1. svm@arch.ethz.ch
SEC
Data Driven Modeling Beyond Idealization
Vahid Moosavi
PhD Student at Chair for Computer Aided Architectural Design (CAAD),
Professor Ludger Hovestadt, ETH Zurich
Researcher at ETH-Singapore Centre, Future Cities Laboratory (FCL)
24 May 2014
1
12. 12
A Normal Distribution with
outlier or a “unique case”?
Fourier Transformation: any
form is a linear combination
of some ideal forms
Each Code Follows On An Ideal Form
13. Each code consists of certain aspects (features) of the natural system
(Models as Pairs of Glasses): Minimalist Idealizations and multi-models
Networks: Structural thinking
Agents (actors): Interactions
between different agencies
System Dynamics: Process
Oriented View
13
14.
15. Challenge: What to do in dealing with Complex Adaptive
Systems? Which glasses (i.e. modeling approach) is
sufficient, when in principle each view is arbitrary?
15
Hypothesis: Majority of current modeling approaches
are fundamentally limited in dealing with complex
systems and what we need is an abstraction from the
concept of “rational modeling”.
Idea: Is it conceptually possible to have all the views at
once?
17. Current Trend: Parametricism (multi-model Idealization) and the
Curse of Dimensionality…
17
…Complicated, but not complex models
Properties of the system for modeling
PossibleRelations
(typesandnumbers)
Complex Systems
Simple Systems
Minimal
idealization
Multi-model
idealization
18. 18
A new realm of modeling?!
Properties of the system for modeling
PossibleRelations
(typesandnumbers)
Multi-Agent
Systems
Urban Cellular
automata
Urban Dynamics
Basic Statistics
(Hypothesis Testing)
Urban Metabolism
Urban Scaling
Social Physics
Fractal Models
19. Toward a new formalism for the concept of modeling
“Models without Ideals or All the Potential Ideals”
19
Third Section
20. How to avoid the curse of dimensionality? (Or How to
Encapsulate all the potentialities?)
Selected Features to Represent the Objects
Objects
Encapsulation RelationalityRationality
Examples:
• Cities
• Streets
• Buildings
• People
• Companies
• Food
• Energy
• Medicine
• Internet
• Words in a text
Abstract Universals (ideal forms)
Concrete Universals
20
21. It is a self-referential Setup
21
Page Rank
..Can be local or global
22. Relational Modeling
22
An Example in Natural Language Modeling
Rational Modeling Relational Modeling
External Reference
We have the Ideal model of the
language
No External Reference
We have A Huge Corpus of
language
Main Assumption
Relational Representation of symbols in a language
Noam Chomsky MarkovHeroes
Based on
Approach
23. However, it took a century…
23
Markov (1907) Shannon(1948) Google 2000-
“For Linguists it is hard to
believe it as a practical
approach”
“Interesting idea, but
Computationally
Expensive”
“Getting Feasible! With Billions
of text documents”
Relational Representation of symbols in a language
Data
Deluge
28. 28
Using GPS tracks of cars within a city:
Taking urban cells as a word in a language,
each individual driver is a unique story teller,
while driving within urban grid cells…
...A Markov Chain Model of traffic dynamics
Can be developed for :
• Simulation
• community detection
• Network Engineering
• Sensitivity Analysis
31. SOM from the Context of (Nonlinear) Transformation:
Dimensionality Reduction
31
32. • Finding an ideal (global) transformation: e.g. PCA
• Observations are instances of an abstract representation
32
Selected Features to Represent the Objects
Objects
First General Approach: Direct Transformation
X TW
33. • Each observation is a dimension itself: e.g. MDS, LLE, ISOMAP,…
• There is always a mechanism to preserve neighborhood topology.
33
Second General Approach: Indirect Transformation
35. Self Organizing Map (SOM) : A generic setup, based on
symbolic indexes
• SOM as a transformation
based on topology
preserving mechanism,
but at the same time
creating an abstraction
from observations.
35
A Primal-Dual Representation
X TSOM
50. 50
1-Median list Price
2-Median sale price
3-Median list price -sq. ft.
4-Median sale price-sq. ft.
5-Sold for loss
6-Sold for gain
7-Increasing values
8-Decreasing values
9-Listings with price cut
10-Median price cut
11-Sold in past year
12-Homes for Rent
13-Homes foreclosed
14-Foreclosure re-sales
15-Sale-to-list price ratio
16-Price to rent ratio
Multi-Dimensional Time Series
Modeling (Real Estate Dynamics)
55. SOM: A Generic Computing Machine Beyond Ideal Forms
55
Democratic Computing Social computing
(Computing with any function)
Observed Data
Resamples of Data by SOM
In terms of representation and inference we can go beyond idealization
Focus on Idealization
In the beginning ---> Simplification, but gradually it has changed to a situation that every phenomenon has an underlying nature?!!
Examples: Point mass, infinite speed, circle, rational choice, rational maximizer Agents,
Aristotelian: Striping away properties (elements) from the concrete case
An example is a classical mechanics model of the planetary system, describing the planets as objects only having shape and mass, disregarding all other properties Galilean idealizations: Simplifying properties (mess points)
Coding is based on an arbitrary Ideal
Coordinatization
Multiple ideals
Numbers onion model
Quantum Mechanics
Similarly for the concept of modeling
Similar to FlatLand
Elephant and the blind men
Don’t say about models, just the reason for the limit
Walkability as an example
Godel’s incompleteness theorem
Hamiltonian Complexity Theory to show the limit of Model-ability!!
Identity
Assumption of objects independence
Sheaf Theory
Self-organizing Map : Mapping as transformation
Sheaf Theory
Localization -Globalization
For the issue of Causality: Prediction, classification, clustering
I am co-organizing a conference on SOM next Month
From an imposed structure to operation of any potential structure
Similar to Quantum Mechanics and abstraction from explicit representation to space of potential explicit representations
Discontinuity