N. Gu, S. Watanabe, H. Erhan, M. Hank Haeusler, W. Huang, R. Sosa (eds.), Rethinking Comprehensive
Design: Speculative Counterculture, Proceedings of the 19th International Conference on Computer-
Aided Architectural Design Research in Asia CAADRIA 2014, 233–242. © 2014, The Association for
Computer-Aided Architectural Design Research in Asia (CAADRIA), Hong Kong
EXPLORING URBAN CONFIGURATIONS FOR A
WALKABLE NEW TOWN USING EVOLUTIONARY
EUGENE CHIAN1 and PATRICK JANSSEN2
1,2National University of Singapore, Singapore
Abstract. Multi-objective evolutionary algorithms have been success-
fully applied within various design domains in order to explore the
trade-offs between conflicting design criteria. This research investi-
gates how evolutionary algorithms can be used to develop urban con-
figurations for walkable new towns, focusing in particular on the
trade-off between travelling time using public transport and accessi-
bility to open space. A population of optimised urban configurations
was evolved and analysed, resulting in the identification of three dif-
fering typologies for walkable new towns.
Keywords. Urban structure; transportation network; urban density;
multi-objective evolutionary algorithms.
Cities and towns today have outgrown simple ‘hub-and-spoke’ planning ty-
pologies to complex and interlinked urban configurations. Developing urban
configurations may involve optimizing multiple conflicting performance cri-
teria. In order to facilitate more experimental and at the same time more sys-
tematic design explorations, this paper proposes an evolutionary approach to
designing an urban configuration for a walkable new town. The walkable
new town refers to a town whereby all people can ride public transport to
within the neighbourhood of their destinations and then walk to final destina-
tions. The urban configuration refers principally to the urban structure that
defines the relationship between transportation networks and land use
(Westerman and Austroads, 1998). It involves strategically placing transport
nodes in relation to the density of built up areas to make highly accessible
234 E. CHIAN AND P. JANSSEN
The scenario explored in this paper is based on the brief of Vertical Cities
Asia Competition 2013 that was organized by National University of Singa-
pore (NUS, 2014). This competition requires high density design proposals
for one hundred thousand people on a site with an area of 1 km2. For this re-
search, the starting point will also be to provide for one hundred thousand
people on a 1 km2 site. However, the focus will be on exploring overall ur-
ban configurations for walkable new towns. In particular, the two main per-
formance criteria that are to be explored are: minimization of travel times
and maximization of accessibility to open space. In order to further constrain
the evolutionary search process, two secondary performance criteria are add-
ed: minimization of transport infrastructure cost and minimization of block
Section 2 describes a number of precedents that have applied evolution-
ary algorithms to evolve urban configurations, section 3 describes the pro-
posed evolutionary exploration method, and section 4 analyses the results
and identifies certain key typologies. Finally, section 5 briefly draws conclu-
sions and suggests avenues for future research.
A number of researchers have applied evolutionary algorithms to urban de-
sign to explore different paradigms.
Balling et al (1999) used evolutionary algorithms to design a future land-
use plan for the city of Provo, Utah. The goal was to search for a set of plans
that satisfied the housing constraints and were Pareto optimal for evaluation
criteria ‘Travel Time’, ‘Cost’, and ‘Change’ by changing the widths of 25
identified major corridors of the city. ‘Travel Time’ was calculated on a
model that routed all trips on streets throughout the city. ‘Costs’ were calcu-
lated as the cost of the corridor upgrades. ‘Change’ was measured by the
product of status quo land value and a degree-of-change factor.
Rakha and Reinhart (2012) presented a new urban analysis workflow that
develops street and massing layouts for new neighbourhoods in hilly terrains.
It consisted of first subdividing the site up according to terrain and, making
street widths offsets and generating massing. Then, the potential walkability
of each design was evaluated using "Street Smart" walk score algorithm that
computes the shortest path to randomly placed amenities such ‘grocery’,
‘shopping’, ‘parks’, and others, and give weighted scores for distances to get
to the respective amenities.
Pedersen and VanMater (2013) generated a city based on branching of
both a salt-water canal system and the road system. He used a formula to de-
fine branches and another power law function formula to determine foot-
URBAN STRUCTURE WITH EVOLUTIONARY ALGORITHMS 235
prints and heights of building envelopes. The genetic algorithms changed the
way of branching and evaluated the fitness level with a ratio comparing
length of waterways generated, number of buildings generated and connec-
tivity of both the road and waterway system.
