OR607

Multi-objective Decision Making

An Introduction
Mohamed Haridy
Nov. 2024
 References and Credit
• Multiple Criteria Decision Making - Mila Zeleny_ 1982
• Kaisa-Miettinen-Nonlinear-Multiobjective-Optimization
• Multiple Criteria Optimization Theory, Computation, and Application - Ralph E. Steuer
• Multiobjective Decision Making Theory and Methodology - Series Volo. 8 - Vira Chankong,
Yacov Y. Haimes, 1983
• Real-world problems normally have more than one objective function, can be
conflicting (competing) with each other
• Examples (MCDM _ Zeleny M.,1982):
• Manufacturing:
• Steady state manufacturing with consistent and reliable demand is
concerned about objectives of low cost and rapid delivery.
• Market-Creating Firm concern is reliable delivery and flexibility.
• Automotive Design: to maximize fuel efficiency and acceleration
performance, with minimizing cost
• Publishing:
• Book sponsor’s editor is concerned about minimizing the number of
unsellable books, so his objective is to print less books
• Publisher objective is to print more books to increase the long-term
expected return
 MOOP
• There are two approaches* to address the solution of MOOP:
• The classical methods of solving such problems were primarily
focused on scalarizing multiple objectives into a single objective
• The evolutionary methods which are to solve a multi-objective
optimization problem as it is.

*Mila Zeleny,1982. “ Multiple Criteria Decision Making”, McGraw Hill,

Notations
Miettinen, K., On the Methodology of Multiobjective Optimization with Applications, Report 60 (doctoral thesis),
University of Jyväskylä, Department of Mathematics, Jyväskylä, 1994. Published in parallel in JYX (open access).
Definitions
• Rn: Decision Space which defines all possible values of decision variables
• S: Feasible Space defines only the feasible values of the decision variables
• F: Objective Space

S
F
Important Definition

gives the
direction and
the rate of
fastest
increase.
For R2: A set S is convex if the line segment between
any two points in S lies in S :
i.e. ∀x1, x2 ∈ S, ∀ t ∈ [0, 1]
tx1 + (1 − t)x2 ∈ S
Pareto Optimal Solutions
Example
• Cantilever Beam Design Example:
• D. V.’s are d mm and l mm
• Beam to carry a load P
• 2 Objectives:
• minimize beam weight and
• Minimize end deflection
•These are contradicting , since minimizing end deflection requires more weight
•2 Constraints
•Developed max. stress ơmax ≤ Sy (where Sy is allowable strength)
•End deflection ơmax
 Solution
• Weight objective: Min.
• End def. objective : Min.
• Subject to:
• ơmax ≤ Sy
•

• P = 1 KN
• E = 207 GB
• Sy = 20 MPa
• D. V. Bounds:
• 10 mm ≤ d ≥ 50 mm
• 100 mm ≤ l ≥ 1000 mm
 Solution d l Weight Deflection
A 18.94 200 0.44 2.04
B 21.24 200 0.5 1.1
C 34.19 200 1.43 0.19
D 50 200 3.06 0.04
E 33.02 362.49 2.42 1.31

1000 mm (5,0)

l (End Deflection, D)
F Z

P. O. F
20 mm (0,0) (0,14)
100 mm
10 mm d 50 mm (Weight, W)
 P.O.Ss
• A is better if we consider weight objective
• D is better if we consider end deflection objective
• P.O solutions are non-dominated solutions such as A&D
• C is better than E in both objectives, so E is not P.O.S W

E
D A

B C D
Pareto-Optimal Pooling
The goal of multi-objective optimization is to obtain the solution space, so
that the derivation of as many Pareto-optimal solutions as possible is
desirable, provided that solutions are not too close to each other ( as
diverted as possible). This gives rise to the Pareto-optimal pooling technique
shown in Fig. 2 where evaluation of the population of search points is conducted
independent of the derivation of Pareto-optimal solutions. The initial Pareto-
optimal solutions are those that are ranked No. 1 in the initial population. When
a new population is created, the population is compared to the current Pareto-
optimal set, and those that are ranked No. 1 become the new Pareto-optimal
set. In the single-objective optimization, the best search point is merely stored
over iterations.
Concept of Dominance
• In single objective problem, the superiority of a solution over other
solutions is easily determined by comparing the values of their
objective functions
• In MOOP, the goodness of a solution is determined by its dominance
over other solutions
X dominates Y , if:
Solution X is not worse than Y WRT all objectives
Solution X is strictly better than Y WRT at least one objective

Then it is said that ,

X dominates Y , or
Y is dominated by X
X is not dominated by Y
 Mathematically :

