Вы находитесь на странице: 1из 55

Using optimtool/ga in Matlab

Optimtool
Optimtool is an optimization tool with several different applications, one of which is genetic algorithms. These slides are to show how to use optimtool with two kinds of chromosomes:
Binary chromosomes Double vector chromosomes

Function to Optimize
The goal of this optimization example is to find the minimum for the following function. Its plot is on the next slide:
f ( x ) ! x 5  12.1x 4  40.59 x 3  17.015 x 2  71.95 x  35.88

Notice that the minimum ~-165 occurs at x=~5.75

Optimtool
Optimtool is invoked by typing optimtool at the Matlab prompt. Once started, you will see the window on the next page: Don t be intimidated by the complexity and the number of parameters of this window. Much of what we will use is the default. Note also that on the right side of the window there are references for the various options. Selecting one of these options will expand the window to give further explanation of the option.
5

Optimtool Double Vector or Binary Chromosome


A double vector chromosome is simply a row vector of n double values. These values can be thought of as genes. Thus 4 genes/chromosome means a double vector of 4 elements. A binary chromosome is a k bit vector of binary values
Note that for this chromosome, genes are not actually delineated. In other words if one needed a chromosome with 4 genes of 2, 8, 9, and 7 bits each, then one would specify a 26 bit chromosome
7

Optimtool Double Vector Chromosome


In order to use Optimtool as a genetic algorithm solver, one must select ga Genetic Algorithm in the solver box. (Next slide)

Optimtool Double Vector Chromosome


When using double vectors, you don t have as much latitude compared to binary in how chromosomes change. Because a chromosome is made up of individual double precision values, crossover only occurs on the boundaries of the individual doubles. Mutation thus takes on more significance in terms of moving in the search space. This is especially true if your chromosome consists of a single double (gene).
10

Optimtool Double Vector Chromosome


Since in this problem, we have only one gene, we will only have a single parameter and thus fill in number of variables window with a 1 Note:
With Double Vectors, a chromosome is a row vector whose length is the number of genes in a chromosome.

11

Optimtool Double Vector Chromosome


Problem Segment (portion of the window labeled Problem
In this segment, one specifies a m-file that is to be used as a fitness function and the number of variables. You must create the m-file in your Matlab workspace or working directory and reference it in optimtool as @m-filename In the definition of the m-file, for this example there will be one passed parameter, namely a row vector of a single double representing the chromosome. The m-file must return a scalar. The Matlab GA s goal is to find the minimum of the fitness function.
12

Question
Since Optimtool only works to find the minimum could you use it if the fitness function was for a maximum?
Yes, simply use 1/fitness If fitness can go to zero, then you might want to use 1/(1+fitness) Of course if fitness can be -1, then that solution won t work, and another must be used.

13

Optimtool Double Vector Chromosome


On slide 3 we defined the fitness function. The following is the m-file (with no documentation) for that fitness function. function [ funval ] = polyVal( X ) % funval= X^5-12.1*X^4+40.59*X^3-17.015*X^2 -71.95*X+35.88; end For the preceding m-file fitness function, you would type @polyVal in the fitness function window.
14

Fitness Function
Note that in optimtool the fitness function is defined without parameters while in the Matlab definition, the parameter X is shown.
This means that in your Matlab workspace there must be only one function with that name. If you have a chromosome with say 10 genes, then polyVal will still be defined as an m-file with one parameter. The difference is that in the function you must recognize that what is passed is really a vector (X) with 10 elements.
15

Optimtool Double Vector Chromosome


Constraints
There are several options for setting constraints. In essence what they allow you to do is restrict the range of values that each of the genes can have. For a three gene chromosome [x1 x2 x3] consider the following constraints:
-5.5 <= x1 <= 7.6 10 <= x2 <= 25 -15.0 <= x3 <= -7.6
16

Optimtool Double Vector Chromosome


In this case, it would be simplest to use the lower and upper bounds boxes. In doing this, you have two choices.
Define two vectors in your work space and refer to them in the windows, or define the vectors in the window. Vectors in the Matlab work space >> low=[-5.5;10;-15]; >> upp=[7.6 ;25; -7.6] And then in the upper and lower windows type low and upp respectively
17

Optimtool Double Vector Chromosome


In this example problem, it would be simplest to use the lower and upper bounds boxes.
To specify the bounds in the windows type [-5.5;10;-15] in the window labeled Lower: and [7.6; 25; -7.6] in the window labeled Upper:

18

Optimtool Double Vector Chromosome


For our problem, we have only one gene/chromosome. Also, looking at the figure of slide 4 you can see that the range is (approximately) -1.5 to 6.75. Thus,
you can simply type [-1.5] and [6.75] in the upper and lower windows.

