-
Ellen Goldberg mentioned the frequently asked question about a
definition of "complexity". She emphasized that instead of having
a generally accepted, rigorous definition it is more adequate for
this young and evolving field of research to agree on some basic
conditions for a system to be complex. For instance that a complex
system consists of simple, interacting subsystems that lead to
new, emergent properties. She gave an overview of work at SFI and
described two ongoing and upcoming projects at SFI: Stefanie
Forest and co-workers were able to design a computer virus
detection program that is structured like the biological immune
system (see video
clip). The system scans bit patterns and establishes rules to
discriminate between "self" (bit strings that are supposed to be
on the computer) and "non-self" (bit strings that are candidates
for computer viruses).
In a new project on networks the Santa Fe Institute will
install a "Beowulf" network computer and study in an
interdisciplinary group different aspects of the complexity of
networks in a number of different manifestations. This project
will be especially important in a world where the Internet becomes
a dominant factor in many areas of public and economic life.
WOULD-BE WORLDS, The Science and the Surprise of Artificial Worlds, John Casti, SFI External Faculty 
Abstract: By their very nature, complex systems resist
analysis by decomposition. It is just not possible to study, say,
the human immune system or a stock market, by breaking it up into
individual parts---molecules or traders---and looking at what
these parts do in isolation. The very essence of the system lies
in the interaction among all its parts, with the overall behavior
of the system emerging from these interactions. So by throwing
away the interactions, one also throws away any hope of actually
understanding the workings of the system. The problem is that
until very recently, there was no way of studying these sorts of
systems as complete entities, since to do experiments with stock
markets, immune systems, rainforest ecosystems and the like was
either too expensive, too dangerous or just plain too difficult.
But the arrival of cheap, powerful, widespread computing
capability over the past decade or so has changed the situation
entirely.
This talk will examine the way in which the ability to create
surrogate versions of real complex systems inside our computing
machines changes the way we do science. In particular, emphasis
will be laid upon the idea that these so-called ``artificial
worlds'' play the role of laboratories for complex systems,
laboratories that are completely analogous to the more familiar
laboratories that have been used by physicists, biologists and
chemists for centuries to understand the workings of matter. But
these are laboratories in which we explore the informational
rather than the material structure of systems. And since the
ability to do controlled, repeatable experiments is a necessary
precondition to the creation of a scientific theory of anything,
the argument will be made that for perhaps the first time in
history, we are now in a position to realistically think about the
creation of a theory of complex systems.
These essentially philosophical points will be illustrated by
on-going work with the world's catastrophe insurance industry, as
well as with a supermarket simulator done for the British chain J.
Sainsburys.
Editor's Comment: Casti gave his own conditions for a
system to be complex: It should be "medium-sized" that is
consisting of more than two and less than 1023 parts or
agents. The agents should be intelligent in the sense that they
follow rules (for instance football players are intelligent
according to this definition). The agents behave according to
local information and no agent knows what all other agents are
doing. Casti's first example was that of a football simulator (see
video
clip). The program has as input the characteristics of real
players and allows simulating games between different teams. As he
discusses in one of his books even if the simulator is very good
it will only give the correct prediction of averaged game results.
The result of individual games can be quite different from the
average outcome.
Understanding Size and Scale in Biology from Molecules and Cells to Whales, Geoffrey B. West, LANL
Abstract: Even though biological systems are the most complex
physical systems known, they satisfy remarkably simple scaling
laws. For example, metabolic rate (the power needed to sustain
life) scales like the 3/4-power of mass over 27 orders of
magnitude ranging from the molecular respiratory complex within
mitochondria up through the smallest unicellular organism
(mycoplasma) to the largest animals (whales) and plants (giant
sequoia). Other scaling laws relate how organismal features change
with size over many orders of magnitude; these include time-scales
(such as lifespan and heart-rate) and sizes (such as the radius of
a tree trunk or the aorta). All of these can be expressed as power
laws with exponents which are typically simple multiples of 1/4.
The systematics of these phenomena will be reviewed and a
quantitative, unified model presented that can explain their
origin, including that of the universal 1/4-power. The model is
based on the fundamental observation that essentially all of life,
regardless of size, is constrained, and ultimately limited, by the
rate at which essential resources that sustain it can be supplied.
General principles that are complementary to the principle of
natural selection are proposed. Assuming that organisms have
evolved a space-filling hierarchical branching network in which
energy dissipated is minimized, the known structural, functional
and scaling properties of many such systems, such as the
cardiovascular, respiratory and that of plants can be
quantitatively understood. In addition, because natural selection
has acted to maximize the area of interface of these hierarchical
systems with their resource environment, they are essentially
fractal-like structures in four spatial dimensions, rather than
three. The possible extension of these ideas to other systems,
such as river networks, transport systems and corporate structures
will be discussed.
Editor's comments: West mentioned a number of curious facts
about familiar complex systems: It seems that for a large variety
of quite different animals (including humans) in spite of very
different life spans and heart rates, the total number of heart
beats is roughly the same and around one (US) billion (= 109).
Surprisingly that is just slightly more than the number of
expected cycles that a car engine performs during it's lifetime.
West tries to formulate fundamental conditions such as the optimal
use of energy by an organism that lead to scaling laws with
universal exponents. The interpretation of his mathematical
explanation of the 1/4 power law ("life is a 5D object") might
still be a challenge in this area of quantitative theoretical
biology.
Allometric
scaling of production and life-history variation in
vascular plants, Brian J Enquist, Geoffrey B West,
Eric L Charnov, James H Brown, Nature, 28 October
1999.
