector W = i 1
Minimize fm {w1,w2 . . .wm} Subject to S and V

Subject to
constraint S
with design
variables V
Single Optimal GA to solve
Solution the problem

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 15 / 56

Schematic of a posteriori MOEA approach

A MOOP problem
Minimize f1 Ideal Multiple Pareto-
Minimize f2 Multiobjective optimal
......... Optimizer solutions
Minimize fm

Subject to
constraint S
with design
Choose one Higher level
variables V
solution Informaiton

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 16 / 56

IDEAL multi-objective optimization

Here, effort have been made in finding the set of trade-off solutions by
considering all objectives to be important.
Steps
1 Find multiple trade-off optimal solutions with a wide range of
values for objectives. (Note: here, we do not use any relative
preference vector information). The task here is to find as many
different trade-off solutions as possible.
2 Choose one of the obtained solutions using higher level
information (i.e. evaluate and compare the obtained trade-off
solutions)

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 17 / 56

Illustration: Higher level information

Consider the decision making involved in buying an automobile car.

Consider two objectives.

minimize Cost
maximize Comfort

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 18 / 56

Illustration: Higher level information

Here, solution 1 and 2 are two extreme cases.

Between these two extreme solutions, there exist many other

solutions, where trade-off between cost and comfort exist.

In this case, all such trade-off solutions are optimal solutions to a

multi-objective optimization problem.

Often, such trade-off solution provides a clear front on an

objective space plotted with the objective values.

This front is called Pareto-optimal front and all such trade-off

solutions are called Pareto-optimal solutions (after the name of
Vilfredo Pareto, 1986)

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 19 / 56

Choosing a solution with higher level information

Knowing the number of solutions that exist in the market with

different trade-offs between cost and comfort, which car does one
buy?

It involves many other considerations

total finance available to buy the car
fuel consumption
depreciation value
road condition
physical health of the passengers
social status
After sales service, vendor’s reputation, manufacturer’s past history
etc.

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 20 / 56

 Preliminaries of MOEA

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 21 / 56

Formal specification of MOEA approach

In the next few slides, we shall discuss the above idea of solving
MOOPs more precisely. Before that, let us familiar to few more basic
definitions and terminologies.
1 Concept of domination
2 Properties of dominance relation
3 Pareto-optimization
4 Solutions with multiple-objectives

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 22 / 56

Solution with multiple objectives : Ideal objective
vector
For each of the M-th conflicting objectives, there exist one different
optimal solution. An objective vector constructed with these individual
optimal objective values constitute the ideal objective vector.

Definition 1: Ideal objective vector

Without any loss of generality, suppose the MOOP is defined as
Minimize fm (x), m = 1, 2, · · · , M
Subject to X ∈ S, where S denotes the search space.
and
fm∗ denotes the minimum solution for the m-th objective functions, then
the ideal objective vector can be defined as

Z ∗ = f ∗ = [f1∗ , f2∗ , · · · , fM∗ ]

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 23 / 56

Ideal objective vector : Physical interpretation

Z*1

f2
f2
Z*

Z* Z*2
Z*

f1 f1
(B) A good solution vector should
(A) Ideal objective vector be as close to ideal solution vector

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 24 / 56

Ideal objective vector : Physical interpretation
Let us consider a MOOP with two objective functions f1 and f2 where
both are to be minimized.
If z ∗ = f ∗ = [f1∗ , f2∗ ] then both f1 and f2 are minimum at x ∗ ∈ S.
(That is, there is a feasible solution when the minimum solutions to
both the objective functions are identical).
In general, the ideal objective vector z ∗ corresponds to a
non-existent solution (this is because the minimum solution for
each objective function need not be the same solution).
If there exist an ideal objective vector, then the objectives are
non-conflicting with each other and the minimum solution to any
objective function would be the only optimal solution to the MOOP.
Although, an ideal objective vector is usually non-existing, it is
useful in the sense that any solution closer to the ideal objective
vector are better. (In other words, it provides a knowledge on the
lower bound on each objective function to normalize objective
values within a common range).
Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 25 / 56
Solution with multiple objectives : Utopian
objective vector
Utopian objective vector corresponding to a solution which has an
objective value strictly better than (and not equal to) that of any
solution in search space.

f2

Z*
Utopian objective vector

Z** Ideal objective vector

f1

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 26 / 56

Solution with multiple objectives : Utopian
objective vector

The Utopian objective vector can be formally defined as follows.

Definition 2 : Utopian objective vector

A Utopian objective vector z ∗∗ has each of its component marginally
smaller than that of the ideal objective vector, that is
zi∗∗ = zi∗ − ∈i with ∈i > 0 for all i = 1, 2, · · · , M

Note :
Like the ideal objective vector, the Utopian objective vector also
represents a non-existent solution.