19

Optimtool Double Vector Chromosome - OPTIONS


Even though the next portion of the optimtool window is the Run solver part, you must first set the options in the right hand side of the optimtool window. Population
Population type: this should be Double Vector Population size: you can leave it at the default of 20 chromosomes or specify a larger population size in its specify: window.
20

Optimtool Double Vector Chromosome - OPTIONS


Creation function
Here you can use constraint dependent or feasible population. Either will operate the same (see the reference)

Initial population
Leave this blank so that optimtool creates the initial population

Initial scores leave this blank so that optimtool will calculate the fitness for the initial population.
21

Optimtool Double Vector Chromosome - OPTIONS


Initial range: - IMPORTANT
If you use the default, then your initial range for the randomly chosen vectors and each gene will be [0;1]. In this case you should specify the initial range as the same as the lower and upper bounds, i.e. [-1.5;6.75]

Scaling: default or experiment Selection: Default or experiment For all of the rest, you can use the default, expect you should also experiment. Note: The generations under stopping criteria should generally be set higher than 100.
22

Optimtool Double Vector Chromosome - OPTIONS


Once parameters, etc. have been set, select start in the Run solver window. In this example, for the plot window I chose a plot interval of 10 and best fitness and best individual. The next two slides shows a run.

23

24

25

Double Vector Example


Although hard to read, in the window of the preceding slide there are three outputs to note:
Current iteration 51 (Since we have stopped, this the number of iterations of the GA required to arrive at this value) In the window we see optimization running objective function value: -165.9 . Optimization terminated, average . In the bottom box labeled final point is the value of the chromosome that gave the best (minimum) fitness of 165.9
26

Double Vector Example


It should be noted that there is a bug in the Matlab Ga/Optimtool. Specifically, if you use the default population size, the routine always stops at 51 generations. The fact that it says it stopped because of no change in the fitness is not necessarily correct. In order to not always stop at 51 iterations you must specify stall generations as other than 50 and a population size other than the default of 100
27

28

Optimtool Binary Chromosome


In general, a binary chromosome allows some additional flexibility in searching the space. For example, crossover is no longer constrained to occur on gene boundaries. It can occur within genes. Binary chromosomes also require a few different initial settings and a significantly different fitness function.
29

Binary Chromosome Fitness Function


As noted on slide 7, the number of bits in a chromosome is the sum of the number of bits in each gene, but how many bits are needed in each gene? The number of bits needed in each gene is dependent on the range of values for that gene, and resolution desired.

30

Binary Strings in Matlab/Optimtool


When you specify a binary vector, you specify the number of bits as the total number of bits in all of the variables.
As far as optimtool is concerned, a chromosome is just a string of bits. It is up to the writer of the fitness function to separate these bits into individual genes and then evaluate the genes, i.e convert them into decimal values.

31

Binary Chromosome Vector Example


Let s say that we have three genes, a resolution for each gene of 0.1, and each gene has a range of values of:
Gene 1: -1 to 1 => 2/0.1 = 20 values or 5 bits Gene 2: 10 to 20 => 10/0.1 = 100 values or 7 bits Gene 3: -100 to 10 => 110/0.1 = 1100 values or 11 bits Thus, total gene size is 23 bits.
32

Binary Chromosome Vector Example


We will assume that the fitness is simply the sum of the values of the individual genes.
What this means is that our optimum will be a gene of all 0 s.

In the m-file that calculates the fitness, we will need to convert the 23 bit string into 3 equivalent decimal values.

33

Optimtool/Matlab Binary Strings


In Optimtool, if you specify, under population type, a binary string(vector), and under the number of variables 23, then each string is a random string of 23 bits, e.g. it is of the form
String=[1 0 1 0 0] % for a total of 23 bits The next two slides show how to extract a portion of such a string to get a particular gene s decimal equivalent value.

34

Optimtool/Matlab Binary Strings


The actual fitness function would call ConvertPortionToBinary for each of the genes in the chromosome. Finally, the fitness function would simply, for the fitness of the given chromosome, return the sum of these individual gene values.
Obviously, for most problems, the fitness function itself would need to perform a more complicated function.

