You are on page 1of 13

European Journal of Operational Research 274 (2019) 1125–1137

Contents lists available at ScienceDirect

European Journal of Operational Research


journal homepage: www.elsevier.com/locate/ejor

Innovative Applications of O.R.

Integer programming for urban design


Hao Hua a,b,∗, Ludger Hovestadt c, Peng Tang a,b, Biao Li a,b
a
Key Laboratory of Urban and Architectural Heritage Conservation (Southeast University), Ministry of Education, China
b
School of Architecture, Southeast University, 2 Sipailou, Nanjing 210096, China
c
ETH, Zürich, Switzerland

a r t i c l e i n f o a b s t r a c t

Article history: We present an integer program (IP) for urban design that (1) maximizes the floor area; (2) fills building
Received 22 March 2018 volume with room templates; (3) encodes translational symmetry in urban layout; and (4) constructs eco-
Accepted 29 October 2018
nomical urban routes. Regardless that integer programming is intensively studied in operational research
Available online 3 November 2018
(OR), its role in solving geometrical and topological problems in urban design was overlooked. Based on a
Keywords: regular grid, our 0–1 IP formulates the sunlight-gain rules, which give urban sites their shapes, especially
Linear programming for residential projects. With predefined plot templates, the gross floor area (volume) within a given site
Urban design can be maximized under various sunlight requirements. Subsequently, the IP fills each building volume
Special ordered set with 2D/3D room templates. Finally, an IP-based algorithm constructs routes that connect all plots and
Translational symmetry the site’s entrances to public transportation. Both the classical Steiner tree model and the latest coverage
Steiner tree network model are extended to create reasonable routes. In addition, this work extends the concept of
special ordered sets (SOS) to encode translational symmetry in urban layouts. Encoding layout symmetry
can benefit from the solvers’ SOS2 mechanism in the Branch-and-Bound search algorithm. The results
indicate that traditional decision making for cities could be partially automated by IP and an abundance
of valid solutions are available for designers.
© 2018 Elsevier B.V. All rights reserved.

1. Introduction 1.1. Urban design

Making cities efficient, sustainable, and livable has become a Urban design arranges building groups, street networks and
critical social issue. Governments have been encouraging and fund- public spaces. Urban designers focus on the layout of urban sites,
ing research projects to design better urban environments with as a composition of buildings, terraces, routes, green spaces, etc.
state-of-art technologies. For example, the Space Syntax Network Urban design is at the interface between architectural design and
in the UK, the computer graphics team at Purdue University, and urban planning but is distinct form both disciplines (Moughtin,
the Future City Laboratory as a Singapore-ETH Zürich corporation 2007). Urban planning emphasizes on large-scale urban structures,
strove to apply computational intelligence to design problems and such as regional land use and transportation, and is less concerned
decision making arising in cities. Sophisticated evaluation models with individual site layouts. Architectural design deals with indi-
(e.g. foot traffic simulation and evaluation), interactive procedural vidual buildings, usually after the building’s envelope is specified
modeling (Smelik, Tutenel, Bidarra, & Benes, 2014; Talton et al., by urban design. Rowley (1994) pointed out that the results of ur-
2011), inverse modeling and reconstruction (Aliaga, Vanegas, & ban design tend to be more relativistic and less deterministic than
Benes, 2008; Bokeloh, Wand, & Seidel, 2010), and layout synthe- architecture, but more definite than planner’s prescription.
sis (Merrell, Schkufza, & Koltun, 2010; Rodrigues, Gaspar, & Gomes, Today’s urban design inevitably involves a multitude of
2013) have lately inspired many designers and architects. rules, stipulations, and design principles. There are both aes-
thetic/idealism principles and performance-based principles.
Schenk (2013) identified the qualitative traits (symmetry, gestalt
cognitive laws, proportions and so on) and various measurable
performances (functions of buildings, accesses to natural resources,

circulation efficiency, etc.). This work deals with the following as-
Corresponding author at: School of Architecture, Southeast University, 2
Sipailou, Nanjing 210096, China. pects of urban design: (1) gross floor area; (2) the percentages
E-mail addresses: hua@arch.ethz.ch, whitegreen@163.com (H. Hua). of distinct building/room types; (3) sunlight gain; (4) routes and

https://doi.org/10.1016/j.ejor.2018.10.055
0377-2217/© 2018 Elsevier B.V. All rights reserved.
1126 H. Hua, L. Hovestadt and P. Tang et al. / European Journal of Operational Research 274 (2019) 1125–1137

circulation; and (5) symmetry in layout. The first four aspects are 1960). However, we discovered a connection between SOS2 and
functional while the fifth aspect emphasizes geometric patterns. translational symmetry in layout.
Architects have been enthusiastic in developing computational We regard SOS2 as both a mathematical concept and a data
models for urban design, especially after the 20 0 0s. For instance, structure. As an abstract model, it represents translational sym-
the Architectural Association School of Architecture (AA School) in metry within the IP framework. As a data structure, SOS2 gives
London pioneered parametric urbanism (Leach, 2009; Schumacher, the B&B algorithm a cleverer way to handle the search procedure
2009); the Berlage Institute at Rotterdam investigated associative (Ernee Filho, 2014). Most commercial solvers have implemented
design and synthetic vernacular (Trummer, 2008); and Computer- the SOS2 constraints.
Aided Architectural Design (CAAD) at ETH Zürich advanced self-
organization methods in urban design (Hovestadt, 2009). However, 1.4. Contributions
to date, few works have been made for applying mathematical
programming to urban design. This paper’s contributions include:

