A Brief Introduction to Genetic

Optimization
S.D. Sudhoff

Fall 2005
 Traditional Optimization Methods

• Newton’s Method

Fall 2005 EE595S 2

 Traditional Optimization Methods

• Steepest Decent
• Conjugate Direction Algorithm
• Conjugate Gradient Algorithm
• Davidon, Fletcher, Powell, Algorithm (DFP)
• Broyden, Fletcher, Goldfarb, and Shanno
Algorithm (BFGS)
• Nelder-Mead Simplex Method
• Take ECE580

Fall 2005 EE595S 3

 Traditional Optimization Methods

• Properties of all of these methods

¾Need an initial guess
¾Very susceptible to finding local minimum
• Properties of most of these methods (not
Nelder Meed Simplex)
¾Take derivatives

Fall 2005 EE595S 4

 An Alternate Approach

Fall 2005 EE595S 5

 Genetic Algorithms in a Nutshell

• Probabilistic Optimization Technique

• Loosely Based in Principals of Genetics
• First Developed By Holland, Late 60’s –
Early 70’s
• Does Not Require Gradients or Hessians
• Does Not Require Initial Guess
• Operates on a Population

Fall 2005 EE595S 6

 A Gene As A Parameter:
Biological Systems

• Each Gene Has Value From Alphabet

{AT,GC,CG,AT} from Nitrogenous
Adenine, Guanine, Thymine, Cytosine
• Each Gene is Located on One of a Number
of Chromosomes
• Chromosomes May Be Haploid, Diploid,
Polyploid

Fall 2005 EE595S 7

 A Gene As A Parameter:
Canonical Genetic Algorithm

• Each Gene Has a Value From Alphabet

(Normally Binary {0,1})
• Each Gene is Located on a Chromosome
(Normally 1)
• Chromosomes Generally Haploid

Fall 2005 EE595S 8

 Evolution in a Canonical GA

START

Initialization

Fitness evaluation

Selection

Crossover

Mutation

STOP?

END

Fall 2005 EE595S 9

 Example Operator:
Single Point Crossover

Gene 1 Gene 1 Gene 1 Gene 1

Gene 2 Gene 2 Gene 2 Gene 2

Gene 3 Gene 3 Gene 3 Gene 3

Gene 4 Gene 4 Gene 4 Gene 4

Gene 5 Gene 5 Gene 5 Gene 5

Parent 1 Parent 2 Child 1 Child 2

Fall 2005 EE595S 10
 Example Operator:
Mutation

Mutation point

Original chromosome 0 1 0 0 0 1 1 0 1 1

Mutated chromosome 0 1 0 0 1 1 1 0 1 1

Fall 2005 EE595S 11

 A Gene As A Parameter:
Non-Encoded Genetic Algorithms
• Each Gene Has A Range (Minimum Value,
Maximum Value)
• Each Gene Has A Type (Mapping)
¾Integer
¾Linearly Mapped Real Number
¾Logarithmically Mapped Real Number
• Each Gene is Located on One of a Number
of Chromosomes
• Chromosomes Are Haploid
Fall 2005 EE595S 12
 Genetic Optimization System
Engineering Tool (GOSET)
• Matlab based toolbox
• Usable with minimum knowledge
• Allows, if desired, high degree of algorithm
control

Fall 2005 EE595S 13

 An Individual in GOSET

• A Collection of
Genes on Gene 1 Gene 6 Gene 9
Chromosomes Gene 2 Gene 7 Gene 10
• Fitness (Measures of Gene 3 Gene 8 Gene 11
Goodness)
Gene 4 Chromosome 2 Gene 12
f = [ f1 fN ]
T
f2 " Gene 5 Chromosome 3

Chromosome 1
• Region

Fall 2005 EE595S 14

 A Population In GOSET

Fi Fi
Region 1 Fi
Fi
Fi
Fi Fi
Fi
Fi

Fi Fi Fi
Fi

Region 2
Fi Fi

Fi Fi
Fi

Fi

Region 3
Fall 2005 EE595S 15
 Evolution in GOSET

Fall 2005 EE595S 16

 Example: Powell’s Problem

• Minimize
f ( x) = ( x1 + 10 x 2 ) 2 + 5( x3 − x 4 ) 2 + ( x 2 − 2 x 3 ) 4 + 10( x1 − x 4 ) 4

• Fitness
−5
fitness( x) = 5 − log 10( f ( x) + 10 )

Fall 2005 EE595S 17

 powell.m

% Optimization of Powell's Problem

% Initialize the parameters

GAP=gapdefault;

% x1 x2 x3 x4
% gene 1 2 3 4
GAP.gd_min =[ -2.1 -2.1 -2.1 -2.1];
GAP.gd_max =[ 2.0 2.0 2.0 2.0];
GAP.gd_type=[ 2 2 2 2 ];
GAP.gd_cid =[ 1 1 1 1 ];

[P,GAS]= gaoptimize(@powell_fit,GAP,[],[],[],[]);

parameters=GAS.bestgenes(:,GAS.cg);

Fall 2005 EE595S 18

 powell_fit.m

function f = powell(x)

% Powell's function taken from Chong & Zak, p. 140.

