Elitist Non-dominated Sorting

Genetic Algorithm: NSGA-II

Tushar Goel
Multi-objective optimization problem

Problems with more than one objectives – typically
conflicting objectives

Cars: Luxury vs. Price

Mathematical formulation
Minimize F(x),
where F(x) = {fi: ∀i = 1, M},
x= {xj: ∀j = 1, N}
Subject to:
C(x) ≤ 0, where C = {Ck: ∀ k = 1, P}
H(x) = 0, where H = {Hl: ∀ l = 1, Q}

tusharg@ufl.edu 2
Pareto optimal front

Many optimal solutions

Min f2

Usual approaches:
weighted sum strategy,
ε-constraint modeling,
Multi-objective GA

Algorithm requirements:
 Convergence
 Spread

Min f1

tusharg@ufl.edu 3
Terminology
f2

Non-domination
criterion

Ranking

f1

tusharg@ufl.edu 4
Terminology
f2

Niching – parametric

Crowding distance
 c=a+b
 Ends have infinite
crowding distance
a

f1

tusharg@ufl.edu 5
Elitism
f2
Parent

Elitism: Keep
Child
the best
individuals
from the
parent and
child
population

f1

tusharg@ufl.edu 6
 Flowchart of NSGA-II
Begin: initialize Evaluate objective Rank
population (size N) functions population

Child population created

Selection

Crossover
Report final
population and
Mutation
Stop
No
Evaluate objective
function
Stopping Elitism
criteria
Yes met? Combine parent and
Select N child populations,
individuals rank population

tusharg@ufl.edu 7
 Elitism Process

Rank 1
Rank 1 Rank 1
population
Child

Rank 2
Rank 3 Combined
Rank 2 Rank 2
population
Rank 4
Rank 3
Rank 3
Rank 1 Elitist selection
population
Parent

Rank 4 New
Rank 2
Rank 3 Rank 5 population

Rank 4 Rank 6
Rank 5+ Rank 7+

tusharg@ufl.edu 8
 Example: Bicycle Frame Design

Objectives
 Minimize area
m
 Minimize max. deflection 10 kN

Constraints
 Component should be a
valid geometry
 Maximum stress < Yield m
stress
( σmax < σ allowed )
 Maximum deflection <
Allowed deflection
Plate thickness = 20 mm
( δmax < δallowed )
tusharg@ufl.edu 9
Problem Modeling

Shapes are represented by binary strings, where ‘0’
represents void region and ‘1’ represents material
region

Example : A typical binary string is
01110 11111 10001 11111

tusharg@ufl.edu 10
Material Properties and GA Parameters

Material Properties
 Yield Stress ( σallowed ) 140 MPa
Max Deflection ( δ 5 mm
allowed )


 Young’s Modulus (E) 80GPa

 Poisson’s Ratio ( ν ) 0.25

GA Parameters
 Binary String Size (L) 14x9
 Population Size 30
 Crossover Probability 0.95
 Mutation Probability 1/L
 # of Generations 150

tusharg@ufl.edu 11
 Pareto Optimal Front

Small increase in
weight leads to large
drop in deflection

Similarly small
change in deflection
allows significant
reduction of the
weight

tusharg@ufl.edu 12
Optimal shapes
Least weight

Least deflection

Different conceptual designs

tusharg@ufl.edu 13
Some More Engineering Applications

Structural designs of mechanical components

Design of turbo-machinery components

Bioinformatics – protein unfolding

VLSI circuit designs

Packaging

…

tusharg@ufl.edu 14
Other related topics of interest

Real-coded genetic algorithms

Other multi-objective evolutionary algorithms
 Pareto archived evolutionary strategies (PAES)
 Strength Pareto evolutionary algorithm (SPEA)
 ε – multi-objective evolutionary algorithm (ε-
MOEA)

Hybrid GAs

Particle swarm algorithms

Ant colony optimization

tusharg@ufl.edu 15