1.2. Integer programming 1. Using an IP to create urban layout that maximizes gross floor
area, subject to desirable percentages of distinct building/room
Linear programming (LP) is a well-developed technique for types and sunlight gain (shadow-free rule and n-hour sunlight
optimizing linear objective functions, subject to linear equal- rule). The IP also packs 2D/3D room templates into the building
ity/inequality constraints (Dantzig, 2016). Given a problem, LP volumes.
completely represents both the goal (objective function) and the Our shadow-free model can be viewed as a variation of facility
constraints only with linear formula. In integer programming (IP) layout programs (Amaral, 2008b; Singh & Sharma, 2006) if the
(Jünger et al., 2009; Schrijver, 1998), the decision variables are re- noon shadows are appropriately transformed into the template
stricted to integers. This paper mostly involves 0–1 binary vari- formulations. While our n-hour model is novel as it develops
ables. The 0–1 IP is a thriving area of optimization because an OR logical treatment of shadows at different times during a
it is closely associated with combinatorial optimization (Wolsey day.
& Rinaldi, 1993). The 0–1 IP is classified as NP-hard and be- 2. Using the IP to construct economical routes that connect all
longs to Karp’s 21 NP-complete problems. Researchers have de- buildings and the site’s entrances to public transportation. Both
veloped efficient algorithms such as Branch-and-Bound (B&B) al- the classical Steiner tree model and the latest coverage net-
gorithm (Conforti, Cornuejols, & Zambelli, 2014) and the cutting work model (Peng et al., 2016) are extended to create reason-
plane method for solving IP. Nowadays, cross-platform commer- able routes. The straightness of routes is also considered.
cial solvers (Meindl & Templ, 2012) such as Gurobi and IBM CPLEX Compared with the coverage network of Peng et al. (2016),
often employ both methods within a Mixed-Integer Programming our formulation involves fewer types of variables/constraints to
(MIP) framework. achieve coverage and avoid islands. Peng et al. (2016) treated
Urban layout bears some similarity with the facility layout zig-zag and T-junction respectively, while our model addresses
(Amaral, 2006; Drira, Pierreval, & Hajri-Gabouj, 2007; Friedrich, the straightness of routes instead.
Klausnitzer, & Lasch, 2018), block layout for enterprises (Li & 3. Employing SOS2 to encode the translational symmetry in the
Smith, 2018), urban facility location (Farahani, SteadieSeifi, & layout. Cyclic SOS2 and composite SOS2 are defined to facilitate
Asgari, 2010), the Very Large Scale Integration (VLSI) layout (Kim the pattern formulation.
& Kim, 2003; Srinivasan, Chatha, & Konjevod, 2006). The facility
Our method has a three-level scheme: (1) the peripheral cod-
layout problem (FLP) involves the arrangement of given facilities
ing such as specifying the plot templates or choosing a regular lat-
so that the total cost is minimized (Heragu & Kusiak, 1991). Ei-
tice (for example, square or triangular). Using different templates
ther continuous variables (e.g., representing the relative position
or grids will not change the IP formulation; (2) the IP formulation;
of each facility) or boolean variables (e.g., representing whether a
and (3) solving the IP with common solvers. Table 1 compares the
facility occupies a specific cell in a regular grid) were employed to
results using Gurobi and IBM CPLEX. This paper mostly focuses on
model FLP (Amaral, 2006; Kothari & Ghosh, 2012). In urban facility
the IP and its references to the elements of urban design.
location problems, a major factor is the desirability of the facility
(Coutinho-Rodrigues, Tralhão, & Alçada-Almeida, 2012). Hammad,
Akbarnezhad, and Rey (2017) formulated a mixed integer non- 1.5. Paper structure
linear programming model to minimize noise pollution and net-
work congestion. Wu, Fan, Liu, and Wonka (2018) developed an Section 2 introduces the IP for shadow-free urban layouts. Var-
MIP-based model to make layout design for building interiors. The ious sunlight-gain rules are modeled with linear expressions. In
MIP has been a common solution approach to layout problems Section 3, we investigate the symmetry in the layout where plot
(Amaral, 20 08a; 20 08b; Hathhorn, Sisikoglu, & Sir, 2013; Singh & overlapping is allowed. We view such layouts as SOS2 layouts. In
Sharma, 2006). Our model employs boolean variables and is solved Section 4, we construct economical routes based on the Steiner
by common MIP solvers. A latest breakthrough in the field of com- tree and the coverage network.
puter graphics is integrating IP with urban street networks and
building floorplans based on a deformable grid (Peng et al., 2016; 2. Shadow-free layouts
Peng, Yang, & Wonka, 2014).
A plot typically presents one building and its surrounding area
1.3. Special ordered set (ground terrace, route, green space, etc.) (Meeda, Parkyn, & Walton,
2007). A site usually has a network of roads or paths that subdi-
In 0–1 IP, the special ordered set (SOS) is an ordered set of vari- vide the site into a set of plots (Lynch, Lynch, & Hack, 1984). So an
ables satisfying particular constraints. This paper works with the essential step in urban design is to place the desirable plots inside
special ordered sets of type 2 (SOS2): at most, two variables in the a given site subject to legal constraints and restrictions. Some re-
set can be non-zero, and if the two are non-zero they must be con- searchers regarded the road network as a major factor in shaping
secutive in the set. A historical motivation for developing SOS is to the urban structure (Peng et al., 2016; Samaniego & Moses, 2008),
model piecewise linear functions (Beale & Forrest, 1976; Dantzig, which will be elaborated in the Section 3.
H. Hua, L. Hovestadt and P. Tang et al. / European Journal of Operational Research 274 (2019) 1125–1137 1127

Table 1
The timing, number of B&B nodes, gap, and value of solving IPs with commercial solvers (Gurobi
& IBM CPLEX). The ∗ symbol means that the solution is illustrated in the image.

Program Time(s) Nodes Gap(%) Value Solver

Packing room templates (Fig. 4) <1 1 0 198 CPLEX



Packing room templates (Fig. 4) <1 1 0 198 Gurobi
Shadow-free layout (Fig. 3left) 5 164 0.2 2473.5 CPLEX

Shadow-free layout (Fig. 3left) 39 247 0 2473.5 Gurobi
Shadow-free layout (Fig. 3right) 180 17568 2.5 1557.9 CPLEX

Shadow-free layout (Fig. 3right) 180 3411 5.9 1497.9 Gurobi

n-Hour sunlight layout (Fig. 4) 30 3311 27.6 111 CPLEX
n-Hour sunlight layout (Fig. 4) 120 2137 29.9 111 Gurobi

n-Hour sunlight layout (Fig. 5) 30 43 48.5 57882 CPLEX
n-Hour sunlight layout (Fig. 5) 60 286 49.3 57882 Gurobi

SOS2 layout (Fig. 7left) 60 607554 199 101 CPLEX SOS2
SOS2 layout (Fig. 7left) 60 3361 0 105 CPLEX linear
SOS2 layout (Fig. 7left) 60 3538702 130 97 Gurobi SOS2
SOS2 layout (Fig. 7left) 60 254 2 105 Gurobi linear
Steiner tree (Fig. 4(e)) 180 11 0.8 142.8 CPLEX

Steiner tree (Fig. 4(e)) 74 158 0 142.8 Gurobi
Steiner tree (Fig. 5(e)) 600 1 1.9 151.3 CPLEX

Steiner tree (Fig. 5(e)) 215 8 1.1 150.3 Gurobi
CST (Fig. 7left) 230 1 0.1 179.1 CPLEX

CST (Fig. 7left) 98 1 0 179 Gurobi

subject to

xmn = 0, ∀(i, j ) ∈ Q (2)
mn∈A(i, j )∩P


xmn ≤ 1, ∀(i, j ) ∈ P (3)
mn∈A(i, j )∩P

The first constraint keeps the void pixels empty while the sec-
ond serves as the non-overlapping principle. The parameter (i, j) in
(2–3) could be a pixel from either P or Q, however, the subscript
of xmn is only defined on P. Thus, the constraint (2) must address
mn ∈ A(i, j) ∩ P. There are |P | + |Q | linear constraints. The boundary
Fig. 1. Left: The T-shape template consists of four pixels. The bottom pixel is se- constraint (2) is violated at position (1,7) in Fig. 1 because
lected as the reference point. Right: The site P is in white and the void pixels 
(boundary) Q are in light grey. Two templates overlap at (4,3). One template over- xmn = x17 + x06 + x07 + x08 = x06 = 1 = 0
laps with the boundary at (1,7). mn∈A(1,7 )∩P

The non-overlapping constraint (3) is violated at position (4,3) in


2.1. 2D/3D packing Fig. 1 because

xmn = x43 + x32 + x33 + x34 = x32 + x34 = 2
An elementary principle of arranging multiple plots in a site is mn∈A(4,3 )∩P
non-overlapping. In other words, the given plots should be com-
pactly packed within the given site without having them overlap. Filling a building volume with 3D room templates can also be
Let P denote the site, namely the pixels of buildable area on a reg- performed if the IP above replaces the pixels with voxels.
ular grid. A rectangular site could be represented by P = {(i, j )|1 ≤
i ≤ I, 1 ≤ j ≤ J} where i and j refer to the ith row and jth column 2.2. Maximize floor area with shadow-free rule
of the site lattice, respectively. The set Q identifies the pixels that
should be kept void. P ∩ Q = ∅ holds. The pixels in Q should en- An urban site often comes across several types of plot tem-
close that in P, see Fig. 1. plates. Inside a plot template, there are distinct parts such as build-
A plot template located at (i, j) is represented by a set of pixels ings and (ground) terraces. Suppose there are L types of plot tem-
A(i, j), namely the plot’s footprint. One pixel is selected as a refer- plates, we can characterize a plot template by
ence point, so A(i, j ) = {(i, j ), (i − 1, j − 1 ), (i − 1, j ), (i − 1, j + 1 )} Al (i, j) : the building’s footprint of the lth type plot, 1 ≤ l ≤ L.
describes the four-pixel template in Fig. 1(left). Each position in A Bl (i, j) : the south rooms which require sunlight (for example,
refers to the reference point’s relative position to every pixel of the the bedroom).
template. For instance, (i − 1, j ) refers to the pixel above the refer- Cl (i, j) describes the building’s shadow at noon. The shadow’s
ence point in the T-shape template (Fig. 1, left), since the reference dimension is conventionally reduced to the gap/height ratio stipu-
point’s relative position to this pixel is (−1, 0 ). lated by the local government.
Let binary variables xij , ∀ij ∈ P represent whether position (i, j) Dl (i, j) identifies the building’s additional extents, such as a
is occupied by the reference point of a template. Then, the layout ground terrace, courtyard, or garden.
problem is to B ⊆ A, A ∩ C = ∅ (the building’s footprint does not intersect with
maximize the building’s shadow), A ∩ D = ∅ hold.
 The height (number of stories) of the building is specified so
xi j (1) the shadows can be pre-calculated. The shadows of low build-
i j∈P ings (no more than six stories) are often officially reduced to a
1128 H. Hua, L. Hovestadt and P. Tang et al. / European Journal of Operational Research 274 (2019) 1125–1137

Fig. 2. A plot template characterizes the extent of the building (orange), the south rooms (green) that require sunlight, the shadows (dark grey), and ground terrace (cyan).
The shadow pixels are determined via ray trace given the sun position at noon. (For interpretation of the references to color in this figure legend, the reader is referred to
the web version of this article.)

gap/height ratio (Fig. 2). For high buildings, one needs to calculate native of (9) is to exclude terrace-terrace overlapping (if the tem-
the exact hour of sunlight duration within a day (midwinter). The plate’s terrace encloses its building):
shape of shadows is a function of the sun’s position within the day.

