Hi Johanna,
another vote for Python. We are using a python script to start ECJ with different parameter files. Therefore all parameter files are first generated (by hand or automatically) and then copied to the same directory. The Python script looks in the directory and starts ECJ sequentially with all of the files. The results are stored in a directory named like the parameter file. This is easy and reliable and you can run large numbers of experiments over the weekend automatically.
Cheers, Stefan

On Wed, Apr 2, 2008 at 3:53 PM, Sean Luke <[log in to unmask]> wrote:
There are a couple of options you can do:

1. Use a scripting language to launch ECJ with different parameters (it accepts "-p param=value" on the command line, and that overrides any parameters defined in the parameter files).  You might try Python.

2. Run ECJ using its rudimentary jobs facility:

       java ec.Evolve -file my.params -p jobs=20000
       [eesh, that's a big number: 400x500]

The idea of the second option is this.  By default your EvolutionState object contains an instance variable called 'job'.  This is an auxiliary variable which can be anything, but in the default form, jobs[0] contains an Integer which stores the current job number.  Thus you could do something like this: before your EvolutionState.setup() method is called, you look up jobs[0] and based on the value stored in there you set various parameters in the ParameterDatabase as you like (just use state.parameters.set(_parameter_, _value_).  These programmatically set parameters override any parameters loaded from files or specified on the command line.  Then when EvolutionState.setup() gets called, it'll set up based on those parameters.  Thus each new job you can have your evolution use different "parameters" on a per-job basis.  Each of those jobs will have a different random seed too (they'll each differ by 1 -- of you can reseed the generator as you like prior to setup()).

So how do you tap into something before setup()?  A couple of ways:

       A. Make a subclass of EvolutionState and override the setup() method to do your stuff and THEN call super.setup()

       B. Make a subclass of EvolutionState and override the startFresh() method to do your stuff and THEN call super.startFresh()  (setup() is called from startFresh)

       C. [probably best]   Modify the Evolve class's main() method.  Note that there's a spot in the main() method, inside the jobs loop, where you can modify parameters in your EvolutionState based on the job number.  This spot is marked prior to state.run(), which in turn calls state.startFresh(), so you know you're modifying stuff before anything shows up.

3. Build your own looper inside your own Evolve.main() code.  This isn't nearly as hard as it sounds; in fact it may be the easiest of your three options.  In the Evolve class, I have (in comments) examples of how you make a very simple main() function rather than the elaborate ones used in Evolve proper.  You can just rewrite main() to repeatedly create EvolutionStates and run them.  What you could do is make a loop which creates EvolutionStates, initializes them (with a new random number seed offset each time -- perhaps your job number -- to make sure each time they have different random number generators), CHANGES THE PARAMETERS right then and there in the parameter database, and THEN runs them.

I'd do #1 if possible, then #2C if you're sick of fiddling with scripting languages.


On Apr 1, 2008, at 2:52 PM, Johanna Rosenbusch wrote:

Dear all,

I finished my coevolutionary algorithm and now I need to run many testruns
each with different parameter combinations. It's probably going to be around
500 different combinations, each one running for maybe 400 jobs.
I was thinking about using some kind of script language for windows. Since I
don't have any experience with scripting I wanted to ask if anyone has done
something similar?

Thanks and best regards,