35

function [ output ] = ConvertPortionToBinary( input_string, position, number_of_bits ) %Takes a binary row vector (input_string) as input and % converts it to its decimal equivalent. % position is where in input to begin the conversion % number_of_bits is the number of bits starting at position to convert % The decimal equivalent is output in the parameter output output=0; % We begin by extracting from the total string the individual bits that are % needed portion = input_string(position:position+number_of_bits-1); % Since the string portion is of the form [1 0 1 1 0 ...], we need to % compact it so the spaces between binary digits is removed s = strrep(int2str(portion),' ',''); % And finally, this parts creates in output the actual decimal equivalent j=1; L=length(s); while(j<=L) output=output * 2 + bin2dec(s(j)); j=j+1; end end
36

Binary Chromosome Bits in Gene for Single Gene Example


In this problem, there is only one gene, and its values range from -1.5 to 6.75. Let s say we want a resolution of at least 0.01
That means there must be at least 8250 values in the range -1.5 to 6.75 6.75 (-1.5) = 8.25 For a resolution of at least 0.01 which is 1/100, we need at least 8.25(100) = 825 values.

37

Binary Chromosome Bits in Gene


What the preceding tells us is that we need a binary string that has at least 825 different values. Thus we need a binary string of at least 10 bits since 210 = 1024 > 825.
Remember that the actual decimal equivalent values of this string will be 0 to 1023

That means our conversion from a binary string to a float will be: dec_val=8.25(DecimalValueofString/1023)-1.5
38

Binary Gene Fitness Function


The fitness function for this example must
Convert the binary string to its floating point equivalent (this will be x in the function on slide 3) Substitute this value into the polynomial on slide 3 and evaluate the polynomial Then return this value as the fitness

39

Binary Chromosome
What if we want to maximize a fitness function f(x)?
Although Optimtool only finds a minimum, it can still be used.
Instead of returning f(x), return 1/f(x) Remember that if f(x) can be 0, you must not allow this to happen, e.g. 1/(1+f(x))

The next slide shows the fitness function for the single gene example.
40

function [ fitness ] = CalculateFitness(Bstring) % Takes a binary row vector (Bstring) as input % This string is 10 bits and hence 0 to 1023 in decimal % Converts the vector to a decimal number in the % range -1.5 to 6.75 6.75-(-1.5) = 8.25 [DecimalValue]=ConvertBinary(Bstring) DecimalValue=8.25*(DecimalValue/1023) - 1.5 fitness= DecimalValue^5-12.1*DecimalValue^4+40.59*... DecimalValue^3-17.015*DecimalValue^2-71.95*DecimalValue+35.88; end function [ output ] = ConvertBinary( input ) %Takes a binary row vector as input %Converts it to its decimal equivalent. output=0; s = strrep(int2str(input),' ',''); j=1; L=length(s); while(j<=L) output=output * 2 + bin2dec(s(j)); j=j+1; end end
41

Fitness Function
Note that the previous fitness function called the function ConvertBinary. The file in which CalculateFitness is stored is called CalculateFitness

42

Example Continuing
Now, as before, the first step is to select ga Genetic Algorithm in the Solver box

43

44

Optimtool Binary Chromosome


Next you must assign a fitness function and specify the number of variables.
In this case the name of the function is CalculateFitness (slide 41) and the number of variables is the number of bits in a chromosome which was calculated on slide 38 as 10 bits Note: don t forget the @ in front of the fitness function name

45

46

Optimtool Binary Chromosome


Because this is a binary chromosome, you don t need to specify any constraints. In the options segments specify:
Population type: Bit string Population size: 50 (as an example) Creation function: uniform Generations: 500 (as an example) Stall generations: 500 (as an example) Other options: default
47

48

Optimtool Binary Chromosome


Now you are ready to run the ga
Click on the Start button in the run solver segment

The result is shown as before on the next slide. Notice that at the bottom of the slide the final (optimum) point is given in decimal.

49

50

Result
From the previous slide, the actual (binary) chromosome value was displayed as: 1101111100. Thus the decimal value of x for the best fitness is: 11011111002 = 89210 8.25(892/1023)-1.5 = 5.69

51

Homework 4
All of what has been shown for the Binary chromosome works for homework 4 with one variation. Note that in this homework the fitness calculation will need to be different.
In the preceding example, the equation for the fitness function was already parameterized and you were looking for a single value of x
f ( x ) ! x 5  12.1x 4  40.59 x 3  17.01x 2  71.95 x  35.88
52

Homework 4
In Homework 4 you have multiple variables, i.e. AF and your function is not parameterized

Also, to calculate the fitness, you need the H4Data table

53

Homework 4
There are multiple ways of addressing this problem:
Embed H4Data as a declared array in your fitness function If H4Data is already a matrix in your Matlab work space, you can sort of pass H4Data directly to your fitness function
Rather than declaring a fitness function in the fitness function window as: @CalculateFitness

54

Homework 4
Rather than declaring a fitness function in the fitness function window as:
@CalculateFitness

Declare it as @(X)CalculateFitness(X,H4Data) The assumption is that H4Data already exists in the Matlab workspace You also now need to alter the CalculateFitness declaration to be CalculateFitness(X,H4Data)
55

Вам также может понравиться