This paper focuses on developing a design methodology that explores ur-
ban configurations at a highly conceptual level. Most of the existing research
focuses on specific scenarios, and relatively complex evolutionary algo-
rithms have been developed specifically for those scenarios. In this research,
the focus is on trying to develop the simplest possible evolutionary algo-
rithms that can still give meaningful insight into the issue at hand. In this
case, the issue is the conflict between travelling time and accessibility to
open space for residents in the town. The performance indicators are purely
geometric (based network analysis and proximity calculations) and do not
take into account system dynamics of the transport systems within the city.
However, it is argued that the results of the exploration provide intriguing
starting points for further analysis using more detailed simulation models.
3. Evolutionary Exploration Method
The exploration method uses evolutionary algorithms to evolve a population
of design variants with respect to certain quantifiable performance criteria.
This method requires the designer to define a development procedure that
generates design variants and a set of evaluation procedures to evaluate de-
sign variants (Janssen et al, 2011b). Lastly, a feedback procedure is used to
ensure that design variants are continuously reproduced and evolved through
the inheritance of desirable traits. For this research, the Dexen distributed
evolutionary system was used (Janssen et al, 2011a), thereby allowing the
evolutionary process to be accelerated.
The development procedure must generate design variants based on only
a small number of genes, and the evaluation procedures must calculate glob-
al performance metrics that represent the performance of the design variant
as a single numeric value that can be either maximized or minimized
(Janssen et al, 2011a). Furthermore, thousands of design will need to be de-
veloped and evaluated, and as a result it is essential that these procedures are
fast to execute. For these reasons, the development and evaluation proce-
dures use highly abstracted models. The development procedure generates
design variants that consist of highly simplified transport networks with ur-
ban density represented as extruded blocks on a square grid. The evaluation
procedures calculate simplified performance indicators for travelling time
for residents, the amount of open space available to each resident, average
block height and the cost of building the transport network.
236 E. CHIAN AND P. JANSSEN
3.1. DEVELOPMENT PROCEDURE
Based on the assumption that each person needs 50 m2 of space (including
residential, commercial, and other facilities), a total gross floor area of 5 km2
is required. A blank 1 km2 site is assumed, with one Mass Rapid Transport
(MRT) station located at the centre, connecting it to other towns and the
Central Business District. Within the site itself, two types of public transport
systems are assumed to be available: a People Mover System (PMS) for
travel over intermediate distances and a bus system for more localised travel.
The site boundary and the position of the main MRT station are prede-
fined and the centre of the town is assumed to be located at the MRT station.
The development procedure generates design variants by first creating the
transport network connecting the town centre to surrounding areas and then
creating urban density around the transport nodes, with decreasing density as
the distance from the transport node increases.
Figure 1. Abstracted model of the transportation networks, consisting of a pedestrian network
at the bottom, a bus network in the middle, and a PMS network at the top.
The transport network is represented as a connected set of lines at differ-
ent heights (see Figure 1). The top level represents the PMS network, the
middle level represents the bus network, and the bottom level represents the
pedestrian network. The PMS and bus networks are defined by a set of
points connected with straight lines. In the case of the PMS, the points repre-
sent stations, while in the case of the bus, the points represent bus stops. The
pedestrian network is represented as an orthogonal grid of streets.
The development procedure starts by defining the position of the PMS
stations. The total number of stations is set to 8 stations based on the area of
the site and the number of residents. The position of each PMS station is de-
fined by a pair of genes that define a position (u and v coordinates) on the
site. In order to generate the connections between the PMS stations, the sta-
tion points are triangulated to form "well-shaped" triangles (roughly similar
sizes without acute angles).
The bus network is created by generating a bus stop node at each PMS
station, and then clustering a set of additional bus stop nodes around each
URBAN STRUCTURE WITH EVOLUTIONARY ALGORITHMS 237
station. The average distance between the bus stop nodes is defined by a
gene, thereby allowing the evolutionary system to generate bus networks of
varying density. The bus stop nodes are triangulated to form the network of
bus connections, and additional bus stops are then inserted along each con-
necting line. In this way, the bus network connects the PMS stations to the
Bus stops are clustered around PMS stations in circular zones. The radius
of each zone is defined by Equation 1. The radii of these zones start small
and get progressively bigger as they get further from the town, with the ex-
ception of the town centre which has a fixed zone radius of 500 meters.
where r is the radius of the bus stop zone, m is the distance from the selected
PMS station to the MRT station , and maxi mi is the maximum of the set of
all m distances.