L 
A building’s gross floor area equals the area of a single floor (the xmnl ≤ 1, ∀(i, j ) ∈ P (10)
same as the footprint’s area in our cases) multiplying the number l=1 mn∈Dl (i, j )∩P
of stories. Typically, wl in the objective function equals (or be pro-
portional to) the number of stories of the lth type building so that The program in Fig. 4 employs this constraint to prevent the
the objective function (4) characterizes the gross floor area of the ground terraces of adjacent plots from overlapping each other.
entire site. Fig. 3 illustrates two examples of the shadow-free layouts con-
A common rule in arranging plots is that the south rooms structed by the IP. The grey areas in the plans (Fig. 3(b)(d)) depict
which require sunlight should not be covered by the shadows of the noon shadows of the allocated buildings. Thus, the amount of
any other buildings. Thus, the IP is to white pixels indicates how much space is left after packing the plot
maximize templates (buildings). The site is 100% used if there is no white
pixel in the generated layout. There is one white pixel in Fig(b),

L 
which indicates that the result is near optimum. The 3D envelopes
wl xi jl (4)
of the buildings are predefined for each plot template. The choices
l=1 i j∈P
on the lattices (rectangular, triangular, and others), the templates,
over binary variables xijl ∈ P. The xijl denotes whether the position and the sites can be separated from the IP.
(i, j) is occupied by the reference point of the lth template.
subject to 2.3. N-hour sunlight rule

xmnl = 0, ∀l, ∀(i, j ) ∈ Q (5)
A more precise way of evaluating the amount of sunlight gain
mn∈Al (i, j )∩P
is to trace the sun’s position within a day. Many countries have
stipulated special requirements for sunlight availability in dense

L  construction and for residential projects. For instance, Zürich em-
xmnl ≤ 1, ∀(i, j ) ∈ P (6)
ploys the so-called two-hour shadow rule, which only permits a
l=1 mn∈Al (i, j )∩P
building to cast shadows on residential buildings for (at most) two
hours in a day (Hovestadt, 2009). A variation around the east coast

L  of China is that the residual building must receive at least two-
xmnl ≤ 1, ∀(i, j ) ∈ P (7)
hour sunlight on the ‘worst’ day, that is, during the winter solstice.
l=1 mn∈[Bl (i, j )∪Cl (i, j )]∩P
The general principle of such n-hour rules is to bind the shadowed
 hours below a threshold.
ξl ≤ xi jl ≤ τl , ∀l (8) The program first pre-calculates the building’s shadows at each
i j∈P hour (or at a smaller time interval) for each plot template, then
There are L|Q | + 2|P | + L constraints. The constraints in (7) pre- constructs Clt (i, j ) which describes the shadowed pixels at a spe-
vent the south rooms from overlapping with shadows. The con- cific time t, 1 ≤ t ≤ T. We used ray trace to determine whether a
straint (8) specifies the minimum (and the maximum) number of pixel is shadowed by the building at time t. First, the sun’s position
each type of plots. As a result, the number of the lth type plot will in the sky at a particular time during a year follows a paramet-
between ξ l and τ l in the generated layout. An alternative of (8) is ric model. Second, each face of the building’s 3D envelope is pro-
to control the relative percentages. jected to the ground according to the sun position. Third, A pixel is
An additional constraint is counted as the lth building’s shadow (at time t) if it is inside any
projected faces (at time t) from the building. Finally, all resultant
 
L  shadowed pixels are saved as the constant arrays Clt (i, j ) before the
xmnl + xmnl ≤ 1, ∀l, ∀(i, j ) ∈ P (9)
IP starts.
mn∈Dl (i, j )∩P l=1 mn∈Al (i, j )∩P
The n-hour sunlight IP contains a set of binary variables ytij that
which excludes (ground)terrace-building overlapping. A ground tells whether position (i, j) is shadowed at time t. If the shadow at
terrace is illustrated in Fig. 2, sometimes a designer wants to pre- time t from any plot template covers the pixel (i, j), we assign 1
vent such terrace from being occupied by other facilities. An alter- to ytij . This logical condition can be written as yti j = OR{xmn |mn ∈
H. Hua, L. Hovestadt and P. Tang et al. / European Journal of Operational Research 274 (2019) 1125–1137 1129

Fig. 3. Fee-shadow layouts constructed by the IP. The grey areas in the plans depict the noon shadows of the buildings. The envelopes of the buildings in the 3D view are
predefined for each plot template. The square cell is 6 × 6 m (a–c). The side length of the triangular cell is 6 m (d–f).

Fig. 4. N-hour sunlight layout on a rectangular grid. The square cell is 7 × 7 m. (a)(c): pre-calculate the building’s shadows at each hour for each type of plot template.
Footprint A is marked in orange, south rooms B green, ground terrace D cyan. (b): The 3D room templates for the building in the plot template(a). (d): The 3D room
templates for the building in the plot template(c). (e): STPSP constructs the walking routes after the layout IP optimizes the buildings’ locations. (f): The shadow conditions
in the optimized layout. (g): 3D view of the layout. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Ct (i, j ) ∩ P }. Therefore, the IP maximizes the objective function 


T 
in (4) yti j + (T − α ) xmn ≤ T , ∀(i, j ) ∈ P (15)
subject to t=1 mn∈B(i, j )∩P

xmn = 0, ∀(i, j ) ∈ Q (11) The logical OR could be transformed into linear inequalities
mn∈A(i, j )∩P

0 ≤ Myti j − xmn ≤ M − 1, ∀t, ∀(i, j ) ∈ P
 mn∈Ct (i, j )∩P
xmn ≤ 1, ∀(i, j ) ∈ P (12)
mn∈A(i, j )∩P
where M = |Ct |.
Constraints in (14,15) mean that the shadowed hours must be
 below the threshold α if the position is occupied by a sunlight-
xmn ≤ 1, ∀(i, j ) ∈ P (13) demanding room. Imposing a slightly stronger constraint, one can
mn∈D(i, j )∩P
replace (14,15) with

yti j = OR{xmn |mn ∈ Ct (i, j ) ∩ P }, ∀t, ∀(i, j ) ∈ P (14) xmn ≤ M, ∀t, ∀(i, j ) ∈ Q (16)
mn∈Ct (i, j )∩P
1130 H. Hua, L. Hovestadt and P. Tang et al. / European Journal of Operational Research 274 (2019) 1125–1137

Fig. 5. n-hour sunlight layout on a triangular grid. The side length of the triangular cell is 7 m. (a)(c): pre-calculate the building’s shadows at each hour for each type of plot
template. Footprint A is marked in orange, south rooms B green, ground terrace D cyan. (e): STPSP constructs the walking paths after the layout IP optimizes the buildings’
locations. (f): The shadow conditions in the optimized layout. (g): 3D view of the layout. (For interpretation of the references to color in this figure legend, the reader is
referred to the web version of this article.)


