Professional Documents
Culture Documents
2014 Lee and Madanat - Joint Optimization of Pavement Design, Resurfacing and Maintenance Strategies With History-Dependent Deterioration Models
2014 Lee and Madanat - Joint Optimization of Pavement Design, Resurfacing and Maintenance Strategies With History-Dependent Deterioration Models
a r t i c l e i n f o a b s t r a c t
Article history: The subject of this paper is the joint optimization of pavement design, maintenance and
Received 26 November 2013 resurfacing (M&R) strategies. This problem is solved for continuous pavement state, con-
Received in revised form 11 June 2014 tinuous time, infinite planning horizon and non-Markovian (history-dependent) pavement
Accepted 13 June 2014
deterioration model. This paper presents a mathematical formulation of the joint optimi-
zation problem to minimize the total discounted lifecycle costs, using a combination of
analytical and numerical tools. The lifecycle costs include both user costs and agency (con-
Keywords:
struction, resurfacing and maintenance) costs. This paper shows that resurfacing schedule
Joint optimization
Pavement design
converges to a steady state after a few resurfacing cycles. The research results should be of
M&R strategies use to developing countries in the process of expanding their highway networks facing
Non-Markovian model multiple constraints.
Convergence 2014 Elsevier Ltd. All rights reserved.
Steady-state
1. Introduction
Governments in developing countries need to construct transportation infrastructure to meet rapidly growing demand,
under strict budget constraints. Insufficient transportation capacity forces heavier traffic onto restricted roadways, which
accelerates the aging of the infrastructure. As a consequence, poor road conditions yield higher user costs. Due to the large
construction and maintenance costs, as well as high user costs, it is necessary to address simultaneously the problems of
pavement design, maintenance and resurfacing.
Pavement design refers to the selection of the thickness of the pavement layers, with consideration of the properties of
the materials, environmental conditions and expected traffic loading. After pavement facilities have been built, it is neces-
sary to allocate limited resources for maintenance and resurfacing (M&R) activities. Pavement deterioration rates are influ-
enced by pavement design. More durably designed pavements require less frequent resurfacings and a lower level of
maintenance. Therefore, pavement design influences future M&R strategies after construction. Furthermore, pavement
design directly affects the construction costs, which are very large and need to be invested all at one time.
Pavement deterioration is influenced by traffic loading and various environmental factors. Therefore, system-level opti-
mization should be solved by bottom-up approaches which account for the heterogeneity of roadway pavement networks.
Bottom-up approaches start by solving for single-facility optimization problems, and this paper focuses on the single-facility
http://dx.doi.org/10.1016/j.trb.2014.06.008
0191-2615/ 2014 Elsevier Ltd. All rights reserved.
142 J. Lee, S. Madanat / Transportation Research Part B 68 (2014) 141–153
optimization problems. The approach used in this research is based on a deterministic and continuous deterioration model.
This model is non-Markovian: it accounts for the influence of M&R history on the deterioration rate, which is physically more
realistic.
In Section 2, the literature related to pavement design and management strategies is reviewed. In Section 3, a method-
ology for solving the joint optimization problem of design and M&R is presented. The solution methodology is utilized in a
set of parametric studies in Section 4. The conclusion is presented in Section 5.
2. Literature review
The review of the literature is organized around the two components of our problem: pavement design and pavement
management.
The most commonly used pavement design approach in practice is the AASHTO design method (AASHTO, 1986). It is
based on a regression analysis with experimental data which was collected from only one subgrade material and from only
one type of climatic region as well. This method is based on statistical analysis of pavement failure in an accelerated loading
experiment. The statistical method used did not appropriately account for censored observations, which led to statistical
bias, as Small and Winston (1988) and Madanat et al. (2002) noted.
Small and Winston (1988) and Madanat et al. (2002) corrected for censoring bias by improved statistical estimation of the
coefficients in the deterioration model based on the Tobit and Weibull models, respectively. These authors also found
solutions for the optimization problem of initial pavement design; given a pre-determined rehabilitation strategy.
Pryke et al. (2006) and Rajbongshi and Das (2008) presented mechanistic-empirical design methods, which included
better representation of the physical properties of materials. McDonald and Madanat (2012) formulated an optimization
problem for mechanistic-empirical pavement design to minimize the total life cycle costs of asphalt/concrete pavements.
The tradeoffs between constructing a more durable pavement with a higher structural number at the start of the time hori-
zon, and more frequent resurfacings, were evaluated quantitatively. Optimal pavement design was determined by nonlinear
mathematical programming. The maintenance method assumed in their analysis is limited only to resurfacings.
There is a rich literature in the field of M&R optimization for pavement. Friesz and Fernandez (1979), Fernandez and
Friesz (1981), Markow and Balta (1985), and Markow et al. (1993) used optimal control theory to solve the single-facility
maintenance optimization problem with continuous state, and continuous time, to minimize life cycle costs. Other research
in pavement management involves discrete states and time (Golabi et al., 1982; Carnahan, 1988) or continuous states and
discrete time (Durango-Cohen and Sarutipand, 2007). In the context of continuous state formulations, discontinuous jumps
in the pavement condition trajectory caused by resurfacing activities, make the use of optimal-control impractical for
optimizing resurfacing strategies.
