Concrete Structure Design Using Mixed Integer Nonlinear Programming with Complementarity Constraints,

Concrete Structure Design Using Mixed Integer Nonlinear Programming with Complementarity Constraints,
SIAM J. OPTIM.
Vol. 21, No. 3, pp. 833–863
© 2011 Society for Industrial and Applied Mathematics
CONCRETE STRUCTURE DESIGN USING MIXED-INTEGER
NONLINEAR PROGRAMMING WITH COMPLEMENTARITY
CONSTRAINTS*
A. GUERRA†, A. M. NEWMAN‡,
AND
S. LEYFFER§
Abstract. We present a mixed-integer nonlinear programming (MINLP) formulation to achieve minimum-cost designs for reinforced concrete (RC) structures that satisfy building code requirements. The objective function includes material and labor costs for concrete, steel reinforcing bars, and formwork according to
typical contractor methods. Restrictions enforce correct geometry of the cross-section dimensions for each
element and relative sizes of cross-section dimensions of elements within the structure. Other restrictions define
a stiffness and displacement correlation among all structural elements via finite element analysis. The design of
minimum cost RC structures introduces a new class of optimization problems, namely, mixed-integer nonlinear
programs with complementarity constraints. The complementarity constraints are used to model RC element
strength and American Concrete Institute code-required safety factors. We reformulate the complementarity
constraints as nonlinear equations and show that the resulting ill-conditioned MINLPs can be solved by using
an off-the-shelf MINLP solver. Our work provides discrete-valued design solutions for an explicit representation of a process most often performed implicitly with iterative calculations.We demonstrate the capabilities of
a mixed-integer nonlinear algorithm, MINLPBB, to find optimal sizing and reinforcing for cast-in-place beam
and column elements in multistory RC structures. Problem instances contain up to 678 variables, of which 214
are integer, and 844 constraints, of which 582 are nonlinear. We solve problems to local optimality within a
reasonable amount of computational time, and we find an average cost savings over typical-practice design
solutions of 13 percent.
Key words. complementarity problems, applications in optimization, mixed-integer nonlinear
programming
AMS subject classifications. 90C11, 90C30, 90C33, 90C90
DOI. 10.1137/090778286
1. Introduction. In this paper, we present a new class of problems of the following
form:
min f ðx; yÞ
subject to gi ðx; yÞ ≥ 0 ∀i ∈ I 1 ;
hi ðx; yÞ ¼ 0 ∀i ∈ I 2 ;
0 ≤ F i ðxÞ ⊥ xi ≥ 0 ∀i ∈ I 3 ;
y∈Y
integer;
where I 1 , I 2 , and I 3 are (distinct) sets containing indices of functions associated with
inequality, equality, and complementarity constraints, respectively; f ð·Þ, gð·Þ, and hð·Þ
are twice continuously differentiable nonlinear functions; and the complementarity constraints in I 3 indicate that F i ðxÞ and xi are nonnegative, and at most one of the two is
*Received by the editors November 25, 2009; accepted for publication (in revised form) April 4, 2011;
published electronically September 1, 2011. This work was supported by the Office of Advanced Scientific
Computing Research, Office of Science, U.S. Department of Energy, under contract DE-AC02-06CH11357
and through grant DE-FG02-05ER25694.
http://www.siam.org/journals/siopt/21-3/77828.html
†
Division of Engineering, Colorado School of Mines, Golden, CO 80401 ([email protected]).
‡
Division of Economics and Business, Colorado School of Mines, Golden, CO 80401 ([email protected]).
§
Mathematics and Computer Science Division, Argonne National Laboratory, Argonne, IL 60439 ([email protected]
mcs.anl.gov).
833
Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
834
A. GUERRA, A. M. NEWMAN, AND S. LEYFFER
positive. As such, we address a mixed-integer nonlinear program with complementarity
constraints. This type of problem arises in engineering design applications, specifically,
in the context of the design of reinforced concrete (RC) structures. RC possesses the
following important properties: (i) significant compressive strength that increases over
time, (ii) low maintenance, (iii) fire resistance, and (iv) constitution of inexpensive local
materials such as sand, gravel, and water. RC consists of large portions of sand and
gravel and smaller portions of cement and steel reinforcing bars. The cement and water
chemically interact to cohere the sand and gravel into a solid mass surrounding the reinforcing bars. Cast-in-place RC construction refers to methods used to fabricate structural elements in the intended design position and location. Wet concrete is placed into a
wooden or steel formwork that holds the concrete in place until it develops sufficient
self-supporting strength. Reinforcement is placed within the formwork before pouring
the concrete so that the concrete hardens around the reinforcement. The combination of
concrete and reinforcing bars provides elements that can withstand large forces. The
design of RC involves selecting dimensions consisting of discrete-valued element width
and depth, reinforcing bar sizes, and the number of bars ensuring structural integrity.
The goal of this paper is to provide a sound mathematical model and solution methodology that yield minimum cost designs of large and complex RC buildings.
The analysis procedures for RC that are typically adopted in practice assume a
structural system with a fixed initial stiffness. The demand on the RC elements in terms
of displacements and forces depends on the applied loads and relative stiffness of elements, where stiffness is a measure of displacement with respect to force. A fixed initial
stiffness distribution is necessary to calculate the demand, or internal forces, in a statically indeterminate structure. Engineers then design element dimensions to resist these
internal forces, which, unfortunately, are inconsistent with the internal forces associated
with the final design dimensions. This inconsistency creates unnecessarily expensive and
overengineered solutions.
An explicit formulation of the RC design problem improves the fidelity of the current-practice structural analysis by resolving inconsistencies between the initial design
assumptions and the final design dimensions. We develop an explicit formulation for the
RC design problem using continuous and continuously differentiable expressions to enforce relationships between decision variables. We represent Boolean expressions with
binary variables whose continuous relaxations yield differentiable expressions. An explicit formulation provides a feasible solution directly and completely computed without
an iterative procedure and, in our case, locally optimal solutions that can be easily evaluated. An explicit formulation also allows for the use of a robust solver on larger
problems that produce better-quality solutions in less computation time.
In our explicit mixed-integer nonlinear optimization model, we use integer variables
to represent the width and depth of each element and the number of reinforcing bars.
We employ binary variables to select discrete reinforcing bar sizes from a set of potential
sizes and to represent discrete decisions for formwork reuse. Continuous decision variables relate to applied forces and resistive capacity of each element. Formulating
the problem with integer variables allows for an objective function including material
and labor costs for concrete, steel reinforcing bars, and formwork in accordance with
typical methods used by contractors. Constraints or restrictions enforce (i) upper
and lower bounds on the variables, (ii) the correct geometry of the cross-section
dimensions for each element, (iii) the relative sizes of cross-section dimensions of
elements within the structure, (iv) finite element analysis equilibrium equations that
define the appropriate relationships between forces, displacements, and stiffness,
Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
CONCRETE STRUCTURE DESIGN
835
(v) the resistive capacities of each element, and (vi) bounds on the applied loads relative
to the resistive capacities of each element.
Our model provides an example of a new class of challenging optimization problems,
namely, mixed-integer nonlinear optimization problems with complementarity constraints. The complementarity constraints are necessary to model the resistive forces
provided by the concrete, elastic-perfectly plastic material response for the steel reinforcement, and the American Concrete Institute (ACI) [1] code-specified strength reduction factor and minimum axial loads for flexural elements. We reformulate the
complementarity constraints as nonlinear inequalities. This approach gives rise to a
degenerate mixed-integer nonlinear programming (MINLP) problem that violates
the Mangasarian–Fromowitz constraint qualification at any feasible point; see, for example, Scheel and Scholtes [41]. We show that the resulting MINLP can nonetheless be
solved reliably using a suitable off-the-shelf MINLP solver. We demonstrate the
capability of our explicit formulation to find lower-cost and more efficient solutions than
currently found in practice, extending RC design optimization.
The contributions of the paper lie in (i) defining a new class of problems; (ii)
demonstrating real-world applicability of this problem class; (iii) proposing an explicit
formulation for the real-world application; (iv) suggesting alternate formulations that
exhibit greater tractability than other common formulations; (v) developing heuristics
to enhance solvability of the reformulated model; and (vi) providing electronically available files (in AMPL format) to enable an interested reader to implement and test our
instances. The AMPL (Fourer, Gay, and Kernighan [22], [2]) model, data, and .nl files
for the examples we present in this paper and for unsolved instances can be found at
http://wiki.mcs.anl.gov/NEOS/index.php/RCMINLP. The remainder of this paper is
organized such that in section 2, we review relevant literature; in section 3 we present
the problem description and formulation. In section 4, we give model instances. Section 5
details the corresponding results, and section 6 concludes the paper.
2. Literature review. The use of optimization in RC design is not new. The first
instances of optimization techniques for RC structures were explicit methods to determine inelastic solutions with fixed element dimensions. Inelastic material behavior incorporates material nonlinearities, whereas elastic material behavior contains linear material
properties. De Donato and Maier [9] were among the first to minimize a quadratic function
subject only to sign constraints, incorporating inelastic material behavior for RC structural analysis. Prior to the problem posed by De Donato and Maier, an inelastic solution of
an RC structure with fixed element dimensions was difficult to obtain, especially for statically indeterminate problems that require discretization into finite elements in order to
accurately capture structural behavior (most RC design problems require discretization).
A study by Corradi, De Donato, and Maier [8] formulates and solves a linear complementarity problem by overrelaxation to find inelastic solutions for multistory frames
discretized into finite elements with fixed element dimensions. Kaneko [29], Maier,
Gianessi, and Nappi [37], and Dinno and Mekha [12] expand the use of complementarity
to perform inelastic analysis of RC structures with fixed element dimensions. When the
element dimensions are fixed, finite element methods model constitutive behavior using
mathematical operations on linear systems of equations. When element dimensions are
variables, finite element analysis becomes increasingly difficult because the system of
equations contains nonlinear functions of the element dimensions. We further discuss
finite element methods for frames in section 3.3.1.
While the methods previously discussed use mathematical programming techniques to model inelastic behavior for structures with fixed element dimensions,
Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
836
A. GUERRA, A. M. NEWMAN, AND S. LEYFFER
Krishnamoorthy and Mosi [30] and Dinno and Mekha [11] include element dimensions as
design variables and search for minimum cost solutions in addition to using complementarity constraints to define inelastic behavior of RC frames. However, to find solutions
for reasonably sized buildings, Krishnamoorthy and Mosi [30] include variables to describe the reinforcing bar areas and determine the width and depth of elements a priori.
Dinno and Mekha [11] include variables to describe the width and depth of elements and
to determine the reinforcing bar areas a priori. The objective functions in these studies
include simplified costs for concrete, formwork, and reinforcement that do not entirely
capture construction practices. Both papers enforce structural stability by ensuring that
resistive forces are greater than applied forces, which are determined by finite element
analysis. Reducing the number of design variables facilitates solving problems with a
larger number of story levels but requires implicit, or iterative, methods to find a feasible
solution because the number and size of reinforcing bars depend on the width and depth
of each structural element.
Ferris and Tin-Loi [18] present an explicit formulation to find minimum-weight solutions for truss-like structures with complementarity constraints that incorporate inelastic
behavior. We consider structures with a greater number of degrees of freedom than trusslike structures have. However, this work demonstrates the use of complementarity constraints in explicit structural optimization problems. Horowitz and Moraes [26] utilize a
mixed-integer nonlinear algorithm, MINLPBB, to solve a problem with complementarity
constraints to determine failure loads (rather than actual dimensions) of a continuous RC
beam with inelastic material properties. Horowitz and Moraes [26] demonstrate the ability
to conduct complex inelastic analysis of RC structures using explicit methods.
