Mathematics VS? Agent Based Modeling

[Parks, Raymond]

Robert Holmes wrote:

> However Parunak makes a stronger claim: that ABMs can be used for
> *automatic* closed-loop control: "...a control scheme that more or less
> automatically regulates the behavior of the entire system." In other words,
> we can use ABMs where the feedback loop is not mediated by a human. And
> (IMHO) this is a big and unsupported claim. So to rephrase my question, are
> there any real-world examples of using an ABM in a *automatic* closed-loop
> control system?

   I don't know of any such systems.

> Now if I'm using equation-based models, I've a fair chance of
> doing that. But if I'm using agent-based models which might (or might not)
> show sensitivity to intial conditions, which might (or might not) be
> stochastic and give multi-valued solutions, what do I do? As Owen once
> succinctly put it, how do I differentiate an ABM?

  This is where my naivete with agent-based modeling will show, but, why not run the model over a range of conditions?  As a long-time software tester, I'm used to the idea that a system can have unexpected results when subjected to unexpected inputs and environment.  In the software testing world, I've long since given up testing all the possible paths through the software.  Instead, I do what amounts to a sensitivity analysis, looking for inputs and environment that causes change in output.  Once I narrow down the range of possible testing scenarios, I filter again for interesting output - if the change is great but not interesting, there's no use in testing it.  ABMs seem like a worse case of that software testing paradigm.

  As an aside, I find it fascinating that "fuzzing" has become a mainstay of the security assessment world.  It's another example of everything old is new again.  Now, security testers "fuzz" the packets they send to a network service.  Back when I tested satellite ground systems, I "fuzzed" the user interface by sweeping my hands across the keyboard.

> And this is the reason I think Parunak's claim is too strong: if you try to
> design an automatic closed-loop control system with an ABM at its heart, you
> don't have access to all the tools and methodologies that enable you to
> assess the controller's stability and robustness. And if the designer can't
> assess a controller's stability and robustness s/he is unlikely to implement
> it in a real-world application.

  I would agree with you completely on this point except that I know there are some critical infrastructure applications that can now be controlled but the control equations can't be developed.  The owners of these critical infrastructures are starting to look at agent-based solutions (not necessarily models).

Ray Parks                   [hidden email]
IDART Project Lead          Voice:505-844-4024
IORTA Department            Fax:[phone]
http://www.sandia.gov/idart Pager:800-690-5288


-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/ms-tnef
Size: 4097 bytes
Desc: not available
Url : /pipermail/friam_redfish.com/attachments/20050316/41ed41dd/attachment.bin

[Robert Holmes]

Raymond Parks writes:

>   This is where my naivete with agent-based modeling will
> show, but, why not run the model over a range of conditions?  

Good point Ray, but given the number of agents in a model you quickly get
combinatorial explosion. Let's say you've got 100 agents and they each only
take two values. That's 2^100 combinations to work through if you want to do
a full sweep of parameter space (and to put this into context the universe
is 2^98 picoseconds old).

One solution is to use 'design of experiments' techniques to cut down
(hugely) on the number of runs without sacrificing too much in the quality
of conclusions that you can draw. However there don't seem to be many in the
ABM community doing this - one of the few papers that uses it is the one by
Steve, Mike & myself (Steve - what's the link?) from last year.

Indeed this points the way to how you might use ABMs in automatic
closed-loop control: use DOE to create some simple parametric representation
of the ABM and put that in the controller. Of course what you've then done
is represent an ABM as an EBM-M (equation based meta-model) :-)

Robert