Tsunokawa and Schofer (1994) solved the resurfacing optimization problem using a ‘‘Trend Curve Optimal Control
Model’’ which is an approximate approach to overcome the discontinuity in the pavement trajectory, with a Markovian dete-
rioration model. Li and Madanat (2002) solved the same problem analytically, formulated with the same deterioration, cost
and performance models that Tsunokawa and Schofer (1994) used, without using optimal control. They showed that the
pavement management strategy reached steady state immediately after the initial resurfacing for an infinite time horizon.
Ouyang and Madanat (2006) derived an analytical methodology using calculus of variations to exactly solve the optimization
problem for the finite time horizon. Both papers (Li and Madanat, 2002; Ouyang and Madanat, 2006) revealed the threshold
structures of the optimal solution, and the robustness of their solutions to model uncertainty for finite (Li and Madanat,
2002) and infinite (Ouyang and Madanat, 2006) time horizons.
Rashid and Tsunokawa (2012) expanded the problem beyond resurfacing to include resealing and reconstruction as treat-
ments also, and formulated the joint optimization problem for a single facility using the trend-curve approach. As shown in
the literature (Ponniah and Kennepohl, 1996; Labi and Sinha, 2003), a routine maintenance activity such as resealing slows
down the deterioration of pavement, but Rashid and Tsunokawa (2012) assumed that it affects the pavement condition
immediately in a way similar to resurfacing. Gu et al. (2012) proposed a methodology which utilizes the optimality condition
found by Ouyang and Madanat (2006) to solve the joint optimization problem for maintenance and resurfacing with a
realistic model where maintenance slows down the deterioration process. Their work suffers from the use of a Markovian
deterioration model, which is usually not realistic.
3. Methodology
This paper is focused on the joint optimization problem of pavement design and M&R strategies. These include frequent
maintenance activities aimed at slowing down the deterioration process, and resurfacing activities aimed at reducing
J. Lee, S. Madanat / Transportation Research Part B 68 (2014) 141–153 143
pavement roughness immediately. The effects from maintenance activities and from resurfacings are assumed to be inde-
pendent. The scope of this paper is focused on the case of continuous time and continuous state of pavement condition.
The pavement deterioration model of Paterson (1987) is non-Markovian. The deterioration rate depends not only on the
current condition of the pavement, but also on its history of resurfacing activity. Thus, a recently resurfaced pavement has a
slower deterioration rate than a pavement resurfaced previously, even though both current surface conditions are identical.
Additionally, pavement deterioration depends on traffic loading and the structural number of the pavement (i.e., heavier
traffic causes faster deterioration and thicker pavements retard deterioration). The expected traffic loading is assumed to
be fixed and exogenous in the optimization problem, and not responsive to pavement condition.
The objective of the optimization is to minimize the net present value of discounted lifecycle costs over an infinite time
horizon, including initial construction costs, agency investments for both maintenance activities and resurfacings, and user
costs as shown in (1a). The decision variables consist of the structural number, N, in terms of pavement thickness, selected at
t = 0 and assumed to be constant throughout the planning horizon, the intensities (i.e. thickness), w = {w1, w2,. . .}, and tim-
ings t = {t1, t2,. . .}, of resurfacings consecutively numbered as i = 1, 2, 3. . ., and the level of maintenance activities, D , which is
expressed as a reduction in the deterioration rate: the deterioration rate b decreases as the intensity of maintenance activ-
ities increases. The assumption of constant pavement thickness is reasonable, because thickness is constant if the agency
applies a resurfacing after removing the top-most layer of asphalt concrete, which is the practice in California and in many
countries. Given the above the objective function and decision variables, the problem can be formulated as following:
1 Z
X t iþ1 X
1
min JðN; t; w; DbÞ ¼ KðNÞ þ ðC ðsðt ÞÞ þ C M ðsðt Þ; DbÞÞert dt þ Mðwi Þerti ð1aÞ
i¼0 tþ
i i¼1
s:t: sðti Þ s tþi ¼ Gðwi ; sðt i ÞÞ; 8i ¼ 1; 2; . . . ð1bÞ
sðt Þ ¼ F s tþi1 ; t t i1 ; b; N ; 8t 2 t þi1 ; t i ; i ¼ 1; 2 . . . ð1cÞ
þ
s t i P s0 ; 8i ¼ 1; 2; . . . ð1dÞ
0 < wi 6 Rðsðti ÞÞ; 8i ¼ 1; 2; . . . ð1eÞ
t0 ¼ 0; t i P t i1 ; 8i ¼ 1; 2; . . . ð1fÞ
b ¼ b 0 Db ð1gÞ
b < b 6 b0 ð1hÞ
Nmin 6 N 6 Nmax ð1iÞ
sð0Þ ¼ snew ð1jÞ
where
J(N, t, w, Db): net present value of lifecycle costs as a function of decision variables
s(t): pavement roughness at time t, and a higher value of s(t) refers to worse pavement condition
K(N): construction costs at the initial point in time, an increasing function of N
C(s(t)): user costs per unit time, an increasing function of s(t)
