Pre-Specific Modeling - Computational Machines in Coexistence with Urban Data Streams
1. PhD Candidate: Vahid Moosavi
Supervisor: Ludger Hovestadt
Co-supervisor: Barbara Hammer
Pre-Specific Modeling
Computational Machines in Coexistence with Urban Data Streams
27th of April 2015
ETH Zurich
Tuesday, 28 April, 15
7. 3
If one scans 5 bn books
from 1900 (Google n-gram)
Tuesday, 28 April, 15
8. And if we just follow the market?!
Google Trends 4
Tuesday, 28 April, 15
9. And if we just follow the market?!
Google Trends
RQ: ButWhat are the underlying paradigms
and abstract concepts behind these patterns?
and
How architects and urban designers can
be proactive contributors in this domain?
4
Tuesday, 28 April, 15
10. Approach of this research
(how to be abstract but operational?)
1.It is not adequate to talk about individual modeling approaches.
Instead, we need abstract and philosophical research in
computational modeling in order to discuss the foundations.
(Toward this, we tried to formalize “pre-specific modeling” as a
way of thinking in computational urban modeling.)
5
Tuesday, 28 April, 15
11. Approach of this research
(how to be abstract but operational?)
1.It is not adequate to talk about individual modeling approaches.
Instead, we need abstract and philosophical research in
computational modeling in order to discuss the foundations.
(Toward this, we tried to formalize “pre-specific modeling” as a
way of thinking in computational urban modeling.)
2.Further, if the developed concepts and ideas are true then
they should work in different applied domains
(Therefore, we implemented the developed concepts in
different applied domains.)
5
Tuesday, 28 April, 15
12. First part
Conceptual Set up of Pre-Specific Modeling
(Toward formalizing the elements of a modeling approach)
6
Tuesday, 28 April, 15
14. The issue of Representation
7
Specific Modeling Pre-Specific Modeling
Tuesday, 28 April, 15
15. The issue of Representation
Selected'Features'to'Represent'the'Objects'
Objects'
Examples:*
• Ci-es*
• Streets*
• Buildings*
• People*
• Companies*
• Food**
• Energy*
• Medicine*
• Internet*
• Words*in*a*text*
Each concrete object is an instance of
a generic object, defined a-priori
7
Specific Modeling Pre-Specific Modeling
Tuesday, 28 April, 15
16. A
B
The issue of Representation
Selected'Features'to'Represent'the'Objects'
Objects'
Examples:*
• Ci-es*
• Streets*
• Buildings*
• People*
• Companies*
• Food**
• Energy*
• Medicine*
• Internet*
• Words*in*a*text*
Each concrete object is a feature for itself
related to other objects based on concrete
relations
Each concrete object is an instance of
a generic object, defined a-priori
7
Specific Modeling Pre-Specific Modeling
Tuesday, 28 April, 15
17. A
B
The issue of Representation
Selected'Features'to'Represent'the'Objects'
Objects'
Examples:*
• Ci-es*
• Streets*
• Buildings*
• People*
• Companies*
• Food**
• Energy*
• Medicine*
• Internet*
• Words*in*a*text*
Each concrete object is a feature for itself
related to other objects based on concrete
relations
Each concrete object is an instance of
a generic object, defined a-priori
Abstract Universals Concrete Universals 7
Specific Modeling Pre-Specific Modeling
Tuesday, 28 April, 15
18. The Issue of Causality
8
Specific Modeling Pre-Specific Modeling
Tuesday, 28 April, 15
19. The Issue of Causality
The Rule
8
Specific Modeling Pre-Specific Modeling
Tuesday, 28 April, 15
20. The Issue of Causality
The Rule Any Rule
8
Specific Modeling Pre-Specific Modeling
Tuesday, 28 April, 15
21. The role of Data
9
Specific Modeling Pre-Specific Modeling
Tuesday, 28 April, 15
22. The role of Data
9
Specific Modeling Pre-Specific Modeling
Classical SpaceSyntax, London
“The social logic of space”1984
Structured and designed data
Tuesday, 28 April, 15
23. The role of Data
9
Specific Modeling Pre-Specific Modeling
Classical SpaceSyntax, London
“The social logic of space”1984
GPS Trajectory of Taxicabs, Beijing, 2012
Structured and designed data Unstructured and indirectly produced data
streams
Tuesday, 28 April, 15
24. Second part
Computational Machines that aligns with the concept of
pre-specific modeling
10
Markov Chains
Self Organizing Maps
Tuesday, 28 April, 15
30. Learned Dictionary of Forms in Co-existence with Data
12
Pre-Specific Modeling
A
B
Tuesday, 28 April, 15
31. Learned Dictionary of Forms in Co-existence with Data
Self Organizing Map (SOM)
12
Pre-Specific Modeling
A
B
Tuesday, 28 April, 15
