You are on page 1of 19

Transportation Research Part A 40 (2006) 725–743

www.elsevier.com/locate/tra

A multi-year pavement maintenance program using a


stochastic simulation-based genetic algorithm approach
Piya Chootinan a, Anthony Chen a,*
, Matthew R. Horrocks b, Doyt Bolling c

a
Department of Civil and Environmental Engineering, Utah State University, 4110 Old Main Hill, Logan, UT 84322-4110, USA
b
Horrocks Engineers, American Fork, UT 84003, USA
c
Utah Local Technology Assistance Program, Utah State University, Logan, UT 84322-4110, USA

Received 19 April 2005; accepted 15 December 2005

Abstract

The objective of this paper is to introduce a multi-year pavement maintenance programming methodology that can
explicitly account for uncertainty in pavement deterioration. This is accomplished with the development of a simula-
tion-based genetic algorithm (GA) approach that is capable of planning the maintenance activities over a multi-year plan-
ning period. A stochastic simulation is used to simulate the uncertainty of future pavement conditions based on the
calibrated deterioration model while GA is used to handle the combinatorial nature of the network-level pavement main-
tenance programming. The effects of the uncertainty of pavement deterioration on the maintenance program are investi-
gated using a case study. The results show that programming the maintenance activities using only the expected pavement
conditions is likely to underestimate the required maintenance budget and overestimate the performance of pavement
network.
Ó 2005 Elsevier Ltd. All rights reserved.

Keywords: Pavement maintenance programming; Simulation-optimization; Genetic algorithms

1. Introduction

One of the major requirements of pavement management system (PMS) is the ability to develop a multi-
year pavement maintenance program for the entire road network under the jurisdiction of the highway
agency. The development of a multi-year pavement maintenance program is highly dependent on the ability
to estimate future pavement conditions (Zimmerman, 1995). Unfortunately, the accuracy of the predicted
pavement conditions (as well as other types of infrastructure) could be influenced by the choice of prediction
model (Durango and Madanat, 2002) as well as the accuracy of inputs of the deterioration model (e.g., future
traffic load, weather condition, etc.). As a result, the predicted pavement conditions could be subject to
substantial uncertainty. Because the uncertainty of the predicted pavement conditions contributes to the

*
Corresponding author. Tel.: +1 435 7977109; fax: +1 435 7971185.
E-mail address: achen@cc.usu.edu (A. Chen).

0965-8564/$ - see front matter Ó 2005 Elsevier Ltd. All rights reserved.
doi:10.1016/j.tra.2005.12.003
726 P. Chootinan et al. / Transportation Research Part A 40 (2006) 725–743