C M ðsðt Þ; DbÞ: agency costs for routine maintenance activities per unit time, an increasing function of both s(t) and Db
M(wi): agency costs for the ith resurfacing, an increasing function of wi
r: discount rate, assumed to be constant
tþ þ
i : time right after the ith resurfacing; t i lim!þ0 t i þ
G(wi, s(ti)): effectiveness of the ith resurfacing as a function of intensity wi and the current pavement roughness, s(ti)
s0: the best
achievable roughness after resurfacing
F s tþ i1 ; t t i1 ; b; N : non-Markovian deterioration model
Rðsðt i ÞÞ: upper bound of the effective resurfacing intensity, a function of current pavement roughness
b0: deterioration rate without maintenance activities
b: minimum achievable level of the deterioration rate by applying maintenance activities
Nmin, Nmin: lower and upper bounds of structural number
snew: pavement roughness achieved by construction; s0 > snew > 0 due to technological reasons
By the definition of tþ i , we can define the pavement roughness as an injective function expanded in the continuous time
horizon. Constraint (1b) describes the resurfacing effectiveness as a reduction in roughness. Constraint (1c) describes the
non-Markovian deterioration process. The term, t ti1, defines the history-dependent properties. The best achievable level
of pavement condition is constrained in (1d). Constraint (1e) describes the range of effective resurfacing intensity, wi. The
improvement in pavement condition, G(wi, s(ti)), does not increase with wi beyond R(s(ti)) (Ouyang and Madanat, 2004). Con-
straint (1f) indicates the sequential order of resurfacings and t0 = 0. The deterioration rate is defined by (1g) and constrained
by (1h). Constraint (1i) describes the range of the structural number. The initial pavement roughness is given in (1j), which is
always lower than the best achievable pavement roughness by resurfacing, s0, because of permanent loss.
144 J. Lee, S. Madanat / Transportation Research Part B 68 (2014) 141–153
In this section, we describe the pavement deterioration, cost and effectiveness models. We use the pavement deteriora-
tion model developed by Paterson (1987), shown in (2). The pavement roughness function is continuous between consecu-
tive resurfacings.
F s t þi1 ; t t i1 ; b; N ¼ s t þi1 þ alðt t i1 Þ½1 þ Nq ebðtti1 Þ ; 8t 2 tþi1 ; ti ; i ¼ 1; 2 . . . ð2Þ
In (2), a and q are constants. From empirical research, it is known that q has negative value, so a higher structural number,
N, slows down the deterioration process. The traffic loading per unit time is denoted as l, and l(t ti1) indicates the cumu-
lative traffic loading from the time of last resurfacing. We assume that l is exogenous, and independent of pavement condi-
tion and M&R activities. In practice, the traffic loading is influenced by pavement condition, as well as the deterioration
process depends on the traffic loading. In addition, an implementation of M&R activities on pavement causes significant user
costs from travel time delay due to the loss in network capacity. These dynamic interactions among the traffic loading, pave-
ment condition and M&R strategies are influenced by the configuration of a pavement network, but this paper is focusing on
a single facility, so we assume the constant traffic loading along the planning horizon to simplify the problem. With the con-
stant traffic loading, al[1 + N]q becomes a function of the structural number; the term X(N) denotes al[1 + N]q. The cost and
effectiveness models are based on empirical studies; these models are:
KðNÞ ¼ k1 N þ k2 ð3Þ
Mðwi Þ ¼ m1 wi þ m2 ; 8i ¼ 1; 2; . . . ð5Þ
g 1 wi sðt i Þ
Gðwi ; sðt i ÞÞ ¼ ; 8i; 0 6 wi 6 Ri ¼ g 2 sðt i Þ þ g 3 ð6Þ
g 2 sðt i Þ þ g 3
1 g
wi ¼ sðt i Þ sðt þi Þ g 2 þ 3 ð7Þ
g1 sðti Þ
C M ðsðtÞ; DbÞ ¼ C M;1 ðDbÞs ti þ C M;2 ðDbÞ ð8Þ
(
aj ebj Db þ cj ; if Db > 0
C M;j ðDbÞ ¼ ; 8j ¼ 1; 2 ð9Þ
0; if Db ¼ 0
The construction cost per unit length of a new pavement section is expressed as (3). It is a linear function of the structural
number, N. This is realistic because the construction cost increases as larger quantities of materials are used to obtain a
higher structural number. The constant term, k1, is the unit cost of pavement materials per unit of N. The constant term,
k2, does not influence the optimal solution, so it is assumed to be 0. The user cost per unit time, defined by (4), is a linear
function of pavement roughness, s(t), and proportional to the traffic loading, where c1 and c1 are positive constant param-
eters. Because it does not influence the optimal solution, we set c2 = 0. The resurfacing cost and effectiveness models, shown
in (5) and (6), are taken from Ouyang and Madanat (2004), where m1, m2, g1, g2 and g3 are positive constant parameters. The
intensity of resurfacings, wi, is shown in (7). The maintenance cost per unit time is taken from Gu et al. (2012) as shown in (8)
and (9), where aj, bj and cj are positive constant parameters, "j = 1, 2. The maintenance cost per unit time is positively related
to Db, and it is a linear function of pavement roughness.
