l13 Msdo Pso

Particle Swarm Optimization:
Method and Applications
Rania Hassan
Post-doctoral Associate
Engineering Systems Division
© Rania Hassan 3/2004

Engineering Systems Division - Massachusetts Institute of Technology 1
Particle Swarm Optimization
A pseudo-optimization method (heuristic) inspired by the
collective intelligence of swarms of biological populations.
Flocks of Birds Colonies of Insects

Particle Swarm Optimization
A pseudo-optimization method (heuristic) inspired by the
collective intelligence of swarms of biological populations.
Schools of Fish Herds of Animals

Inventors
Introduced in 1995: Kennedy, J. and Eberhart, R., “Particle Swarm
Optimization,” Proceedings of the IEEE International Conference on Neural
Networks, Perth, Australia 1995, pp. 1942-1945.
James Kennedy
Social Psychologist
US Department of Labor
Russell Eberhart
Dean of Engineering Research
Indiana Univ. Purdue Univ. Indianapolis

Swarming in System Design
• Swarming theory has been used in system design.

Examples of aerospace systems utilizing swarming theory
include formation flying of aircraft and spacecraft.
Flocks of birds fly in V-shaped formations to reduce drag

and save energy on long migrations.

Weimerskirch, H. et al. "Energy saving in flight formation." Nature 413,

(18 October 2001): 697 - 698.
A study of great white pelicans has found that birds

flying in formation use up to a fifth less energy than
those flying solo (Weimerskirch et al.).

• A space system in formation flying combines data from several
spacecraft rather than flying all the instruments on one costly satellite.
• It also allows for the collection of data unavailable from a single
satellite, such as stereo views or simultaneously collecting data of the
same ground scene at different angles.
Image taken from NASA's website. http://www.nasa.gov.

Lecture Overview
• Introduction
• Motivation
• PSO Conceptual Development
• PSO Algorithm
– Basic Algorithm
– Constraint Handling
– Discretization
• Applications
– PSO Demo
– Benchmark test problems
• Unconstrained
• Constrained
– System Problem: Spacecraft Design
• Final Comments on PSO
• References

“Personal” Motivation
PSO
– is a zero-order, non-calculus-based method (no gradients are needed).
– can solve discontinuous, mutlimodal, non-convex problems.
– includes some probabilistic features in the motion of particles.
– is a population-based search method, i.e. it moves from a set of points
(particles’ positions) to another set of points with likely improvement in
one iteration (move). [is that good or bad ??]
Does it remind you of another heuristic?
The Genetic Algorithm (GA)
– The GA is inherently discrete (in terms of handling design variables)
– PSO is inherently continuous (in terms of handling design variables)
– Some researchers report that PSO requires less function evaluations

than the GA (most problems studied are continuous).
– In Compindex: there are 18150 hits for the GA from 1990 to 2004,
whereas there are only 105 hits for PSO - many version of PSO are
likely to appear.
PSO Conceptual Development
• Social Behavior Simulation Optimizer
• The social model was intended to answer the following
questions:
– How do large numbers of birds (or other populations
exhibiting swarming behavior) produce seamless, graceful
flocking choreography, while often, but suddenly changing
direction, scattering and regrouping?
– Are there any advantages to the swarming behavior for an
individual in a swarm?
– Do humans exhibit social interaction similar to the swarming
behavior in other species?
• Keep these questions in mind while watching clips from

the French documentary “Winged Migration” by Jacques
Perrin.
• How do large numbers of birds produce seamless, graceful

flocking choreography, while often, but suddenly changing
direction, scattering and regrouping?
– “Decentralized” local processes.
– Manipulation of inter-individual distances (keep pace and avoid collision).
• Are there any advantages to the swarming behavior for an

individual in a swarm?
– Can profit from the discoveries and previous experience of other swarm
members in search for food, avoiding predators, adjusting to the
environment, i.e. information sharing yields evolutionary advantage.
• Do humans exhibit social interaction similar to the swarming

behavior in other species?
– Absolutely, humans learn to imitate physical motion early on; as they grow
older, they imitate their peers on a more abstract level by adjusting their
beliefs and attitudes to conform with societal standards.