Much of the remaining work in optimal design of RC structures excludes complementarity constraints and evaluates structural behavior with elastic material properties
to find solutions for buildings with a larger number of design variables. Models using
elastic (rather than inelastic) material properties generally preclude the need for complementarity constraints. Fadaee and Grierson [16], Balling and Yao [5], and Balling [4]
develop multistep approaches in conjunction with nonlinear programming (NLP) techniques to find minimum-cost solutions with elastic material properties. Fadaee and
Grierson [16] develop implicit methods for three-dimensional RC frames with fewer than
25 continuous-valued variables. Balling and Yao [5] and Balling [4] use postprocess
rounding operations to obtain discrete-valued, constructible design solutions for twodimensional RC frames. To find solutions for problems with a larger number of story
levels, Balling and Yao [5] reduce the number of variables that require rounding. However, their simplifying assumptions diminish solution quality. More recently, Guerra and
Kiousis [25] used sequential quadratic programming to determine optimal design solutions of two-dimensional RC structures. They implement a rounding heuristic that requires the solution of a secondary optimization problem in which many of the decision
variable values are fixed based on the solution of the monolith. Guerra and Kiousis find
solutions for problems with a similar number of story levels as those demonstrated by
Balling and Yao [5]. The objective function presented by Guerra and Kiousis advances
previous studies by including changing unit costs as a function of the decision variables
to more accurately incorporate construction practices. Lee and Ahn [31] and Camp,
Pezeshk, and Hansson [7] implement genetic algorithms that search for discrete-valued
solutions of beam and column elements in RC frames. Like many others, the authors
implement elastic material properties. The search for discrete-valued solutions using
genetic algorithms is difficult because of the nonlinearities in the model, as well as
the large number of combinations of possible element dimensions.
Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
CONCRETE STRUCTURE DESIGN
837
FIG. 1. Three-story structure with a story height of three meters, beam lengths of five meters, and vertical
and horizontal loads, including node number labels, element number labels, and degree of freedom number
labels. Node numbers are given inside the circles, element numbers are given in the squares, and the degree
of freedom numbers are placed next to the arrows.
3. Problem description and formulation. We begin by describing the model in
general terms, then present and justify the mathematical formulation, and finally
describe the formulation in detail.
3.1. Parameters. We consider multistory RC buildings with rectangular beam
(horizontally oriented) and column (vertically oriented) elements (see Figure 1).
Parameters in the formulation describe structural geometry, element geometry, material
properties, applied loads, construction costs, available reinforcement sizes, and upper
and lower variable bounds. For each element e, we use parameter values for beam
lengths and story heights; these values are fixed throughout our numerical study at
5 meters and 3 meters, respectively. We also introduce typical parameter values for vertical and horizontal applied loading, ue and vje , respectively. (See Figure 1.) The discretization shown includes finite elements with end points identified as nodes, which
occur only at the ends of each column and beam. Degrees of freedom represent possible
directions of displacement due to loading at each node. We illustrate typical node number labels in circles, element number labels in squares, and we place degree of freedom
labels next to corresponding arrows. In Figure 2, we show parameters that define (i) the
ACI code-specified distance from the concrete edge to the centroid of the reinforcing
bars, dp , (ii) the strain in the most compressive concrete fiber, ϵcu , (iii) the ACI
code-specified strength reduction factor for computing resistive capacity of concrete,
β1 , and (iv) the compressive strength of the concrete, f c 0 .
3.2. Variables. Figure 2 illustrates the discrete decision variables that describe
the cross-section dimensions for the eth element: integer-valued cross-section width
and depth, be and he , respectively, and discrete-valued compressive and tensile reinforcement cross-sectional areas, As1e and As2e , respectively. The reinforcement crosssection areas depend on the size of bar, which we represent with a binary variable,
#
yem , and number of bars, which we represent with integer variables, As#
1e and As2e .
Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
838
A. GUERRA, A. M. NEWMAN, AND S. LEYFFER
FIG. 2. We show (i) reinforced concrete cross-section element dimensions, (ii) linear strain distribution
including the location of the neutral axis, ce , measured from the concrete fiber with crushing strain ϵcu , and (iii)
resistive capacities for compressive and tensile reinforcement and concrete. A denotes a profile of the cross
section from the leftmost figure.
Note that the depth of each element is parallel to the plane of the structure shown in
Figure 1. We also define continuous-valued auxiliary variables.
3.3. Objective and constraints. Our formulation seeks to find cross-section dimensions for each beam and column element that result in the lowest labor and material
cost while meeting building code requirements for safety and serviceability for axial
forces, bending forces, and constructibility. The constraints include restrictions on
(i) upper and lower bounds on the variables, (ii) the geometry of the cross-section dimensions for each element, (iii) relative sizes of cross-section dimensions of elements
within the structure, (iv) a finite element analysis based on the relative stiffness of
the elements in the structure, (v) the resistive capacity of each element, and (vi) bounds
on the applied axial forces and bending moments in relation to the axial and bending
moment resistive capacities of each element. We restrict the upper and lower bounds
with linear constraints. We enforce the restrictions for the geometry of the cross-section
dimensions for each element. The formulation contains nonconvex constraints for the
relative sizes of cross-section dimensions of elements within the structure and for the
finite element analysis. We model the resistive capacities of each element and bounds
on the applied axial forces with complementarity constraints.
3.3.1. Finite element analysis. The magnitude of axial and bending forces resulting from the applied loads in RC buildings depends on the relative stiffness of elements,
which is a function of the width and depth of each element. We enforce restrictions for the
finite element analysis that ensure equilibrium of forces, stiffnesses, and displacements for
the three degrees of freedom at each node: (i) parallel to the beam length, (ii) perpendicular to the beam length, and (iii) rotational about an axis perpendicular to the plane of
the structure. In order to maintain equilibrium at nodes, applied forces that span nodes
must be transformed into equivalent nodal forces. For each element in the structure,
we define the relative stiffness between the two nodes of each element in matrix form.
Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
CONCRETE STRUCTURE DESIGN
839
Based on the degree of freedom labels, we assemble all element stiffness matrices into a
global stiffness matrix, which defines the relative stiffnesses between elements with a common node. We assume a rigid, or infinitely stiff, connection between the structure and
foundation, which we consider to be fixed degrees of freedom that contain zero displacements. Degrees of freedom one through six in Figure 1 are fixed degrees of freedom. The
discretization specifies the structural response with variables that describe axial forces
(parallel to the length of an element), shear forces (perpendicular to the length), and
bending moments (rotation about an axis perpendicular to the plane of the structure).
We design the cross-section dimensions of beam and column elements for a combination of
axial and bending forces but not shear forces because the influence of shear forces is not
significant in long, slender elements (Inel and Ozmen [28]). In the formulation we show all
constraints for the finite element analysis but provide only that the relative stiffnesses
between elements are functions of the element dimensions. The specific equations for
the relative stiffnesses of elements are given in Chapter 5 of Guerra [24].
3.3.2. Resistive capacity of RC. The capacity to resist applied forces, or resistive capacity, of an RC element is a function of the element width and depth as well as
the reinforcement areas and strain. Recall that the magnitude of applied forces is also a
function of the element width and depth. Strain is a unitless variable that describes the
change of length relative to the effective length over which the displacements occur.
Axial and bending forces in RC buildings result in cross sections that experience both
elongating and shortening strains. We enforce restrictions to develop a linear strain distribution as shown in Figure 2 using variables that describe the location of the neutral
axis, ce , and strain at the location of the compressive and tensile reinforcement, ϵse and
ϵte , respectively. The neutral axis defines the location of zero strain. Figure 2 illustrates
one instance of a linear strain distribution in which compressive strain above the location of the neutral axis is associated with shortening, and tensile strain below the neutral
axis is associated with elongation. Concrete provides large resistance for compressive
strains but cracks when elongated and provides no resistance to tensile strains.
Compressive reinforcement resists forces associated with shortening an element, and
tensile reinforcement predominately resists forces associated with elongating an element. Tensile reinforcement sometimes provides compressive resistance when the location of the neutral axis is below the location of the tensile reinforcement.
The resistive forces provided by the concrete, compressive reinforcement, and tensile reinforcement are a function of the location of the neutral axis, the strain in the
compressive reinforcement, ϵse , and strain in the tensile reinforcement, ϵte , respectively.
Figure 2 illustrates the resistive forces of the concrete, compressive reinforcement, and
tensile reinforcement in relation to the linear strain distribution. The concrete resistive
force is a function of the neutral axis reduced by an ACI code-specified factor, β1 , 85 percent of the concrete compressive strength, 0.85 · f c 0 , and the element width, be . For
instances in which the location of the neutral axis occurs outside of the concrete cross
section, we use complementarity constraints to enforce the resistive capacity to be a
function of the element depth rather than the location of the neutral axis. While it
is necessary to allow the neutral axis to occur outside the cross section, appropriate
resistive capacity of the concrete includes only cross-section extents.
Resistive forces provided by the compressive and tensile reinforcement are a function of the strain in the reinforcement, the modulus of elasticity of the reinforcement, E s ,
and the cross-sectional area of reinforcement. We incorporate elastic-perfectly plastic
(i.e., inelastic) reinforcement material behavior using complementarity constraints that
Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
840
A. GUERRA, A. M. NEWMAN, AND S. LEYFFER
enforce a limit on resistive force when strains are greater than the yield strain of the
reinforcement.
In RC elements, the locus of combinations of axial and bending forces that results in
failure defines the resistive capacity and is termed the interaction diagram in structural
engineering. Structural stability is maintained by enforcing that the demand, or applied
bending and axial forces, be less than the resistive capacities defined by an interaction
diagram. While we enforce structural stability based on ACI code requirements for axial
and bending forces, we assume that the cross-section dimensions are not sensitive to
connection design between elements or displacements of the structural elements.
For structures in seismic design categories A, B, and C, as classified in the American
Society of Civil Engineers 7 standard (SEI/ASCE 7-98) [43], these assumptions are
acceptable.
3.3.3. Problem notation. We use single, lowercase Roman letters for the index
names and single, capital Roman letters with and without superscripts for set names.
Parameter names are Greek and Roman letters following structural engineering notation. The notation of the problem and the associated formulation follow. All variables
except the binary ones follow structural engineering notation.
Sets.
e ∈ E set of all elements e in the structure
e ∈ E b set of all elements e that are beams
e ∈ E c set of all elements e that are columns
e ∈ E d set of elements e with distributed loads
e ∈ E x set of all elements e subject to strength-related constraints
d ∈ De set of global degrees of freedom d for each element e
d ∈ Df set of free degrees of freedom d
d ∈ Dx set of fixed structural degrees of freedom d in each element
j ∈ D # set of number of degrees of freedom j in each element ð1 : : : 6Þ
m ∈ M set of reinforcement sizes ¼ f#13; #16; #19; #22g
#
ACIðbe ; he ; yem ; As#
1e ; As2e Þ set of all elements e that conform to ACI code requirements for reinforcement spacing and percentages
SYMMðbe ; he ; As1e ; As2e Þ set of all symmetrically located elements e that conform to equal cross-section dimension requirements
Parameters.