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 27 / 56

Solution with multiple objectives : Nadir objective
vector
The ideal objective vector represents the lower bound of each
objective in the entire feasible search space. In contrast to this, the
Nadir objective vector, denoted as z nadir , represents the upper bound
of each objective in the entire Pareto-optimal set (note: not in the
entire search space).

(f1max,f2max)
*
Z2
2 Znadir

Z* Z1*

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 28 / 56

Solution with multiple objectives : Nadir objective
vector

Note :
z nadir is the upper bound with respect to Pareto optimal set. Whereas,
a vector of objective W found by using the worst feasible function
values fimax in the entire search space.

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 29 / 56

Usefulness of Nadir objective vector

In order to normalize each objective in the entire range of

Pareto-optimal region, the knowledge of Nadir and ideal objective
vectors can be used as follows.
fi −zi∗
f¯i = nadir
zi −zi∗

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 30 / 56

Concept of domination

Notation

Suppose, f1 , f2 , · · · , fM are the objective functions

xi and xj are any two solutions
The operator
between two solutions xi and xj as xi
xj to
denote that solution xi is better than the solution xj on a particular
objective.
Alternatively, xi  xj for a particular objective implies that solution
xi is worst than the solution xj on this objective.
Note :
If an objective function is to be minimized, the operator
would mean
the ”<” (less than operator), whereas if the objective function is to be
maximized, the operator
would mean the ”>” (greater than operator).

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 31 / 56

Concept of domination

Definition 3 : Domination
A solution xi is said to dominate the other solution xj if both condition I
and II are true.
Condition : I
The solution xi is no worse than xj in all objectives. That is
fk (xi ) ⋫ fk (xj ) for all k = 1, 2, · · · , M

Condition : II
The solution xi is strictly better than xj in at least one objective. That is
fk̄ (xi )
fk̄ (xj ) for at least one k̄ ∈ {1, 2, · · · , M}

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 32 / 56

Illustration 1

Consider that f1 and f2 are two objectives to be minimized.

x3
Minimize f1
f2 Minimize f2
x1 ≤ x2
x2
x1
x1 ≤ x3 but x3 ≤ x1
x2 ≤ x3 as well as x3 ≤ x2

f1

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 33 / 56

Illustration 2

x3 Minimize f1
f2 Maximize f2

x2 x1 ≤ x2 or x2 ≤ x1 ?
x1
x1 ≤ x3 or x3 ≤ x1 ?
x2 ≤ x3 or x3 ≤ x2 ?

f1

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 34 / 56

Points to be noted

Note :

If either of the condition I and II is violated then the solution xi

does not dominate the solution xj .

If xi dominates the solution xj (it is also mathematically denoted as

xi ≤ xj .
The domination also alternatively can be stated in any of the
following ways.
xj is dominated by xi
xi is non-dominated by xj
xi is non-inferior to xj

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 35 / 56

Illustration 3

2
5
4
4
f2 minimize

3
1 5
2
3
1

2 6 10 14 18
f1 maximize

Here, 1 dominates 2, 5 dominates 1 etc.

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 36 / 56

Properties of dominance relation

Definition 3 defines the dominance relation between any two

solutions.

This dominance relation satisfies four binary relation properties.

Reflexive :
The dominance relation is NOT reflexive.
Any solution x does not dominate itself.

Condition II of definition 3 does not allow the reflexive property to

be satisfied.
Symmetric :
The dominance relation also NOT symmetric
x ⪯ y does not imply y ⪯ x.

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 37 / 56

Properties of dominance relation

Antisymmetric :

Dominance relation can not be antisymmetric

Transitive :
The dominance relation is TRANSITIVE
If x ⪯ y and y ⪯ z, then x ⪯ z.

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 38 / 56

Properties of dominance relation

Note :
1 An interesting property that dominance relation possesses is : If
solution x does not dominate solution y, this does not imply that y
dominates x.
2 In order for a binary relation to qualify as an ordering relation, it
must be at least transitive. Hence, dominance relation qualifies as
an ordering relation.
3 A relation is called partially ordered set, if it is reflexive,
antisymmetric and transitive. Since dominance relation is NOT
REFLEXIVE, NOT ANTISYMMETRIC, it is NOT a PARTIALLY
ORDER RELATION
4 Since, the dominance relation is not reflexive, it is a STRICT
PARTIAL ORDER.

