Lesson 2 -
Tutorial
main page -
Algorithm-Based - Component-Based
- Programming hints -EO
documentation
Tutorial: Lesson 1
This lesson will let you
-
run your first Evolutionary Algorithm written within
EO Library, choosing between evolving bitstrings,
or evolving vectors of real numbers,
-
browse through the code of these algorithms, or
-
follow the guided tour.
Later you will be asked to
-
write your own fitness function,
-
check that EO let you separate the representation
from the evolution engine.
-
use different kinds of selection
procedures in the framework of a generational GA evolution
engine,
I want to run
an Evolutionary Algorithm
now
You can choose to run a standard bitstring Genetic
Algorithm (as defined in Goldberg's book) or a standard real-valued
genetic algorithm, as proposed in Micahlewicz's book.
If you have not already done what was recommended in the Tutorial
main page , do it NOW. Then go
to the Lesson1 sub-dir of the tutorial dir, and simply type at the system
prompt
(myname@myhost) EOdir/Tutorial/Lesson1 % FirstRealGA
or
(myname@myhost) EOdir/Tutorial/Lesson1 % FirstBitGA
and something should happen.
What is happening?
At the moment, the FirstBitGA maximizes the
number of ones in the bitstring (also calls the OneMaxfunction,
whose solution is, as you can guess,
11111111),
and the FirstRealGA is maximizing (the default
action for EO is to maximize) the inverse of the sum (i.e. minimizing the
sum) of the square of its variables (also called the sphere
function, whose solution is ... all zeroes).
And what you see on the screen when running one of these two programs
is, in each case, the initial and final population of an Evolutionary run,
one individual per line, its fitness first, then the number of items (bits
or real numbers) of the genotype, and the genotype itself. The final population
hopefully contains the solution in the discrete case, and is close to it
in the continuous case.
Browsing
the code:
Now you need to take a look at the program codes, either by browsing
alone through the sources for FirstBitGA
and FirstRealGA, or by following the guided
tour below. You might prefer to go directly to the exercises.
Guided tour:
-
General includes:Like
all C-like code, the file starts with include directives (Bit
- Real). Apart from standard includes,
the specific file to include is eo:
this is a file that contains the list of the all important representation-independent
EO files.
-
Representation:
you then have to declare the type of individuals you will be handling.
All evolution-related objects you will need are templatized w.r.t. the
type of individuals.
-
Fitness function:
the code for the fitness function is included in the file. It must take
as argument a reference to an individual (at the moment).
-
Bit This function simply computes
the number of ones of the bitstring (it's called the OneMax function).
The optimum is of course the all-ones bitstring.
-
Real This function simply computes
the inverse of the sum of the squares of all variables (also called the
sphere function). The optimum is of course the all-zeroes vector.
-
Parameters:
all parameters of the algorithm are declared here (Bit
- Real), and their values and
assigned. Of course, this means that you will need to recompile to change
these values - see Lesson 3 to get rid of that heavy requirement.
-
Random seeding:
Random numbers play an important role in Evolutionary Algorithms. See in
EO
programming hints more details about how this is simulated in EO -
but as far as you are concerned now, remember that the global
Random Number Generator is called rng
and should be used everywhere you need a realization of a random variable
of known law. Moreover, this RNG requires a seed,
which is set here (Bit - Real):
every time you run the algorithm with the same
seed, you will get the same
result. Hence, to test the robustness of your
algorithm, you should run it with different seeds. This is rather time
consuming in the present programs, so we suggest that you wait until Lesson
3 to do so.
-
Fitness function encapsulation: EO
is based on the notion of functors
- hence you now need to encapsulate your fitness function into a functor
object. This is what is done here (Bit
- Real).
-
Initialization:
to initialize the population, first declare an empty object of class eoPop<Indi>,
which is basically an STL vector<Indi>,
then fill it with Indi's. And remember that
v.push_back
simply appends its argument at the end of STL
vector v.
-
Output: take
a snapshot at the initial population (Bit
- Real). Sort it first, so the best
individuals are first, and display it. Note that an eoPop has a <<
method, which means that a simple os
<< pop streams the pop
onto the ostream os.
This is true for all objects of of class eoPrintable
(most EO objects) through the method printOn
(which is then called by the <<
operator).
-
Evolution engine:
The selection/replacement mechanism (Bit
- Real) is a simple generational
GA here: a simple selector, and a generational replacement. The eoDetTournamentSelect
has been chosen as a robust selection, and the generational replacement
(all parents are replaced by the offspring) is hard-coded in the eoSGA
algorithm.
-
Variation operators:
in the simple algorithm considered here, individuals undergo crossover
and mutation.
In EO, these operators are (functor)
objects of class eoQuadOp
(binary operator that modifies both its arguments) and eoMonOp
(unary operator). These operators are applied in turn to all selected
parents, according to user-defined probabilities. These probabilities
are defined with all other parameters, and will
be passed to the eoSGA algorithm.