ḇ lower bound for the width, be , and depth, he , of all elements [cm]
b̄ upper bound for the width, be , and depth, he , of all elements [cm]
dp distance from the centroid of the compressive and tensile reinforcement to the
top and bottom of each element, respectively [cm] (see Figure 2)
le length of structural element e ∀e ∈ E [m]
ld development length of reinforcement for all elements based on typical
practice [m]
x̱e near node x-coordinate of element e ∀e ∈ E [m]
As lower bound for the compressive reinforcement, As1e , and tensile reinforcement, As2e , for all elements [cm2 ]
As upper bound for the compressive reinforcement, As1e , and tensile reinforcement, As2e , for all elements [cm2 ]
#
As lower bound for the number of bars for the compressive reinforcement, As#
1e ,
and tensile reinforcement, As#
,
for
all
elements
2e
C C material and placement cost of concrete [$∕ m3 ]
Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
CONCRETE STRUCTURE DESIGN
CR
CF
CT
βam
β1
f c 0
fy
Es
ϵcu
ρs
vje
841
material and installation cost of reinforcement [$∕ kg]
cost to build formwork [$∕ square meter contact area]
cost to build and install formwork [$∕ square meter contact area]
area of the mth reinforcing bar size ∀m ∈ M ¼ f1.29; 1.99; 2.84; 3.87g [cm2 ]
concrete strength reduction factor for rectangular stress block [unitless] (see
Figure 2)
compressive strength of concrete [MPa] (see Figure 2)
yield stress of reinforcement [MPa]
modulus of elasticity of reinforcement [MPa]
crushing strain of concrete [unitless] (see Figure 2)
density of reinforcement [kg∕ m3 ]
equivalent horizontal earthquake load applied to the jth degree of freedom
for element e ∀e ∈ E d [kN and kN-m]
Decision variables.
Integer variables.
As#
1e number of compressive reinforcing bars in element e ∀e ∈ E [unitless]
As#
2e number of tensile reinforcing bars in element e ∀e ∈ E [unitless]
Discrete variables (see Figure 2).
be width of element e ∀e ∈ E [cm]; be ∈ f20; 25 : : : 200g
he depth of element e ∀e ∈ E [cm]; he ∈ f20; 25 : : : 200g
As1e total compressive area of reinforcement in element e ∀e ∈ E [cm2 ];
As1e ∈f1.29 · 2; 1.29 · 3; : : : 1.99 · 2; 1.99 · 3; : : : 3.87 · 2; 3.87 · 3; : : : 3.87 ·10g
(depends on bar size, βam , and number, As#
1e )
As2e total tensile area of reinforcement in element e ∀e ∈ E [cm2 ]; As2e ∈ f1.29 · 2;
1.29 · 3; : : : 1.99 · 2; 1.99 · 3; : : : 3.87 · 2; 3.87 · 3; : : : 3.87 · 10g (depends on
bar size, βam , and number, As#
2e )
Binary variables.
yem ¼ 1 if the mth rebar size is used for element e ∀e ∈ E x , ∀m ∈ M , 0 otherwise
we ¼ 1 if the eth element formwork is already built ∀e ∈ E x , 0 otherwise
zcee 0 ¼ 1 if column element e is the same size as column element e 0 ∀e,
e 0 ∈ E x ∩ E c ∋ e ≤ e 0 , 0 otherwise
b
zee 0 ¼ 1 if beam element e is the same size as beam element e 0 ∀e,
e 0 ∈ E x ∩ E b ∋ e ≤ e 0 , 0 otherwise
Auxiliary variables.
ce distance from the most compressive concrete fiber to the neutral axis for element e ∀e ∈ E x [cm] (see Figure 2)
c^ e reduced distance from the most compressive concrete fiber to incorporate appropriate concrete resistive strength of element e ∀e ∈ E x [cm]
f ed applied axial force, shear force, and bending moments at the dth degree of
freedom for element e ∀e ∈ E, ∀d ∈ D e [kN and kN∕ m]
kedd 0 stiffness between degrees of freedom d and d 0 for element e ∀e ∈ E, ∀d, d 0 ∈
D e [kN and kN∕ m]
K dd 0 stiffness between the free degrees of freedom d and d 0 ∀d, d 0 ∈ Df [kN
and kN∕ m]
qed equivalent nodal load for the dth degree of freedom for element e with distributed loads ∀e ∈ E d , ∀d ∈ De [kN and kN-m]
Q d nodal load for the dth free degree of freedom ∀d ∈ Df [kN and kN-m]
sþ
e positive slack variable for determining strain in the tensile reinforcement in
element e ∀e ∈ E [unitless]
Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
842
A. GUERRA, A. M. NEWMAN, AND S. LEYFFER
s−
e negative slack variable for determining strain in the tensile reinforcement in
element e ∀e ∈ E [unitless]
T e slack variable for minimum applied axial force requirements in element e
∀e ∈ E [MPa]
Δd nodal displacement for the dth free global degree of freedom ∀d ∈ Df [m]
δed nodal displacement at the dth global degree of freedom for element e ∀e ∈ E
∀d ∈ De [m]
ϕe ACI strength reduction factor for element e ∀e ∈ E x [unitless]
ϵse strain in the compressive reinforcement for element e ∀e ∈ E x [unitless] (see
Figure 2)
ϵte strain in the tensile reinforcement for element e ∀e ∈ E x [unitless] (see
Figure 2)
ϵ^ se strain in compressive reinforcement of element e that remains within yield
strain limits ∀e ∈ E x [unitless]
t
ϵ^ e strain in tensile reinforcement of element e that remains within yield strain
limits ∀e ∈ E x [unitless]
ϕ
ϵ^ e strain in tensile reinforcement of element e to formulate the strength reduction factor, ϕe [unitless]
ude factored distributed load including the self weight of element e
∀e ∈ E d [kN∕ m]
x̆e location of applied loading measured from the concrete fiber with crushing
strain ϵcu ∀e ∈ E [cm]
We now state the full model before commenting on some of its aspects in detail.
Problem statement.
X
be þ 2he
2be þ 2he
le ðC T − C F we Þ
þ
100
100
e∈E c
e∈E b
X
X
b h
As1e
As
þ
le · C C e e2 þ
ρs · C R
ðl − ld Þ þ 2e2 ðle þ ld Þ
2 e
100
100
100
e∈E
e∈E b
X
As1e þ As2e
þ
ρs · C R
le
1002
c
e∈E
ðC Þ: min
X
le ðC T − C F we Þ
subject to
(see section 3.4.2)
ð1aÞ
ḇ ≤ be ≤ b̄
∀e ∈ E;
ð1bÞ
ḇ ≤ he ≤ b̄
∀e ∈ E;
ð1cÞ
As ≤ As1e ≤ As ∀e ∈ E;
ð1dÞ
As ≤ As2e ≤ As ∀e ∈ E;
ð1eÞ
As# ≤ As#
1e
∀e ∈ E x ;
ð1fÞ
As# ≤ As#
2e
∀e ∈ E x ;
Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
CONCRETE STRUCTURE DESIGN
843
(see section 3.4.3)
ð2aÞ
be ≤ he
∀e ∈ E x ;
ð2bÞ
he ≤ 5 · be
ð2cÞ
#
As#
1e ≤ As2e
∀e ∈ E x ;
∀e ∈ E x ;
ðbe ; he ; As1e ; As2e Þ ∈ SYMM ∀e ∈ E;
ð2dÞ
X
ð2eÞ
yem ¼ 1
∀e ∈ E x ;
m∈M
As1e ¼ As#
1e
ð2fÞ
X
βam · yem
∀e ∈ E x ;
βam · yem
∀e ∈ E x ;
m∈M
As2e ¼ As#
2e
ð2gÞ
X
m∈M
#
ðbe ; he ; yem ; As#
1e ; As2e Þ ∈ ACI
ð2hÞ
∀e ∈ E;
(see section 3.4.4)
ð3aÞ
ðbe − be 0 Þ · zcee 0 ¼ 0
∀e; e 0 ∈ E x ∩ E c ∋ e 0 > e; xe ¼ xe 0 ;
ð3bÞ
ðhe − he 0 Þ · zcee 0 ¼ 0
∀e; e 0 ∈ E x ∩ E c ∋ e 0 > e; xe ¼ xe 0 ;
ð3cÞ
ðbe − be 0 Þ · zbee 0 ¼ 0
∀e; e 0 ∈ E x ∩ E b ∋ e 0 > e; xe ¼ xe 0 ;
ð3dÞ
ðhe − he 0 Þ · zbee 0 ¼ 0
∀e; e 0 ∈ E x ∩ E b ∋ e 0 > e; xe ¼ xe 0 ;
X
ð3eÞ
zcee 0 ≥ we 0
∀e 0 ∈ E x ∩ E c ∋ e 0 > e; xe ¼ xe 0 ;
zbee 0 ≥ we 0
∀e 0 ∈ E x ∩ E b ∋ e 0 > e; xe ¼ xe 0 ;
e∈E x ∩E c
X
ð3fÞ
e∈E x ∩E b
(see section 3.4.5)
ð4aÞ
Qd ¼
X
qed
e∈E d ∋d∈D
ð4bÞ
kedd 0 ¼ f ðbe ; he Þ
ð4cÞ
K dd 0 ¼
e
∀e ∈ E; d; d 0 ∈ De ;
X
e∈E∋d;d 0 ∈D
∀d ∈ Df ;
kedd 0
∀d; d 0 ∈ Df ;
e
Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
844
A. GUERRA, A. M. NEWMAN, AND S. LEYFFER
X
ð4dÞ
K dd 0 Δd 0 ¼ Q d
∀d ∈ Df ;
d 0 ∈D f
ð4eÞ
δed ¼ Δd
∀e ∈ E; d ∈ Df ∋ d ∈ D e ;
δed ¼ 0 ∀e ∈ E; d ∈ D x ∋ d ∈ De ;
X
kedd 0 δed 0 − qed ¼ f ed ∀e ∈ E; d ∈ De ;
ð4fÞ
ð4gÞ
d 0 ∈D e
(see section 3.4.6)
ð5aÞ
ϵcu ðce − dp Þ ¼ ϵse ce
ð5bÞ
ϵcu ðhe − dp − ce Þ ¼ ϵte ce
ð5cÞ
c^ e ≤ β1 · ce ⊥ c^ e ≤ he
ð5dÞ
ce ≥ dp
ð5eÞ
c^ e ≥ β1 · dp
ð5fÞ
ϵ^ se ≤ f y∕ Es ⊥ ϵ^ se ≤ ϵse
ð5gÞ
ϵ^ se ≥ 0
ð5hÞ
−
^ te − ϵte
sþ
e − se ¼ ϵ
ð5iÞ
−f y∕ Es ≤ ϵ^ t ⊥ sþ
e ≥0
ð5jÞ
ϵ^ t ≤ f y∕ Es ⊥ s−
e ≥0
∀e ∈ E;
ð5kÞ
ϵ^ ϕe ≥ ϵte ⊥ ϵ^ ϕe ≥ 0.002
∀e ∈ E;
ð5lÞ
ϕe ≤ 0.90 ⊥ ϕe ≤ 0.65 þ ð250∕ 3Þð^ϵϕe − 0.002Þ
∀e ∈ E;
∀e ∈ E;
∀e ∈ E;
∀e ∈ E;
∀e ∈ E;
∀e ∈ E;
∀e ∈ E;
∀e ∈ E;
∀e ∈ E;
∀e ∈ E;
(see section 3.4.7)
ð6aÞ
f ed − 0.1 · f c 0 · be · he ∕ 1002 ≥ −T e
∀e ∈ E b ∋ ordðdÞ ¼ 1;
ð6bÞ
f ed − 0.1 · f c 0 · be · he ∕ 1002 ≥ −T e
∀e ∈ E c ∋ ordðdÞ ¼ 2;
ð6cÞ
Te ≥ 0
∀e ∈ E;
ð6dÞ
ð0.004 − ϵte ÞT e ≤ 0
∀e ∈ E;
(see section 3.4.7)
b
As
As
f ed ¼ 0.85 · f c 0 · c^ e · e 2 þ ϵ^ se · Es · 1e2 − ϵ^ te · Es · 2e2 · ϕe
100
100
100
ð7aÞ
x
b
∀e ∈ E ∩ E ∋ ordðdÞ ¼ 1;
b
As
As
f ed ¼ 0.85 · f c 0 · c^ e · e 2 þ ϵ^ se · Es · 1e2 − ϵ^ te · Es · 2e2 · ϕe
100
100
100
ð7bÞ
∀e ∈ E x ∩ E c ∋ ordðdÞ ¼ 2;
Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
845
CONCRETE STRUCTURE DESIGN
ð7cÞ
∀e ∈ E;
b
ðx̆ − c^ e Þ
0.85 · f c 0 · c^ e · e 2 · e
maxð−f ed ; f ed 0 Þ ≤
100
100
p
As1e
ðx̆e − d Þ
As2e
ðhe − dp − x̆e Þ
s
t
− ϵ^ e · Es ·
þ ϵ^ e · Es ·
· ϕe
·
·
100
100
1002
1002
ð7dÞ
x̆e ¼ he ∕ 2
∀e ∈ E x ∩ E b ∋ ordðd; d 0 Þ ¼ 6; 3
ð7eÞ
ð7fÞ
d 2
ðue · le Þ ðf ed − f ed 0 Þ ðf ed þ f ed 0 Þ2
−
þ
−
8
2
2 · ude · l2e
be
ðx̆e − c^ e Þ
As1e
ðx̆e − dp Þ
0
s
^
þ
ϵ
·
Es
·
≤
0.85 · f c · c^ e ·
·
·
e
100
100
1002
1002
p
As
ðh − d − x̆e Þ
− ϵ^ te · Es · 2e2 · e
· ϕe ∀e ∈ E x ∩ E b ∋ ordðd; d 0 Þ ¼ 6; 3
100
100
be
ðx̆ − c^ e Þ
0 ^
f ed ≤
0.85 · f c · ce ·
· e
2
100
100
p
As
ðx̆ − d Þ
As
ðh − dp − x̆e Þ
− ϵ^ te · Es · 2e2 · e
þ ϵ^ se · Es · 1e2 · e
· ϕe
100
100
100
100
∀e ∈ E x ∩ E c ∋ ordðdÞ ¼ 3; 6
ð7gÞ
ðbe · he − As1e − As2e Þ
1002
∀e ∈ E x ∩ E c ∋ ordðdÞ ¼ 2:
