Swarmfest 2000 Conference Program and Abstracts (preliminary)

[lee]

Swarmfest 2000 Conference Program and Abstracts
===============================================

Saturday, March 11, 2000

09:00-10:00 am  Workshop, and Conference Registration

10:00-05:00 pm  Swarm Workshop: A Hands-on Workshop for those already
                familiar with Swarm taught by Paul Box of Utah State 
                University

05:00-06:30 pm  Conference Registration

06:00-10:00 pm  Opening Reception at Logan Country Club



Sunday March 12, 2000 -- Technology

08:00-09:00 am  Conference Registration and Breakfast, 
                Eccles Conference Center

09:00-09:10 am  Introduction to Utah State University
                F. E. Busby, Dean of the College of Natural Resources
                Utah State University

09:10-09:20 am  Welcome and Introduction to the Swarm Development Group
                Irene Lee, Swarm Development Group

09:20-10:00 am  Life, The Universe, and Everything
                Glen E. Ropella, The Swarm Corporation

Abstract:
As computational resources become more ubiquitous, we must develop the 
ability to delegate responsibility to automata.  The current state of 
computational design prohibits this delegation because we just can't 
develop the trust that is a necessary prerequisite for such delegation.
Therefore, we need to examine "trust" in the context of our computation-
al artifacts.  Trust is built on expectations. Expectations are built 
by abduction from an often dynamic set of observables.  Simulation, 
particularly the kind used by agent-based modelers, provides a powerful 
context in which to experiment with the process by which expectations 
and trust are built.  And if we can develop this "estimation theory" 
beyond mathematical modeling, we should be able to advance the state-
of-the-art not only in simulation and modeling, but in parallel systems 
analysis and design, as well.


10:00-10:10 am  Break 

10:10-11:10 am  RePast: An Agent Based Simulation Framework in Java
                Nick Collier, Social Science Research Computing
                University of Chicago

Abstract:
RePast is a software framework for agent based simulation created by 
Social Science Research Computing at the University of Chicago. It 
provides a library of classes for creating, running, displaying, and 
collecting data from an agent based simulation. This talk will focus 
primarily on demonstrating RePast's features and capabilities. In 
addition, I will also briefly discuss RePast's internal architecture, 
design goals, and future direction.


11:10-11:20 am  Break 

11:20-12:20 pm  Ascape: Abstracting Complexity
                Miles Parker, Center on Social and Economic Dynamics
                Brookings Institution

Abstract:
Software tools used in science typically take a kitchen-sink approach 
to design. From statistics to mathematics to engineering to agent 
modeling, even those tools that have a strong organizing theme tend 
towards supporting every contingency and methodology. This impulse 
towards generalization and breadth is laudable and necessary. But there 
is a harmonious case to be made for the discipline of abstraction, 
parsimony, and depth, and that is the case I make for Ascape.

I discuss a few key abstractions enforced in Ascape, and the opportu-
nities they create for expressibility and simplicity. While these 
abstractions seem especially suited to the domain of social and econo-
mic systems, they are not limited to it. By demonstrating recent work 
in Ascape and building a simple Ascape model, I show how these appar-
ently constraining abstractions benefit the Ascape user and developer 
experience.


12:30-01:30 pm  Luncheon, Taggart Student Center Walnut Room

01:40-02:40 pm  Kenge: Swarm GIS-CA Libraries
                Paul Box, Geography, Utah State University

Abstract:
This is a presentation on Kenge, a library for using raster GIS data 
in multi-agent and cellular models.  Kenge is designed to facilitate 
incorporation of GIS data into landscape-level simulations. It is 
intended for simulations of scenarios where dynamic, heterogeneous 
landscapes are interacting with dynamic, heterogeneous populations 
within them. The intended use is simulations where the populations 
affect the landscape in which they live, the landscape affects the 
behavior of the population(s) in it, the individuals of the populations 
are affecting each others' behaviors, and the landscape is continually 
evolving by its own rules.  

In Kenge, the modeler provides rules at the level of the cell (raster), 
and landscape processes are simulated as emergent processes. Agent-based
models can be incorporated into the Kenge landscape, with agents 
"running around" on top of the cells.  As with the cells, the modeler 
gives rules to the individual agents, and observes their emergent 
collective behavior as the simulation progresses.  As with many agent-
based and CA simulations, simple rules generate complex landscape 
patterns.

The basic design of Kenge will be presented, as will a number of 
applications that have been developed around the libraries.


02:40-02:50 pm  Break 

02:50-03:50 pm  GeoGraphic Smallworlds: Agent Models on Graphs
                Catherine Dibble, University of California 
                Santa Barbara 

Abstract:
Structured geographical or organizational environments often mediate 
agent interactions in profound ways.   Both existing geographical or 
organizational systems and the agent-based simulation models that 
represent them may exhibit path-dependent co-evolution and multi-scale
feedback effects, which are difficult to examine except under laboratory
conditions.   Yet richly-structured landscapes for agent simulations 
have been difficult to develop, with many models still constrained to 
aspatial soups, isotropic planes, or at best to relational networks 
among individual agents, but not yet as landscapes on which hetero-
geneous mobile agents interact.   This talk introduces a new prototype 
for a general class of network-based landscapes for the Swarm simulation
platform. 