• Phase I:
– Randomly generated particle positions in 2-d space.
– Randomly generated velocity vectors for each particle in 2-d space.
– For each swarm movement (iteration), each particle (agent) matches

the velocity of its nearest neighbor to provide synchrony.
– Random changes in velocities (craziness) are added in each iteration

to provide variation in motion and “life-like” appearance – artificial.

• Phase II:
– Heppner’s simulation used a roost (cornfield), that the birds
flocked around before landing there – eliminating the need for
“craziness “.
– The mathematical formulation for the roosting model included:
• each particle evaluates its current fitness by comparing its position to
the roost position (100, 100).
ith particle
kth time unit f ki = (xki − ) + (yki − )
• each particle remembers its best ever fitness value, f ,i and the
position associated with it, p i = > x i y i @ , and adjusts its velocity
accordingly.
LI xki > x i ∴Vxki + = Vxki − crand Uniformly distributed

LI xki < x i ∴Vxki + = Vxki + crand random number
update the velocity in the y-direction in the same manner

• Phase II:
• each particle knows the position of the best particle in the current
swarm, p g = > x g y g @ . This is the particle with the lowest fitness
k k k
value, f g (closest to the roost).
k
LI xki > xkg ∴Vxki + = Vxki − c rand
global best
LI xki < xkg ∴Vxki + = Vxki + c rand
• in the simulation, when c and c are set high, the flock is sucked
violently into the cornfield; whereas if they are set low, the flock
swirls around the cornfield and slowly approach it.
– The simulation with the known cornfield position appeared real
and raised the question of how do birds find “optimal” food
sources in reality without knowing the “cornfield” location?
• Example 1: scientists proved that parks in affluent neighborhoods
attract more birds than parks in poor neighborhoods.
• Example 2: put a bird feeder in your balcony and see what happens
in few hours.
PSO Algorithm

Basic PSO Algorithm
• Phase III
– The swarming behavior of the birds could be the reason for finding optimal
food resources.
– The model developed in Phase II could be used (with minor modifications)
to find optimal solutions for N-dimensional, non-convex, multi-modal,
nonlinear functions.
– In this current basic version of PSO, craziness and velocity matching are
eliminated.
Algorithm Description
• Particle Description: each particle has three features

i
– Position x k (this is the ith particle at time k, notice vector notation)
i
– Velocity v k (similar to search direction, used to update the position)
i
( )
– Fitness or objective f x k (determines which particle has the best value in
the swarm and also determines the best position of each particle over time.

Basic PSO Algorithm
• Initial Swarm
– No well established guidelines for swarm size, normally 15 to 30.
– particles are randomly distributed across the design space.
x i = x PLQ + rand (x PD[ − x PLQ )
where x PLQ and x PD[ are vectors of lower and upper limit values
respectively.
– Evaluate the fitness of each particle and store:

• particle best ever position (particle memory p i here is same as x i )
g
• Best position in current swarm (influence of swarm p )
– Initial velocity is randomly generated.

x + rand (x PD[ − x PLQ ) SRVLWLRQ
v i = PLQ =
ǻt WLPH
Basic PSO Algorithm
• Velocity Update
– provides search directions
– Includes deterministic and probabilistic parameters.
– Combines effect of current motion, particle own memory,
and swarm influence.
New velocity
v ik + = wv ik + c rand
( p i
− x ik ) + c rand
§¨ p g − x i ·¸
© k k
¹

∆t ∆t
current motion particle memory swarm influence

influence
inertia factor self confidence self confidence

0.4 to 1.4 1.5 to 2 2 to 2.5

Basic PSO Algorithm
• Position Update
– Position is updated by velocity vector.
x ik +
pg
i k
x ik + = x ik + v ik + ∆t p
x ik v ik
• Stopping Criteria
– Maximum change in best fitness smaller than specified
tolerance for a specified number of moves (S).
( ) ( )
f p kg − f p kg− q ≤ ε q = , ,..S