f ed ≤ 0.80 · ϕe · 0.85 · f c 0 ·
þ fy ·
ðAs1e þ As2e Þ
1002
3.4. Problem discussion. Next we describe the objective function and constraints in detail.
3.4.1. Objective function. The objective function includes, in order of appearance, formwork cost for all elements, concrete cost for all elements, and reinforcement
cost for all elements. We separate formwork costs into terms for beams and columns because columns require formwork on all four sides, whereas beams require formwork on
only three sides. Similarly, we develop reinforcement costs with terms for beams and columns because the length of reinforcement relative to the element length is different for
beams and columns. We utilize binary variables, we , in the objective function, as
explained in section 3.4.4, to develop a cost model for formwork reuse that follows the
estimating methods of RC construction companies. The objective function contains
the following mathematical structure: (i) linear terms, i.e., those associated with the parameter C T ; (ii) bilinear terms, i.e., those associated with the parameters C F and C C ; and
(iii) nonlinear terms, i.e., those associated with the parameter C R . The bilinear terms can
be convexified as discussed in section 5.1. The reinforcement cost terms are nonlinear
because we determine the variables As1e and As2e using nonlinear special ordered set
constraints as described in section 3.4.3; in the same section, we study a linear reformulation for defining As1e and As2e . Thus, the entire objective function could be linearized.
3.4.2. Lower and upper variable bounds. Constraints (1a) through (1d) define
lower and upper bounds on the variables that describe the cross-section dimensions of
Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
846
A. GUERRA, A. M. NEWMAN, AND S. LEYFFER
each element e. Constraints (1e) and (1f) incorporate lower bounds for the number of
#
bars for the compressive and tensile reinforcement, As#
1e and As2e .
3.4.3. Geometry restrictions. While constraint (2a) ensures that the width is less
than the depth, constraint (2b) prevents the creation of tall, slender elements and maintains dimensions corresponding to typical behavior for beams and columns. Constraint
(2c) ensures that the area of tensile reinforcement is always greater than or equal to the
area of compressive reinforcement to maintain flexible elements. Constraint (2d) restricts
all symmetric elements with respect to horizontal location to contain the same width,
depth, and reinforcement area, which reduces the number of integer variables in the problem. Constraints (2e) through (2g) define a special ordered set for selecting a discrete
reinforcement bar size from the set of jM j sizes, where the parameter βam is the crosssectional area for the mth bar size. We study a convex reformulation of the nonlinear
equality constraints (2f) and (2g) by replacing them with the following linear constraints:
−2 · As · ð1 − yem Þ ≤ As1e − βam · As#
1e ≤ 2 · As · ð1 − yem Þ
∀e ∈ E x ;
∀m ∈ M ;
∀e ∈ E x ;
∀m ∈ M ;
ð8aÞ
−2 · As · ð1 − yem Þ ≤ As2e − βam · As#
2e ≤ 2 · As · ð1 − yem Þ
ð8bÞ
which is a big-M formulation.
Constraint (2h) consists of five nonlinear sets of inequality constraints with cardinality equal to E x and enforces ACI specifications for reinforcing bar spacing and the ratio
of reinforcement to the concrete cross-sectional area. Additionally, constraint (2h) ensures that columns have the same amount of tensile and compressive reinforcement
(As1e ¼ As2e ∀e ∈ E x ∩ E c ) to ensure stability for horizontal loads applied from either
direction: left-to-right or right-to-left. Note that the solutions that satisfy the ACI
requirements are implicitly defined and that it is therefore difficult to estimate the cardinality of the ACI set.
3.4.4. Formwork reuse. Nonlinear equality constraints (3a) and (3b) restrict the
formwork reuse binary variables for columns, zcee 0 , to equal 1 if and only if the column
element e has the same width and depth as a column element, e 0 , on a higher story level,
indicated by a larger element number, with the same horizontal ordinate (x̱e ). The same
convention follows for beam elements in constraints (3c) and (3d). While constraints
(3a) through (3d) compare the dimensions of elements on each story level, constraints
(3e) and (3f) define the binary variable, we , in terms of the values of the binary variables
zcee 0 and zbee 0 . Essentially, constraints (3e) and (3f) allow we to equal 1 if a beam or
column element, respectively, on a higher story level contains the same dimensions
as a beam or column on any lower story level. When columns or beams, respectively,
on a higher story level contain the same width and depth as columns or beams on a lower
story level, formwork costs are substantially lower because formwork is reused.
We can convexify the nonlinear equality constraints (which are also nonconvex,
bilinear constraints) (3a)–(3d) and replace them with an equivalent convex hull formulation. Unfortunately, this reformulation does not result in faster solve times or improvements to the objective; we give this reformulation in Appendix A.
3.4.5. Finite element analysis. Constraints (4a) through (4g) define the relations between stiffness, applied forces, and displacements for finite element analysis
Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
CONCRETE STRUCTURE DESIGN
847
with beam and column elements. Constraint (4a) defines the equivalent nodal loads for
all elements in the structure based on equivalent nodal loads for each element.
Equation (4b) defines the element stiffnesses as a function of element widths and depths.
Element stiffness, in general, contains terms equal to ðbe · he Þ∕ le , l2e ðbe · h3e Þ∕ 12, and
l3e ðbe · h3e Þ∕ 12. Constraint (4c) assembles the global stiffnesses at each degree of freedom
d and d 0 from element stiffnesses, kedd 0 . Constraint (4d) defines equilibrium of forces,
displacements, and stiffnesses for the finite element analysis. Constraint (4e) defines
the nodal displacements for each element by extracting the appropriate values from
the displacement vector, Δd . Constraint (4f) sets the displacements at each fixed degree
of freedom to zero. Constraint (4g) defines internal forces, f ed , at the dth degree of
freedom for element e. The internal forces include the shear, axial, and bending moments
at the two nodes in each beam and column element. Further details about the finite
element analysis can be found in Chapter 5 of Guerra [24].
3.4.6. Resistive forces: Complementarity constraints. While the finite element analysis provides demand in terms of bending and axial forces that must be resisted in each element, constraints (5a) through (51) express the capacity in terms of
bending and axial forces that each element can resist. In order to meet structural stability, which is discussed in section 3.4.7, the resistive capacity must be greater than or
equal to the demand in each element. Constraints (5a) and (5b) define a linear strain
distribution across the depth of each element in order to determine the resistive capacity
of the reinforced cross section. Complementarity constraints (5c) through (5e) ensure
appropriate concrete compressive resistance when the entire cross section contains compressive strains and when only a portion of the cross section contains compressive
strains. (See the left-hand side of Figure 3.) Complementarity constraints (5f)
and (5g) define elastic-perfectly plastic material response for the resistive capacity of
the compressive reinforcement when the strain in the compressive reinforcement is less
FIG. 3. The left figure illustrates the relationship between the concrete resistive force (y-axis) and the
location of the neutral axis reduced by the strength reduction factor (x-axis), which defines the amount of
concrete that contains compressive strains; see constraints (5c)–(5e). When the entire cross section contains
compressive strains, the amount of concrete in compression is equal to he ∕ 100; when only a portion of the cross
section contains compressive strains, the amount of concrete in compression is equal to β1 · ðce ∕ 100Þ. The
right figure illustrates the relationship between the stress in the tensile reinforcement (y-axis) and the strain in
the tensile reinforcement (x-axis), see constraints (5f) and (5g), when the stresses and strains lie in the first
quadrant and illustrates constraints (5h) through (5j) when the stresses and strains lie in the first and third
quadrants,
Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
848
A. GUERRA, A. M. NEWMAN, AND S. LEYFFER
than and greater than the yield strain. Elastic-perfectly plastic material response ensures
that the maximum resistive capacity corresponds to the yield strain. (See the first quadrant of the right-hand side of Figure 3.) Complementarity constraints (5h) through (5j)
define elastic-perfectly plastic material response for the resistive capacity of the tensile
−
reinforcement using positive and negative slack variables, sþ
e and se , respectively.
(See Figure 3 in its entirety.) Complementarity constraints (5k) and (5l) define the appropriate value for the strength reduction factor according to ACI code requirements as
a function of the strain in the tensile reinforcement. The strength reduction factor incorporates factors of safety into the design. (See the left-hand side of Figure 4.) Typical
computer methods for incorporating RC strength use if-then-else statements, which are
not algebraic expressions and are not suited for explicit optimization approaches. For
example, constraint (5c) would be expressed as follows: if β1 · ce ≤ he , then c^ e ¼ β1 · ce ,
else c^ e ¼ he . We use complementarity to provide an algebraic expression to explicitly
incorporate RC strength with individual constraint qualifiers over the set of e elements.
3.4.7. Restrictions for structural stability. In general, in order to maintain
structural stability, the capacity of each element must be greater than the demand. Constraints (6a) through (6d) enforce ACI code requirements for minimum eccentricity of
applied forces using conditional complementarity depending on the value of a slack variable, T e . For situations in which the strain in the tensile reinforcement is greater than or
equal to 0.004 (an ACI code-specified value), there is sufficient bending in the element,
the value of the slack variable must only be greater than or equal to zero, conditional
complementarity is not invoked, and requirements for minimum eccentricity do not apply. For situations in which the strain in the tensile reinforcement is less than 0.004, the
value of T e must equal zero and the applied axial force must be greater than 10 percent
of the maximum axial concrete resistance. (See the right-hand side of Figure 4.)
Constraints (7a) and (7b) restrict the resistive axial force of the cross section to be
equal to the applied axial force in each beam and column element, respectively. Because
the bending resistive force can increase with increasing applied axial forces, equality of
FIG. 4. The left figure illustrates the relationship between strain in the tensile reinforcement and the
strength reduction factor; see constraints (5k) and (5l). Note the code-specified transition points of 0.002
and 0.005 in the strain of the tensile reinforcement. The right figure illustrates the relationship between
the strain in the tensile reinforcement and the applied axial force; see constraints (6a)–(6d).
Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
CONCRETE STRUCTURE DESIGN
849
the resistive and applied forces ensures that the magnitude of the resistive bending force
corresponds to the appropriate applied axial force. Constraint (7c) defines the location
of applied loading measured from the most compressive concrete fiber, which represents
the point about which the bending resistance is computed. Constraints (7d) and (7e)
enforce that the corresponding applied bending forces at the ends and in the span of the
beams, respectively, are less than or equal to the resistive bending force. Constraint (7f)
enforces that the applied bending forces at both ends of columns are less than or equal to
the resistive bending force. Constraint (7g) ensures that the applied axial force in column elements is less than an ACI-prescribed maximum axial load, which is a function of
the strength of the cross section. The equality of resistive and applied axial forces, (7a)
and (7b), and inequality of resistive and applied bending forces, (7d)–(7f), ensure that
the applied axial and bending forces fall within the locus of failure for any feasible combination of axial and bending forces. Constraint (7g) enforces an ACI code-specified
upper bound on the applied axial force.
3.4.8. Fixed case of formwork reuse constraints. We can reduce the combinatorial complexity of the problem by adding logical constraints that force a particular
formwork reuse. We do so by fixing the binary variables zcee 0 and zbee 0 to a value of 0 or 1
to force a particular formwork reuse distribution. We include formwork distributions
such that all columns contain the same dimensions and all beams contain the same dimensions by forcing zcee 0 ¼ 1 ∀e, e 0 ∈ E x ∩ E c ∋ e 0 > e, xe ¼ xe 0 and zbee 0 ¼ 1 ∀e,
e 0 ∈ E x ∩ E b ∋ e 0 > e, xe ¼ xe 0 . Forcing values for these binary variables eliminates
the associated nonlinear equality and inequality constraints (3a)–(3f) in the formulation. For the fixed case, we could solve all possible formwork reuse distributions;
i.e., for eight stories, we force all to be one size, one to be one size and seven to be another;
two to be one size and six to be another; three to be one size and five to be another, etc.
However, in practice we solve a relatively small number of possible combinations because for our problem instances, solutions with two or more form sizes possess a larger
objective function value than those with one form size. The costs associated with only
one form size are minimal because (i) formwork costs contribute a relatively large portion of the total cost, (ii) formwork distributions restrict the relative sizes of elements,
and (iii) constraints for stability of the RC structure restrict the sizes of elements. In
section 5, we compare the algorithm performance for the fixed case of the formulation,
which contains additional restrictions, and the free case of the formulation, which does
not contain additional restrictions on the binary variables.
4. Model instances. We include two-dimensional design examples for structures
with one to eight story levels. These structures consist of one bay, i.e., one span supported by a column element at each end. We use beam lengths of five meters, typical for
RC structures, and a load case with both horizontal and vertical loads to demonstrate
optimal solutions for combinations of applied axial force and bending moment magnitudes that cover a wide range of values typically observed in RC elements.
4.1. Parameter values. The parameters in this study contain common values
used in the design of RC structures. Table 1 presents material properties, code specified
values, and unit prices for concrete and reinforcement. Note that we give unit prices for
formwork per square meter contact area (SMCA). All structures are loaded by their self
weight, wG , an additional gravity dead load, wD ¼ 30kN ∕ m, and a gravity live load,
wL ¼ 30 kN∕ m, which are typical values for office buildings. We calculate the parameters that describe horizontal seismic forces, vje , for the jth degree of freedom for
Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
850
A. GUERRA, A. M. NEWMAN, AND S. LEYFFER
TABLE 1
Material properties, code specified values, and unit prices for concrete and reinforcement.
Description
Parameter name
Value
Units
Lower bound for the width and depth of elements
Upper bound for the width and depth of elements
Concrete cover
Length of structural element e
Development length of reinforcement
Near node x-coordinate of element e
Lower bound for reinforcement
Upper bound for reinforcement
Lower bound for the number of reinforcing bars
Material and placement unit price of concrete
Material and installation unit price of reinforcement
Unit price to build formwork
Unit price to build and install formwork
ACI section 10.2.7.3 reduction factor
Concrete compressive strength
Reinforcement yield stress
Reinforcement modulus of elasticity
Crushing strain of concrete
Density of reinforcement
ḇ
b̄
dp
le
ld
xe
As
As
As#
CC
CR
CF
CT
β1
f c 0
fy
Es
ϵcu
ρs
20
200
7
3–10
1.11
0–20
2.58
77.4
2
192.80
1.55
32.60
38.00
0.85
28
420
200,000
0.003
7870
[cm]
[cm]
[cm]
[m]
[m]
[m]
[cm2 ]
[cm2 ]
[unitless]
[$∕ m3 ]
[$∕ kg]
$∕ SMCA
$∕ SMCA
[unitless]
[MPa]
[MPa]
[MPa]
[unitless]
[kg∕ m3 ]
TABLE 2
Horizontal loads for spans of five meters.
Horizontal load on story level
Total
number of
story levels
1
2
3
4
5
6
7
8
One
(kN)
15.6
10.4
7.8
6.3
4.8
3.3
2.3
1.9
Two
(kN)
Three
(kN)
Four
(kN)
Five
(kN)
Six
(kN)
Seven
(kN)
Eight
(kN)
—
20.8
15.6
12.5
9.9
6.9
5.0
3.8
—
—
23.4
18.8
14.9
10.7
7.9
6.2
—
—
—
25.0
20.0
14.5
10.9
8.6
—
—
—
—
25.1
18.4
13.2
11.1
—
—
—
—
—
22.3
17.0
13.6
—
—
—
—
—
—
20.2
16.3
—
—
—
—
—
—
—
19.0
element e using the ASCE 7 (SEI/ASCE 7-98) equivalent lateral load procedure for a
structure in Denver, Colorado, the failure of which would result in a substantial public
hazard. Buildings subjected to gravity and seismic loads contain factored loads of
1.2wG þ 1.2wD þ 1.0wL þ 1.0vje ¼ 1.2wG þ 66 kN∕ m þ 1.0vje . Table 2 summarizes
the values of the seismic horizontal forces for the multistory design examples with beam
lengths of five meters. For all examples, the horizontal loads increase for each story level
so that the top story level contains the largest horizontal load.
4.2. Initialization of variables and typical solution. Initial variable values
provide a starting point for the algorithm and strongly influence the performance.
We set an initial value for all variables in the optimization formulation based on a
Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
851
CONCRETE STRUCTURE DESIGN
TABLE 3
Problem size for the one- through eight-story structures with beam lengths of five meters and vertical and
horizontal loads when the binary variables for formwork reuse are free and fixed.
Number of
story levels
1
2
3
4
5
6
7
8
Number of variables
(integer)
Fixed
Free
76 (18)
152 (36)
228 (54)
304 (72)
380 (90)
456 (108)
532 (126)
608 (144)
—
156 (40)
238 (64)
322 (90)
408 (118)
496 (148)
586 (180)
678 (214)
Number of constraints
(nonlinear)
Fixed
95
188
285
386
491
600
713
830
(64)
(122)
(180)
(238)
(296)
(354)
(412)
(470)
Number of complementarity
constraint pairs
Free
Fixed and free
—
190 (126)
289 (192)
392 (262)
499 (336)
610 (414)
725 (496)
844 (582)
15
30
45
60
75
90
105
120
typical design practice that meets all the constraints in the problem formulation. This
“typical solution” provides a feasible starting point and can be compared with solutions
from MINLPBB to determine cost savings potential for the structures we consider. We
develop the typical solution with a method often used in practice: (i) approximate an
initial width and depth of beam and column elements using applied loads and beams
lengths, (ii) determine the applied forces and moments on each element using finite element analysis, (iii) determine the number and size of reinforcing bars to resist applied
forces and moments, and (iv) determine whether the number and size of reinforcing bars
meet the required spacing and concrete cover. If the reinforcement does not meet the
requirements, then element widths and/or depths are increased in five-centimeter increments and we repeat steps (ii)–(iv). While we could use various initial variable values
and compare the corresponding objective function values to find the best locally optimal
solution, we use only the initial variable values from the typical solution.
4.3. Problem size and structure. The mathematical structure of the RC design
problem is highly nonlinear and nonconvex. Discrete-valued variables describe constructible design solutions and continuous-valued variables model material response. Nonlinear equality and complementarity constraints make the problem nonconvex. We
present information about the size of the problems in terms of the number of story levels,
number of variables (including integer variables), number of constraints (including nonlinear constraints), and number of complementarity constraint pairs for the fixed and
free cases of our model runs (see section 3.4.8). We solve problems with up to 678 variables, of which 214 are integer, 844 constraints, of which 582 are nonlinear, and with 120
complementarity constraint pairs (see Table 3).
4.4. Choice of MINLP algorithm. A range of different MINLP solution techniques exists, and we briefly comment on their suitability for our problem instances.
Outer approximation (Duran and Grossmann [15]) and generalized Benders decomposition (Geoffrion [23]) solve a sequence of mixed integer linear programming (MILP)
master problems and NLP subproblems. The extended cutting plane method (Westerlund and Pettersson [45]) solves only a sequence of MILP problems. We considered
branch-and-reduce optimization navigator (BARON) (Tawarmalani and Sahinidis
[44]), which can solve nonconvex MINLP problems to global optimality using global
underestimators and branch-and-reduce techniques. However, the size and complexity
of the nonlinear expressions in the finite element analysis make it unlikely that BARON
Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
852
A. GUERRA, A. M. NEWMAN, AND S. LEYFFER
can succeed in solving our problem (Sahinidis [40]). All four methods rely heavily on the
convexity of the problem functions or the ready availability of underestimators. Unfortunately, our models do not easily admit underestimators, and we cannot apply these
techniques.
An alternative to these techniques is nonlinear branch-and-bound. This method
solves a sequence of nonlinear problems at every node of a tree. One advantage of this
approach is that nonlinear solvers often find good solutions even to nonconvex problems.
The MINLPBB solver is attractive because the underlying SQP method has been shown
to be an efficient and robust solver for optimization problems with complementarity
constraints (Fletcher and Leyffer [20]).
We solve the RC design problem with a software package, MINLPBB (Leyffer [32]),
on Linux-based servers with dual core Advanced Micro Devices Opteron processors with
a CPU speed of 2.41 GHz and random access memory of approximately one gigabyte.
MINLPBB uses a branch-and-bound framework with a sequential quadratic programming technique to solve the continuous relaxations of the problem (Fletcher and Leyffer
[19], Leyffer [32], and Leyffer [34]). The branch-and-bound method in MINLPBB uses a
depth-first search and solves the NLP relaxations using filterSQP (Fletcher and Leyffer
[19]). The user can influence branching by supplying priorities for integer variables as
well as by choosing node selection strategies and branching rules (Leyffer [32]). We use a
depth-first search strategy, branching on the variable with the highest user-defined
priority. MINLPBB guarantees global optimality only for convex problems. However,
it also provides a more robust solution technique for nonconvex MINLP problems compared to outer approximation and Benders decomposition (Leyffer [32]). Outer approximation and Benders decomposition often reduce the size of the feasible region by using
cuts, which can eliminate optimal solutions in nonconvex problems like our RC design
problem. Fletcher and Leyffer [20] have demonstrated the abilities of the NLP solver in
MINLPBB, filterSQP, to solve over 150 optimization problems with complementarity
constraints.
4.5. Formulation of complementarity constraints. The complementarity
constraints (5a) through (5l) turn our problem into a mathematical program with equilibrium constraints (MPEC) and add another level of difficulty to the existence of integer variables. MPECs arise in a broad range of applications—see the survey by Ferris
and Pang [17], the monographs by Luo, Pang, and Ralph [36] and Outrata, Kocvara,
and Zowe [38], and the sets of test problems in Leyffer [33] and in Dirkse [13]. However,
the combination of both integer variables and complementarity constraints is new as far
as we know.