The pedestrian network is created by generating a regular grid on the
ground for all areas within a 100 meter distance from any bus stop. The rela-
tively small distance of 100 meters was set in order to ensure that compact
towns would be generated where all residents could easily access public
Finally, urban density is generated in all areas directly accessible from
the pedestrian network. Urban density is visually represented by extruding a
square block vertically, with the height of the extrusion set to vary in relation
to the proximity to the closest PMS station. The heights are also represented
by vertical lines (lift lines) connected to the transportation network model
(see Figure 2). One key constraint is the total floor area, which is fixed at 5
km2. The extrusion height is therefore determined in a two-step process. In
the first step, the height factor is calculated using Equation 2 (see Figure 2).
The equation causes the height of extrusion to decrease exponentially as the
distance to the closest PMS station increases. The power value, denoted by g,
is defined by a gene in the range 0.1 to 0.9. This allows the evolutionary sys-
tem to vary the rate at which the density falls off.
where hf is the height factor, p is the distance from the selected block to the
closest PMS station, and maxi pi is the maximum of the set of all p distances.
The extrusion height is then calculated using Equation 3. In order to
achieve 5 km2 of floor area, and assuming a typical floor plate of 400m2, a
total block height of 37,500 m is required. The extrusion height for a particu-
500max/ 5.0 ii mmr
ii pphf max/
238 E. CHIAN AND P. JANSSEN
lar block is calculated by multiplying this total block height by the fraction
of the height factor relative to the sum of all height factors.
where eh is the extrusion height, hf is the height factor of the selected block,
and ∑ hfi is the sum of all the height factors.
Figure 2. Lift lines in transportation network model (a), examples of a g value 0.8 (b) and a g
value of 0.4(c), distances used to define extrusion height (d).
3.2. EVALUATION PROCEDURES
Four indicators are used to evaluate and optimise the urban configuration.
They are travelling time indicator, open space indicator, transport infrastruc-
ture cost indicator and average building height indicator.
3.2.1 Travelling Time Indicator
The travelling time indicator refers to an indication of times needed to get
around the town. Two types of trips are considered: traveling to the town
centre and travelling across town (to visit friends). The indicator will take
the sum of both travelling times and is set to be minimized so that the town
is optimized for travelling.
In order to calculate the approximate travelling time between two points,
the shortest path (in terms of travelling time) through the transport network
is calculated. This calculation takes into account the travelling speeds for the
different transport modes and waiting times when switching between
transport modes. For example in Figure 3, the shortest path from A to B was
to get down from residential block at point A, walk to a nearest PMS station,
URBAN STRUCTURE WITH EVOLUTIONARY ALGORITHMS 239
wait 3 minutes for an PMS to arrive and ride the PMS to reach the station
nearest to B and continue the journey on foot to B. The horizontal and verti-
cal lines of the PMS network are encoded with a train speed of 40 km/h and
a waiting time of 3 minutes respectively; horizontal and vertical lines of the
bus network encoded with bus speed of 30 km/h and a waiting time of 4
minutes respectively; and the horizontal lines of the walking grid are encod-
ed with a walking speed of 30 m/min.
In order to calculate an overall travelling time for a particular town, 100
trips to the town centre and 100 trips to visit friends were calculated. For the
trips to the town centre, starting points were chosen randomly. For the trips
to across town, both starting points and destination points were chosen ran-
domly. The travelling time indicator is calculated as the sum of both types of
Figure 3. Calculation of shortest path between two points, A and B.
3.2.2. Open Space Indicator
The open space indicator calculates the amount of open space directly adja-
cent to the built up areas, as this is the area that is assumed to be used by res-
idents in the town. The adjacency threshold is set to 100 meters. Open space
is set to be maximised.
3. 2. 3. Infrastructure Cost Indicator
The transport infrastructure cost indicator calculates the sum of the weighted
lengths of both the bus network and the PMS network. The PMS network is
assumed to be twice as costly as the bus network for a given length. The cost
is set to be minimised.
3. 2.4. Block Height Indicator
The block height indicator calculates the average of the block heights, meas-
ured in terms of the number of floors. This indicator gives an approximation
of average density and is inversely related to site coverage. Block height is
set to be minimized.