X1 X2 Reads X2 is better than x1 on a particular objective

X1 X2 Reads X 1 is worse than x2 on a particular objective

X 1 dominates X2 if fj(X1) fj(X2) ꓯ j = 1……M

fj(X1) fj(X2) for at least objective j = 1……M

Represented as : X2 ≻ X1
2≻1
1≻5
3 & 5 do not dominate each other
Procedure:
• Set i= 1, P.O.S = P = Φ (P is Pareto O. Set)
• For j P , set j ≠ i
• if j dominates I , increment I ( I = i+1)
• If j does not dominate I, increment j ( j=j+1)
• Example:
• Set i = 1, P = Φ
• Set j = 2
• X1 dominates X2
• j=j+1 = 2+1= 3
• X3 dominates X1
• i = i+1 = 1+1= 2 and j = 1
• X1 dominates X2
• i = i+1 = 2+1=3, j=1
• X1 does not dominates X3
• j = 1+1 =2
• X2 does not dominates X3
• j=4
• X4 does not dominates X3
• j = 4+1 =5
• X5 does not dominates X3
• X3 IS NOT DOMINATED
• P = { 3}
• Exclude X3 and repeat procedure
• First tier is P = {3, 5}
• Second tier P = {1, 4}
• Third tier P = {2}
Formulation Example
USBRP
Problem Characteristics
• The goal of the urban school bus routing problem, USBRP, can be specified as follows :
• Find a set of school bus routes that provide transportation to and from school for all eligible students, n urban areas
• the school bus transportation system operates as follows
• Students are picked up at a bus stop that is within walking distance of their residence
• Their school bus completes visiting the rest of the bus stops remaining on its route and then goes directly
to the school.
• In this system, each school bus route serves only one school. In the afternoon, the process is reversed,
and the students are dropped off at the bus stops where they were picked up in the morning.
• The suitability of a site for being a school bus stop is influenced by characteristics such as traffic density,
proximity to corners and adjacency to public property. Because of the complicated nature of these
criteria, we assume that the potential bus stop sites have been selected by an analyst such as a school
board transportation planner.
 Problem Characteristics
• The USBRP, as just above, involves two interrelated problems.
• One problem is the assignment of students to their respective bus stops, and
• The second problem is the routing of the bus to the bus stops.

• Problems with these characteristics are known as location-routing problems(LRPs). One important
characteristic of USBRP is that they are organized into a series of layers.
• With this formulation, the LRPs is organized into three layers:
• Layer 1 is the schools,
• layer 2 is the bus stops, and
• layer 3 is the students.
• School bus routes interact between layer 1 and layer 2 (the school and the bus stops), whereas the students walking
to their bus stops in the morning and back home in the afternoon causes the interaction between layers 2 and 3
Optimization Criteria
1. Number of routes. Because the capital cost is significantly larger per bus than the incremental cost over the
year, the number of routes generated should be held to a minimum.
2. Total bus route length. This criterion reduces the total length of the school bus routes.
3. Load balancing, load balancing involves minimizing the variation in the number of students transported along
each route.
4. Length balancing, this criterion involves reducing the variation in route lengths.
5. Student walking distance. This criterion balances the total distance that students walk from home to and from
their bus stops against route length.
Routing constraints
• 1. An upper bound on the number of students on each route (bus capacity)
• 2. An upper bound on the length (or travel time) on each route
• 3. An upper bound on the total travel time (both pick-up and delivery) that students can travel
Problem Formulation
•Sets:
S = set of cardinality 1 , representing the schools
B = set of all potential bus stop sites
I = S U B, all potential routing points
J = set of all students
K = set of school buses
•Parameters:
W = school bus capacity

Sj = maximum walking distance for student j € J to a bus stop

vj = load of student j J (2/3 if student j is in early primary grades and 1 otherwise)

n = the number of routing points, equivalent to | I |
Problem Formulation
•Variables:
Problem Formulation
•Measures:

Total bus route length

Student walking distance
Loan balancing

Length balancing
Mathematical Model
• Minimize: (f1, f2, f3 and f4)
• subject to:
cijzij Sj , i B; j J (1)

zij ui , i B,jJ (2)

, jJ (3)
(4)
(5)
Mathematical Model
, j I, k (6)
, i I, k (7)
i,k (8)
fijk ≤ nxijk , i, j,k (9)

fijk ≥ 0 , i, j,k (10)

Mathematical Model
ui {0,1} , (11)

xijk {0,1} , i, j,k (12)

yik {0,1} , i I, k (13)

zij {0,1} , i B, j (14)

References
ZelenyM.,1982. “ Multiple Criteria Decision Making”, McGraw Hill”
Miettinen, K., ”On the Methodology of Multiobjective Optimization with Applications”, Report 60
(doctoral thesis), University of Jyväskylä, Department of Mathematics, Jyväskylä, 1994.
Published in parallel in JYX (open access).
Robert Bowerman, Brent Hall, Paul Calamai,”A multi-objective optimization approach to urban
school bus routing: Formulation and solution method", Transportation Research Part A: Policy and
Practice,Volume 29, Issue 2,1995,Pages 107-123,ISSN 0965-8564,https://doi.org/10.1016/0965-
8564(94)E0006-U.