T  
xmn + (MT − α ) xmn ≤ MT , ∀(i, j ) ∈ P (17)
t=1 mn∈Ct (i, j )∩P mn∈B(i, j )∩P

where 1 ≤ M ≤ |Ct | (usually set M to an estimated value, e.g., 2).


 mn ∈ B(i, j) xmn is limited to 0 or 1 by extents-extents exclusion. No-
tice that the constraints in (16,17) no longer involves the auxiliary
variable ytji .
Therefore, the n-hour sunlight IP for multiple plot templates is
to
maximize

L 
wl xi j (18)
l=1 i j∈P

where the weight wl typically equals Hl |Al |. Hl denotes the number


of stories in the lth plot template.
subject to (5)(6)(10) and Fig. 6. Left: the ordered elements of the Knight8 template. Within its SOS2 layout,
any overlapped pair of templates must be associated with a horse-jump translation.

L 
xmnl ≤ M, ∀t, ∀(i, j ) ∈ Q (19) Right: the ordered elements of the Lemniscate16 template. Within its CS2 layout,
any overlapped pair of templates must be associated with a two-step diagonal or
l=1 mn∈Clt (i, j )∩P vertical translations.


T 
L  
L 
xmnl + (MT − α ) xmnl ≤ MT , ∀(i, j ) ∈ P
t=1 l=1 mn∈Clt (i, j )∩P l=1 mn∈Bl (i, j )∩P
of social ranks, and the behavior of citizens. In terms of model-
(20)
ing techniques, shape grammar, L-system, and procedural models
Constraints (19–20) are multi-type extension of constraints have dominated the landscape of symmetry modeling (Aliaga et al.,
(16–17). 2008; Bokeloh et al., 2010; Hua, 2017). However, we discovered a
Figs. 4 and 5 illustrate two examples of layout construction un- “hidden” mechanism to encode translational symmetry in SOS2-
der the n-hour sunlight rule, on a rectangular grid and on a trian- based IP. What we need to do is to manipulate the order in which
gular grid, respectively. The illustrated results involve three IPs: the the elements (pixels) of a template are presented to the IP.
layout IP (n-hour sunlight rule), the 3D room templates packing IP, Modeling urban layouts with SOS2 allows two plots to overlap.
and the circulation IP (Section 4). The 3D packing and circulation As a result, a position (pixel) in the site can be (1) unoccupied,
are performed after the layout IP optimizes the buildings’ locations. (2) occupied once, and (3) occupied twice by the plot templates.
How the three situations correspond to the presence of building,
3. Symmetry and SOS2 (ground) terrace, or other spaces is left to who designs the plot
templates. Two examples of such mapping are given in Fig. 7.
Symmetry is a constant topic of urban and architectural design Since the SOS2 constraint only allows two consecutive binary
in both theory and in practice. Schenk (2013) regarded the sym- variables to be 1, the rules for overlapping templates are implic-
metry as an essential quality when evaluating or designing an ur- itly encoded by the order in which the template’s pixels are listed
ban area. The urban symmetry has a powerful psychology effect in A. In contrast to the IP in Section 2 which is invariant to the
and was thought invincible in reality (Lynch, 1984). The symme- predefined order of elements of a template, such order defined in
try may reinforce the organization of government, the disposition SOS2 IP informs the displacements of plot templates in the layout.
H. Hua, L. Hovestadt and P. Tang et al. / European Journal of Operational Research 274 (2019) 1125–1137 1131

Fig. 7. Urban layouts constructed with Knight8 (left) and Hexa9 (right). The vehicle route network is constructed via CST before the arrangement of plot templates.

As a result, bringing a specific symmetry into the layout can be symmetry. Take the eight-element template in Fig. 6(top left) for
achieved at no additional expense. example. We arrange the elements in the sequence of the knight
move (as in chess). We named this plot template with ‘Knight8’:
3.1. Generalizing SOS2
A(i, j ) = {(i, j − 1 ), (i − 1, j + 1 ), (i + 1, j ), (i − 1, j − 1 ),
In IP formalism, one role of SOS2 is addressing constraints: no (i, j + 1 ), (i + 1, j − 1 ), (i − 1, j ), (i + 1, j + 1 )}
more than two variables in the set may be non-zero, with the fur-
The SOS2 layout program
ther condition that if there are two they must be adjacent (Beale &
maximizes
Forrest, 1976; Ernee Filho, 2014). Another role is offering a special 
data structure that facilitates the B&B search algorithm. This work xi j (21)
focuses on the confluence of the first role with translational sym- i j∈P
metry, as most MIP solvers have implemented SOS2 algorithms.
subject to
The SOS2 model was developed for modeling piecewise lin-

ear functions. One way to formulate nonlinear functions is by xmn = 0, ∀(i, j ) ∈ Q (22)
adding binary variables and new inequalities to the model, yield- mn∈A(i, j )∩P
ing a mixed-integer program (MIP) (Keha, de Farias, & Nemhauser,
2004). The well-known MIP formulations of piecewise linear func-
SOS2{xmn |mn ∈ A(i, j )}, ∀(i, j ) ∈ P (23)
tions include the incremental model and the convex combination
model (Croxton, Gendron, & Magnanti, 2003; Vielma, Ahmed, & If (m, n) ∈ P occurs for some (m, n) ∈ A(i, j), one can place a free
Nemhauser, 2010). The SOS2 formulation is similar to convex com- binary variable at the corresponding position in the ordered set.
bination, however, the nonlinearity is implemented in the B&B This program yields the layout in Fig. 7(left) with the MIP
algorithm of solver. Given a set of binary variables {xi }, the nota- solver’s built-in SOS2 function. We can see that the relative dis-
tion SOS2{xi } means that the set {xi } satisfies the SOS2 constraints. placement between every pair of overlapping templates complies
In terms of conventional linear expressions, SOS2{xi } means with the horse-jump pattern, because
(Dantzig, 1960)
⎧ 1. Only two consecutive positions can be simultaneously activated
⎪x ≤y
⎪x1 ≤ y1 + y

(which causes overlapping) by SOS2.

⎨ 2 1 2 2. All the consecutive positions are predefined in a horse-jump se-
xi ≤ yi−1 + yi quence.
⎪ ···


⎪ ≤ yI−1
⎩xI I−1 The following example explains the horse-jump mechanism. Let
i yi = 1 (i, j ) = (0, 0 ) in (23), then the SOS2 constraint says that only the
consecutive pair in
where yi are auxiliary binary variables. It is natural to introduce a
cyclic version CS2{xi } {x0,−1 , x−1,1 , x1,0 , x−1,−1 , x0,1 , x1,−1 , x−1,0 , x1,1 }

xi ≤ yi + y(i+1)%I , i = 1, 2, . . . , I can equal 1. All the possible pairs are
I
i yi = 1
1. x0,−1 = x−1,1 = 1.
where % denotes the modulo operation. I is the set’s cardinality. 2. x−1,1 = x1,0 = 1.
One can further extend the SOS2 concept to a set of sets. 3. x1,0 = x−1,−1 = 1.
CS2{{x11 , x12 , . . .}, {x21 , x22 , . . .}, . . .}, namely composite cyclic SOS2, 4. x−1,−1 = x0,1 = 1.
imposes the following constraints 5. x0,1 = x1,−1 = 1.