GeoGraphic Smallworlds have the advantage that landscapes are repre-
sented as formal graphs with realistic structures.   While they can 
represent isotropic planes as regular lattices, they are most useful 
when the landscapes are most naturally formalized as one or more 
interlocking parameterized families of irregular graphs.   Separate 
landscape and agent random number seeds allow us to run many agent 
simulations on any given GeoGraphic structure.   Similarly, we can 
generate many distinct families of GeoGraphic landscapes that differ 
in their particular structural details yet share common graph charac-
teristics that are relevant to the behavior of the model.   Richer 
simulation landscapes provide controlled environments in which to 
build and test formal models grounded in explicit spatial structures, 
diverse distributed mobile agents, and context-specific behavior. 


03:50-04:00 pm  Break 
 
04:00-05:00 pm  Execution & Performance of Swarm Models: How to write 
                fast models in Swarm.
                Marcus Daniels, Swarm Development Group

Abstract:
Agent-based models are by nature computationally intensive. Finding 
interesting behaviors in a large space can involve parameter sweeps 
over numerous variables and these variables may drive the agents in a 
model for hundreds of thousands of steps.  Many questions are costly 
to ask and there is no getting around it.

Choosing programming techniques that fully-exploit available computer 
resources can make a real difference in determining whether or not a 
given agent-based model's dynamics can be considered to an adequate 
extent.

In my talk, I'll trace through the process of improving the performance
of a Swarm model.  Topics that will be covered include:

  o Constructs to avoid and their algorithmic cost

  o Advanced fine-grained profiling techniques with the GNU C Library

  o Java profiling techniques

  o Demo of ahead-of-time native code generation: 
     The GCC Java frontend and Kaffe

  o Common bottlenecks for Java and Objective C Swarm models

  o Cache behavior

  o Technologies (within Swarm and external) for facilitating parallel
     parameter sweeps, and a demo


06:00-08:00 pm  Keynote Address: Chris Langton, 
                Welcome Dinner and Entertainment, 
                Zanavoo Restaurant in Logan Canyon


Monday. March 13, 2000 -- Applications

08:00-09:00 am  Breakfast, Eccles Conference Center

09:00-10:00 am  Modeling Smolt Migration with Swarm
                Jim Anderson, School of Fisheries
                University of Washington

Abstract:
SWARM was used to model the movement and interactions of salmon smolts 
and their predators under reservoir and natural river conditions. We 
first modeled smolt migration through the existing 40-km reservoir 
behind Little Goose Dam on the Lower Snake River.  Output from a 2-
dimensional hydraulics model was used to define detailed flow fields 
for both the existing reservoir and a free flowing river.  Smolts 
movement was described by flow and random movements that decrease when 
smolts were in their preferred depth habitat.  Predator movements were 
random and independent of the flow.  Smolt migration was characterized 
by adjusting their random movement to fit observed smolt travel times 
through the reservoir and the predator-prey dynamics were characterized 
by adjusting the predator-prey reactive distance to fit the observed 
smolt survival.  The impacts of dam breaching were then explored by 
using the free flowing bathymetry and hydraulics with the smolt migra-
tion and predator-prey factors fit for the reservoir.  In this manner, 
we explored the effect of changing the environment without changing the 
behavior. From this state, we then explored how changes in behavior 
affected smolt survival. In developing this model we learned that a 
realistic formulation of the boundary processes, that is the river edge,
was critical to obtaining meaningful results.  We also found our inter-
pretation of the model was best guided by an analytical solution of the 
migration process. This analytical model identified how to view and 
interpret the output from the SWARM.


10:00-10:10 am  Break 

10:10-11:10 am  Getting "Results": The Pattern-oriented Approach to 
                Analyzing Complex Systems with Agent-based Models
                Steve Railsback, Lang, Railsback & Associates

Abstract:
A critical question commonly heard by Swarm modelers is whether agent-
based models can produce "results"- general concepts instead of just 
noisy stochastic simulation output. This question results largely from 
the very questionable assumption that conventional differential equation
models are more general because they use aggregated measures of the 
population being modeled, but this issue still must be addressed by 
agent-based modelers. Pattern-oriented modeling is a framework for 
ecological analysis but is widely applicable to complex systems. This 
approaach involves designing a mode specifically to simulate observed 
patterns of system-level behavior, and testing the model by whether it 
can reproduce those patterns. Alternative formulations for agent 
behavior can be posited as hypotheses and tested by whether they 
reproduce observed system-level response patterns. This provides a 
hypothesis-testing approach that allows inferences about agent behavior.
We applied this approach to a model of how fish select habitat in a 
stream. The model was designed to predict how individual fish select 
among alternative habitats that vary in food intake and mortality risks.
>From the fish literature, we a priori identified six patterns of 
observed habitat shifts in response to known stimuli, then tested the 
model to see if these patterns emerged from individual behaviors. Three 
alternative rules for making habitat choices were compared, and two of 
the rules were rejected as unable to reproduce the observed patterns. 
This analysis produced "results": we showed that some assumptions 
commonly made by ecologists are incapable of explaining population-
level patterns.