Constraint Handling
• Side Constraints
– Velocity vectors can drive particles to “explosion”.
– Upper and lower variable limits can be treated as regular constraints.
– Particles with violated side constraints could be reset to the nearest
limit.
• Functional Constraints
– Exterior penalty methods (linear, step linear, or quadratic).
penalty multipliers
fitness function N con
f (x ) = φ (x ) + ¦ ri (PD[[,gi ])
i =
objective function penalty function
– If a particle is infeasible, last search direction (velocity) was not

feasible. Set current velocity to zero.
v ik + = c rand
( p i
− x ik ) + c rand
§¨ p g − x i ·¸
© k k
¹

∆t ∆t
Discretization
• System problems typically include continuous, integer,

and discrete design variables.
• Basic PSO works with continuous variables.
• There are several methods that allows PSO to handle

discrete variables.
• The literature reports that the simple method of rounding

particle position coordinates to the nearest integers
provide the best computational performance.

Applications

PSO Demo
• Demo provided by Gerhard Venter of Vanderplatts R&D.
• Smooth function with three local minima and a known global
minimum of [0,0] and an optimal function value of 0.

Courtesy of Gerhard Venter. Used with permission.

PSO Demo
• Demo provided by Gerhard Venter of Vanderplatts R&D.
• Noisy function with many local minima and a known global
minimum of [0,0] and an optimal function value of 1000.

Courtesy of Gerhard Venter. Used with permission.

Unconstrained Benchmark Problems
Eggcrate Function
Minimize (
f (x ) = x + x + VLQ x + VLQ x )
• Known global minimum of [0,0] and an optimal function value of 0.
• PSO parameters:
– Swarm Size = 30
– Inertia, w = 0.5 (static)
– Self Confidence, c = 1.5
– Swarm Confidence, c = 1.5
– Stopping Tolerance, ε = 0.001

Eggcrate Function
Results for move0 Results for move5
6 2.5
2
4
Statistics 1.5 Statistics
2 1
Min = 15.1 0
0.5
Min = 0.25
x2
x2
0
-2
Mean = 49.9 -0.5
Mean = 26.5
-4 -1
-1.5
Max = 92.6 -6
-2
Max = 86.6
-8 -2.5
-6 -4 -2 0 2 4 6 -8 -6 -4 -2 0 2 4 6 8 10
x1 x1

2.5 1
2
0.8
Statistics 1.5 Statistics
1 0.6
Min = 0.02 0.5

0.4 Min = 0.3E-5
x2
x2
0
0.2
Mean = 13.6 -0.5
Mean = 4.69
-1 0
Max = 66.5
-1.5
-2
-0.2 Max = 41.9
-2.5 -0.4
-5 -4 -3 -2 -1 0 1 2 3 4 -6 -5 -4 -3 -2 -1 0 1 2
x1 x1

Banana (Rosenbrock) Function
Minimize f (x ) = x − (
x ) + ( − x )
• Known global minimum of [1,1] and an optimal function value of 0.
• PSO parameters:
– Swarm Size = 30
– Stopping Tolerance, ε = 0.001

Banana (Rosenbrock) Function
5 4
Statistics 3
3
Statistics
2
2
Min = 1.03 1
Min = 0.26
x2
x2
0 1
Mean = 1.4E4 -1
0
Mean = 5.8E3
-2
-3
Max = 5.3E4 -4
-1
Max = 6.9E4
-5 -2
-5 0 5 -4 -2 0 2 4 6
x1 x1