∀i, j 6. x1,−1 = x−1,0 = 1.
ij ≤
x yi j + yi,( j+1)%J , 7. x−1,0 = x1,1 = 1.
i j yi j = 1

where yij are binary variables. Here we assume that the cardi- The relative displacement in each pair follows the horse-jump
nalities the of subsets are the same (equal J). SOS2{{}} has the pattern, e.g. (0, −1 ) against (−1, 1 ), or (−1, 1 ) against (1, 0). One
same structure as CS2{{}} except for the cyclic feature. The clas- can see that the resultant displacement between overlapping tem-
sical SOS2{xi } is a special case of composite SOS2{{xi }}. plates is a function of the predefined order of the elements in the
template.
3.2. Encoding translational symmetry If the horse-jump sequence is a closed tour, then we should use
CS2 (cyclic) constraints instead of the standard SOS2. How to find
The SOS2 mechanism and the user-defined order of elements such a knight-move sequence that exactly fills an arbitrary pixe-
in a template lead to an implicit formulation of translational lated template? This is the classic Knight’s tour problem.
1132 H. Hua, L. Hovestadt and P. Tang et al. / European Journal of Operational Research 274 (2019) 1125–1137

The Lemniscate16 template (Fig. 6 top right) illustrates the ap-


plication of the composite SOS2. The template consists of 16 pixels
organized into four groups. Lemniscate16 is defined by

Ak (i, j ) = {(i − p, j − q ), (i − 2 − p, j − q ),
(i − p, j − 2 − q ), (i − 2 − p, j − 2 − q )},
p = (k − 1 )%2, q = (k − 1 )/2
Within each subset Ak (i, j) of four pixels, the sequence follows
a lemniscate-shape pattern. Then the layout program employs

CS2{{xmn |mn ∈ A1 (i, j )}}, {xmn |mn ∈ A2 (i, j )}, . . .}, ∀(i, j ) ∈ P
which allows two templates to overlap (Fig. 6 bottom right) with
either a diagonal or a vertical translation.
The SOS2 encoded symmetry allows for a certain degree of free-
Fig. 8. Both Steiner trees have the same length while the one on the right prefers
dom in regulating the pattern of covering. Firstly, a plot’s position
straight routes. The source is marked in green and sinks in blue. (For interpretation
is not SOS2 constrained if it does not overlap others. Secondly, a of the references to color in this figure legend, the reader is referred to the web
SOS2 structure may prescribe multiple transitional rules for over- version of this article.)
lapped plots (for example, the Lemniscate16 permits relative trans-
lations at 45°, 90°, and 135°). Finally, what the overlapped parts
2|E| integer variables dij (for coverage network), directed, de-
mean is an essential part of the template design. The layout au-
notes the distance from the edge (i, j) to a terminal. The lower
tomation program still leaves much of the design decisions up to
bound of dij is 0, the upper bound is the longest among the short-
designers.
est distances between any pair of nodes.
2K|E| binary variables fikj (for Steiner tree), denotes the flow of
4. Route network
commodity k from ith node to jth node.
Notice that zij corresponds to the original graph (undirected)
The circulation between the plots (buildings) is planned before
while yij renders all edges as bi-directional. Our flow formalism
or after the layout composition is computed by the IP in Section 2.
imposes
The grid P, excluding the parts occupied by buildings, presents
a graph from which the routes will be identified to connect all yi j + y ji = zi j , ∀(i, j ) ∈ E (24)
buildings (terminals). This task raises classical topics such as the which
minimum spanning tree (MST) and the network flow problems.
Another significant development, namely network design from 1. prevents two flow goes through an edge with opposite di-
functional specifications (or coverage network), was recently pro- rections, because yi j = y ji = 1 leads to yi j + y ji = 2 = zi j (notice
posed by Peng et al. (2016). We treat both the Steiner tree and the that zij ≤ 1).
coverage network as a flow formulation of network. The major dif- 2. imposes zi j = 1 when either yij or yji equals 1, i.e., makes an
ference between the Steiner tree and the coverage network is that undirected edge active if there is any flow goes through it.
the former connects a set of predefined destinations (nodes) while
the latter makes any node within a certain range of the routes. 4.1. Straightness
They are suited to different scenarios in urban design.
An urban designer is usually concerned with two practical cri- Urban design prefers straight paths because people do not like
teria. First, the routes should allow the inhabitants to conveniently detours - whether a segment of a route is straight on a rectangular
walk from one building to other buildings (interior pairwise con- grid can be identified by znI + znIII and znII + znIV (see Fig. 8, bottom
nections). Second, the routes should provide easy access to urban left). zij refers to an edge by the indices of the two end nodes;
traffic by reaching certain boundaries of the site (interior–exterior while znI , znII , . . . refers to an edge by a node and the directions. The
connections). Differing from classical graph-based problems, a de- IP program does not create variables for znI , znII , . . . , as they refer to
sirable circulation demands both topological and geometric fea- the same set of variables {zij }.
tures: If znI + znIII equals 1, there is a right-angle turn or a dead end;
otherwise (0 or 2), the path is straight or there is no path. It is the
1. The walking path network must connect all buildings and cer- same for the term znII + znIV . To count the number of turns in the
tain predefined boundaries of the site. tree, the IP introduces the binary variables
2. The routes should be as straight as possible. This is equivalent
to minimizing the number of the routes’ turns on a regular grid. hn = znI XOR znIII
This geometric property is beyond the scope of graph theory.
∀n ∈ V
vn = znII XOR znIV
First, a graph (V, E) is constructed from the grid P (see examples If the solver does not support the XOR logical constraint, one
in Figs. 4(e) and 5(e)). People can either regard the cells (pixels) can use the following equalities instead
as the nodes (like a chess board) or view the intersection points
as the nodes (like the Go game). Given an undirected graph (V, E) znI + znIII = hn + 2mn
and a subset of K nodes as the terminals, a flow formulation of the ∀n ∈ V (25)
znII + znIV = vn + 2on
route network typically includes:
|E| positive weights wi j , denotes the Euclidean length of the where mn , on are auxiliary binary variables.
edge (i, j). Here we let wi j = 1, ∀(i, j ) ∈ E. The triangular grid pertains to a similar formulation with a dif-
|E| binary variables zij , undirected, denotes whether the edge (i, ferent index method (see Fig. 8, top left).

j) is in the route network. To impose straightness, one can add a term n∈V (hn + vn ) in
2|E| binary variables yij , directed, denotes whether the directed the objective function. This technique is employed in our exten-
edge (i, j) is in the route network. sions of the Steiner tree and the coverage network.
H. Hua, L. Hovestadt and P. Tang et al. / European Journal of Operational Research 274 (2019) 1125–1137 1133

4.2. Steiner tree and extensions

Given an undirected graph and a subset of vertices called termi-


nals, the Steiner tree is a minimum weight subgraph that contains
all terminals. We choose an arbitrary terminal as the source and
other terminals as sinks. The terminals are arranged as the first
K + 1 nodes of the graph. Index 0 denotes the source, then the
sinks are indexed with k, 1 ≤ k ≤ K. In an urban layout, one build-
ing is the source, while other buildings (and the site’s entrances)
serve as sinks. One can imagine there are a group of people walk-
ing from the source towards each sink. The total foot travelling dis-
tance should be minimized.
The flow formulation renders the edges bi-directional. If E∗ de-
notes the set of directed edges, |E ∗ | = 2|E | holds. The multicom-
modity flow IP (Chopra & Tsai, 2001; Polzin & Daneshmand, 2001)
minimizes

wi j zi j (26)
i j∈E

subject to (24) and




−1 if j = 0
fikj − f ji =
k
1 if j = k ∀k, ∀ j ∈ V (27)
i j∈E 0 otherwise

fikj ≤ yi j ∀k, ∀i j ∈ E ∗ (28)


Fig. 9. The Steiner trees in (a) and (b) have the same length while (b) keeps the
(27) presents K|V| constraints. (28) presents 2K|E| constraints source-sink distances below a threshold so that the distance from the source to A
(edges are bi-directed). (24) presents |E| constraints. and B are significantly reduced. The tree in (c) controls straightness and source-
The straight Steiner tree sink distances. Adding pairwise shortest paths (C to C’, D to D’, E to E’) leads to the
minimizes network in (d).
 
