ec.multiobjective.nsga2.README Maven / Gradle / Ivy

Go to download

Show more of this group Show more artifacts with this name
Show all versions of ecj Show documentation

ECJ, A Java-based Evolutionary Computation Research System. ECJ is a research EC system written in Java. It was designed to be highly flexible, with nearly all classes (and all of their settings) dynamically determined at runtime by a user-provided parameter file. All structures in the system are arranged to be easily modifiable. Even so, the system was designed with an eye toward efficiency. ECJ is developed at George Mason University's ECLab Evolutionary Computation Laboratory. The software has nothing to do with its initials' namesake, Evolutionary Computation Journal. ECJ's sister project is MASON, a multi-agent simulation system which dovetails with ECJ nicely.

The newest version!

This package contains an implementation of the Non-Dominated Sorting 
Genetic Algorithm 2 (NSGA-II).

Details of this approach can be found in the following paper:

Kalyanmoy Deb, Amrit Pratap, Sameer Agarwal, and T. Meyarivan. A Fast and 
Elitist Multiobjective Genetic Algorithm: NSGA-II. In IEEE Transactions 
On Evolutionary Computation.  6(2).  2002.

The ec.app.moosuite package contains common multiobjective test cases
to test NSGA-II against.



How ECJ implements NSGA-II
--------------------------

NSGA-II is an "archive" elitist method which uses Pareto front ranks 
instead of composite fitness, and sparsity along the front rank when 
there is not enough remaining room to fill the archive.

We extend SimpleBreeder merely to build the next-generation population 
by breeding children from the previous population (the archive) and then
appending the current archive to the newly-generated children.  This 
is essentially a version of (mu+mu).

Most of the actual work is done in a special version of SimpleEvaluator,
where the Individuals are evaluated, then broken into Pareto Front Ranks,
then loaded into the archive.  When there's not enough room left in the
archive for another full rank, sparsity is used to decide who gets to be
included.  The archive then forms the next-generation population.

NSGA-II has fixed archive size (the population size), and so ignores the 'elites'
declaration.  However it will adhere to the 'reevaluate-elites' parameter
to determine whether to force fitness reevaluation.

SPEA2 and NSGA-II have different approaches to building archives which impacts
on how you need to set your population sizes in order to achieve the same number
of evaluations.  In ECJ, NSGA-II does not include the archive as part of the basic 
population size.  Rather, it builds the archive separately, then builds the
population by breeding from the archive, then gloms the two together.  On the
other hand, SPEA2 uses a (tunable) portion of its population as the archive, and
breeds individuals into the remainder of the population.  This means that to have
a "population size" for SPEA2 that's the same as NSGA-II, you need to increase
SPEA2's population size to NSGA-II's population size plus SPEA2's archive (elites)
size.  For example, if you have an NSGA-II population size of 100, and SPEA2 is using
an archive size of 50, to be fair you should make SPEA2's population size be 150.

If you set things so that SPEA2 and NSGA-II must reevaluate the fitness of 
their archives (which is rare), things are different.  Now you should set SPEA2
so that its archive size and population size is equal to twice NSGA-II's population
size (because NSGA-II's archive size is the size of its population).

The classes in question:


ec.multiobjective.nsga2.NSGA2Breeder

The SimpleBreeder subclass.


ec.multiobjective.nsga2.NSGA2Evaluator

The SimpleEvaluator subclass.


ec.multiobjective.nsga2.NSGA2MultiObjectiveFitness

A special subclass of MultiObjectiveFitness which adds auxillary fitness 
measures special to NSGA-II (notably rank and sparsity).