For more details on these classes, go to the algorithm-based
corresponding pages, or to their respective documentation pages.
-
Bit The crossover eo1PtBitXover
is the standard 1-point crossover, and eoBitMutation
is the standard bit-flip mutation that randomly
flips all bits with a given probability P_MUT_PER_BIT.
Warning: the P_MUT_PER_BIT
probability is an internal parameter of the
eoBinMutation,
it is NOT the probability of mutation
at the individual level. EO corrects what can be viewed as an inconsistency
in Holland's original work, further used in Goldberg's book by separating
the probability of mutation for each individual (independent of the type
of mutation that will be applied) from the probability of flipping each
bit, which is specific of the bit-flip mutation. Hence, to run the
same algorithm as Goldberg's SGA, the mutation probability (at individual
level) is 1, and the probability of flipping each bit is P_MUT_PER_BIT.
-
Real The crossover eoSegmentCrossover
is the standard segment crossover for real-valued
vectors, that chooses a point randomly on the segment between both parents
(also termed BLX-0). eoUniformMutation
is the uniform mutation for real-valued vectors
that chooses a new value for each variable uniformly on an interval centered
on the parent value. The width of the interval is an internal
parameter of the object, here called EPSILON.
-
Stopping criterion:
Specify a maximum number of generations
to run (Bit - Real):
the simplest of all stopping criteria at the moment, using an object of
a sub-class of class eoContinue.
-
The algorithm: the
simple algorithm that is used here, called eoSGA
requires
as parameters a selector,
a crossover and
the associated crossover rate,
a mutation and
the associated mutation rate,
and a stopping criterion.
Take a look at the corresponding
constructor
of the class eoSGA:
it only initializes its private data
with the parameters. Now look at the operator()
method - the one that is called in the code for FirstBitGA
or FirstRealGA - and you'll find
out that is is as simple as it sounds.
-
Output: After
running the algorithm, output the sorted final population (Bit
- Real) - and look at the best
individual: this is the result of the algorithm.
-
Main body: for
technical reasons (intercepting the exceptions), we need a main like this
one (Bit - Real).,
and you should not touch it unless you know what you are doing. Simply
note that this main calls the function main_function, which we have been
discussing up to now!
Exercise
1: maximize your own function
This is very easy - if your search space is that of bitstring or of unbounded
real numbers.
-
Go to the tutorial directory, and copy the
program you want to modify onto mytest.cpp.
-
Edit mytest.cpp
with any text editor:
-
Modify the fitness function itself (binary_value
in FirstBitGA,real_value
in FirstRealGA)
-
Compile the program by typing make
mytest at system prompt
-
Run the new program by entering the command
mytest
at system prompt.
Exercise
2: check the differences between both programs
Go and take a look at the code for these programs (Bit
- Real). Use the symbolic representation
of an Evolutionary Algorithm (you should understand that figure now, otherwise
go there and come back) to understand how each
part of the EA is coded. Try to spot the differences between both codes:
there are not so many!
After you've tried that alone, take a look at the solution
:-)
Exercise
3: change the selection procedure
This is rather straightforward ... if you know what other types of selection
are available!
At the moment, let's only consider only the following simple ones:
-
You already know the tournament selection
Syntax: eoDetTournamentSelect<Indi>
select(T_SIZE); // T_SIZE in [2,POP_SIZE)
-
Try the well-known roulette wheel
Syntax: eoProportionalSelect<Indi>
select;
-
Or the stochastic binary tournament
Syntax: eoStochTournamentSelect<Indi>
select(RATE);
// RATE in ]0.5,1]
-
and of course the random selection should
give bad results!
Syntax: eoRandomSelect<Indi>
select;
Note that all these classes of eoObjects are derived from the abstract
class
eoSelectOne.
To find out exactly how each procedure selects the individuals, read the
corresponding component-based page.
Lessons learned:
-
in EO, all actions are performed by functor
objects (this section is the last time in this tutorial that there
is a direct link to the EO Programming hints
page - though the link at top and bottom of all pages will remain there).
-
in EO, all object you will usually need to manipulate are templatized
w.r.t. the type of the individual you are handling.
-
The type of the individual is itself templatized
w.r.t. the type of fitness (double by default).
-
In EO (actually, in EC!) initialization and variation
operators are representation-dependent, while
the evolution engine is representation-independent
(well, like any rule, this one does have some exceptions).
-
Changing the fitness function, or the selection
procedure inside the generational GA evolution engine is straightforward.
-
remember, all solutions to exercises are in
the same sub-dir of dir Tutorial than the lesson itself (see here).
Lesson 2 -
Tutorial
main page -
Algorithm-Based - Component-Based
- Programming hints - EO
documentation
Marc Schoenauer
Last
modified: Fri Nov 3 18:49:12 CET 2000