zi j + s ( hn + vn ) (29)
i j∈E n∈V 3. For each pair in {(i, j)}, Do
a. run a shortest path IP to find a short circuit between the ith
subject to (25). s is a constant weight for imposing straightness. and jth nodes.
Source-sink distances. The Steiner tree characterizes the length b. add the new path into T (the resulting edge network).
of the entire tree but ignores the lengths of pairwise paths. As the
distance from the source to the kth sink is simply f ikj + f jik , one The shortest path IP is a special case of the Steiner tree IP with
strategy for controlling source-sink distances is (1) choosing one one source and one sink. Given a Steiner tree T, the pairwise IP
terminal around the site’s center as the source, and (2) imposing minimizes

  
fikj + f jik ≤ Mk , ∀k (30) zi j − t zi j + s ( hn + vn ) (31)
i j∈E i j∈T n∈V
i j∈E
subject to (24,25,27,28,30).
where Mk is a positive threshold for each sink. If the threshold is
where t is the constant weight for the coincidence between the
too small, the IP is infeasible; if the threshold is too big, this IP
new path and the given tree T. A typical value of t is 0.1. The
is reduced to the standard Steiner tree. The consequence of regu-
second term tends to overlap the added path with the existing
lating source-sink distances are illustrated in Fig. 9. Combining the
Steiner tree. An illustration of adding pairwise paths are shown in
constraints on straightness and on source-sink distances leads to
Fig. 9(d). The routes in Figs. 4 and 5 are also constructed by the
more reasonable routes in urban layouts.
STPSP method.
Steiner tree with pairwise shortest paths (STPSP). An essential
Coupled Steiner trees (CST). If there are two flows travelling
property of the Steiner tree is that it excludes cycles. Our program
through distinct paths from the source to a sink, then there is a
transforms certain dead ends (the leaves of the Steiner tree) into
cycle through the source and the sink. Based on this insight, we
cycles by adding pairwise shortest paths after the Steiner tree is
can construct a couple of Steiner trees to create cycles instead of
constructed. The program identifies a pair of terminals whose Eu-
dead-end branches. We termed this model with Coupled Steiner
clidean distance is small but the travelling distance through the
trees (CST). To distinguish the two trees in the formulation, we de-
tree is large. The added paths should coincide with the Steiner tree
note the variables of the male tree with zij and the female with
as much as possible, so the number of added segments is mini-

mized. zi j .
Given a Steiner tree T as a set of edges, the pseudo-codes of The CST program
adding pairwise paths are minimizes
 
1. Construct a set of node pairs {(i, j)} where dij > lij β . lij denotes (zi j + zi†j + ρ ci j ) + s ( hn + vn ) (32)
the Euclidean distance between the ith and jth nodes, while dij i j∈E n∈V

denotes the length of i − j path through the tree. β is a con- subject to (24,25,27,28,30) and the auxiliary variables cij satisfy
stant threshold.
2. Sort the pairs by dij in descending order. ci j = zi j ∧ zi†j , ∀i j ∈ E (33)
1134 H. Hua, L. Hovestadt and P. Tang et al. / European Journal of Operational Research 274 (2019) 1125–1137

Fig. 10. Two series of Coupled Steiner trees (CST) tests. One tree is marked in black, the other in red. Negative ρ enforces coincidences between the two trees, while positive
ρ prevents coincidences. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

The first term in the objective function governs the coincidence The set N(l) denotes the edges which should be covered by
of the two trees. A negative ρ enforces coincidences between the the lth node. So (37) requires that the topological distance of any
two trees, while a positive ρ will reduce coincidences. Two exam- node to active edges should be within a threshold. This is why we
ples are shown in Fig 10. One tree is marked in black, the other in named this model with coverage network. The value of the cover-
red. If ρ =0, the program results in two independent Steiner trees. age range (distance threshold) is critical (see Fig. 11).
Compared with original formulation of Peng et al. (2016), our
4.3. Coverage network program involves fewer variables (the original involves five types
of variables while our program has three: zij , yij and dij ) and fewer
The coverage network takes a distinct approach to networking constraints (the original has six kinds of constraints while our pro-
planning. It constructs a connected subgraph which reach all nodes gram has four) to achieve (1) the no-island goal, and (2) the cov-
within a predefined range (Peng et al., 2016). So the network can erage goal. To make the routes as straight as possible, one can use
evenly cover the entire site even with a single terminal. The objec- constraints in (25) and
tive prefers a network with (1) smaller total length; and (2) short minimize
distances to the terminals (similar to the constraints in (30) of our   
zi j + ρ di j + s ( hn + vn ) (38)
Steiner tree model).
i j∈E i j∈E ∗ n∈V
Given an undirected graph (V, E) and the terminals indexed as
the first K nodes, the coverage network is to
4.4. Comparison
minimize
 
zi j + ρ di j (34) Both the coverage network and the Steiner tree have been
i j∈E i j∈E ∗ treated as a flow problem defined in a graph. The coverage net-
where E∗ denotes the set of directed edges. work uses a distance indicator dij (integer) for each directed edge
subject to (24) and to achieve global connectivity (i.e., no islands) while Steiner tree
employs a binary variable fikj for each directed edge (and for each
IFyi j = 1, di j − d jk ≥ 1, ∀i j ∈ E ∗ ∧ j > K (35) sink) to achieve global connectivity.
The coverage network is more suitable for the design tasks
 where the goal is to roughly cover the entire site and few spe-
y jk ≥ yi j , ∀i j ∈ E ∗ ∧ j > K (36) cific terminals are known. In this situation, the designer assumes
that the buildings can reach the route network via ground terraces
 within a distance threshold.
yi j ≥ 1, ∀l ∈ V (37) Comparing STPSP with CST, the former is more flexible since
i j∈N (l )
the designer may choose particular pairs to add the short circuits,
where (35) and (36) present roughly 2|E| constraints, receptively. while the latter systematically create cycles. In terms of the ur-
While (37) presents V constraints. The condition j > K means that ban/street circulations, STPSP which probably still contain dead
the node j is not a terminal. ends is more suitable for pedestrians, while CST which facilitates
The IF condition in (35) can be directly implemented by the round-trips is more appropriate for automobiles.
indicator function of MIP solvers, or transformed into inequality
by big-M method (Peng et al., 2016). If the edges are not of uni- 5. Results and discussions
form length, one should replace the inequality di j − d jk ≥ 1 with
di j − d jk ≥ wi j and replace the term  ij ∈ E zij in the objective func- We proposed an IP treatment of urban design, including

tion with i j∈E wi j zi j . shadow-free layout synthesis, room templates allocation, and cir-
The constraints in (36) mean that at least one of outward edges culation planning. Various design criteria and restrictions are com-
of the jth node (except for the terminals) should be active if there pletely modeled with linear expressions. Our method consists of
is an active inward edge (i, j). In other words, a non-terminal node three levels: (1) specification of the site and the plot templates; (2)
should allow a flow travels through it. the IP formalism; and (3) IP solving by professional solvers. Hence,
H. Hua, L. Hovestadt and P. Tang et al. / European Journal of Operational Research 274 (2019) 1125–1137 1135

Fig. 11. Coverage networks with distinct coverage ranges. Top row: without straightness objective. Bottom row: with straightness objective.