reliability of the pavement maintenance plan, these uncertainties (e.g., inputs and prediction model) should be
carefully considered when making pavement maintenance decisions (Ben-Akiva et al., 1993). Durango and
Madanat (2002) suggested using a family of models to represent the physical deterioration of infrastructure
(e.g., pavement). Each model contributes to the predicted value of pavement condition differently according
to the belief (weight or probability) of decision makers that the model would be a good representation of the
deterioration process at each stage of pavement condition. This belief can be updated when the additional
information regarding the actual pavement deterioration becomes available.
Although a large number of sophisticated deterioration models have been developed during the past several
decades, the Markov transition probability matrix (TPM) approach appears to be the only model explicitly
developed to address the uncertainty issue of the predicted pavement condition. In the Markov process, a pave-
ment deteriorates by moving from one state to another state according to the state transition probability spec-
ified for each category of pavement condition and type of maintenance undertaken. As opposed to the
deterministic deterioration models, the Markov TPM approach forecasts the distribution of future pavement
conditions (i.e., the proportion of pavements in each group of pavement condition) rather than the future con-
dition of an individual pavement section (e.g., a single numerical value). Therefore, it is usually applied for the
network-level management process. Additional information such as subjective opinions (or rules) of the high-
way agency is generally required to implement the TPM approach at the project level (Robert et al., 2002). In
addition, the requirement of comprehensive historical performance data for developing the reliable TPMs
appears to limit its application in programming maintenance activities (Wang et al., 1994; Li et al., 1997). To
reduce the amount of data required, Li et al. (1996) suggested using a Monte Carlo simulation (MCS) combined
with a deterministic deterioration model to develop TPMs for a variety of pavement categories and/or individual
pavement sections. In fact, the idea of using MCS allows the stochastic behavior (uncertainty) of the pavement
deterioration to be incorporated explicitly in the modeling process of PMS without the employment of TPMs.
Another challenge in developing a long-term maintenance plan at a network level is the ability to identify
proper maintenance activities for individual pavement sections at the project level that are consistent with
the network-level recommendation (Chan et al., 1994; Mbwana and Turnquist, 1996). This type of problem
is usually formulated using integer variables to represent maintenance activities selected for individual pave-
ment sections (i.e., integer program). It is also known as the combinatorial problem, which is very difficult to
solve due to a vast solution space (i.e., the number of possible solutions increases exponentially as the problem
size increases). Chan et al. (1994) introduced the genetic algorithm (GA), which is one of the evolutionary com-
puting techniques that have shown considerable success in solving a number of complex large-scale problems in
many disciplines, to handle this difficulty in the pavement maintenance programming. In this paper, we present
a simulation-optimization framework for programming the maintenance activities over a multi-year planning
period. The combination of stochastic simulation and GA allows the development of a project-based network-
level maintenance plan that can explicitly take into account the uncertainty of future pavement conditions in the
decision-making process. In other words, the risk that the maintenance plan would fail to fulfill the required
pavement standard is explicitly considered when selecting maintenance activities. The stochastic simulation
is used to simulate the uncertainty of future pavement conditions based on the calibrated deterioration model
(for the evaluation of objective function and constraints of the mathematical program) while GA searches for a
good solution (i.e., pavement maintenance plan) for a given funding level and a required pavement standard.
The organization of this paper is as follows. Section 2 presents the development of mathematical formula-
tions of two pavement maintenance programs, which aim to minimize the maintenance cost and to maximize
the pavement performance, respectively. The stochastic formulations in which the uncertainty of future pave-
ment conditions is explicitly modeled are presented along with their deterministic counterparts. In Section 3,
the solution procedure, which is the stochastic simulation-based GA approach, is described. Section 4 presents
the numerical results of a case study conducted using the proposed framework. Finally, the findings and con-
clusions of the study are reported and discussed in Section 5.

2. Methodology

The problem of pavement maintenance programming is one of maintaining the serviceability of the entire
pavement network with the available funds and resources. To fulfill these requirements, highway agencies can
P. Chootinan et al. / Transportation Research Part A 40 (2006) 725–743 727

utilize a variety of maintenance goals, such as maximizing network performance, maximizing the cost-effec-
tiveness of maintenance activities, minimizing road user cost, minimizing the present worth of the total main-
tenance cost, etc., with a certain set of constraints (e.g., budget, pavement standard, manpower, equipment,
etc.). The development of a multi-year maintenance plan for the entire network requires a systematic approach
to optimally select the appropriate maintenance activities. In addition, the prediction of future pavement con-
ditions and the impacts of maintenance activities on the deterioration of pavement are also very crucial. For
convenience, the notation used in this study is provided in Table 1.

2.1. Pavement performance prediction and associated uncertainties

Pavement performance prediction models are required to perform a long-term maintenance programming.
These prediction models assist highway agencies to determine what has happened to pavements over the past
years and to predict what may happen in the future (Peterson, 1987) so that the maintenance works could be
performed cost-effectively. To simplify the pavement condition analysis and the communication to the higher-
level (administrative) management, the composite performance indices representing the overall pavement con-
dition are often used. In general, the composite performance indices are a function of pavement type, traffic
loading, age of pavement, severity level of individual distress, etc. (Haas and Hudson, 1982).
In this study, present serviceability rating (PSR), developed by Lee et al. (1993) for the highway pavement
monitoring system (HPMS), is used. However, it should be noted that the framework developed in this study
could be applied using any deterioration models with a similar structure. PSR is a surface-condition rating
scheme based on a numeric scale between 0 and 5 (where 4.5 is always used in practice), 0 indicating extremely
poor condition and 5 indicating a perfect pavement.
The condition of pavement section s at any time period t, as given in Eq. (1), is a function of the initial
pavement condition after construction (P0), pavement structural number (STRst), age of pavement (Yst),
and the cumulative 18-kip axle loadings on pavement section s at year t(Dst). The calibrated parameters
(i.e., a, b, c, and d) of this model for different types of pavement can be found in the study of Lee et al.
(1993). The factor AF adjusts the deterioration rate of a pavement in a particular climate zone and functional
group.
PSRst ¼ P 0  AF  ða  STRbst  Y cst  Ddst Þ. ð1Þ