1.2 1.6
1 1.4
Statistics 1.2
Statistics
0.8
Min = 0.36 0.6

1
Min = 0.6E-3
x2
x2
0.8
0.4
Mean = 2.6E4 0.6 Mean = 1.3E3
0.2
0.4
Max = 7.2E5 0 0.2

Max = 2.7E4
-0.2 0
-4 -2 0 2 4 6 8 10 -5 -4 -3 -2 -1 0 1 2 3 4
x1 x1

Constrained Benchmark Problems
Golinski Speed Reducer
• Full problem description can be found at
http://mdob.larc.nasa.gov/mdo.test/class2prob4.html
• This problem represents the design of a simple gear box such
as might be used in a light airplane between the engine and
propeller to allow each to rotate at its most efficient speed.
• The objective is to minimize the speed reducer weight while
satisfying a number of constraints (11) imposed by gear and
shaft design practices.
• Seven design variables are available to the optimizer, and
each has an upper and lower limit imposed.
• PSO parameters:
– Swarm Size = 60
– Stopping Tolerance, ε = 5 kg
Image taken from NASA's website. http://www.nasa.gov.
• Known solution
x = [3.50 0.7 17 7.3 7.30 3.35 5.29]
f (x ) = 2985 kg
• PSO solution
x = [3.53 0.7 17 8.1 7.74 3.35 5.29]
f (x ) = 3019 kg x 10
7
PSO Statistics
5
4.5
Min, Mean, and Max Fitness

3.5
2.5
1.5
0.5
0
0 20 40 60 80 100 120
Move

7 7
x 10 Results for Move 0, # of Feasible Particles (out of 60) = 0 x 10 Results for Move 5, # of Feasible Particles (out of 60) = 24
4.5 4.5
4 4
3.5 3.5
3 3
Fitness Value
Fitness Value
2.5 2.5
2 2
1.5 1.5
1 1
0.5 0.5
0 0
0 10 20 30 40 50 60 0 10 20 30 40 50 60
Particle number Particle number
7 6
3 14
12
2.5
10
2
Fitness Value
Fitness Value
8
1.5
6
1
4
0.5
2
0 0
0 10 20 30 40 50 60 0 10 20 30 40 50 60

System Problem
Communication Satellite Design
Value Acquisition
Cost
Reliability
Payload / Launch
Mass
Mission
Requirements

System Problem
Performance Optimization
Estimation Algorithm
Tools
Challenges
Multidisciplinary Framework Combinatorial Problem
¾ many design variables ¾ discrete
¾ computationally-expensive ¾ integer
disciplinary design tools ¾ continuous
Solutions
Focus on the conceptual Use non-calculus based
design stage (less variables, search methods like the
simpler models) Genetic Algorithm (GA)
System Problem
Design Variables
There are 27 design variables that can be combined into 2.2 trillion different designs.
Design
Description and discrete values
parameter
1 to 8 HPA type (TWTA or SSPA)
Solar array cell type (GaAs single junction, GaAs multi-junction, Si
9
thin, Si normal, or hybrid Si with GaAs multi-junction)
10 Battery cell type (NiCd or NiH2)
11 N/S thermal coupling (no coupling or coupling)
N/S STK thruster technology (xenon plasma, arcjets, bi-propellant,
12
and hydrazine monopropellant)
13 E/W STK thruster technology (bi-propellant or hydrazine)
14 Launch vehicle choice from 8 options

System Problem
Design Variables continued …
Design
Description and discrete values
parameter
Redundancy level for Ku-band transponder that has 12 operating
15
HPAs (0, 1, 2, or 4 redundant HPAs)
Redundancy levels for C-band transponders, each has 12
16 to 22
operating HPAs (0, 1, 2, or 4 redundant HPAs)
Redundancy levels for last C-band transponder that has 2
23
operating HPAs (0 or 1 redundant HPA)
24 Propulsion subsystem redundancy level (0 or full redundancy)
25 ADCS subsystem redundancy level (0 or full redundancy)
26 TCR subsystem redundancy level (0 or full redundancy)
Solar array area redundancy level (0, 1.02, 1.04 or 1.06 × Area
27
redundancy level)

Fairing
Diameter
Stowed Solar Stowed Sidewall

Array Antenna
Stowed Satellite Configuration in Launch

Vehicle Fairing
System Problem
• PSO parameters:
– Swarm Size = 60
– Stopping Tolerance, ε = 5 kg
• Known solution
x = [1 0 1 1 1 1 1 1 4 2 1 1 0 0 1 1 0 1 1 1 0 0 1 1 0 0 3]
f (x ) = 3443 kg
• PSO solution
x = [1 1 1 0 1 1 1 1 4 2 1 1 0 1 0 2 1 1 0 1 1 0 1 0 1 1 2]
f (x ) = 3461 kg