a particular urban design can be implemented without altering the (curve a), 216 meter2 (b), 288 meter2 (c), and 324 meter2 (d), re-
IP. spectively. We also tested the 2-hour IP for the high buildings (6-
story to 18-story) in a 49,980 meter2 site in Shanghai, using 294
5.1. Applications meter2 (e), 392 meter2 (f), 441 meter2 (g), 588 meter2 (h), and
784 meter2 (k) footprint, respectively.
The increasing conviction of AI now makes the automation of Fig. 12 (left) shows that the densities of the most generated lay-
urban/architectural design an eager expectation among architects. outs are between the maximum and the minimum values from
Our IP solves a number of common problems in urban design, the real world. Fig. 12 (right) indicates that (1) the FAR of the
such as maximizing floor area under sunlight demand, endow- generated layouts are around the empirical maximum, probably
ing the urban layout with symmetric patterns, filling the building because the IPs maximize the FAR while the real projects may
volumes with room templates, and planning interior routes. Man- choose a medium value of FAR to reconcile other criteria; and (2)
ual arrangement of these elements requires expertise and is time- The larger the building’s footprint is, the higher the FAR is.
consuming. The parameterization and automation of these tasks The designer may manipulate the settings of the plot template
can quickly report the statistics of a given site and visualize the to enhance the design space for a given project. For instance in the
near-optimal layouts. The interaction between investors, govern- shadow-free program (Section 2.2), one can design a template as
ments, inhabitants, and designers could be reshaped by the au- the void region (as a sunken courtyard) on the basement level. And
tomation programs. subsequently these sunken courtyards can interact with the build-
The IP of n-hour sunlight rule fits residential projects that em- ings above the ground. Fig. 13 shows such an application where the
phasize on the sunlight gain of individual rooms. We can also sunken courtyards do not overlap with the buildings’ footprints.
define proper ground terrace for each plot template (working
with constraints in (9) or (10)) to regulate the distances between 5.2. Implementation and timings
general-type buildings in high-density regions. The symmetry-
encoding mechanism based on SOS2 allows designers to manip- One appeal of modeling urban design problems with IP lies in
ulate translational rules when allocating plots/buildings. Although the clear separation of formulation and implementation. The latest
shape grammar and procedural models have been immensely em- professional solvers usually support a multitude of computer plat-
ployed for modeling geometric symmetry (Bokeloh et al., 2010; forms and programming languages, which makes the implementa-
Hua, 2017), this work presents an alternative approach using pure tion of IP easy. We implemented our programs using Gurobi and
0–1 IP. IBM CPLEX through their Java APIs and often adopted the default
Cheng, Steemers, Montavon, and Compagnon (2006) investi- configuration of their MIP functionality. The solvers take seconds
gated the complex relationship between high density, urban form, or minutes to find the optimal solution for many programs pre-
and solar potential. The randomness was elaborated: sented in this paper. For larger-scale IPs, we opt for deriving a rel-
atively good solution within minutes. Table 1 lists the timing of
1. Horizontal randomness (against uniform distribution of build-
some cases, on a 64-bit Windows computer with 3.6 Gigahertz i7-
ings in the layout) is preferable. Our generated layouts (Fig.
4790 CPU and 16 Gigabyte RAM.
3,4,5,7,13) are non-uniform.
The performances of Gurobi and IBM CPLEX seem similar. Both
2. Vertical randomness (buildings with different heights) is prefer-
solvers support SOS2 constraints. The performances of using SOS2
able. This is implemented in our IPs by employing plot tem-
and linear inequalities are similar.
plates with buildings of different heights (Fig. 3,4,5,7,13).
We compared the statistics of our generated layouts with the 5.3. Further work
empirical data from a survey of residential buildings in Shang-
hai (Yang & Chen, 2005). Fig. 12(left) compares the density and All models proposed in this paper are based on regular grid,
Fig. 12(right) compares the floor area ratio (FAR), namely the ra- which is a fundamental limitation of our IP approaches to ur-
tio of total floor area to the site area. The shadow-free IP is tested ban design. Although many current projects of urban design still
for the low buildings (2-story to 6-story) in a 45,288 meter2 site commence with a regular grid, our future work will integrate de-
in Shanghai, using uniform building with footprint of 114 meter2 formed grid (Peng et al., 2016). It would be interesting to consider
1136 H. Hua, L. Hovestadt and P. Tang et al. / European Journal of Operational Research 274 (2019) 1125–1137

Fig. 12. Left: comparison of the density of empirical data from Shanghai (the maximum and minimum curves in the figure) and that of the generated layouts. Right:
comparison of the FAR of empirical data from Shanghai and that of the generated layouts. Curve a-d: shadow-free layouts. Curve e-k: 2-hour sunlight layouts.

Fig. 13. Layout with sunken courtyards. Four terminals (marked with green dots) are pre-defined on the site’s boundary. The sunken courtyards (i.e., void space on the
basement level) marked in green do not overlap with buildings’ footprints. The buildings’ shadows are marked in grey in the plan. The coverage network program (coverage
range = 2) performs after the buildings and the courtyards are optimized. (For interpretation of the references to color in this figure legend, the reader is referred to the
web version of this article.)

the inverse engineering of exiting cities based on IP. Successful Hagihara, 2017) and simplex type algorithms (Ploskas & Samaras,
works on inverse modeling (Kalogerakis & Chaudhuri, 2012; Pauly, 2015) has recently become advanced. It would be rewarding to
Mitra, Wallner, Pottmann, & Guibas, 2008; Št’ava, Beneš, Měch, study the parallelization of SOS2 constraints. The SOS-based op-
Aliaga, & Krištof, 2010) train the parameter values of a paramet- erations that are native to the hardware are expected.
ric model from given samples. The inverse modeling of street net-
works (Aliaga et al., 2008; Fiser et al., 2016) is probably most close 6. Conclusions
to this purpose. However, encoding the hidden rules in terms of IP
(or other models of mathematical programming) is still very much This paper demonstrates that a series of tasks in urban design
a work in progress. Another future work is integrating GPU par- can be fully modeled by IP and subsequently solved by state-of-art
allel computing with algorithms for solving 1–0 IP. Classical algo- solvers. The multitude of design criteria, stipulations, and physical
rithms such as B&B were not tailored for special hardware, such rules have long delayed the mathematical programming of urban
as GPUs or quantum computers. The parallelization of B&B algo- design. However, our work indicates that both functional objec-
rithms (Borisenko, Haidl, & Gorlatch, 2017; Shen, Shigeoka, Ino, & tives (e.g. n-hour sunlight rule) and aesthetic principles (e.g. trans-
lational symmetry) could be transformed into linear expressions.
H. Hua, L. Hovestadt and P. Tang et al. / European Journal of Operational Research 274 (2019) 1125–1137 1137