Table 1
Mathematical formulation nomenclature
Variable Description
xst Decision variable, type of treatment applied to pavement section s at year t
T Number of years in the planning horizon
S Total number of pavement segments
c(xst) Unit cost function related to type of treatment
P(xst) Condition of pavement section s after applied treatment at year t
r Discount rate
Bt Budget constraint at year t
Ls Length of pavement segment s
Ws Width of pavement segment s
Pt Required network performance standard at year t
P0 Maximum possible pavement condition rating
qst(xst, Yst) Actual improvement in pavement condition as a function of treatment type
AF Adjustment factor regarding climate zone
STRst Structural number of segment s at year t
Yst Age of pavement at year t since the initial construction or last major rehabilitation
Dst Cumulative 18-kip equivalent single-axel loads, EASL (millions) carried by the pavement segment s in year t
a, b, c, d Pavement performance coefficients for pavement performance prediction model
Pr(Æ) Probability of event
xst Predicted (expected) traffic load on segment s at year t
est Prediction error of traffic load on segment s at year t
728 P. Chootinan et al. / Transportation Research Part A 40 (2006) 725–743

As different maintenance activities are implemented, the performance of pavement is affected differently.
Routine maintenance activities (e.g., chip seal) minimize the effects of deterioration and provide minor
improvements in pavement condition (i.e., PSR), some types of treatment, such as minor rehabilitations
(e.g., thin overlay) may add additional strength to the pavement (i.e., increase the thickness of pavement)
and provide a new riding service. The major rehabilitations or reconstruction will reset the pavement to the ini-
tial stage (i.e., zero cumulative load, age of zero). Accordingly, the characteristics of pavement not only change
over time, but they are also affected by the maintenance activities applied. Mathematically speaking, pavement
characteristics (e.g., STRst, Yst, Dst, etc.) become a function of the maintenance activity undertaken (xst).
In this study, the uncertainty of predicted pavement conditions is assumed to be associated with the pre-
diction error (est) of future traffic loads. Although other sources of uncertainty are present, when predicting
future pavement conditions, it is assumed here that the uncertainties of future traffic loads are several magni-
tudes larger than other sources of uncertainty. The accumulative traffic load on pavement section s at year t is
modeled as a random variable and can be expressed as:
~ st ðeÞ ¼ D
D ~ st1 ðeÞ þ xst þ est ; ð2Þ
where xst is the predicted (expected) traffic load on pavement section s at year t, and est is the prediction error,
which could be drawn from any probability distribution (e.g., a normal distribution). It should be noted that
the effect of prediction error could be accumulated over time. Since the predicted annual traffic load is a ran-
dom variable, the predicted pavement performance (Eq. (3)), which takes the effect of maintenance activity
into account, becomes a random variable as well. Here, it is assumed that the pavement structural number
(STRst) remains unchanged regardless of the maintenance activity.
  Xt
~ d ðxst ; eÞ þ
PSRst ðxst ; eÞ ¼ P 0  AF  a  STRbst  Y cst ðxst Þ  D qsi ðxsi ; Y si Þ; ð3Þ
st
i¼1

where qsi is the actual improvement in the condition of pavement section s at year i, as a result of the treatment
option applied (xsi) (Eq. (4)). Since the maintenance activity cannot improve the condition of pavement higher
than P0 (the upper limit), the actual increment is the minimum between the effect of the treatment option (see
Table 2, q
~si ) and the maximal possible increment (i.e., the difference between P0 and the pavement condition
before applying treatment).
qsi ðxsi ; Y si Þ ¼ MinfP 0  PSRsi ð; eÞ; q
~si ðxsi ; Y si Þg. ð4Þ
Because a pavement performance is a function of traffic load, environment, age, and previous maintenance
activities, the improvement in the condition of pavement section s at year i ð~
qsi Þ regarding the same mainte-
nance activity is not consistent between different ages of pavement (Al-Suleiman et al., 1991). For instance, a
routine maintenance could be very effective when applied within the first few years of a new pavement seg-
ment, but its effectiveness reduces when pavements approach their design life. A reconstruction always returns
the pavement to the initial condition regardless the existing pavement condition.

