RE: Simulating Individual Behavior

[Steve Emsley]

----------
From:   Stephen C. Upton[SMTP:address@hidden
Sent:   Friday, April 18, 1997 05:04
To:     address@hidden
Subject:        Re: Simulating Individual Behavior

This is a great thread!!!

Steve Upton's problem domain seems well suited to a hierarchical swarm
approach. In my domain each bucket of seawater can contain more particles
than my computer has bytes of RAM and there is no obvious aggregation
rule. However, it seems inherent in military modelling that there is a
hierarchical structure and a chain of command.

So, the agents are: soldiers -> platoon -> company -> regiment ->
division -> army corresponding to a chain of command: lieutenant -> 
captain -> major -> colonel -> general. The lieutenant polls his grunts and
passes the data up the line to the captain. The captain reports his synopsis
of the data from the lieutenants to the major and so on to the general who,
far from the conflict, passes directives down the chain of command. 
(Apologies for my ignorance of military organisation, whether UK or US!).

"How much detail is required?  (More detail, More detail is the current
rallying cry)"

My point is: As far as the general is concerned there are no individual 
soldiers.
However, individual soldiers could be modelled - but only a few, and less 
(than actually involved in a conflict) platoons, and even less regiments etc. 
Whereas modelling a general would require a context-dependent rule-based
system I should imagine that, the furthur down the chain of command, rules
become more fuzzy until, at the level of the individual platoon, one may as
well use stochastic differential equations. Of course, one might argue that 
engagements are won or lost not because of the 'normal' distribution
but due to the effect of the few on the many. Since heroism is probably as
elusive an AI concept as creativity I imagine that the aim of modelling warfare
is more an exercise in risk assessment than understanding.

Anyway, the main point of this posting is:

(1) SWARM has been designed with hierarchicies of subswarms in mind
although, apart from a few posting, this feature isn't being exploited. Perhaps
the computational overheads of running individual swarms has been a
disincentive. If that is the case, roll on parallel swarm!

(2) Just because SWARM is suited for ABM does not preclude the inclusion
of ODEs. Despite my last post, which may have suggested that I was
denigrating mean-field modelling in respect of ABM, I use ODEs in my
model. My ModelSwarm sends a step message to a ModelState class that
polls my LightSpace, OceanSpace and PhytoPlankton objects for variables and
Runge-Kutta's them for 24 hours to produce the next day's state. More of an
Infinite State Machine than a FSM! 

(3) Assuming hierarchical order then a "surrogate experimental system" is 
a WELL-DEFINED "descriptive model" of the level above.  In addition, the 
 "surrogate system" requires a  "descriptive model" of the level 
below to satisfy the criteria of testability with real world data. Before
disappearing down the hole with diameter = Planck's length the hopeful aim of
the exercise is to define the ranges of the parameters used in the level above
... if only to prevent mean-field modellers from tuning their systems to their
heart's conten... Sorry, [diatribeMode: OFF].

Excuse the musings of a lapsed Type 4 lurker but I couldn't resist this thread -
hail patch dynamics, there's noting like an invasion of territory (ie the
mailing-list) to bring out a spurt of evolution. Maybe it was an experiment!?

Steve Emsley,
Ecosystems Analysis & Management Group,
University of Warwick, UK.

<<application/ms-tnef>>