Last Update 2.12.2013
Download a demo collection of stochastic optimizers.
This collection has been designed by Ch. Hafner and J.
Fröhlich for the optimization in an N-dimensional, real parameter space.
Fortran 90 code, compiled with Compaq Visual Fortran 6.1. Optimizers and fitness
functions are contained in DLLs.
- The parameters are searched within the interval 0...1. For other search
intervals, the user should write appropriate transformations.
- Simple Evolutionary Strategy (SES) or Real Genetic Algorithm (RGA) are our
first choice in most situations
- Complicated Evolutionary Strategy may be more efficient than SES in
complicated situations, but this algorith is much more memory- and
- Particle Swarm Optimization (PSO) was worse than RGA for all our test
cases, except in very simple situations.
- SIMulated annealing (SIM) is similar to the most simple (1+1) SES. Since
we were unable to find good temperature functions, SIM was always
outperformed by (n,m) SES.
- Downhill SiMPlex (SMP) with random initialization and restart is very
quick when the fitness function is smooth enough and exhibits not too many
maxima. For typical engineering problems, the fitness function is often to
tricky for SMP.
- RAndom Search (RAS) would be optimal for random fitness functions. For
usual fitness functions, RAS is extremely slow.
For more information, download a Word document
that contains the transparencies used for PIERS'2000, Cambridge.
Lecture notes (in German, zipped
VGA a Visual GA written using Intel Visual Fortran,
compiled for Windows, without much information, i.e., you might have to read
source files to understand how it works... More info when I find time...