2.2. Deterministic formulations

For comparison purposes, the deterministic models will be presented first and followed by their stochastic
counterparts. In this paper, two maintenance goals commonly used in the network-level pavement mainte-

Table 2
Available treatment type and effect on pavement condition (Utah LTAP, 2004)
Code Treatment type Unit cost ($/m2) Age of pavement applicable (< year)
20 19 16 13 10 7 4 1
0 – 0.00 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
1 Routine maintenance 0.20 0.000 0.000 0.000 0.225 0.225 0.450 0.450 0.450
2 Surface treatment 0.74 0.000 0.450 0.675 1.125 1.575 1.575 1.575 1.575
3 Overlay 4.67 0.000 0.900 1.575 1.800 1.800 1.800 1.800 1.800
4 Major rehabilitation 7.74 4.500 4.500 4.500 4.500 4.500 4.500 4.500 4.500
P. Chootinan et al. / Transportation Research Part A 40 (2006) 725–743 729

nance optimization (Irrgang and Maze, 1993; Zimmerman, 1995), which are pavement-performance maximi-
zation and maintenance-cost minimization, are considered. The development of a multi-year pavement main-
tenance plan is constrained by the available maintenance budget and the minimum acceptable pavement
condition. The formulations presented below use integer variables to represent the maintenance options
(xst) selected for pavement section s at year t.

2.2.1. Pavement performance maximization


X
T
Maximize: Z1 ¼ H t ðxÞ; ð5aÞ
t¼1
X
S
subject to: cðxst ÞLs W s 6 Bt ; 8t; ð5bÞ
s¼1

P ðxst Þ P P t ; 8s; t; ð5cÞ


xst 2 f0; 1; . . . ; mg; 8s; t; ð5dÞ
  X
t
where P ðxst Þ ¼ P 0  AF  a  STRbst  Y cst ðxst Þ  Ddst ðxst Þ þ qsi ðxsi ; Y si Þ; 8s; t; ð5eÞ
i¼1

xst ; if xst ¼ m
Dst ðxst Þ ¼ ; 8s; t; ð5fÞ
Dst1 þ xst ; otherwise

1; if xst ¼ m
Y st ðxst Þ ¼ ; 8s; t; ð5gÞ
Y st1 þ 1; otherwise
H t ðxÞ ¼ MinfP ðxst Þg; 8t. ð5hÞ
8s

Under this formulation, the objective function (Eq. (5a)) aims to maximize the performance of the entire
pavement network, which is defined by the summation of the minimum PSR among all pavement sections
in each year during the planning period given in Eq. (5h). Eq. (5b) ensures that the annual maintenance expen-
diture does not exceed the available budget. The available treatment options and their associated costs are
presented in Table 2. Eq. (5c) maintains that the performance of all pavement sections after treatment is above
a predefined condition. Eq. (5d) constrains the decision variables (treatment options) to be integer. Eq. (5e)
models the performance of pavement considering the actual deterioration and the improvement provided
by the treatment options. Eqs. (5f) and (5g) model the changes of pavement properties (i.e., age and accumu-
lative load) over time. Pavements are triggered to the initial condition (i.e., zero cumulative load and age of
zero) whenever the reconstruction (xst = m) is applied.

2.2.2. Maintenance cost minimization


In addition to the maximization of network-level performance, decision makers often require a mainte-
nance plan that minimizes the present worth of the total maintenance cost over the planning horizon, yet pro-
vides the desired network-level performance. The objective function in Eq. (6) is used in this case. In Eq. (6),
future maintenance costs are converted to the present value by using the conversion factor of (1 + r)t+1. Sim-
ilarly, the maintenance cost minimization model considers the same set of constraints as the performance
maximization model (Eqs. (5b)–(5g)).
X
T X
S
Minimize: Z2 ¼ ð1 þ rÞtþ1  cðxst ÞLs W s ;
t¼1 s¼1 ð6Þ
subject to: Eqs. (5b)–(5g).