3.3. Solution
The optimization problem formulated in this paper is a mixed-integer (binary) nonlinear program (MINLP), and it has four
categories of decision variables: N, t, w and Db. It is difficult to solve this problem by using commercial solvers such as
CPLEX, because the formulations of this problem do not follow the standard forms, which commercial solvers are applicable.
In the solution methodology, we defines the problem as a bi-level optimization problem. We find the lower-level optimal
solution of resurfacing policies, (t, w), with fixed (N, Db), according to the proposed algorithm shown in Section 3.3.1. If
we find the lower-level optimal resurfacing policies given (N, Db) by this algorithm, N and Db are left as the decision vari-
ables of the upper-level optimization problem.
because Db and N are fixed in the lower-level problem. The optimal times and intensities of resurfacings will vary with dif-
ferent traffic loading, and the values of parameters. Two cases are considered:
1. The case where local optimum of pavement roughness achieved by resurfacing i is not the best achievable pavement
roughness, sðtþ Þ>s .
i 0
2. The case of s t þ
i ¼ s0 .
and an alternative resurfacing policy ½S A B00 C00 is constructed. A first order necessary condition for the original curve
to be local optimal is
dJ ¼ 0; if wi 2 ð0; Ri Þ
jgi ¼0 ð10Þ
dgi 6 0; if wi ¼ Ri
According to the Proposition 2 of Ouyang and Madanat (2006), ddJgi 6 0 for all resurfacings. In our problem, Eq. (11) is the first
i
order derivative of J with respect to the infinitesimal gi at gi = 0.
Z siþ1
dJ d d
¼ c0 F ðF ðsðt i Þ; si Þ Gðwi þ gi ; sðt i ÞÞ; uÞeru du þ Mðwi þ gi Þ
dgi 0 dg i d gi
0 ! )
d Gw wi ; s t i
þ M wiþ1 gi F 0sðtþ Þ s tþi ; tiþ1 t i 0 i ersiþ1 ert1 ð11Þ
dg i i G wiþ1 ; s t
wiþ1 iþ1
where
si ti ti1 ; 8t ¼ 1; 2; . . .
dJ
dgi
in (11) has a similar expression to dJ
dgi
in Ouyang and Madanat (2006). Only C 0sðtþ Þ ðsðtÞÞ used in Ouyang and Madanat
i
0
(2006) is replaced by c , but this does not affect the validity of the proposition. That is, the optimal intensities of all resur-
facings, wi, are the maximum intensities, Ri, as shown in (12).
wi ¼ Ri ; 8i ¼ 1; 2; . . . ð12Þ
0 0 0
Suppose ti is perturbed to ti + ei. An alternative resurfacing policy [S–A –B –C ] is constructed, and wi and wi+1 are per-
turbed simultaneously. Because we suppose that [S–A–B–C] is part of an optimal resurfacing policy, so wi at ti is g2s(ti) + g3,
dF ðsðt þ
i Þ iÞ
;e dF ðsðt þ
i1 Þ i
;s þei Þ
and wi at ti + ei is g2s(ti + ei) + g3. The point B0 is not on the curve [B–C], because de
> ð1 g 1 Þ de
. At the i + 1th
i i
resurfacing, the resurfacing policy changes to [C0 –D] from [C–D]. In (7), sðtþ
iþ1 Þ is fixed, but only s(ti+1) is perturbed. A first
necessary condition of the locally optimal curve [S–A0 –B0 –C0 ] is (13) where ei = 0. Except the case of ddJe ¼ 0; the optimal resur-
i
dJ dJ
facings, constructed with dei
P 0 or dei
6 0; occur only when the initial condition is poor or at and the end point of the finite
planning horizon. This cannot happen in the case of an infinite time horizon and a newly constructed facility. Therefore,
dJ
dei
¼ 0 is the local optimality condition. We obtain the first order necessary condition for optimality of the ith resurfacing as
2
(14). ddJe is a function of sðtþ d J
i1 Þ (point ‘‘S’’ in Fig. 1), si and si+1 (see (14) and Appendix A). The second order condition, de2 P 0, is
i i
Steady state conditions are defined as sðt i Þ ¼ sðtiþ1 Þ and si = si+1. We define two new terms: s⁄ is the optimal time between
two consecutive resurfacings and s⁄ is the optimal trigger roughness when we assume that the optimal resurfacing policy
reaches a steady state.
Proposition 1. On an optimal roughness resurfacing policy, si P siþ1 P s and sðti Þ 6 sðtiþ1 Þ 6 s ; 8i if sðtþ1 Þ > s0 .