Unlike many existing approaches that mix mathematical program- Hovestadt, L. (2009). Beyond the grid–architecture and information technology: appli-
ming with other optimization methods (Singh & Sharma, 2006), cations of a digital architectonic. Walter de Gruyter.
Hua, H. (2017). A bi-directional procedural model for architectural design. In Com-
our approach mostly employs 0–1 IP to handle various layout puter graphics forum: 36 (pp. 219–231). Wiley Online Library.
problems. Eventually, current MIP solvers will allow us to obtain Jünger, M., Liebling, T. M., Naddef, D., Nemhauser, G. L., Pulleyblank, W. R.,
near-optimal layouts for middle-scale sites (about 30,0 0 0 meter2 ) Reinelt, G., et al. (2009). 50 Years of Integer Programming 1958–2008: from the
early years to the state-of-the-art. Springer Science & Business Media.
within seconds or minutes. Kalogerakis, E., Chaudhuri, S., K. , V., & Koller, D. (2012). A probabilistic model for
We employ the plot template as the essential element in layout component-based shape synthesis. ACM Transactions on Graphics (TOG), 31(4).
synthesis. A plot can represent a building and its neighborhood; doi:10.1145/2185520.2185551.
Keha, A. B., de Farias Jr, I. R., & Nemhauser, G. L. (2004). Models for representing
the overlap between two plots may represent a building or a ter-
piecewise linear cost functions. Operations Research Letters, 32(1), 44–48.
race. Such flexibility constitutes a design space outside of the IP Kim, J.-G., & Kim, Y.-D. (2003). A linear programming-based algorithm for floorplan-
formalism. Therefore, the designers still play an essential role in ning in VLSI design. IEEE Transactions on Computer-Aided Design of Integrated
Circuits and Systems, 22(5), 584–592.
devising the plot template. The interplay between a designer and
Kothari, R., & Ghosh, D. (2012). The single row facility layout problem: state of the
the IP may yield novel results without modifying the IP algorithm. art. Opsearch, 49(4), 442–462.
Leach, N. (2009). Digital cities. Architectural Design, 79(4), 6–13.
Acknowledgment Li, J., & Smith, A. E. (2018). Block layout for attraction-based enterprises. European
Journal of Operational Research, 266(3), 1100–1112.
Lynch, K. (1984). Good city form. MIT press.
Thank Zifeng Guo at ETH Zürich for the introduction to Gurobi. Lynch, K., Lynch, K. R., & Hack, G. (1984). Site planning. MIT press.
This work was supported by the National Natural Science Foun- Meeda, B., Parkyn, N., & Walton, D. S. (2007). Graphics for urban design. Thomas
Telford.
dation of China [51778118, 51478116, 51578123]; the Ministry of Meindl, B., & Templ, M. (2012). Analysis of commercial and free and open source
Housing and Urban-Rural Development of China [UDC2017020212]. solvers for linear optimization problems p. 20. Eurostat and Statistics Netherlands
within the project ESSnet on common tools and harmonised methodology for
References SDC in the ESS
Merrell, P., Schkufza, E., & Koltun, V. (2010). Computer-generated residential build-
ing layouts. ACM Transactions on Graphics, 29(6), 485–501.
Aliaga, D. G., Vanegas, C. A., & Benes, B. (2008). Interactive example-based urban
Moughtin, C. (2007). Urban design: street and square. Routledge.
layout synthesis. In ACM transactions on graphics (tog): 27 (p. 160). ACM.
Pauly, M., Mitra, N. J., Wallner, J., Pottmann, H., & Guibas, L. J. (2008). Discovering
Amaral, A. R. S. (2006). On the exact solution of a facility layout problem. European
structural regularity in 3D geometry. ACM Transactions on Graphics (TOG), 27(3),
Journal of Operational Research, 173(2), 508–518.
43.
Amaral, A. R. S. (2008a). An exact approach to the one-dimensional facility layout
Peng, C.-H., Yang, Y.-L., Bao, F., Fink, D., Yan, D.-M., Wonka, P., & Mitra, N. J. (2016).
problem. Operations Research, 56(4), 1026–1033.
Computational network design from functional specifications. ACM Transactions
Amaral, A. R. S. (2008b). A MIP model for the layout problems with depart-
on Graphics (TOG), 35(4), 131.
ments of unequal area. In Xl sbpo simpósio brasileiro de pesquisa operacional
Peng, C.-H., Yang, Y.-L., & Wonka, P. (2014). Computing layouts with deformable
(pp. 2017–2028).
templates. ACM Transactions on Graphics (TOG), 33(4), 99.
Beale, E., & Forrest, J. J. (1976). Global optimization using special ordered sets. Math-
Ploskas, N., & Samaras, N. (2015). Efficient GPU-based implementations of simplex
ematical Programming, 10(1), 52–69.
type algorithms. Applied Mathematics and Computation, 250, 552–570.
Bokeloh, M., Wand, M., & Seidel, H.-P. (2010). A connection between partial symme-
Polzin, T., & Daneshmand, S. V. (2001). A comparison of Steiner tree relaxations.
try and inverse procedural modeling. In ACM transactions on graphics (tog): 29
Discrete Applied Mathematics, 112(1–3), 241–261.
(p. 104). ACM.
Rodrigues, E., Gaspar, A. R., & Gomes, Á. (2013). An evolutionary strategy enhanced
Borisenko, A., Haidl, M., & Gorlatch, S. (2017). A GPU parallelization of
with a local search technique for the space allocation problem in architecture,
branch-and-bound for multiproduct batch plants optimization. The Journal of
part 1: Methodology. Computer-Aided Design, 45(5), 887–897.
Supercomputing, 73(2), 639–651.
Rowley, A. (1994). Definitions of urban design: The nature and concerns of urban
Cheng, V., Steemers, K., Montavon, M., & Compagnon, R. (2006). Urban form, density
design. Planning Practice and Research, 9(3), 179–197.
and solar potential. In PLEA 2006. LESO-PB-CONF-20 06-0 08
Samaniego, H., & Moses, M. E. (2008). Cities as organisms: Allometric scaling of
Chopra, S., & Tsai, C.-Y. (2001). Polyhedral approaches for the Steiner tree problem
urban road networks. Journal of Transport and Land Use, 1(1), 21–39.
on graphs. In Steiner trees in industry (pp. 175–201). Springer.
Schenk, L. (2013). Designing cites: basic - principles - projects. Birkhäuser.
Conforti, M., Cornuejols, G., & Zambelli, G. (2014). Integer programming. Springer.
Schrijver, A. (1998). Theory of linear and integer programming. John Wiley & Sons.
Coutinho-Rodrigues, J., Tralhão, L., & Alçada-Almeida, L. (2012). A bi-objective mod-
Schumacher, P. (2009). Parametricism: A new global style for architecture and urban
eling approach applied to an urban semi-desirable facility location problem. Eu-
design. Architectural Design, 79(4), 14–23.
ropean Journal of Operational Research, 223(1), 203–213.
Shen, J., Shigeoka, K., Ino, F., & Hagihara, K. (2017). An out-of-core branch and
Croxton, K. L., Gendron, B., & Magnanti, T. L. (2003). A comparison of mixed-integer
bound method for solving the 0–1 knapsack problem on a GPU. In International
programming models for nonconvex piecewise linear cost minimization prob-
conference on algorithms and architectures for parallel processing (pp. 254–267).
lems. Management Science, 49(9), 1268–1273.
Springer.
Dantzig, G. (2016). Linear programming and extensions. Princeton university press.
Singh, S. P., & Sharma, R. R. (2006). A review of different approaches to the facility
Dantzig, G. B. (1960). On the significance of solving linear programming problems
layout problems. The International Journal of Advanced Manufacturing Technology,
with some integer variables. Econometrica, Journal of the Econometric Society,
30(5–6), 425–433.
30–44.
Smelik, R. M., Tutenel, T., Bidarra, R., & Benes, B. (2014). A survey on procedural
Drira, A., Pierreval, H., & Hajri-Gabouj, S. (2007). Facility layout problems: A survey.
modelling for virtual worlds. Computer Graphics Forum, 33(6), 31–50.
Annual Reviews in Control, 31(2), 255–267.
Srinivasan, K., Chatha, K. S., & Konjevod, G. (2006). Linear-programming-based tech-
Ernee Filho, K. (2014). Valid inequalities and computational results for SOS1-, SOS2-,
niques for synthesis of network-on-chip architectures. IEEE Transactions on Very
and cardinality-constrained linear programs, Ph.D. thesis. Texas Tech University.
Large Scale Integration (VLSI) Systems, 14(4), 407–420.
Farahani, R. Z., SteadieSeifi, M., & Asgari, N. (2010). Multiple criteria facility location
Št’ava, O., Beneš, B., Měch, R., Aliaga, D. G., & Krištof, P. (2010). Inverse procedural
problems: A survey. Applied Mathematical Modelling, 34(7), 1689–1709.
modeling by automatic generation of L-systems. Computer Graphics Forum, 29(2),
Fiser, M., Benes, B., Galicia, J. G., Abdul-Massih, M., Aliaga, D. G., & Krs, V. (2016).
665–674. doi:10.1111/j.1467-8659.2009.01636.x.
Learning geometric graph grammars. In Proceedings of the 32nd spring conference
Talton, J. O., Lou, Y., Lesser, S., Duke, J., Měch, R., & Koltun, V. (2011). Metropolis
on computer graphics (pp. 7–15). ACM.
procedural modeling. ACM Transactions on Graphics (TOG), 30(2), 11.
Friedrich, C., Klausnitzer, A., & Lasch, R. (2018). Integrated slicing tree approach for
Trummer, P. (2008). Engineering ecologies. Architectural Design, 78(2), 96–101.
solving the facility layout problem with input and output locations based on
Vielma, J. P., Ahmed, S., & Nemhauser, G. (2010). Mixed-integer models for nonsep-
contour distance. European Journal of Operational Research, 270(3), 837–851.
arable piecewise-linear optimization: Unifying framework and extensions. Oper-
Hammad, A. W., Akbarnezhad, A., & Rey, D. (2017). Sustainable urban facility lo-
ations Research, 58(2), 303–315.
cation: Minimising noise pollution and network congestion. Transportation Re-
Wolsey, L., & Rinaldi, G. (1993). Integer programming and combinatorial optimization.
search Part E: Logistics and Transportation Review, 107, 38–59.
CIACO.
Hathhorn, J., Sisikoglu, E., & Sir, M. Y. (2013). A multi-objective mixed-integer pro-
Wu, W., Fan, L., Liu, L., & Wonka, P. (2018). MIQP-based layout design for building
gramming model for a multi-floor facility layout. International Journal of Produc-
interiors. Computer Graphics Forum, 37(2), 511–521.
tion Research, 51(14), 4223–4239.
Yang, S., & Chen, W. (2005). Study on reasonable density of residence [j]. City Plan-
Heragu, S. S., & Kusiak, A. (1991). Efficient models for the facility layout problem.
ning Review, 3, 014.
European Journal of Operational Research, 53(1), 1–13.

You might also like