One obvious approach to handling the complementarity constraints is to replace the
disjunction 0 ≤ x1 ⊥ x2 ≥ 0 by a set of nonlinear inequalities: x1 ≥ 0, x2 ≥ 0, and
X 1 x2 ≤ 0, where X 1 is the diagonal matrix with x1 along its diagonal. Unfortunately,
it has been shown that this approach violates the Mangasarian–Fromowitz constraint
qualification (MFCQ) (Scheel and Scholtes [41]). We note that we do not require
X 1 x2 ¼ 0, and, in fact, this formulation can be shown to have worse properties that
preclude fast convergence for SQP methods (Fletcher et al. [21]).
The failure of MFCQ has motivated scientists to develop alternative approaches
that remove or relax the source of the failure of MFCQ, namely, X 1 x2 ≤ 0. Common
approaches include branch-and-bound (Bard [6]), implicit nonsmooth methods
(Outrata, Kocvara, and Zowe [38]), piecewise SQP methods (Luo, Pang, and Ralph
[36]), and perturbation and penalization approaches (Dirkse, Ferris, and Meeraus
[14]; Raghunathan and Biegler [39]; DeMiguel et al. [10]; Leyffer, Lopez-Calva, and
Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
853
CONCRETE STRUCTURE DESIGN
Nocedal [35]; Anitescu [20]; and Scholtes [42]). The latter approaches are among the
most successful at present.
Despite the failure of MFCQ, it has been noted (Fletcher and Leyffer [20] and
Anitescu [3]) that SQP methods perform reliably and robustly for large classes of
MPECs, and we adopt this approach in our work. We explicitly add the nonlinear
inequalities X 1 x2 ≤ 0 to the formwork MINLP and apply a standard branch-and-bound
solver that uses filterSQP (Fletcher and Leyffer [19]) to solve the NLP subproblems.
In addition, we add the following safeguard to the branch-and-bound solver, following ideas that are standard in MILP solvers such as CPLEX (IBM [27]). Whenever we
encounter an NLP node where the violation of MFCQ appears to cause convergence
problems with filterSQP, we select an integer variable and branch, moving to a new
starting point, potentially resolving the difficulty. Our heuristic is motivated by the fact
that the MPECs violate MFCQs, and the lack of MFCQs can cause NLP solvers to
terminate unsuccessfully. In particular, we “branch-through-the-problem” if the solver
terminates with any of the following conditions: (i) if the trust-region becomes too small,
(ii) if an iteration limit (1000) is reached, (iii) if there exists constraint violation less than
the tolerance but linearizations are inconsistent, and (iv) if there exist excessively large
Lagrange multipliers (larger than tolerance−1 ). All these cases refer to potential failures
due to a lack of MFCQ. By branching on an integer variable, we are likely to enter a
different area of convergence and may resolve this issue, though we admit this remains a
heuristic.
5. Numerical results. We present design examples to illustrate three key aspects:
(i) MINLPBB algorithm performance for the RC design problem with respect to various
reformulations and heuristics, (ii) cost savings over typical practice, and (iii) characteristics of the optimal solution, namely, optimal formwork reuse and stiffness distributions. We detail these aspects and the corresponding results in the following three
subsections.
5.1. Algorithm performance. Recall that we solve two cases of the optimization
formulation: the original formulation that allows solutions with any formwork reuse distribution, termed the free case, and the formulation that contains restrictions that enforce a particular formwork reuse distribution, the fixed case. While we find solutions for
# of NLPs Solved
1000000
100000
10000
fixed
free
1000
100
0
50
100
150
200
250
Number of Integer Variables
FIG. 5. Number of nonlinear programs solved to reach the optimal solution versus number of integer
variables for structures using the fixed case of the optimization formulation.
Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
854
A. GUERRA, A. M. NEWMAN, AND S. LEYFFER
TABLE 4
Number of nonlinear programs and the CPU time (seconds) for the fixed and free cases. Also reported are
the percent savings over the typical solution (see section 4.2) as given by the root node solution and the rounded
solution (a heuristic solution) and the local optimal solutions for the fixed case (a heuristic solution) and the free
case obtained with the MINLP solver.
Number
of
story
levels
1
2
3
4
5
6
7
8
Number of NLPs
CPU time (seconds)
Percent savings with respect to typical solution
NLP
MINLP
Fixed
Free
Fixed
Free
Root
node
Rounded
solution
198
239
576
1166
2014
7707
12611
18318
—
284
1637
5133
4776
11038
55622
842945
1
4
20
84
215
1218
2651
5244
—
5
56
357
562
1527
14486
196606
13.1
19.1
22.0
21.4
20.6
18.9
19.4
18.0
2.2
−0.1
0.8
2.6
4.9
−4.3
−2.4
0.7
Fixed
solution
9.4
13.7
16.0
14.2
13.9
11.6
12.2
10.6
Free
solution
—
13.7
16.0
14.9
13.9
11.6
11.6
10.2
up to an eight-story structure for the original problem formulation, we can find solutions
in less CPU time when a particular formwork reuse distribution is enforced. Figure 5
illustrates on a logarithmic scale the CPU time as a function of the number of integer
variables for both the fixed and free cases.
Table 4 summarizes the number of nonlinear programs solved and the algorithm
performance (in terms of solution time) for the fixed and free cases for structures with
up to eight story levels. In each case, we run MINLPBB until termination or until we
reach a memory limit. The fixed cases require significantly less CPU time, on average,
than do the free cases. As expected, the CPU time increases exponentially as the problem size increases in both cases. In the fixed case, however, the amount of CPU time is
less than two hours for the eight-story scenario, whereas for the free case, the largest
scenario requires days of CPU time. Note that in Table 4 the number of NLPs solved
for the free case decreases from 5133 to 4776 for the four-story and five-story example,
respectively, even though the five-story example contains a larger number of variables
and constraints than the four-story example. This characteristic is unexpected and can
also be seen in Figure 5 where the dashed line shows a peak at 90 integer variables and an
inflection point at 118 integer variables.
Convexifying the bilinear objective function terms (see section 3.4.1) results in an
increased number of continuous and integer variables and linear constraints. Note that
the formulation for the convexified bilinear objective function terms contains the same
mathematical structure as that of the convex hull reformulation for formwork reuse
shown in Appendix A. However, to find a solution for the formulation with convexified
objective function terms with computer memory of one gigabyte, we must reduce the
number of potential sizes for both be and he from 37 to 17 for even the smallest problem
instance, the one-story example. Reducing the number of potential sizes eliminates beam
widths and depths between 105 centimeters and 200 centimeters, which does not affect
the optimal solutions for most of the smaller examples but would result in suboptimal
solutions for the larger problem instances. In particular, for the three-story example with
vertical and horizontal loads and beam lengths of five meters, we find the same objective
Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
CONCRETE STRUCTURE DESIGN
855
function value using the convexified bilinear terms as with the fixed case formulation,
but the CPU time for the convexified formulation is about 40 times longer than that of
the fixed case formulation. The objective function value for the NLP relaxation of the
formulation with convexified bilinear objective terms is slightly lower than the NLP
relaxation of the fixed case formulation, indicating a looser lower bound. Convexifying
the bilinear objective function terms results in a problem with 2302 variables, of which
240 are integer variables, and 5795 constraints, of which 180 are nonlinear constraints.
We also study a convex formulation of constraints (2f) and (2g), which contain bilinear equations involving a special ordered set. The formulation in (8a) and (8b) uses a
big-M construct, which is linear. The objective function value is the same for both formulations. However, the CPU time when implementing (8a) and (8b) is about 100 times
longer than that resulting from implementing constraints (2f) and (2g) for the threestory structure with vertical and horizontal loads. Additionally, we study the convex
formulation of constraints (3a) through (3f), which are nonlinear equality constraints,
using a convex hull formulation (see Appendix A). The objective function value for
the NLP relaxation of the convex hull is slightly lower than the NLP relaxation of
the original formulation, indicating a looser lower bound. The objective function value
for the convex hull formulation is the same as that of the original formulation. However,
the CPU time for the three-story case with the convex hull formulation is about 375
times longer than that of the original formulation with nonlinear equality constraints.
While these alternate formulations are certainly of interest to practitioners, we argue that the strategies we outline in this section can be of equal value to researchers
interested in theory and algorithms; algorithms are much more immediately relevant
if it can be demonstrated that they work well in practice. The strategies we espouse
help developers to identify promising formulations and, hence, algorithms that may initially appear not to be promising in practice but that are, in fact, efficient when used
correctly. We also demonstrate how heuristics (e.g., the “fixed” formulation) can help to
expedite solution times. Such heuristics help algorithmic developers with local search
strategies, for example.
5.2. Objective function values. To justify the use of integer variables, we compare rounded solutions from the root node found with MINLPBB and discrete-valued
solutions from MINLPBB for the one- through eight-story structures with beam lengths
of five meters and vertical and horizontal loads. In both rounding and MINLP methods,
the initial variable values equal those of the typical solution. Additionally, rounding
and branching operations both begin from a root node in which the binary variables
for formwork reuse are fixed such that all beams contain the same dimensions and
all columns contain the same dimensions. We also enforce this formwork distribution
in rounding operations. We allow the solver to determine discrete reinforcing bar sizes
before rounding.
We obtain an integer feasible solution using engineering insight by rounding as
follows: (i) if element width and depth, be and he , are continuous-valued, then round
be and he up to the nearest five-centimeter increment; (ii) if the number of compressive
#
#
and tensile reinforcing bars, As#
1e and As2e , are continuous-valued, then round up As1e
#
and As2e to the nearest integer value; (iii) if the number and size of reinforcing bars do
not fit with appropriate cover and spacing, then round up be to the next five-centimeter
increment; (iv) if minimum reinforcement ratios are violated, then round up the bar size
to the next largest size or round up the number of bars if we already use the largest bar
size; (v) if maximum reinforcement ratios are violated, then round up be one increment
and, if needed after rounding be one increment, round up he one increment; (vi) if the
Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
856
A. GUERRA, A. M. NEWMAN, AND S. LEYFFER
location of the neutral axis is less than the lower bound of seven centimeters, then round
up the reinforcing bar size one increment, round up the number of bars one increment, if
needed, and then round up be one increment if needed; (vii) if the bending force resistive
capacity contains a smaller magnitude than the applied bending force, then round
up the bar size one increment and then he one increment if needed; and (viii) repeat
steps (i) through (vii) until a feasible discrete solution is obtained. This rounding method produces feasible solutions that represent those used in engineering practice.
To evaluate the relative quality of our solutions given various methods of increasing
difficulty in obtaining them, we compare the following relative to the objective function
value from a typical solution: (i) the objective function value at the root node of the fixed
case of the NLP relaxation; (ii) the objective function value from a feasible solution
rounded from the root node solution to the fixed problem; (iii) the objective function
value found using MINLPBB with fixed formwork; and (iv) the objective function value
found using MINLPBB allowing formwork reuse to vary. The last four columns of
Table 4 show these results. Because the typical solution represents engineering practice,
we report percent savings with respect to it to indicate the relative quality of our solutions.
On average, the objective function value at the root node is 19 percent better (lower)
than that provided by the typical solution. We expect this marked contrast because the
root node solution is not necessarily integer feasible. On average, the solutions rounded
from the root node show no improvement from the typical solutions; some cases show a
slight degradation, whereas others show a slight improvement. These results demonstrate that rounding is not an effective means of improving on typical practice and underscores the necessity of solving an integer nonlinear program to achieve savings
over typical practice. The integer feasible solution given by the fixed formwork case
is 13 percent better than the typical solution, with no clear trend in solution quality
improvement as the number of story levels increases. The fixed solution generally gives
the same quality objective as the free solution, though the latter requires considerably
more CPU time to obtain a solution because of the increased number of binary variables.
