[Octave-bug-tracker] [bug #63326] [octave forge] (ga) ga function output unbounded results

[Liang Tang]

Follow-up Comment #4, bug #63326 (project octave):

After I reported this bug, I spent more time to find out the available
options.  I am writing this note to document what I know now in case others
experience the same issue later.  

I always appreciate the packages kindly provided by Octave community.  The
only purpose of this note is to state my feedback/observations on ga package
with the intention that this note might help to make the package more useful
and benefit other people who share my needs. 

Optimization algorithms in general can be categorized into methods require
objective function derivatives and methods require only objective function.  


One characteristics of non-derivative optimization algorithm is its ability to
control numbers of function evaluations
through population size and number of iterations/generation.  For problems
with expensive function evaluation, this approach
may provide great relief. For my own purpose, lowing computation expense has
always been a main focus point when I use Octave for more intensive
computation.  Global optimization 
and neural network are two examples.     

To summarize, I have four main points: 
(1) Octave ga is capable of unconstrained optimization.  It appears to me that
the original author was aware of discrepancies of the ga syntax and the
functionalities it actually provides.   
(2) There are other open source ga package/toolbox available.  I eventually
chose to use the one below due to its functionalities and documentation. It is
for unconstrained optimization
http://uos-codem.github.io/GA-Toolbox/ 
(3) Octave sqp illustrates an excellent formulation for constructing
constrained optimization solution. I follow sqp's equation construction to
build up constraints/bounds.  
(4) Penalty method is probably the most popular method for ga or pso to handle
constraint violation. There are still many recent technical papers on this
subject, which implies that no method is found to work universally.  With very
small modifications, the penalty formulation that works for me is found at 
https://github.com/sdnchen/psomatlab  The conventional penalty form such as
augmented lagrangian method did not help.  Convergence checking for exiting
also speeds up the solution.  

Thanks, 


    _______________________________________________________

Reply to this item at:

  <https://savannah.gnu.org/bugs/?63326>

_______________________________________________
Message sent via Savannah
https://savannah.gnu.org/