commons-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Phil Steitz" <>
Subject [math] Genetic Algorithms
Date Sun, 30 Mar 2008 00:53:47 GMT
I would like to start work on a GA framework.  This has been on the
WishList for some time and I now have a TLS type problem that I am
working on that is enough motivation for me personally to get things
rolling.  Here are some rough ideas on what I have in mind.  As
always, feedback / suggestions / patches appreciated.

0) Scope - a framework for implementing genetic algorithms, as
described in e.g. [1], [2].
1) Usage / what the framework provides - user classes should implement
minimally encoding and objective functions.  Stock mutation, crossover
and selection functions operating on binary encodings should be
provided by the framework, as well as execution of the algorithm, but
user classes should be able to plug in custom encoding, mutation,
crossover, and selection.  Objective and other functions should also
be allowed to have parameters, which the framework will somehow gather
and pass in to them.
2) Other - the framework should not require that populations be stored
in memory and it should support parallelization of population
processing functions
3) Package - o.a.c.math.genetic  (I am open to putting this inside
optimization, but think flatter might be better here).
4) Structure straw man:
Chromosomes - Chromosome interface includes mutate(),
crossover(Chromosome) and fitness().  AbstractBinaryChromosome
includes stock implementations of mutate and crossover for binary
Populations - Population interface provides a Chromosome Iterator,
select(long), add(Chromosome), delete(Chromosome).  Abstract classes
provide Roulette, Tournament, other select implementations for
Populations backed / not backed by in memory collections and I/O /
storage management.
GeneticAlgorithm - implements GA (following "canonical algorithm" as
described in [2]) given initial Population and configured sampling
rate and stopping criteria (maximum iterations, convergence criteria).
 Population and Chromosome classes determine the heuristics.  Should
(eventually at least) support parallel execution by worker threads of
breeding operations.

This may be naive and I may find that out as I start playing with real
code.  As I said, feedback / patches welcome. One thing I don't have a
clean idea for is how to make optional parameters available to all
methods mentioned above.



To unsubscribe, e-mail:
For additional commands, e-mail:

View raw message