3.3.1.2. Case of sðt þ i Þ ¼ s0 . Suppose we have a part of an optimal resurfacing policy, [S–A–B–C], as shown in Fig. 2. If ti is per-
turbed into ti + ei, then we obtain an alternative resurfacing policy [S–A0 –B0 –C0 ]. Note that the inequality constraint (1d) is
binding as sðt þ i Þ ¼ s0 , and an optimal resurfacing intensity, wi, is less than Ri. The first order derivative of wi with respect
dwi
to s(ti) is expressed as (18), which is different from the previous case of sðt þ i Þ > s0 . By Substituting (18) for both dsðt i Þ and
dwiþ1
dsðt iþ1 Þ
into the first order necessary condition for optimality of the i^{th} resurfacing, shown in (19), we can formulate the
first order necessary condition as a function of sðtþ i1 Þ, si and siþ1 .
!
dwi 1 s0 g 3
¼ g2 þ ð18Þ
dsðti Þ g 1 sðt i Þ2
( (Z Z siþ1 )
ei ru
dJ 0 d þ ru
¼ c F s t i1 ; siþ1 þ u F ðs0 ; uÞ e du þ ðF ðs0 ; u ei Þ F ðs0 ; uÞÞe du
dei dei 0 ei
dwi dsðti Þ dwiþ1 0
þ M 0wi ðwi Þ rM ðwi Þ þ M0wiþ1 ðwiþ1 Þ F ei ðs0 ; siþ1 ei Þersiþ1 erti ¼ 0 ð19Þ
dsðt i Þ dei dsðtiþ1 Þ
We can find the same properties of convergence for an optimal solution in the previous section: si P siþ1 P s and
sðt i Þ 6 sðtiþ1 Þ 6 s for all i. We can also find an optimal solution in a steady state numerically based on (19) when (1d) is
binding.
3.3.2. Finding the optimal solution of (t,w) for given ðN; DbÞ
In this section, we propose a simple method to obtain the optimal solution of (t, w) based on the local optimality pre-
sented in Sections 3.3.1.1 and 3.3.1.2. Eqs. (14) and (19) have two unknown variables, si and si+1 and there is a unique
si+1 correspondent to a certain si, where si ; siþ1 P s . By sðtþ0 Þ ¼ snew , we can find a narrow range of s1 satisfying
s1 P s2 P s . For each possible value of s1 in the feasible range, we can find s2, s3,. . . successively. We omit a candidate
of s1 when si and s(ti), derived from the candidate value of s1, violate the properties of convergence. We assume that a steady
state starts from the nth resurfacing when (20) is satisfied.
s sn1 s sn
>h> ð> 0Þ ð20Þ
s s
where h is defined as a positive allowable error level.
We use the bisection method to find s1 which satisfies si P siþ1 P s ; 8i < n:s1;k and Ds1,k indicate s1 and the interval
respectively for the kth iteration in the bisection method. For the first iteration, s1,1 is the mid-point of the range, satisfying
s1 P s2 P s , and the interval of this range is Ds1,1. For each iteration before reaching a steady state, if si,k < si+1,k, go to the
next iteration and s1,k+1 = s1,k Ds1,k+1. If si+1,k < s⁄, go to the next iteration and s1,k+1 = s1,k + s1,k+1. If convergence (20) is sat-
isfactory, return s1 and stop iterating. Finally we obtain the lower-level optimal resurfacing plan (t, w) for given ðN; DbÞ. The
cost functions and performance functions described in the previous section are substituted into the objective function (1a).
The upper-level objective function, J ðN; DbÞ, is given as (21).
X
n
ertn
J ðN; DbÞ ¼ k1 N þ erti1 J i þ J ð21Þ
i¼1
1 ers
where
Z si
Ji ¼ ðc1 l þ C M;1 ÞF s t þi1 ; u; b; N þ C M;2 eru du þ M wi ersi
0
Z s
J ¼ ðc1 l þ C M;1 ÞF sþ ; u; b; N þ C M;2 eru du þ Mðw Þers
0
sþ ¼ max ðð1 g 1 Þs ; s0 Þ
4. Parametric study
This section demonstrates the proposed algorithm to solve the joint optimization problem with pavement design and
M&R strategies for different traffic loading by a parametric study. The parameters, b0 and b, in the non-Markovian
148 J. Lee, S. Madanat / Transportation Research Part B 68 (2014) 141–153
Table 1
Parameters used in the parametric study.
–, dimensionless or omitted; IRI, international roughness index (m/km); ESAL, equivalent single axle load; SN, structural number.
Table 2
Optimal pavement design and M&R strategies obtained by the proposed methodology.
Fig. 3. Total discounted lifecycle costs versus structural number when maintenance is not applied (l = 0.8).
deterioration model are selected to have a similar deterioration speed to the Markovian models used in previous works
(Ouyang and Madanat, 2006; Gu et al., 2012). The parameters of maintenance costs are taken from Gu et al. (2012). We
use parameter values for a single pavement with 1 km long of one lane highway as shown in Table 1. The solution algorithms
are programmed in MATLAB, and the optimization problems are solved on a Windows 7 Professional OS with a 2.93 GHz
processor and 3 GB RAM.