System Problem
5 5
16 9
14 8
7
12
6
10
Fitness Value
Fitness Value
5
8
4
6
3
4
2
2 1
0 0
0 10 20 30 40 50 60 0 10 20 30 40 50 60
5 4
14 1.8
12 1.6
1.4
10
1.2
Fitness Value
Fitness Value
8
1
6
0.8
4
0.6
2 0.4
0 0.2
0 10 20 30 40 50 60 0 10 20 30 40 50 60

Final Comments on PSO
• This is a method “in the making” - many versions

are likely to appear.
• Poor repeatability in terms of:

– finding optimal solution
– computational cost
• More robust constraint (side and functional) handling

approaches are needed.
• Guidelines for selection of swarm size, inertia and

confidence parameters are needed.

References
• Kennedy, J. and Eberhart, R., “Particle Swarm Optimization,”

Proceedings of the IEEE International Conference on Neural
Networks, Perth, Australia 1995, pp. 1942-1945.
• Venter, G. and Sobieski, J., “Particle Swarm Optimization,”

AIAA 2002-1235, 43rd AIAA/ASME/ASCE/AHS/ASC Structures,
Structural Dynamics, and Materials Conference, Denver, CO.,
April 2002.
• Kennedy, J. and Eberhart, R., Swarm Intelligence, Academic

Press, 1st ed., San Diego, CA, 2001.


l13 Msdo Pso

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

l13 Msdo Pso

Uploaded by

Copyright:

Available Formats

Particle Swarm Optimization:

Method and Applications

© Rania Hassan 3/2004

Flocks of Birds Colonies of Insects

© Rania Hassan 3/2004

Schools of Fish Herds of Animals

© Rania Hassan 3/2004

© Rania Hassan 3/2004

• Swarming theory has been used in system design.

Flocks of birds fly in V-shaped formations to reduce drag

© Rania Hassan 3/2004

Weimerskirch, H. et al. "Energy saving in flight formation." Nature 413,

A study of great white pelicans has found that birds

© Rania Hassan 3/2004

Image taken from NASA's website. http://www.nasa.gov.

© Rania Hassan 3/2004

– Some researchers report that PSO requires less function evaluations

• Keep these questions in mind while watching clips from

• How do large numbers of birds produce seamless, graceful

• Are there any advantages to the swarming behavior for an

• Do humans exhibit social interaction similar to the swarming

© Rania Hassan 3/2004

– Randomly generated particle positions in 2-d space.

– Randomly generated velocity vectors for each particle in 2-d space.

– For each swarm movement (iteration), each particle (agent) matches

– Random changes in velocities (craziness) are added in each iteration

© Rania Hassan 3/2004

LI xki > x i ∴Vxki + = Vxki − crand Uniformly distributed

update the velocity in the y-direction in the same manner

© Rania Hassan 3/2004

• Particle Description: each particle has three features

© Rania Hassan 3/2004

– Evaluate the fitness of each particle and store:

– Initial velocity is randomly generated.

current motion particle memory swarm influence

inertia factor self confidence self confidence

© Rania Hassan 3/2004

© Rania Hassan 3/2004

objective function penalty function

– If a particle is infeasible, last search direction (velocity) was not

• System problems typically include continuous, integer,

• Basic PSO works with continuous variables.

• There are several methods that allows PSO to handle

• The literature reports that the simple method of rounding

© Rania Hassan 3/2004

© Rania Hassan 3/2004

© Rania Hassan 3/2004

Courtesy of Gerhard Venter. Used with permission.

© Rania Hassan 3/2004

Courtesy of Gerhard Venter. Used with permission.

© Rania Hassan 3/2004

Results for move10 Results for move28

Min = 0.02 0.5

© Rania Hassan 3/2004

• Known global minimum of [1,1] and an optimal function value of 0.

© Rania Hassan 3/2004

Results for move25 Results for move53

Min = 0.36 0.6

Max = 7.2E5 0 0.2

© Rania Hassan 3/2004

Min, Mean, and Max Fitness

© Rania Hassan 3/2004

© Rania Hassan 3/2004

© Rania Hassan 3/2004

LI xki > x i ∴Vxki + = Vxki − crand Uniformly distributed