(In the largest two cases, the objective degrades slightly, indicating that an inferior local
optimum is found.) Recall that we solve all nondominated distributions of formwork
reuse for the fixed case as described in section 3.4.8 in order to find the best distribution
of formwork reuse. Optimal solutions from the free case contain the same distribution of
formwork reuse that we found to be optimal in the fixed case. Therefore, these experiments demonstrate empirically that the free case formulation, although potentially more
likely to converge to local optima, results in solutions consistent with those found with
the more tractable but more restricted fixed case formulation. These results indicate
that it is generally sufficient for these examples (with the corresponding cost structures)
to consider just the fixed case, though, in general, this would not be true.
In Figure 6, we show the distribution of reinforcement, concrete, and forming costs
for the structures with beam lengths of five meters and vertical and horizontal loads for
the fixed case of the formulation. In Figure 6, the large initial percentage of cost attributed to forming decreases to nearly 25 percent for the eight-story structure. Figure 6
illustrates an asymptotic tendency for the percentage of cost attributed to forming, indicating that there is a limit at which the optimal solution contains more than one form
size for columns, for beams, or for both.
5.3. Solution characteristics. We observe that, for all our examples, the lowest
cost solution contains exactly one formwork size for columns and one formwork size
for beams if we use the typical formwork cost parameter values, as defined in Table 1.
Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
CONCRETE STRUCTURE DESIGN
857
100%
percent cost
80%
60%
reinforcement
cost
40%
concrete cost
20%
forming cost
0%
1 2 3 4 5 6 7 8
number of stories
FIG. 6. Reinforcement, concrete, and forming cost percentage for structures with beam lengths of five
meters and vertical and horizontal loads.
To ensure the validity of this result, we conduct an exhaustive search with both the fixed
and free cases of the formulation to find solutions with two formwork sizes for beams
and/or columns. For the fixed and free cases, we use a wide range of initial variable
values such that the starting point is a structure with one and two form sizes, including
the collective set of potential distributions of two formwork sizes as described in section 3.4.8. For the fixed case, we force formwork reuse distributions with one and
two form sizes for the collective set of potential distributions of two formwork sizes.
These collectively exhaustive searches show that one formwork size always results in
a lower objective function value for the typical formwork cost parameter values. As
prices can vary greatly depending on the year and location of construction, we demonstrate optimal solutions with more than one form size with a parametric study of the
forming cost parameters for a six-story structure with vertical and horizontal loads.
To illustrate how changing the cost parameters affects the choice of formwork reuse,
we conduct the following experiment. Recall that the parameters C T and C F are the
cost of building and installing formwork and the cost of building formwork, respectively.
Hence, the difference between the two is the cost of installing formwork. We change the
formwork cost parameters, C T and C F , such that the installation cost remains constant
but the benefit for reusing formwork decreases. We change the values from $38.00/
square meter of contact area (SMCA) and $32.60/SMCA, respectively, to values of
$21.60/SMCA and $16.20/SMCA, respectively, in order for the optimal solution for
the six-story structure to contain two form sizes for columns and two form sizes for
beams. The original formwork costs result in a ratio of installation to building and installation of approximately seven, while the new values result in a ratio of four. We
illustrate the optimal design dimensions, axial force distribution, and bending moment
distribution in Figure 7.
Consider the relative stiffness distribution of the six-story structure with two formwork sizes for columns and two for beams. Recall that the demand on the RC elements in
terms of displacements and forces depends on the applied loads and relative stiffness of
elements, where stiffness increases for increasing cross-sectional element dimensions. For
the six-story example presented in Figure 7, the optimal column width and depth equal
25 and 45 centimeters, respectively, for the bottom three story levels and equal 20 and
40 centimeters, respectively, for the top three story levels. The optimal beam width and
depth equal 30 and 55 centimeters, respectively, for the bottom three story levels
Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
858
A. GUERRA, A. M. NEWMAN, AND S. LEYFFER
FIG. 7. Optimal design dimensions, axial force distribution of the optimal solution, and bending moment
distribution of the optimal solution for the six-story structure with beam lengths of five meters and horizontal
and vertical loads when C T equals $21.60∕ SMCA and C F equals $16.20∕ SMCA.
and equal 20 and 75 centimeters, respectively, for the top three story levels. The column
dimensions decrease for the top three story levels compared to the bottom three story
levels. For the beam dimensions, the width decreases while the depth increases for the
top three story levels compared to the bottom three story levels. Decreases and increases
in element dimensions result in decreases and increases in stiffness, respectively.
The distribution of forces depends on the relative stiffnesses of elements such that elements with smaller or larger stiffness contain relatively smaller or larger forces.
In the six-story structure shown in Figure 7, the axial force distribution of the optimal solution shows the largest forces applied to columns on the bottom story level
(1224 kN) and a linear decrease for higher story levels. Horizontal loads that push
the structure from right to left cause larger axial forces in the left-hand side columns
compared to the right-hand side columns. The bending moment distribution of the optimal solution illustrates moments with the largest magnitude occur at the connection
between the beams and columns on all but the top story level in which the largest moment occurs in the span of the beam. The beam stiffness must be significantly larger than
the column stiffness in order for the largest magnitude moment to occur in the span of
the beam. On the first five story levels, the stiffnesses provided by two columns at the
end of each beam (one below each beam and one above each beam) cause the largest
moment to occur at the ends of the beams. On the top story level in which the beam is
attached only to one column at the end points, the relative stiffness is such that the
largest moment occurs in the span of the beam. The moment distribution remains
Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
CONCRETE STRUCTURE DESIGN
859
the same for the first five story levels because the relative stiffness of the beams and
columns is about the same for these story levels. This six-story example with two formwork sizes demonstrates the range of stiffness distributions that we find for all other
examples.
6. Conclusions. We present, for the first time, an explicit MINLP formulation of
the RC design problem. We find optimal solutions for RC structures that result in an
average of 13 percent cost savings over the typical solution, and we demonstrate the
ability of MINLPBB to solve a new class of hard optimization problems: mixed integer
nonlinear problems with complementarity constraints. For the size of the problems considered in this study, we find locally optimal solutions that contain one form size for
columns and one form size for beams.
Efficient designs for RC structures contain a large number of same-size elements to
substantially reduce formwork cost. For typical formwork unit prices we find optimal
solutions with one form size and must drastically decrease the unit price to find solutions
with two form sizes. Further exploration of the transition between one and two form
sizes would provide a significant contribution to this research topic. Solving systems
with a larger number of story levels is necessary to explore such a transition. We find
that the performance of MINLPBB is affected by the number of variables and constraints and by the number of different size cross sections that must be determined.
We find a decrease in the number of NLPs solved and CPU time for the fixed case
of the optimization formulation in comparison to the free case. We also find that
the free and fixed cases result in nearly the same quality of solutions. While the fixed
case contains a smaller number of integer variables, number of NLPs solved, and CPU
time than does the free case, we find that the fixed case quickly reaches a very large
number of NLPs and CPU time. We show the limits of attainable results when using
MINLPBB for the RC design problem and find that we can solve one-bay structures
with up to eight story levels. To find solutions for structures with a larger number of
story levels, we suggest reducing the number of integer and binary variables and improving the branching scheme. One method to reduce the number of integer and binary
variables in the RC design problem without losing model fidelity would be to describe
the cross-section width and depth of a group of elements with only two variables so that
all elements in the group contain the same dimensions. Although the method would
require a dynamic group of elements in order to find the optimal number of elements
that should contain the same dimensions, the large number of same-size elements indicates that there is always some formwork reuse. Improvements to the branching
scheme could also increase tractability for larger problems. Any improvements to
the branching scheme in reference to the behavior of the RC design problem could substantially increase the tractability of the RC design problem and allow the solution of
instances with a larger number of degrees of freedom. A sensitivity study of the variables
in the problem could be used to determine the best types of improvements for the
branching scheme specific to the RC design problem. A better understanding of the
algorithm performance for the RC design problem could also help improve the branching scheme.
Appendix A. Convex hull reformulation. We replace the nonlinear equality
constraints given in (3a)–(3f) with the convex hull formulation for formwork reuse, as
shown below:
k ∈ K set of sizes for element width and depth; be and he ; respectively:
Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
860
A. GUERRA, A. M. NEWMAN, AND S. LEYFFER
Binary variables.
zbek ¼ 1 if the kth size is used for the width of element e ∀e ∈ E x , ∀k ∈ K , 0
otherwise;
zhek ¼ 1 if the kth size is used for the depth of element e ∀e ∈ E x , ∀k ∈ K , 0
otherwise.
Continuous variables.
be width of element e ∀e ∈ E [cm],
he depth of element e ∀e ∈ E [cm],
we ¼ 1 if the eth element formwork is already built ∀e ∈ E x , 0 otherwise,
zcb
ee 0 k ¼ 1 if width of column element e is the same kth size as the width of column
element e 0 ∀e, e 0 ∈ E x ∩ E c , ∀k ∈ K ∋ e ≤ e 0 , 0 otherwise,
ch
zee 0 k ¼ 1 if the depth of column element e is the same kth size as the depth of
column element e 0 ∀e, e 0 ∈ E x ∩ E c , ∀k ∈ K ∋ e ≤ e 0 , 0 otherwise,
bb
zee 0 k ¼ 1 if the width of beam element e is the same kth size as the width of beam
element e 0 ∀e, e 0 ∈ E x ∩ E b , ∀k ∈ K ∋ e ≤ e 0 , 0 otherwise,
bh
zee 0 k ¼ 1 if the depth of beam element e is the same kth size as the depth of beam
element e 0 ∀e, e 0 ∈ E x ∩ E b , ∀k ∈ K ∋ e ≤ e 0 , 0 otherwise,
c
yee 0 ¼ 1 if column element e contains the same width and depth as column
element e 0 ∀e, e 0 ∈ E x ∩ E c ∋ e ≤ e 0 , 0 otherwise,
b
yee 0 ¼ 1 if beam element e contains the same width and depth as beam element
e 0 ∀e, e 0 ∈ E x ∩ E b ∋ e ≤ e 0 , 0 otherwise,
X
ð9aÞ
k · zbek ∀e ∈ E x ;
be ¼
k∈K
ð9bÞ
he ¼
X
k · zhek
∀e ∈ E x ;
k∈K
X
zbek ¼ 1
ð9cÞ
∀e ∈ E x ;
k∈K
X
zhek ¼ 1
ð9dÞ
∀e ∈ E x ;
k∈K
ð9eÞ
b
b
0
x
c
0
zcb
ee 0 k ≥ zek þ ze 0 k − 1 ∀e; e ∈ E ∩ E ; ∀k ∈ K ∋ e ≤ e ;
ð9fÞ
b
zcb
ee 0 k ≤ zek
∀e; e 0 ∈ E x ∩ E c ; ∀k ∈ K ∋ e ≤ e 0 ;
ð9gÞ
b
zcb
ee 0 k ≤ ze 0 k
∀e; e 0 ∈ E x ∩ E c ; ∀k ∈ K ∋ e ≤ e 0 ;
ð9hÞ
h
h
0
x
c
0
zch
ee 0 k ≥ zek þ ze 0 k − 1 ∀e; e ∈ E ∩ E ; ∀k ∈ K ∋ e ≤ e ;
∀e; e 0 ∈ E x ∩ E c ; ∀k ∈ K ∋ e ≤ e 0 ;
ð9iÞ
h
zch
ee 0 k ≤ zek
ð9jÞ
h
0
x
c
0
zch
ee 0 k ≤ ze 0 k ∀e; e ∈ E ∩ E ; ∀k ∈ K ∋ e ≤ e ;
X
ch
0
x
c
0
≥
ðzcb
ee 0 k þ zee 0 k Þ − 1 ∀e; e ∈ E ∩ E ; ∋ e ≤ e ;
ð9kÞ
ycee 0
k∈K
ð9lÞ
ycee 0 ≤
X
zcb
ee 0 k
∀e; e 0 ∈ E x ∩ E c ; ∋ e ≤ e 0 ;
k∈K
ð9mÞ
ycee 0
X
≤
zch
ee 0 k
∀e; e 0 ∈ E x ∩ E c ; ∋ e ≤ e 0 ;
k∈K
Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
CONCRETE STRUCTURE DESIGN
∀e; e 0 ∈ E x ∩ E b ; ∀k ∈ K ∋ e ≤ e 0 ;
ð9nÞ
b
b
zbb
ee 0 k ≥ zek þ z e 0 k − 1
ð9oÞ
b
zbb
ee 0 k ≤ zek
∀e; e 0 ∈ E x ∩ E b ; ∀k ∈ K ∋ e ≤ e 0 ;
ð9pÞ
b
zbb
ee 0 k ≤ ze 0 k
∀e; e 0 ∈ E x ∩ E b ; ∀k ∈ K ∋ e ≤ e 0 ;
ð9qÞ
h
h
zbh
ee 0 k ≥ zek þ ze 0 k − 1
ð9rÞ
h
zbh
ee 0 k ≤ zek
ð9sÞ
h
0
x
b
0
zbh
ee 0 k ≤ ze 0 k ∀e; e ∈ E ∩ E ; ∀k ∈ K ∋ e ≤ e ;
X
bh
0
x
b
0
≥
ðzbb
ee 0 k þ z ee 0 k Þ − 1 ∀e; e ∈ E ∩ E ; ∋ e ≤ e ;
ð9tÞ
ybee 0
k∈K
ð9uÞ
ybee 0 ≤
861
∀e; e 0 ∈ E x ∩ E b ; ∀k ∈ K ∋ e ≤ e 0 ;
∀e; e 0 ∈ E x ∩ E b ; ∀k ∈ K ∋ e ≤ e 0 ;
X
zbb
ee 0 k
∀e; e 0 ∈ E x ∩ E b ; ∋ e ≤ e 0 ;
k∈K
ð9vÞ
ybee 0
X
≤
zbh
ee 0 k
∀e; e 0 ∈ E x ∩ E b ; ∋ e ≤ e 0 ;
k∈K
X
ð9wÞ
e∈E x ∩E c ;∋e≤e 0
X
ð9xÞ
ycee 0 ≥ we 0
∀e 0 ∈ E x ∩ E c ;
ybee 0 ≥ we 0
∀e 0 ∈ E x ∩ E b .