The optimal pavement designs and M&R strategies for different l are summarized in Table 2. An allowable error, h, is set as
102. In every case, an optimal resurfacing policy reaches the steady state at most in five resurfacings in the lower-level opti-
mization within 10 iterations by using the bisection method, and it is not significantly sensitive to the magnitude of the
allowable error. The table shows that higher structural number, lower trigger roughness of resurfacing and higher mainte-
nance level are jointly needed as the traffic loading increases, in general. However, as shown in the cases of l = 0.5 and l = 0.6,
lower structural number and longer resurfacing period are optimal in heavier traffic, because routine maintenance is applied
only when l = 0.6. In lighter traffic situations (l = 0.4 and 0.5), applying maintenance is suboptimal.
Fig. 3 shows that the structural number of 10.0 is the optimal pavement design when l = 0.8 and maintenance is not
applied, and the minimum total discounted lifecycle costs are 6.39 105 dollars. Fig. 4 exhibits the contour lines of total dis-
counted lifecycle costs versus maintenance level and structural number when l = 0.8 and maintenance is applied. The opti-
mal pavement design is 9.8, and the maintenance level, Db, is 0.0074. In this case, the minimum discounted lifecycle costs
J. Lee, S. Madanat / Transportation Research Part B 68 (2014) 141–153 149
11
6.5
10
8.5
9
9.5
8
7.5
6.5
10.5
7
10
Structural number (
9.5
6.5
10
8.5
9.5
7.5
8
9
6.5
9
7
8.5
6.5
8 7
10
9. 5
9
8.5
7.5
7. 5
8
7
7.5 7
7
0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016 0.018 0.02
Maintenance level ( )
Fig. 4. Total discounted lifecycle costs (105$) when maintenance is applied (l = 0.8).
11
3.5
10.5
3.4
3.3
3.6
10
3.7
3.8
3.9
Structural number (
9.5
3.5
9
4
3.4
3.6
3.7
8.5
3.
8
3.
9
8 4
4.1
3.5
7.5
3.
9
3.7
3.
3.6
4.2
8
4
7
0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016 0.018 0.02
Maintenance level ( )
are 6.31 105 dollars, so applying maintenance is optimal. In heavier traffic loading (l = 0.6, 0.7 and 0.8), the desirable
pavement roughness after resurfacing in a steady state reaches the best achievable state, s0. Figs. 5a–5c are contour plots
of user costs, resurfacing costs and maintenance costs respectively. Fig. 5a shows how maintenance level and pavement
design influence user costs. More durably designed pavement, which is managed higher level of routine maintenance, keeps
its average roughness in lower level, so user costs decrease. Figs. 5b and 5c reveal how resurfacing and maintenance costs
change with pavement design and maintenance level. Maintenance costs slightly decrease with structural number but these
are not very sensitive. As traffic goes heavier, the agency should perform resurfacings more frequently with lower overlay
thicknesses and apply a higher level of maintenance to avoid significant user costs, because user costs proportionally
increase with traffic loading.
In light traffic (l = 0.4), Figs. 6 and 7 show the influence of the decision variables on total discounted lifecycle costs for both
cases: maintenance is applied and not applied. When maintenance is not applied, the minimum total discounted lifecycle
costs are 4.11 105 dollars with structural number, N = 8.5. With maintenance, the minimum total discounted lifecycle costs
are 4.17 105 dollars with maintenance level, Db = 0.0070, and the lower structural number, N = 8.2. In the case of light
traffic, applying maintenance is suboptimal.
150 J. Lee, S. Madanat / Transportation Research Part B 68 (2014) 141–153
11
10.5
0. 8
0. 4
0.4
10
0.6
Structural number (
9.5
0.
9
0.4
0.4
8.5
0.
6
1.
2
8 1.4 1 0.8
6
0.6
0.
7.5 1.6
1.2
1.8
1.4 1 0.8
8
2
0.
7
0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016 0.018 0.02
Maintenance level ( )
11
0.5
1.5
2.5
4.5
1
4
5
3.5
55
10.5
10
Structural number (
9.5
4.5
0.5
1.5
2.5
4
5
1
9
3.5
5.5
8.5
8
0.5
2
2.5
3
4
4.5
5
1
1.5
7.5
3.5
5.5
7
0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016 0.018 0.02
Maintenance level ( )
Fig. 6. Total discounted lifecycle costs versus structural number when maintenance is not applied (l = 0.4).
J. Lee, S. Madanat / Transportation Research Part B 68 (2014) 141–153 151
11
6.25
5.75
7.75
8.25
8.75
6.75
7.25
5.25
4.75
10.5
8.5
6.5
7.5
5.5
4.5
4.5
8
7
5
10
9
Structural number (
9.5
5
4.2
9
5.25
7.75
8.25
8.75
6.75
7.25
5.75
6.25
4.75
5.5
8.5
6.5
7.5
4.5
4.25
8
7
8.5
4. 25
8
7.5
4.5
4.75
5.25
4.2
5.75
6.25
6.75
7.25
7.75
8.25
8.75
5.5
6.5
7.5
8.5
4.5
9
5
7
8
7
0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016 0.018 0.02
Maintenance level ( )
Fig. 7. Total discounted lifecycle costs ( 105$) when maintenance is applied (l = 0.4).