11:10-11:20 am  Break 

11:20-12:20 pm  Lifecycle Structures: Using Life History to Express 
                Shared Models 
                Roger Burkhart, VantagePoint Network

Abstract:
The lifecycle model is a set of ideas for building and sharing con-
ceptual models.  Such models are expressed at a logical level of 
abstraction that is independent of any concrete computational 
implementation.  They may express any range of selected content, 
from a passive record of external input to the autonomous action of 
a living structure.  A conceptual model of lifecycle structure starts 
with a formal, logical core to capture the content of some selected 
domain, but this domain includes the process by which the model itself 
is defined and detailed.  Every model consists of a progressive 
elaboration from an initial controlling structure through an ongoing 
selection out of a remaining potential.  Internal structures can 
include an executable representation of distributed, concurrent 
activity like that of the Swarm simulation system.  These internal 
plans can add a variety of forms of processing as inherent working 
parts of any model, from direct response to external events to con-
struction and management of internal models.  Any model can incorporate 
procedural or declarative levels of cognitive interpretation and 
response over some real or hypothetical domain.  A framework for 
accessing, sharing and recombining any fragment of a lifecycle process 
from any other provides an underlying support for interaction and a 
short-cut for development of any specific individual.  A set of 
conventions for a standardized form of lifecycle structure is itself 
published as a particular example of the generic structure.  This 
initial structure encodes the essential features of the lifecycle 
model and provides a starter set of universal elements out of which 
to build any other model.


12:30-01:30 pm  Lunch on your own

01:40-02:40 pm  Opinion Formation and the Resilience of Diversity
                Paul Johnson, Political Science, University of Kansas

This paper explores agent-based models of the formation of individual 
opinions and collective judgments. Models of information transmission 
between individuals through political communication are contrasted with 
models in which individuals adjust their opinions through observations 
of social aggregates. Several variants of the culture model originally 
proposed by Robert Axelrod are offered. All models are built with the 
Swarm Simulation System, an open source toolkit written in Objective-C.


02:40-02:50 pm  Break 

02:50-03:50 pm  Stochastic Birth/Death Models as a Bridge Between 
                Differential Equation Models and Individual Based Models
                Doug Donalson, Ecology, Evolution and Marine Biology
                University of California, Santa Barbara

Abstract:
Cell-based and discrete (synchronous) time models by definition intro-
duce quantization errors into simulations. In the case of cell-based 
space, individuals are required to reside in the center of a cell. 
This requires that individuals or subpopulations must be separated by 
a minimum distance. (In population ecology this might represent model 
imposed competition). In the case of synchronous time models, if the 
time step is longer than rates of change of the state variables, esti-
mates must be calculated for the number of state changes for each state 
variable in the interval. The maximum number of state changes in the 
interval and the ordering of those changes can have a strong effect on 
the system dynamics.

I will talk about an alternate formulation for IBM's that allows for a 
continuous time and continuous space representation of the dynamics. 
Asynchronous schedules (also called event driven schedules) are based 
on the idea that no two events (state changes) happen simultaneously. 
It should therefore be possible to order events in such a way that there
is no ambiguity. I will provide a handout on asynchronous schedules with
my talk that is part of a chapter in my dissertation, comments are 
encouraged. (Please, beat me up before my committee does!) 

The use of asynchronous schedules allows development of the Stochastic 
Birth-Death (SBD) model. When trying to link equation based models to 
IBMs, there are a number of factors changed in the dynamics. One of 
these is the change from a continuous (often density based) represent-
ation of model dynamics to one where state changes are discrete. For 
instance, in a density based model, a population can change from 100 
to 100.01, in an individual representation, the change (increase) must 
be to 101. An SBD model can be used two ways. First, if the results of 
an IBM are significantly different than its equation based counterpart,
an SBD model can factor out the changes resulting from demographic 
stochasticity (requiring individuals to reproduce in whole units) from 
other factors in the IBM such as space and behavior. SBD's can also be 
used to analyze equation based models for sensitivity to demographic 
stochasticity. 

Finally, and if there is time, I will introduce the some tricks I use 
in my latest model to allow a continuous space representation of a 
mussel bed while still using Swarm's great graphic potential. They 
also allow significant speed up of global search procedures.


03:50-04:00 pm  Break

 
04:00-05:00 pm  Discussion Group on Managing Experiments 

                Paul Johnson:   Drone demo  
                Steve Jackson:  Scenario/replicate infrastructure
                Marcus Daniels: Argument parsing,
                       loading parameters with the Lisp Archivers, and
                       using HDF5 output in R for analysis

05:00 pm        Conference Adjourns (dinner on your own)



Tuesday, March 14, 2000 

10:00-05:00 pm  Kenge (Swarm/GIS) Hands-on Tutorial, taught by 
                Paul Box, Utah State University

12:00-01:00 pm  Lunch on your own


--