A Self-Organizing Neural System
For Urban Design
Brian Lonsway
Rensselaer Polytechnic Insitute
Director, Informatics and Architecture
lonsway@rpi.edu
Ajith Rao Mulky
Rensselaer Polytechnic Institute
PhD candidate
mulkya@rpi.edu
Abstract
The dynamics of urban systems are characterized by complex non-linear relationships
between socio-economic attributes of land use and spatial interactions. A myriad of
theoretical and computational models have been offered to simulate these
complexities, but we have found that, within the realm of design practice, the
application of these models tends to be limited by their ostensibly objective and
predictive natures. This paper examines an alternative to predictive simulations of
urban systems in order to seek greater conceptual consonance with the highly
subjective and interpretive processes of design. In this work, we acknowledge that
proposals, context models, and simulations are at best only subjectively described,
with incomplete, and often inconsistent, data. In response to this, we propose an
approach to model the “design state” of a system rather than its historical or
evolutionary state. The system incorporates Teuvo Kohonen’s self-organizing map
algorithms within an existing GIS application to simulate the interaction between a
design proposal and an existing urban context.
2
A Self-Organizing Neural System For Urban Design
Introduction
The dynamics of urban systems are characterized by complex non-linear relationships
between socio-economic attributes of land use and spatial interaction. Urban theorists
and modelers have grappled with this complexity for decades, and have proposed a
plethora of models which attempt to embed such complexity within their frameworks.
These models, however, are only as “good” as the data upon which they rest and the
skills of the users who interpret and apply their results. Within the realm of design,
where collaboration between empirical understanding and subjective interpretation is
necessary, these models often fail to find an audience because of their limited
application to the factual or predictive end of exploration. The objective desire of
quantitative models, in other words, tends to limit their use to less subjective
processes. We see this empirically in the area of urban design as a component of the
discord between research development and application. The vast amount of work on
urban modeling seems to have found little place in the applied realm of urban design,
in particular when urban design meets its point of realization in architectural and
landscape design and engineering.
We believe this is rooted in the phenomenon of objectivity. Faith in the success of a
particular simulation effort is philosophically based on the assumption that there
exists a universally (or at least, generally) “correct” model. This model is achieved
when the “correct” data is manipulated in the “correct” way to achieve a perfectly
validated simulation of a system. All of this relies in the end on the ability of a
complex system to be accurately represented by quantitative data. In addition,
computational simulation is of necessity based on this presumption of the quantitative
reducibility of complex systems. The presumptions of an ostensible objectivity of
scientific computation, and of an ostensible veracity of numeric data belie the
subjective side of computational simulation: that every aspect of a simulation, from
its conceptual framework, base data, method of visualization, and implementation
framework, carries with it a subjective decision of one or another author.
Nevertheless, the primary goal of simulation remains the objective prediction of
outcomes.
From a cultural studies perspective, the truth of any predictive system can never be
more than a fantasy. Could there not be a form of design simulation which
acknowledges the subjective biases of its underlying processes for productive effect?
This simulation, we theorize, would model the design state of a given system (for the
purpose of contextualization) rather than its historical or evolutionary state (for the
purpose of prediction). The question takes on great significance in the area of design,
as the processes of design are far from objective. An individual’s intuition, personal
experience, cultural context, and social relationships all inform the kinds of decisions
which are made. And in turn, these decisions impact social and cultural situations
which are far from objectively quantifiable. The development of a simulation system
which both supports an individual’s subjective engagement with a design problem
and accommodates the subjective nature of complex systems modeling seems to be
A Self-Organizing Neural System For Urban Design
3
missing from the designer’s toolkit. We have attempted to address at least a fraction
of this problematic with the conceptual formation of what we call Subjective
Simulation, and have explored it through an experimental implementation of a system
for creative design problem solving.
Proposition
Fundamentally, a computational simulation system is a visualized model of data,
temporally extrapolated. The nature of the data, the type of visualization, and the
mode of extrapolation all contribute to the predictive value of the simulation in a
typical simulation system. In our case, we are interested in looking at how these three
aspects can be understood simultaneously as scientific parameters and social
constructions, and how this broader understanding can contribute to the simulation
discourse. Specifically, we believe that a live simulation of the current state of a
complex system can better provide a designer with an operational model within which
design can take place. Typically, as predictive simulations are most useful in the
evaluation of possible outcomes of a resolved design, they are seldom integrated
smoothly within the evolving design process, from early conception to realization.