5. Conclusion
The solution methodology presented in paper is a promising approach to the joint optimization problem of pavement
design and M&R strategies when pavement deterioration is non-Markovian. The method uses the calculus of variations to
find optimal resurfacing policies for both cases: when the constraint of the best achievable pavement roughness after resur-
facing, (1d), is not binding, and when it is binding. It is observed from the results of the parametric study that the optimal
solution, under heavier traffic loading, generally requires a higher structural number for pavement design, higher levels of
maintenance and a lower pavement-roughness trigger. Another finding is that resurfacing planning converges to the steady
state even if the pavement deterioration model has non-Markovian properties. In previous works, the ‘steady state in opti-
mality’ was guaranteed by the use of a Markovian deterioration model. A non-Markovian deterioration model is more real-
istic, because it recognizes that deterioration is influenced by the history of previous resurfacing plans. In practice,
convergence to the steady state gives us a simple form for the optimal solution in the case of non-Markovian deterioration
model, similar to the optimal solution in the Markovian deterioration model case.
In this paper, a solution methodology for the optimization problem for a single pavement was proposed. The optimal
solution contains both the pavement design for a facility and its roughness trajectory, determined by the M&R strategy.
The next step is to extend the single facility problem to the system level problem under budget constraints, with necessary
modifications to ensure its general applicability, and to check the robustness of the solution to the uncertainty of the dete-
rioration models.
Appendix A.
dJ
dei
consists of four parts formulated at ei = 0: (A.1)–(A.3), (and) (A.4). All these parts are the functions of sðtþ
i1 Þ; si and siþ1 .
Z ei
d
ðFðsðt þi1 Þ; siþ1 þ uÞ Fðsðtþi Þ; uÞÞeru du ¼ g 1 sðt i Þ ¼ g 1 fsðt þi1 Þ þ si XðNÞgebsi ðA:1Þ
dei 0
Z siþ1 (
d XðNÞ
br
fð1 eðbrÞsiþ1 Þ r
br
þ ebsi ð1 g 1 Þ þ b iþ1 eðbrÞsiþ1 g; ifb–r
s
ðFðð1 g 1 ÞFðsðt þi1 Þ; si þ ei Þ; u ei Þ Fðsðt þi Þ; uÞÞeru du ¼ ðA:2Þ
dei ei XðNÞsiþ1 f1 e bsi
s
ð1 g 1 Þg 12 2iþ1 bXðNÞ; if b ¼ r
d
fMðg 2 Fðsðtþi1 Þ; si þ ei Þ þ g 3 Þerei g ¼ m1 g 2 fXðNÞebsi þ bðsðt þi1 Þ þ si XðNÞÞebsi g rfm1 ðg 2 ðsðt þi1 Þ þ si XðNÞÞebsi
dei
þ g 3 Þ þ m2 g ðA:3Þ
d
fMðg 2 Fðð1 g 1 Þ Fðsðt þi1 Þ; si þ ei Þ; siþ1 ei Þ þ g 3 Þersiþ1 g ¼ m1 g 2 XðNÞeðbrÞsiþ1 f1 ð1 g 1 Þebsi þ bsiþ1 g ðA:4Þ
dei
152 J. Lee, S. Madanat / Transportation Research Part B 68 (2014) 141–153
Appendix B.
The initial roughness is known as snew ð¼ sð0Þ ¼ sðtþ0 ÞÞ, so (14) becomes an equation which has two unknown variables, s1
and s2, because (14) for the ith resurfacing is the function of sðt þ
i1 Þ, si and si+1. Iteratively, we have a general equation which
has two unknown variables as si and si+1, because sðtþ þ
i1 Þ is function of si1 and sðt i2 Þ for all i P 2. From the first order deriv-
dsiþ1
ative of (14) with respect to si, we can find d si
1 by calculation with the empirical parameters generally used in practice.
This relation shows that si+1 is significantly sensitive to si. The optimal solution of si = si+1 = s⁄ and s(ti) = s(ti+1) = s⁄ for all i is
only for the case of snew = s0 = (1 g1)s⁄ where there is no permanent loss.
We need to show:
Proof of (i):
For a given sðtþ
i1 Þ, we define s0i and s1i as
s0i si jFðsðtþi1 Þ; si Þ ð1 g 1 Þ ¼ sðtþi1 Þ; si P 0;
dJ
s1i si j ¼ 0; si ¼ siþ1 P 0;
dei
ds1
From (14), dsðtþi Þ < 0. In the case of the first resurfacing, i ¼ 1; sðtþ
i1 Þ ¼ snew < ð1 g 1 Þs . For a resurfacing given
i1
sðtþ
i1 Þ < ð1 g 1 Þs
; s 1
i > s0
i is found. Assume that there exists a resurfacing i, sðt þ
i1 Þ P ð1 g þ
1 Þs and sðt i1 Þ < ð1 g 1 Þs .