e∈E x ∩E b ;∋e≤e 0
We note that the reformulated MINLP problem is nonconvex because of the presence of nonlinear equations describing the finite element analysis.
Acknowledgments. The authors are grateful to Professor Panos Kiousis in the
Engineering Division at the Colorado School of Mines for introducing this problem
to them.
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
(ACI), Committee 318, 318-05/318R-05: Building Code Requirements for
Structural Concrete and Commentary, Technical reports ACI 318-05 and ACI 318R-05, Detroit, MI,
2005.
AMPL OPTIMIZATION LLC, AMPL Version 10.6.16, www.ampl.com.
M. ANITESCU, Global convergence of an elastic mode approach for a class of mathematical programs with
complementarity constraints, SIAM J. Optim., 16 (2005), pp. 120–145.
R. J. BALLING, Decomposition methods in structural optimization, in ASCE Recent Adv. in Optimal
Structural Des., S. A. Burns, ed., ASCE, New York, 2002, pp. 41–53.
R. J. BALLING AND X. YAO, Optimization of reinforced concrete frames, J. Struct. Eng., 123 (1997),
pp. 193–202.
J. F. BARD, Convex two-level optimization, Math. Program., 40 (1988), pp. 15–27.
C. V. CAMP, S. PEZESHK, AND H. HANSSON, Flexural design of reinforced concrete frames using a genetic
algorithm, J. Struct. Eng., 129 (2003), pp. 105–115.
L. CORRADI, O. DE DONATO, AND G. MAIER, Inelastic analysis of reinforced concrete frames, J. Struct.
Div., 100 (1974), pp. 1925–1942.
O. DE DONATO AND G. MAIER, Mathematical programming methods for the inelastic analysis of reinforced
concrete frames allowing for limited rotation capacity, Internat. J. Numer. Methods Engrg., 4
(1972), pp. 307–329.
AMERICAN CONCRETE INSTITUTE
Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
862
A. GUERRA, A. M. NEWMAN, AND S. LEYFFER
[10] V. DE MIGUEL, M. P. FRIEDLANDER, F. J. NOGALES, AND S. SCHOLTES, A two-sided relaxation scheme
for mathematical programs with equilibrium constraints, SIAM J. Optim., 16 (2005),
pp. 587–609.
[11] K. S. DINNO AND B. B. MEKHA, Optimal design of reinforced concrete frames based on inelastic analysis,
Comput. Struct., 47 (1993), pp. 245–252.
[12] K. S. DINNO AND B. B. MEKHA, A developed algorithm for the inelastic analysis of reinforced concrete
frames, Comput. Struct., 57 (1995), pp. 227–232.
[13] S. P. DIRKSE, (MPECLib Coordinator), MPEC World, (2001).www.gamsworld.org/mpec/.
[14] S. P. DIRKSE, M. C. FERRIS, AND A. MEERAUS, Mathematical programs with equilibrium constraints:
Automatic reformulation and solution via constraint optimization, in Frontiers in Applied General
Equilibrium Modeling, T. J. Kehoe, T. N. Srinivasan, and J. Whalley, eds., Cambridge University
Press, Cambridge, UK, 2005.
[15] M. DURAN AND I. E. GROSSMANN, An outer-approximation algorithm for a class of mixed integer nonlinear
programs, Math. Program., 36 (1986), pp. 307–339.
[16] M. J. FADAEE AND D. E. GRIERSON, Design optimization of 3D reinforced concrete structures, Struct.
Multidiscip. Optim., 12 (1996), pp. 127–134.
[17] M. C. FERRIS AND J. S. PANG, Engineering and economic applications of complementarity problems, SIAM
Rev., 39 (1997), pp. 669–713.
[18] M. C. FERRIS AND F. TIN-LOI, On the solution of a minimum weight elastoplastic problem involving displacement and complementarity constraints, Comput. Methods Appl. Mech. Engrg., 174 (1999),
pp. 107–120.
[19] R. FLETCHER AND S. LEYFFER, User Manual for Filter SQP, Numerical analysis report NA/181, Dundee
University, Dundee, Scotland, 1998.
[20] R. FLETCHER AND S. LEYFFER, Solving mathematical programs with complementarity constraints as nonlinear programs, Optim. Methods Softw., 19 (2004), pp. 15–40.
[21] R. FLETCHER, S. LEYFFER, D. RALPH, AND S. SCHOLTES, Local convergence of SQP methods for mathematical programs with equilibrium constraints, SIAM J. Optim., 17 (2006), pp. 259–296.
[22] R. FOURER, D. GAY, AND B. KERNIGHAN, AMPL: A Modeling Language for Mathematical Programming,
Thompson Learning, Pacific Grove, CA, 2003.
[23] A. M. GEOFFRION, Generalized benders decomposition, J. Optim. Theory Appl., 10 (1972), pp. 237–260.
[24] A. GUERRA, Design of Reinforced Concrete Using Mixed Integer Nonlinear Programming, Ph.D. dissertation, Colorado School of Mines, Golden, CO, 2008.
[25] A. GUERRA AND P. D. KIOUSIS, Design optimization of reinforced concrete structures, Comput. Concr., 3
(2006), pp. 313–334.
[26] B. HOROWITZ AND G. R. MORAES, Mathematical programming with complementarity constraints applied to
limited ductility analysis of concrete beams, in Proceedings of the 6th Official Congress of the
International Society for Stuctural and Multidisciplinary Optimization, Rio De Janeiro, Brazil,
2005.
[27] IBM, ILOG CPLEX, Incline Village, NV, 2009.
[28] M. INEL AND H. B. OZMEN, Effects of plastic hinge properties in nonlinear analysis of reinforced concrete
buildings, Eng. Struct., 28 (2006), pp. 1494–1502.
[29] I. KANEKO, A mathematical programming method for the inelastic analysis of reinforced concrete frames,
Internat. J. Numer. Methods Engrg., 11 (1977), pp. 1137–1154.
[30] C. S. KRISHNAMOORTHY AND D. R. MOSI, Optimal design of reinforced concrete frames based on inelastic
analysis, Eng. Optim., 5 (1981), pp. 151–167.
[31] C. LEE AND J. AHN, Flexural design of reinforced concrete frames by genetic algorithm, J. Struct. Eng., 129
(2003), pp. 762–774.
[32] S. LEYFFER, User Manual for MINLPBB, University of Dundee numerical analysis report, NA/XXX,
Dundee University, Dundee, Scotland, 1999.
[33] S. LEYFFER, MacMPEC: AMPL Collection of MPECs, www.mcs.anl.gov/~leyffer/MacMPEC/ (2000).
[34] S. LEYFFER, Integrating SQP and branch-and-bound for mixed integer nonlinear programming,
Comput. Optim. Appl., 18 (2001), pp. 295–309.
[35] S. LEYFFER, G. LOPEZ-CALVA, AND J. NOCEDAL, Interior methods for mathematical programs with complementarity constraints, SIAM J. Optim., 17 (2006), pp. 52–77.
[36] Z. Q. LUO, J. S. PANG, AND D. RALPH, Mathematical Programs with Equilibrium Constraints, Cambridge
University Press, Cambridge, UK, 1996.
[37] G. MAIER, F. GIANNESSI, AND A. NAPPI, Indirect identification of yield limits by mathematical programming, Eng. Struct., 4 (1982), pp. 86–98.
Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
CONCRETE STRUCTURE DESIGN
863
[38] J. OUTRATA, M. KOCVARA, AND J. ZOWE, Nonsmooth Approach to Optimization Problems with Equilibrium
Constraints, Kluwer Academic Publishers, Boston, MA, 1998.
[39] A. RAGHUNATHAN AND L. T. BIEGLER, An interior point method for mathematical programs with complementarity constraints (MPCCs), SIAM J. Optim., 15 (2005), pp. 720–750.
[40] N. V. SAHINIDIS, private communication, 2006.
[41] H. SCHEEL AND S. SCHOLTES, Mathematical programs with complementarity constraints: Stationarity, optimality, and sensitivity, Math. Oper. Res., 25 (2000), pp. 1–22.
[42] S. SCHOLTES, Convergence properties of regularization schemes for mathematical programs with complementarity constraints, SIAM J. Optim., 11 (2001), pp. 918–936.
[43] STRUCTURAL ENGINEERING INSTITUTE (SEI) AND AMERICAN SOCIETY OF CIVIL ENGINEERS (ASCE), SEI/ASCE 7-98:
Minimum Design Loads for Buildings and Other Structures, Technical report, SEI and ASCE, Reston, VA, 2000.
[44] M. TAWARMALANI AND N. V. SAHINIDIS, Convexification and Global Optimization in Continuous and
Mixed-Integer Nonlinear Programming: Theory, Algorithms, Software, and Applications,
Nonconvex Optim. Appl. 65, Kluwer Academic Publishers, Boston, MA, 2002.
[45] T. WESTERLUND AND F. PETTERSSON, An extended cutting plane method for solving convex MINLP
problems, Computers Chem. Eng. Sup., 19 (1995), pp. 131–136.
Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement