parameter files

[Ginger Booth]

Thanks to Jan for neatly summarizing the Gecko approach!

I think it might be edifying, though, to mention how Drone and Gecko came
into being as two separate approaches to the "same" problem, because
they're not, quite.

Drone preceded Gecko's parameter file scheme (though not Gecko) by a
short while, both during the beta days of Swarm. Both the Drone crew and
myself were pushing for the Swarm team solve this instead of us. Given
the effort, I think all parties were willing to take a close-enough
solution and make it work, if only they could skip doing it themselves.
The Swarm crew had neither the time, nor probably the confidence that
any of us were sure what the right solution would look like.

Drone, at least in its incarnation when I first looked at it to borrow/
steal for Gecko, was set up to be a batch job runner for parameter
sweeps. Very useful, if you have a few parameters to explore, with
more or less reasonable ranges to sweep.

But that's not what I needed. I needed:

 a. A control file which completely specified the experiment to be
    run. The same control file, output from each run, must be good
    input to reproduce that same run. Legibility is good. But more
    importantly, it must be ONE: ONE control file, completely
    specifying the experiment, the statistics observations, the
    algorithmic parameter search, if any, the whole shebang. ONE.

 b. Even the kinds of agents that exist in a Gecko run are specified
     in the parameter file. It needs to make 'em, set parameters on
    'em, create observables on 'em, etc. There are a couple global
    objects, too. Those are also parameterized.

    So the "parameter file parser" (the protocol reader) is actually
    building the entire simulation tableau, and setting parameters,
    and setting what passes in Gecko for a "probe".

    I never did (probably never will) create any mechanism for
    building Gecko runs from a user interface. The only way to drive
         it is from the protocol file.

 c. Parameter sweeps are lovely in R^2 or R^3 space. Gecko, alas,
    may have 20-50 parameters. Any attempt at a sweep or global
    optimization approach to that merits a weak giggle. So the
    protocol marks which parameters are to be algorithmically tuned,
    and where to start (and optionally another parameter to vary as
    an offset from), and from there, may God go with you. No sweeps.
    The only God (tuning algorithm) Gecko has so far is an Evolution
    Strategy engine, fwiw. (Ingo Rechenberg, if you're interested--
    it's an approach that works well with real-valued R^n spaces if
    n >> 1.)

What Drone and Gecko both need, and do, is to specify:
 - the random seed, so the same experiment can be reproduced (with
   more observables, for instance)
 - a number of iterations, since they're stochastic processes and a
   single run doesn't say much.

(I hope none of the Drone crew feels misrepresented here. Y'all did a
great job, and were kind enough to do it in a more generic way! :)

Anyway, for what it's worth...

Best,
 Ginger



                  ==================================
   Swarm-Support is for discussion of the technical details of the day
   to day usage of Swarm.  For list administration needs (esp.
   [un]subscribing), please send a message to <address@hidden>
   with "help" in the body of the message.
                  ==================================