Because si1 < s < si1 , it is clear that si < si1 < s < si1 < si . Therefore, the roughness threshold increases rapidly and
1 1 1
dJ
dei
¼ 0 is not solvable, at least with generally used parameters. It cannot be optimal, and the assumption turns to be false.
Proof of (ii) and (iii):
ds1i
Assume that si 6 siþ1 . Therefore, sðtþ þ þ
i Þ > sðt i1 Þ, when sðt i1 Þ < ð1 g 1 Þs . From dsðt þ Þ < 0, we can find si > siþ1 . However,
1 1
i1
s1iþ1 < s1i 6 siþ1 and ddssiþ2
iþ1
1; so s1iþ1 siþ2 . Therefore, the roughness threshold increases rapidly and ddJei ¼ 0 is not solvable, at
least with generally used parameters. The assumption of si 6 siþ1 is not true, so si > siþ1 and sðti Þ < sðt iþ1 Þ; 8i.
Proof of (iv):
From (i), (ii) and (iii), si > s⁄, "i.
References
AASHTO, 1986. Guide for Design of Pavement Structures. American Association of State Highway and Transportation Officials, Washington, DC.
Carnahan, J., 1988. Analytical framework for optimizing pavement maintenance. Journal of Transportation Engineering 114 (3), 307–322.
Durango-Cohen, P., Sarutipand, P., 2007. Capturing interdependencies and heterogeneity in the management of multifacility transportation infrastructure
systems. Journal of Infrastructure Systems 13 (2), 115–123.
Fernandez, J.E., Friesz, T.L., 1981. Influence of demand-quality interrelationships on optimal policies for stage construction of transportation facilities.
Transportation Science 15 (1), 16–31.
Friesz, T.L., Fernandez, J.E., 1979. A model of optimal transport maintenance with demand responsiveness. Transportation Research Part B 13 (4), 317–339.
Golabi, K., Kulkarni, R., Way, G., 1982. A statewide pavement management system. Interfaces 12 (6), 5–21.
Gu, W., Ouyang, Y., Madanat, S., 2012. Joint optimization of pavement maintenance and resurfacing planning. Transportation Research Part B 46 (4), 511–
519.
Labi, S., Sinha, K.C., 2003. The effectiveness of maintenance and its impact on capital expenditures. Joint Transportation Research Program, Paper 208.
Purdue University.
Li, Y., Madanat, S., 2002. A steady-state solution for the optimal pavement resurfacing problem. Transportation Research Part A 36 (6), 525–535.
Madanat, S., Prozzi, J., Han, M., 2002. Effect of performance model accuracy on optimal pavement design. Computer-Aided Civil and Infrastructure
Engineering 17 (1), 22–30.
Markow, M., Balta, W., 1985. Optimal rehabilitation frequencies for highway pavements. In: Transportation Research Record 1035. Washington, DC:
Transportation Research Board, National Research Council, pp. 31–43.
Markow, M.J., Madanat, S.M., Gurenich, D.I. 1993. Optimal rehabilitation times for concrete bridge decks. In: Transportation Research Record 1392.
Washington, DC: Transportation Research Board, National Research Council, pp. 79–89.
McDonald, M., Madanat, S., 2012. Life-cycle cost minimization and sensitivity analysis for mechanistic-empirical pavement design. Journal of
Transportation Engineering 138 (6), 706–713.
Ouyang, Y., Madanat, S., 2004. Optimal scheduling of rehabilitation activities for multiple pavement facilities: exact and approximate solutions.
Transportation Research Part A 38 (5), 347–365.
Ouyang, Y., Madanat, S., 2006. An analytical solution for the finite-horizon pavement resurfacing planning problem. Transportation Research Part B 40 (9),
767–778.
Paterson, W.D.O., 1987. Road Deterioration and Maintenance Effects: Models for Planning and Management. Johns Hopkins University Press, Baltimore, MD.
Ponniah, J.E., Kennepohl, G.J., 1996. Crack sealing in flexible pavements: a life-cycle cost analysis. In: Transportation Research Record 1529. Washington, DC:
Transportation Research Board, National Research Council, pp. 86–94.
Pryke, A., Evdorides, H., Abu Ermaileh, R., 2006. Optimization of pavement design using a genetic algorithm. In: Proc., IEEE Congress on Evolutionary
Computation. IEEE, New York, pp. 1095–1098.
Rajbongshi, P., Das, A., 2008. Optimal asphalt pavement design considering cost and reliability. Journal of Transportation Engineering 134 (6), 255–261.
J. Lee, S. Madanat / Transportation Research Part B 68 (2014) 141–153 153
Rashid, M.M., Tsunokawa, K., 2012. Trend curve optimal control model for optimizing pavement maintenance strategies consisting of various treatments.
Computer-Aided Civil and Infrastructure Engineering 27 (3), 155–169.
Small, K., Winston, C., 1988. Optimal highway durability. Academic Economic Review 78 (3), 560–569.
Tsunokawa, K., Schofer, J.L., 1994. Trend curve optimal control model for highway pavement maintenance: case study and evaluation. Transportation
Research Part A 28 (2), 151–166.