You can use genetic algorithms, using individuals that are a vector of integers. You create a template settings files and to evaluate each individual you instantiate that template using the parameter values from your individual and get the results of the evolution, lets say, from a statistics file. I think you can directly manipulate the parameter database programatically, but I always used the settings file, so I don't know how.

On Wed, Oct 28, 2009 at 11:08 AM, Claes Gyllenswärd <[log in to unmask]> wrote:
Would it be reasonably complicated and/or meaningful, to get ecj to evolve some sub-set of the ecj settings?

I'm working on a project, where I'd like to determine what effects different settings have on the outcome. Settings such as population sizes,
mutation frequencies, migration and such. Instead of manually setting up a thousand runs with different settings, I'm wondering if ECJ could
be tasked with handling this?

The fitness for the meta-ecj could be calculated from for example the average time it takes the "normal" ecj to reach a certain fitness level.

The main question in my mind right now is how one would express a settings-file in a evolvable fashion.

Any comments, pratical or theoretical, would be interesting.