2.2.3. Performance maximization and cost minimization


It is natural to combine the previous two optimization problems to the bi-objective model in which both
maintenance goals are simultaneously optimized. That is, the maintenance plan that costs less and provides
730 P. Chootinan et al. / Transportation Research Part A 40 (2006) 725–743

higher pavement performance is more preferable. In this study, both objectives are given a weight (e.g., pri-
ority) and combined into a single objective value as shown in Eq. (7).
Maximize : w1  Z~ 1 þ w2  Z~ 2 ;
ð7Þ
subject to : Eqs. (5b)–(5h),
where w1 and w2 are weights (priorities) given to the performance of pavement and the maintenance cost sav-
ing. Z~ 1 and Z~ 2 are respectively, the standardized versions of Eqs. (5a) and (6). The standardization (e.g., nor-
malized between 0 and 1) is required because different objectives are usually defined in incomparable scales
(e.g., money and time). In addition, the direct usage of the original objective values may cause the domination
of one objective over another. The original objective values can be normalized as follows.
Z 1  Z min
Z~ 1 ¼ max 1 min ; ð8Þ
Z1  Z1
Z max  Z2
Z~ 2 ¼ max
2
; ð9Þ
Z 2  Z min
2

where Z max
i and Z min
i are respectively, the maximal and minimal possible values of objective i.

2.3. Stochastic formulations

In the preceding section, pavements are assumed to deteriorate deterministically (i.e., no uncertainty
involved in the predicted pavement condition). By assuming that the prediction error (est) of future traffic
loads follows a known probability distribution, it can be incorporated into the existing pavement performance
models (Eqs. (5e)–(5h)) to account for the uncertainty of pavement deterioration as follows.

2.3.1. Expected performance maximization


" #
XT
Maximize: Z 3 ¼ E H t ðx; eÞ ; ð10aÞ
t¼1

X
S
subject to: cðxst ÞLs W s 6 Bt ; 8t; ð10bÞ
s¼1

PrðP ðxst ; eÞ P P t Þ P ast ; 8s; t; ð10cÞ


xst 2 f0; 1; . . . ; mg; 8s; t; ð10dÞ
  Xt
where ~ d ðxst ; eÞ þ
P ðxst ; eÞ ¼ P 0  AF  a  STRbst  Y cst ðxst Þ  D qsi ðxsi ; Y si Þ; 8s; t; ð10eÞ
st
i¼1

xst þ est ; if xst ¼ m
~ st ðxst ; eÞ ¼
D ; 8s; t; ð10fÞ
~
Dst1 þ xst þ est ; otherwise

1; if xst ¼ m
Y st ðxst Þ ¼ ; 8s; t; ð10gÞ
Y st1 þ 1; otherwise
H t ðx; eÞ ¼ Min fP ðxst ; eÞg; 8t. ð10hÞ
8s

As a result of the uncertainty of future traffic loads, the future pavement conditions become random vari-
ables. In addition, all relations associated with pavement performance (e.g., Eqs. (5a), (5c), (5e), etc.) become
probabilistic functions. Since the value of these functions is not a single numerical value but a probability dis-
tribution, the expectation and/or the probability of a certain event are usually used to formulate the stochastic
optimization problem (Liu, 1999). Now, the first problem considered in the previous section becomes one of
maximizing the expected value of pavement performance (Eq. (10a)) for a given maintenance budget. While
accounting for several possibilities of future pavement conditions, the conditions of individual pavements
maintained in each year must satisfy the required standard with a certain confidence level (ast)—a chance
P. Chootinan et al. / Transportation Research Part A 40 (2006) 725–743 731

constraint. In other words, the probability measure of the event {P(xst, e) P Pt} is at least ast, which can be
referred to as the confidence that the requirement (e.g., constraint) will not be violated. The remaining deter-
ministic formulations can be modified to account for the uncertainty of future pavement conditions in the
same manner.

2.3.2. Maintenance cost minimization


X
T
tþ1
X
S
Minimize: Z2 ¼ ð1 þ rÞ  cðxst ÞLs W s ;
t¼1 s¼1 ð11Þ
subject to: Eqs. (10b)–(10g).

2.3.3. Expected performance maximization and cost minimization

Maximize: w1  Z~ 3 þ w2  Z~ 2 ;
ð12Þ
subject to: Eqs. (10b)–(10h),
where Z~ 3 is determined similarly to Z~ 1 shown in Eq. (8).
As shown above, these formulations involve non-linear functions, non-differentiable functions, step func-
tions, and integer variables. Although the step functions can be generalized to the linear forms, the transfor-
mation will require additional variables, which will increase the problem size. In addition, the pavement
performance model (deterioration model) is certainly non-linear. Moreover, when the stochastic elements
are incorporated into the formulation, these characteristics all together are incompatible with the traditional
optimization techniques. This motivates the usage of GA for solving pavement maintenance programs pro-
posed in this study.