32. Functions of Self Organizing Map (SOM)
•Nonlinear function approximation
•Time series prediction
•Classification
•Clustering
•Representation Learning
•Dimensionality Reduction
•Topological Data Analysis (TDA)
•Nonparametric Probability Density Estimation
13
Tuesday, 28 April, 15
33. Functions of Self Organizing Map (SOM)
•Nonlinear function approximation
•Time series prediction
•Classification
•Clustering
•Representation Learning
•Dimensionality Reduction
•Topological Data Analysis (TDA)
•Nonparametric Probability Density Estimation
An Open Source Python Library
based on parallel processing
13
Tuesday, 28 April, 15
35. Applications
•Urban Traffic Modeling Using Data Streams
•Urban Air Pollution Modeling
•Contextual Mapping
•Modeling the Dynamics of Real Estate Market
•Other experiments in different domains:
•Financial Markets
•Biological Systems
•Atmospheric Environments
•Natural Language Processing and Text Modeling
15
Tuesday, 28 April, 15
36. First Application
Urban Traffic Modeling Using Data Streams
Moosavi, Hovestadt 2013 16
Selected'Features'to'Represent'the'Objects'
Objects'
Tuesday, 28 April, 15
37. First Application
Urban Traffic Modeling Using Data Streams
Moosavi, Hovestadt 2013
State of the art:Agent Based Modeling
Needs the grammars and rules of transport
in advance like a language
To be validated by the classical data such as
surveys
16
Selected'Features'to'Represent'the'Objects'
Objects'
Tuesday, 28 April, 15
38. First Application
Urban Traffic Modeling Using Data Streams
Moosavi, Hovestadt 2013
State of the art:Agent Based Modeling
Needs the grammars and rules of transport
in advance like a language
To be validated by the classical data such as
surveys
16
Selected'Features'to'Represent'the'Objects'
Objects'
Tuesday, 28 April, 15
39. First Application
Urban Traffic Modeling Using Data Streams
Moosavi, Hovestadt 2013
State of the art:Agent Based Modeling
Needs the grammars and rules of transport
in advance like a language
To be validated by the classical data such as
surveys
Using GPS tracks of cars as data streams
In an analogy to spoken language, taking urban
cells as a word in a language, each individual
driver is a unique story teller, while driving within
urban grid cells…
16
Selected'Features'to'Represent'the'Objects'
Objects'
A
B
Tuesday, 28 April, 15
40. First Application
Urban Traffic Modeling Using Urban Data Streams
GPS$Trajectory$of$Taxicabs,$Beijing$
Road$communi;es$based$on$
movement$of$the$cars$not$just$the$
physical$road$network$
highly$cri;cal$areas$for$the$whole$traffic$flow$
Areas$with$high$poten;al$of$traffic$jam$
Moosavi, Hovestadt 2013 17
Tuesday, 28 April, 15
41. Second Application
Urban Air Pollution Modeling
•State of the art: CFD models
•Based on idealized models of the
fluid movements and numerical
methods (Given Priors)
•With many simplifying
assumptions on the ambient level
•Limited in resolution and
integration
•Computational complexity
Moosavi, Aschwanden, Velasco 2015 18
Tuesday, 28 April, 15
42. Second Application
Urban Air Pollution Modeling
•State of the art: CFD models
•Based on idealized models of the
fluid movements and numerical
methods (Given Priors)
•With many simplifying
assumptions on the ambient level
•Limited in resolution and
integration
•Computational complexity
Moosavi, Aschwanden, Velasco 2015 18
Tuesday, 28 April, 15
43. Second Application
Urban Air Pollution Modeling
•State of the art: CFD models
•Based on idealized models of the
fluid movements and numerical
methods (Given Priors)
•With many simplifying
assumptions on the ambient level
•Limited in resolution and
integration
•Computational complexity
•Using only 3% of direct measurements
at the ground level
•Plus more than 80 urban parameters
using SOM, we were able to estimate the
aerosol levels of the whole region with a
reasonable accuracy (Cross Validated)
•Further, we were able to detect the
most informative candidate locations for
monitoring stations (which are
expensive)Moosavi, Aschwanden, Velasco 2015 18
Tuesday, 28 April, 15
47. Third Application
Contextual Mapping
20
Top down or the bottom up Map
“The image of the city”
Any Map
Individualized and context dependent Maps
“Any image of the city”
The Map
Tuesday, 28 April, 15
49. How to look at a city with huge amount of data? “Digital Opacity?”
21
Tuesday, 28 April, 15
50. How to look at a city with huge amount of data? “Digital Opacity?”
Is it possible that instead of a top-down map, we provide a
simple mapping process?
“Toward designing the process of any map”
21
Tuesday, 28 April, 15
52. Method: Learn a one dimensional dictionary and take its elements as an
ordered set of numbers. (Computing with contextual numbers)
23
Tuesday, 28 April, 15
53. Method: Learn a one dimensional dictionary and take its elements as an
ordered set of numbers. (Computing with contextual numbers)
23
Tuesday, 28 April, 15
58. Contextual Mapping
Redistricting and the will of the people
How polarized is Colorado in different elections?
Is there any bottom up approach to define the electoral borders?
27
Tuesday, 28 April, 15
60. Monthly Median Price per sq.ft of sold
houses of each Zip code in the state of
Florida 29
Tuesday, 28 April, 15
61. One
of
the
Basics
of
Spa0al
Modeling
is
Spa0al
Correla0ons
30
Tuesday, 28 April, 15
62. Classical Notion of Space
(First Law of Geography)
One
of
the
Basics
of
Spa0al
Modeling
is
Spa0al
Correla0ons
30
Tuesday, 28 April, 15
63. As an alternative: to create a space of temporal patterns
SOM$
31
Tuesday, 28 April, 15
64. As an alternative: to create a space of temporal patterns
SOM$
31
This space can be based on any
other aspects too.
(It is a TimeVarying Random Field)
Tuesday, 28 April, 15
70. Modeling the Dynamics of Real Estate Market
Results'
Predic-ng:'2005406401'from'12'months'in'advance' 36
Tuesday, 28 April, 15
71. Modeling the Dynamics of Real Estate Market
Results'
Predic-ng:'2006404401'from'12'months'in'
advance' 37
Tuesday, 28 April, 15
72. Modeling the Dynamics of Real Estate Market
26#
Results'
Predic-ng:'2013509501'from'12'months'in'advance'
38
Tuesday, 28 April, 15
73. Modeling the Dynamics of Real Estate Market
Most of the times (>80%),
we get better results with
looking to the new space!
39
Tuesday, 28 April, 15
74. Modeling the Dynamics of Real Estate Market
Spatial Distribution of Error
40
Tuesday, 28 April, 15
77. Summary of this research
42
•Since computational technologies have
become a dominant factor in urban design
and modeling...
Tuesday, 28 April, 15
78. Summary of this research
42
•Since computational technologies have
become a dominant factor in urban design
and modeling...
•In order to have a pro-active role in this
domain, the designers needs to be
computationally literate, not just (expert)
users of the computational products.
Tuesday, 28 April, 15
79. Summary of this research
42
•Since computational technologies have
become a dominant factor in urban design
and modeling...
•In order to have a pro-active role in this
domain, the designers needs to be
computationally literate, not just (expert)
users of the computational products.
•Toward this goal as a first step, we tried to formalize
the abstract notion of pre-specific modeling as a
conceptual alternative to the state of the art in
computational urban modeling and simulation.
Tuesday, 28 April, 15
80. Summary of this research
42
•Further, we investigated the applicability of a certain
category of generic computational methods that
encapsulate the complex real phenomena in a co-existence
with urban data and depend less on prior domain specific
theories.
•Since computational technologies have
become a dominant factor in urban design
and modeling...
•In order to have a pro-active role in this
domain, the designers needs to be
computationally literate, not just (expert)
users of the computational products.
•Toward this goal as a first step, we tried to formalize
the abstract notion of pre-specific modeling as a
conceptual alternative to the state of the art in
computational urban modeling and simulation.
Tuesday, 28 April, 15
81. Summary of this research Publications
•5 conference paper presentations
•1 journal paper accepted
•1 journal paper submitted
•3 working papers
•1 forthcoming book (Co-editor)
42
•Further, we investigated the applicability of a certain
category of generic computational methods that
encapsulate the complex real phenomena in a co-existence
with urban data and depend less on prior domain specific
theories.
•Since computational technologies have
become a dominant factor in urban design
and modeling...
•In order to have a pro-active role in this
domain, the designers needs to be
computationally literate, not just (expert)
users of the computational products.
•Toward this goal as a first step, we tried to formalize
the abstract notion of pre-specific modeling as a
conceptual alternative to the state of the art in
computational urban modeling and simulation.
Tuesday, 28 April, 15