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 39 / 56

Pareto optimality

2
5
4
f2 minimize 4
3
1 5
2
3
1

2 6 10 14 18

f1 maximize

Non-dominated front

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 40 / 56

Pareto optimality

Consider solution 3 and 5.

Solution 5 is better than solution 3 with respect to f1 while 5 is

worse than 3 with respect to f2 .

Thus, condition I (of Definition 3) is not satisfied for both of these

solutions.

Hence, we can not conclude that 5 dominates 3 nor 3 dominated

5.

In other words, we can not say that two solutions 3 and 5 are
better.

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 41 / 56

Non-dominated set

2
5
4
4
f2 minimize

3
1 5
2
3
1

2 6 10 14 18

f1 maximize

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 42 / 56

Non-dominated set

From the figure it is evident that

There are a set of solutions namely 1, 2, 3, 4 and 5.

1 dominates 2; 5 dominates 1 etc.

Neither 3 dominates 5 nor 5 dominates 3

We say that solution 3 and 5 are non-dominated with respect to
each other.

Similarly, we say that solution 1 and 4 are non-dominated.

In this example, there is not a single solution, which dominates all

other solution

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 43 / 56

Non-dominated set: A counter example

2
5
4
4
f2 minimize

3
1 5
2
3
1

2 6 10 14 18 22

f1 maximize

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 44 / 56

Non-dominated set

Definition 4 : Non-dominated set

Among a set of solutions P, the non-dominated set of solutions P ′ are
those which are not dominated by any member of the set P.

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 45 / 56

Non-dominated set

2
5
4
4
f2 maximize

3 Here P = {1,2,3,4,5}
1 5
2 Non-dominated set
3 P’ = {3, 5}
1

2 6 10 14 18

f1 maximize

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 46 / 56

How to find a non-dominated set ?

For a given finite set of solutions, we can perform all pair-wise

comparisons.
Find which solution dominates
Find which solutions are non-dominated with respect to each other.

Property of solutions in non-dominated set

∃xi , xj ∈ P ′ such that xi ⪯̸ xj and xj ⪯̸ xi
A set of solution where any two of which do not dominate each
other if
/ P ′ then xi ⪯̸ xj where xj ∈ P ′ for any solution outside
∃xi ∈ P and xi ∈
of the non-dominated set, we can always find a solution in this set
which will domnaite each other.

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 47 / 56

Some important observations

The above definition does not applicable to ideal situation.

2
5
4
4
f2 miniimize

f2 (minimize)
3
1 5
2
3
1

2 6 10 14 18 22
f1 (minimize)
f1 maximize

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 48 / 56

Some important observations

The non-dominated set concept is applicable when there is a trade-off

in solutions.

2
5
4
4
f2 minimize

3
1 5
2

F2 (minimize)
3 f2 (maximize)
1

2 6 10 14 18

f1 maximize

f1 (maximize) F1 (minimize)

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 49 / 56

Pareto optimal set

Definition 5: Pareto optimal set

When the set P is the entire search space, that is P = S, the resulting
non-dominated set P ′ is called the Pareto-optimal set.

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 50 / 56

Examples: Pareto optimal sets
Following figures shows the Pareto ooptimal set for a set of feasible
solutions over an entire search space under four different situations
with two ojective functions f1 and f2 .

F2 (maximize)
F2 (miniimize)

F1 (miniimize) F1 (miniimize)

F2 (maximize)
F2 (miniimize)

F1 (maximize) F1 (maximize)

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 51 / 56

Pareto optimal fronts

In visual representation, all Pareto optimal solutions lie on a front

called Pareto optimal front, or simply, Pareto front.

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 52 / 56

Examples

f2 (min)

f2 (min)
f2 max)

f1 (min) f1 (min) f1 (min)

F2 max)

f2 (min)
F2 (min)

f1 (min) f1 (max) f1 (min)

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 53 / 56

Examples

f2 (min)

f2 (min)
f2 (min)

f1 (min) f1 (max) f1 (min)

f2 (max)
F2 (max)

F2 max)

f1 (min) f1 (max) f1 (max)

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 54 / 56

Few good articles to read.
1 ”An Updated Survey of GA Based Multi-objective Optimization
Techniques” by Carles A Coello Coello, ACM Computing Surveys,
No.2,Vol. 32, June 2000.

2 ”Comparison of Multi-objective Evolutionary Algorithm : Empirical

Result” by E. Zitzler, K.Deb, Lother Thiele, IEEE Transaction of
Evolutionary Computation, No.2, Vol.8, Year 2000.

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 55 / 56

 Any questions??

Debasis Samanta (IIT Kharagpur) Soft Computing Applications 01.04.2024 56 / 56