Our focus has been on the use of artificial neural networks, specifically SelfOrganizing Maps (SOM’s) as introduced by Teuvo Kohonen [Kohonen 1984]. These
forms of neural networks distinguish themselves in the field by (1) requiring no
manual training (thus, their self-organizing nature), and (2) representing not only the
dendritic structure of neural connections but the topological function of neural
arrangement (thus their map-like nature). Their logic allows them to function
particularly well at discovering patterns among highly multivariate datasets of
unknown structure or context. While SOM’s are not typically used as the core of a
simulation system, we have focused our explorations on them for two complimentary
reasons: their topologically-based ‘storage’ function represents a complex dataset
without actually storing the data from the dataset, and their biologically-influenced
learning process models the kind of knowledge that can be gained from the
integration of radically diverse data sources.
Both of these characteristics can be seen as opportunities for engaging subjective
processes. First, we believe that data is at best a vague approximation of the subject
which it represents, and that a simulation system which relies on such approximations
cannot escape their subjective nature. The SOM’s formation of abstractions or
interpolations of data is itself a rather subjective process, and one which is more
conceptually aligned with interpretation over validation. And second, as the human
process of learning is a vague process which is as much social construction as it is
biological process, a database system which does not constrain its input to a
predetermined type or classification supports processes of learning and model
formation which value the context of information as much as its content.
4
A Self-Organizing Neural System For Urban Design
Implementation
We have developed a simulation system which models the current state of a
designer’s problem (in our example, an urban design scenario), as the designer
chooses to frame it, and works in conjunction with the designer to evaluate,
contextualize, and visualize design solutions within this chosen context. It relies on
an unconventional application of the Self-Organizing Map algorithm (implemented in
MatLab using the SOM Toolkit [Helsinki University of Technology 2004]),
employing the SOM both to model the designer’s context and to simulate the
interactions of a given proposal with it. These interactions are then visualized for the
designer as real-time feedback which can in turn be interactively manipulated (see
figure 1). This latter step is accomplished using MapInfo’s MapBasic scripting
extension to their MapInfo software.
We chose a popular, simple-to-use
Geographical Information System software in order to constrain our interface research
to a commonly accepted framework. We have also built a preliminary interface to the
SOM core using Macromedia’s Flash, which affords substantially greater creative
play in interface design than does MapInfo. These concerns are paramount for the
design community, which desires to work as fluidly as possible between creatively
designed and simple to use modeling, simulation, and visualization applications.
Figure 1. System Implementation.
A Self-Organizing Neural System For Urban Design
5
The system takes input from a variety of sources which the designer can subjectively
choose to frame their particular approach. In an urban design scenario, this data could
consist of current base data from a given city, comparative data from other cities, or
even data from outside the urban context which a designer finds relevant to the task.
It is the application’s goal to allow a designer to easily cull from available datasets or
to define and enter a dataset in order to commence the design process. In the
meanwhile, we are working with a rich dataset of Troy and Albany New York which
includes traffic pattern, land use, zoning, environmental toxicity, and landmark
proximity data. Our design sessions assume this data is the subjectively defined
context for the design problem, and then allow the designer to define the particular
design solutions within this data space. The SOM network is then trained on this
dataset (see Figure 2).
Figure 2. The core interface.
One of the strengths of the SOM algorithm is the representation of patterns both
within and between datasets, thus facilitating the discovery of relationships which
might not be apparent between wildly diverse sets of data. This process is very much
akin to the memory function of the human brain, where information from multiple
sensory inputs, various temporal periods, and distracting mental processes are
embedded in nearly every thought we have. It is our argument that this subjective
process of modeling our world through the integration of diverse mental processes is
what affords our understanding of complexity and our ability to cope with complex
scenarios. The data model formed in this way by our SOM implementation, then, is
seen as a robust yet iconoclastic model of a designer’s problem space – the urban
context of the design problem.