Agent based Simulations: An Overview of BT's Research and Applications, Iqbal Adjali, BT Labs
Abstract: This talk will review work done in BT in the areas of
business modeling and intelligent systems design, following novel
approaches inspired from complex systems and using tools like
agent-based simulations and evolutionary game theory. We will
present our work on the modeling and consumer behavior using
agent-based simulations, and introduce the agent-based software
tool ZEUS. The agent based consumer model uses real data and
preserves the characteristics of customers. It allows interaction
between groups of customers to be investigated. ZEUS is an
agent-based software toolkit which enables the rapid and
systematic engineering of agent systems both for research and the
development of real-scale applications.
Editor's comments: While SWARM is a sophisticated research and
prototyping tool, ZEUS is more application oriented with
considerable details in the description of the agents. It is
written in Java code but there are plans to transfer it to faster
programming languages. Current applications include a personal
travel agent, an agent based electronic market place, home
shopping agents, and enhanced network management. Future
developments will include the ability of the agents to learn.
Economic and Business Applications of Agent-Based Modeling: The Swarm Tool Kit, Benedikt Stefansson, CASA
Abstract: Agent Based Models (ABMs) of economic and
business systems are slowly gaining ground both in academia and
industry. One of the stumbling blocks in developing such models is
the lack of programming libraries which enable the users to
construct complex ABMs with less effort. In 1995 the Santa Fe
Institute (SFI) launched the Swarm project with the
goal of providing scientists and programmers with a standard, open
source' ABM tool kit. Last fall the Swarm Development
Group moved from SFI and formed its own, separate, non-profit
organization. I will discuss the history and lessons from the
Swarm project, give a brief overview of the tool kit and some
examples of its use in academic and business applications. The
Swarm libraries are distributed under the GNU Library Public
License and freely available on the SFI website.
Editor's comments: Swarm is object oriented where the
scheduling is also done by objects. Hierarchical models in the
form of nested swarms are possible. Swarm is not yet able to run
as parallel code but the infrastructure is there already. For an
easier real time analysis of the simulation probe objects can be
attached to agents. In one application farmers in Brasil used
swarm based modeling to develop agricultural practices with less
detrimental impact on the environment
-
Abstract: In this talk I distinguish between difficulty -- hard
to solve isolated problems -- and complexity -- dynamic adaptive
systems with feedback. A difficult problem, such as designing an
efficient engine, does not change over time, though the approaches
to solving it may. In contrast, complex problems, such as
designing an effective marketing plan, change constantly. A good
solution one day may be a poor one the next.
The diversity of human agents helps us to cope with both
difficulty and complexity. I will discuss how diversity can help
lead us to optimal solutions to difficult problems. I will also
comment on the role of diversity in reducing complexity and in
creating robust system level performance.
Editor's Comments: Page discussed the role of diversity in
solving difficult and complex problems. In complexity there is a
strategic interdependence between the agents: other agents'
decisions affect every problem. Generally a difficult problem is
the large number of possibilities for solutions. Page discussed a
two-step approach. 1. Perspective: how to encode the problem, for
instance in a formal math language. 2. Heuristics: how to search
for the solution within this framework. He illustrated this
process with the help of the "Ben and Jerry & ice cream
landscape". When the two ice-cream manufacturer searched for the
best ice cream they created a table full of ice cream samples for
which ingredients changed along the two different axes: In one
direction the size of the chocolate chunks would increase and in
the other direction the number of chunks per serving. In this
array they searched for the combination with the best chance to
sell.
Page then claimed that their simulations indicated that groups
of experts often get stuck in sub-optimal solutions and that
external consultants might help to find a solution in a completely
different direction. Their simulations seem to indicate that under
the assumption that there is only a countable set of points at
which people get stuck, non-expert consultants will always find
the better solution: a random collection of agents as group is
better than a smart group. He didn't mention, however, how long it
would take the group to find the solution. He also assumed that
the agents communicated efficiently so probably his results are
better not interpreted as a recommendation to have consultants run
the companies. He saw global chains of restaurants like McDonalds
or toys like Pokemon as a threat to diversity that change a local
culture if there are no sufficiently high thresholds for replacing
traditional businesses.
-
Work with agent-based models (abm) have a long tradition in
Japan, They have translated Axelrod's work, worked on pocket
pagers, and collaborated with universities to construct interview
systems. They analyzed trade-offs and used Monte Carlo simulations
to analyze risks. Now they want to develop practical tools to
teach about non-linear concepts like the whole is more than the
sum of the parts and how to assess emergence. These new ABM
simulations are different from systems dynamics simulations. They
used StarLogo and Swarm (with the Swarm manual translated into
Japanese). The speaker also mentioned that artificial societies
are often best expressed in books and novels. For instance "Ring,
Spiral, Loop (refers to SFI) has been turned into a movie. Their
agent based simulator is focused on educational applications in
universities and social science studies. They formulated three
objectives: 1. express the system in Japanese, 2. ease of use that
only requires visual basic background, 3. open system, scale up,
incorporate other programs in Windows environment.
Hattori proceeded to give demos of Japanese versions of some
agent-based simulation programs like "turtles", "sugar-scape", all
programs written in visual basic and described with objects (see
videoclip).
Although current versions are two- dimensional there is no problem
to generalize them to multi-dimensional versions. They all have
drag-and-drop capability and log files that allow data mining.
Other demos showed an international politics model that allows
alliance formation between 92 countries and a forest fire model.
In the future they want to use genetic algorithms and link to
other programs for parallel operation and speed. Future
applications include: area marketing, traffic, evacuation,
forecasting business sales. So far their models failed to come up
with accurate forecasts but they trigger deeper thinking. He
concluded by mentioning that they will have a seminar with Josh
Epstein, author of "Growing artificial societies". In the
discussion section Ellen Goldberg mentioned that SFI tries to work
with schools to teach children about Star Logo.
Kozo
Keikaku Engineering Inc. Agent
Based Simulator Website