240 E. CHIAN AND P. JANSSEN
The evolutionary exploration process generated a total of 20,000 urban con-
figurations, resulting in a final population of 100 optimised configurations.
Figure 4 show selected designs from the final population.
Figure 4. Visualisations of a set of evolved designs
4. 1. GENERAL TRENDS
In order to better understand the relationships between the performance indi-
cators, a number of graphs were plotted for pairs of indicators. From the
graphs, certain general trends can be identified.
The Travelling Time versus Open Space graph in Figure 5 (top) shows a
general trend of travelling time increasing together with open space. For ex-
ample, comparing ID2080 and ID16563, higher block heights resulted in
higher open space but with longer travelling times. The Traveling Time ver-
sus Infrastructure Cost graph in Figure 5 (bottom) shows a general trend of
travelling times increasing as the cost of transportation network decreases.
For example, comparing ID1251 and ID15998, the very dense and extensive
bus network resulted in higher infrastructure costs and lower travelling times.
4. 2. URBAN TYPOLOGIES
Analysis of the final population of evolved design resulted in the identifica-
tion of three main urban typologies, referred to as ‘compact town’, ‘segre-
gated town’ or ‘stretched town’, as shown in Figure 6.
A compact town has a single compact built-up area surrounded by open space.
It tends to favour travelling time over open space.
A segregated town has open spaces separating clusters of built-up areas. It
tends to favour open space over travelling time.
A stretched town has a built-up area that meanders through the site. It tends
to find a balance between travelling time and open space.
URBAN STRUCTURE WITH EVOLUTIONARY ALGORITHMS 241
Figure 5. Graphs of Traveling Time versus Infrastructure Cost (top)
and Traveling Time versus Open Space (bottom)
Figure 6. Compact Town (left), Segregated Town (middle), Stretched Town (right)
242 E. CHIAN AND P. JANSSEN
This paper proposed an evolutionary approach to exploring urban configura-
tions. The exploration process resulted in a population of urban configura-
tions, where each configuration represents alternative trade-off between a set
conflicting performance criteria. Within this population, clusters of solutions
were identified with common typologies. Three differing typologies were
identified as possible starting points for more detailed design development
taking into account a broader set of social, environmental, and economic is-
In order to be able to effectively test the potential of the proposed ap-
proach, a highly simplified urban planning scenario was used. Future re-
search will start to add further complexity to this scenario in three main ways.
First, with regards to the urban configurations being evolved within the site,
the evolutionary procedures will be enhanced to be able to manipulate a
richer set of programs, focusing in particular on facilities that play a signifi-
cant role in people’s daily lives, such as schools and malls. Second, with re-
gards to the conditions within the site, the evolutionary procedures will be
enhanced to take into account site features that impose constraints on the
proposed urban configurations, starting with topography and water-bodies.
Third, with regards to existing conditions adjacent to the site, the evolution-
ary procedures will be enhanced to take into account surrounding urban are-
as such as neighbouring towns and villages.
Balling, R., Taber, J., Brown, M. and Day, K..: 1999, Multiobjective Urban Planning Using
Genetic Algorithm in Journal of Urban Planning and Development 125 (2): 86–99.
Janssen, P., Basol, C. and Chen, K. W.: 2011, Evolutionary Developmental Design for Non-
Programmers, proceedings of 29th eCAADe Conference, University of Ljubljana, Slove-
Janssen, P., Chen, K. W. and Cihat, B.: 2011, Iterative Virtual Prototyping: Performance
Based Design Exploration, proceedings of 29th eCAADe Conference Proceedings, Uni-
versity of Ljubljana, Slovenia, 253-260.
NUS: 2009, "Vertical Cities Asia". Available from: Open Source Repository <
http://www.verticalcitiesasia.com/> (accessed 17 February 2014).
Pedersen, J, and VanMater, A..: 2013, Resource Driven Urban Metabolism – How Can Met-
abolic Scaling Be Used in Urban Design?, Proceedings of the 18th International Confer-
ence on Computer-Aided Architectural Design Research in Asia CAADRIA 2013, Singa-
pore , 561-570.
Rakha, T. and Christoph R.: 2012, GenerativeUrban Modeling: A Design Work Flow for
Westerman, H. and AUSTROADS: 1998, Cities for Tomorrow Resource Document: Integrat-
ing Land Use, Transport and the Enviroment, Austroads Incorporated.