% The global miniizer is powell(0,0,0,0) = 0

x1=x(1);
x2=x(2);
x3=x(3);
x4=x(4);
f1 = (x1 + 10*x2)^2 + 5*(x3 - x4)^2 + (x2 - 2*x3)^4 + 10*(x1-x4)^4;
f=5-log10(f1+0.00001);

Fall 2005 EE595S 19

 Powell’s Problem: Conclusion

1
Normalized Value

0.5

parameters =
0
1 2 3 4
Parameter Number

10
-0.0098
0.0014
f:B(b) Md(g) Mn(r)

8
-0.0097
6
-0.0105
4

2
10 20 30 40 50 60 70 80 90 100
Generation

Fall 2005 EE595S 20

 IGBT Conduction Loss

• Fit IGBT Conduction Loss Data To:

v = ai + (bi) c

Fall 2005 EE595S 21

 igbt.m
% load experimental data
% voltage is first column
% current is second column
load igbt_data.mat
Data.v=v_data;
Data.i=i_data;

% gafit
GAP=gapdefault;
GAP.rp_lvl=1;
GAP.mg_nreg=4;
GAP.mg_tmig=20;
GAP.mg_pmig=0.05;
GAP.dp_fp=0.5;
GAP.dth_alg=2;
Fall 2005 EE595S 22
 igbt.m (cont.)

% declare the types of all the parameters

GAP.gd_min = [1e-8 1.0e-6 1.0e-3];
GAP.gd_max = [1e+2 1.0e+3 1.0e+0];
GAP.gd_type= [ 3 3 3 ];
GAP.gd_cid = [ 1 1 1 ];

% do the optimization
[fP,GAS]= gaoptimize(@igbt_fit,GAP,Data,[],[],[]);

% plot the best fit

bestparameters=GAS.bestgenes(:,GAS.cg,1)
igbt_fit(bestparameters,Data,3);
Fall 2005 EE595S 23
 igbt_fit
% this rountine returns the fitness of the IGBT conduction
% loss parameters based on the mean percentage error
function fitness = IGBTfitness(parameters,data,fignum)

% assign genes to parameters

a=parameters(1);
b=parameters(2);
c=parameters(3);

vpred=a*data.i+(b*data.i).^c;
error=abs(1-vpred./data.v);
fitness=1.0/(1.0e-6+mean(error));
Fall 2005 EE595S 24
 igbt_fit (cont)
if nargin>2

fitness

figure(fignum);
Npoints=200;
ip=linspace(0,max(data.i),Npoints);
vp=a*ip+(b*ip).^c;
plot(data.i,data.v,'bx',ip,vp,'r')
title('Voltage Versus Currrent');
xlabel('Current, A');
ylabel('Voltage, V');
legend({'Measured','Fit'});

figure(fignum+1);
Npoints=200;
plot(data.i,data.v.*data.i,'bx',ip,vp.*ip,'r')
title('Power Versus Currrent');
xlabel('Current, A');
ylabel('Power, V');
legend({'Measured','Fit'});

end

Fall 2005 EE595S 25

 IGBT Results

1
Normalized Value

0.5
bestparameters =
0
1 2 3
Parameter Number 0.0000
10 4.5288
0.2588
f:B(b) Md(g) Mn(r)

0
10 20 30 40 50 60 70 80 90 100
Generation

Fall 2005 EE595S 26

 IGBT Results (Cont.)
Voltage Versus Currrent
4

3.5

3 Measured
Fit
2.5
Voltage, V

1.5

0.5

0
0 5 10 15 20 25 30
Current, A

Fall 2005 EE595S 27

 IGBT Results (Cont.)
Power Versus Currrent
120
Measured
Fit
100

80
Power, V

60

40

20

0
0 5 10 15 20 25 30
Current, A

Fall 2005 EE595S 28

 Your Problem

• Consider a machine for which

λ m = 0.3 VS

Lq = 15 mH

L d = 10 mH

P=4

Fall 2005 EE595S 29

 Your Problem (Cont)

• Develop an approximate expression for the

optimal q-axis current command (in MTPA
sense) as a function of desired torque

• Deliverables: short write up with your

expression, all *.m files etc used.

• NOTE: GOSET2.2 is on the web. See

software distribution.
Fall 2005 EE595S 30
 Comments on Approach

Fall 2005 EE595S 31