6
A Self-Organizing Neural System For Urban Design
Typically, this map or model is the end goal of a Self Organizing Map system. It is
used to extract patterns and correlations within mapped data. While this is a highly
useful result of the SOM process, we see much greater potential of the SOM. As
human learning never stops with a discrete map of knowledge gained, our “training”
process is a continuum. By extension, then, if the training process of the SOM is a
valid (although highly approximate) model of one aspect of the human learning
process, the temporal continuation of this training cycle represents the ability of the
SOM network to continue to learn. In other words, if this discrete map of the SOM
process is seen as a ‘complete’ set of knowledge about a condition, then future
training cycles can be seen as a form of query into this knowledge.
Our approach builds upon conventional search algorithms which use SOM’s, but
takes the nature of the query further. Once a data model is formed, the designer then
formulates a proposal and presents it to this model as if it were a new piece of data for
the network to learn. Three options are possible at this point: the proposal can be
presented as a search, where the network returns a representation of the data in the
network which most closely matches the proposal; the proposal can be presented as
new training data with which the network can be trained, and a new state of the model
is formed; or the proposal can be seen as new training data, a copy of the network can
be trained on it, and the results from the copy and the original networks can be
compared. Where the first of these is limited to a straightforward search capability,
the latter two form the core of our simulation system. In effect, the difference
between the ‘pre-proposal’ and ‘post-proposal’ states of the network can be seen as a
simulation of the impact the design proposal has on the current state of the data
model. By exploring the creative representation of this difference, it is possible to
develop a richly interactive simulation of a design problem. And, by including
arbitrary design proposals as part of the training dataset, the network can observe
patterns among evolving proposals and contextualize them collectively within the
urban context itself.
Both of these propositions are highly experimental, and have not yet been thoroughly
tested within an application framework. Many concerns remain. How valid is a
training cycle if only one piece of data is used as the training dataset (as is done for
proposal evaluation)? What is the best way to visualize the difference between “preproposal” and “post-proposal” states of the neurode’s data vectors (should vector data
be componentized and compared piecewise, or not)? What is the temporal impact of
using hypothetical design proposal data as training data (does the network degenerate
into meaninglessness, or is an intelligence gathered which engages the designer’s
subjective experimentation)?
A Self-Organizing Neural System For Urban Design
7
Conclusion
These and other questions are dominating our current work. Fundamentally, we are
examining the majority of these questions as problems of visualization. Our
theoretical framework proposes that any data model, because it is incomplete,
contains inaccuracies, and that one’s engagement with any such model must
acknowledge both the subjective nature of the data representation and the subjective
nature of one’s interaction with it. As long as there is a meaningful desire to conceive
and build a given model, then, it is valid and useful within the construct of that desire.
Certain questions of validation drop out, then, in our evaluation of our system. As
long as the network is originally well-trained with enough input samples, then, a
retraining of the network with a single data vector (i.e., a design proposal) should not
invalidate the experiment. (Recall the instructor’s aphorism: “The only dumb question
is the one that isn’t asked.”) A clear visualization of the process will make it clear to
the designer if something unexpected appears, and the designer may choose to
proceed as appropriate to the task at hand (e.g., undo, restart, investigate the
unexpected result further, etc.)
Like the underlying data model and method of extrapolation, the visualization should
challenge the ostensible objectivity of representation in order to afford open-ended
interpretation on the part of the designer. Specifically, we believe this visualization
should be highly interactive such that the simulation can operate near real-time,
allowing a designer to modify a design on the fly and immediately understand the
evolving relationship the visualization has to the design problem.
This is a highly experimental application at this point. Since it operates outside of the
predictive arena, quantitative validation is impossible – and anathema to our
theoretical framework. Yet, because it is highly reliant on user interaction, a rigorous
subject testing is required. To date, we have developed the functional core of the
system and are currently experimenting with visualization and user interface
frameworks. Nonetheless, this work is primarily a theoretical inquiry. The
implementation is seen as an experimental testing ground for a more culturally
situated inquiry, meant to be a critical prod into the arena of social simulation at large.
References
Helsinki University of Technology, Laboratory of Computer and Information Science. The
SOM ToolBox. http://www.cis.hut.fi/projects/somtoolbox. Accessed February 13, 2004.
Kohonen, Teuvo: 1984, Self-organization and Associative Memory, Springer-Verlag (Berlin).