3. Simulation-optimization framework

3.1. Stochastic simulation

A simulation-based GA procedure that combines the stochastic simulation and the GA to solve the sto-
chastic programs is developed in this study. Traditionally, simulation is the process of replicating reality based
on a set of assumptions and conceived models of reality (Ang and Tang, 1984). To handle the uncertainty of
future pavement conditions, the stochastic simulation is used to simulate the uncertainty of future traffic loads
used in the prediction model based on a probability distribution with the predefined mean and standard devi-
ation. Latin hypercube sampling (LHS) technique, one of the stratified sampling techniques that has shown to
outperform the simple Monte Carlo sampling technique (McKay, 1988), is employed in this study. LHS par-
titions the distribution of future traffic loads into several equal intervals according to the number of samples
required. Only one random variate is sampled from each interval. This sampling technique significantly
reduces the number of samples yet delivers a reasonable level of accuracy. LHS is incorporated into the
GA to evaluate the values of objective function and constraints corresponding to the maintenance decision
under a stochastic environment (i.e., pavement deterioration). That is, the objective function and constraints
of the stochastic formulations presented earlier are evaluated several times (i.e., the number of samples) using
the simulation.

3.2. Genetic algorithm (GA)

GA is an evolutionary computing technique that, in principal, mimics the mechanism of natural selection
process. According to Goldberg (1989), GA differs from the classical, calculus-based optimization techniques
in the following ways: (i) instead of using a point-to-point search method, as in the traditional optimization
techniques, GA simultaneously searches from a population of points, known as chromosomes, to explore the
solution space; (ii) GA uses probabilistic transition rules (for its operators) as a guide to search the solution
732 P. Chootinan et al. / Transportation Research Part A 40 (2006) 725–743

1st Section 2nd Section

X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15 X16 X17 X18 X19 X20 …

10th yea r

10th ye a r
2nd year

2 nd ye ar
3rd year

3rd year
8th yea r
9th ye ar

8th y ear
9th yea r
1st y ear

1 st y ear
. . . . . . . .

Xi = type of treatment ∈ {0,1,2,3,4}

Fig. 1. Representation of decision variables.

space with likely improvement; (iii) GA can work with continuous and discrete parameters, differentiable and
non-differentiable functions, uni-modal and multi-modal functions, as well as convex and non-convex feasible
regions.

3.2.1. Representation of decision variables


One important aspect of applying GA to any problem is the representation of the decision variables in the
genetic fashion or as a chromosome. For the pavement maintenance problem studied here, the chromosome is
coded as a series of T-year maintenance activities for all pavement segments S, as shown in Fig. 1. As depicted
in Fig. 1, each chromosome consists of a series of maintenance actions represented by numerical values:
0, 1, 2, . . . , m (see Table 2 for treatment options). For each pavement section s, there are T genes representing
the maintenance treatments for T years. For example, the first T genes represent the maintenance treatments
for the first pavement section over T years; the second set of T genes is for the second pavement section, and so
on.

3.2.2. Constraint handling


Because the traditional GA operators are blind to constraints of an optimization problem, the special-pur-
pose constraint-handling methods are usually required to ensure the feasibility of solutions. As with any opti-
mization procedure, improper constraint handling will result in a considerable amount of effort wasted in
evaluating infeasible solutions. To ensure that GA performs the search effectively, the budget constraint
and the pavement performance constraint are handled by the special constraint-handling procedure. This pro-
cedure utilizes a combination of the penalty and repair methods. The penalty method converts a constrained
problem into an unconstrained problem by including a penalty value in the objective function (Goldberg,
1989). The repair method attempts to repair the infeasible solutions by using a special solution mapping to
ensure the feasibility of the solutions (Liu, 1999).
For the pavement maintenance problem studied here, the repair method is applied first by attempting to
obtain a feasible or near-feasible solution. Under the repair method, both budget and pavement performance
constraints are examined simultaneously to determine the positions of genes in the chromosome that cause
the solution to be infeasible. The positions of the ‘‘bad’’ genes indicate when and where (i.e., which
pavement segment and at what year) the budget and/or pavement performance constraints are violated. Once
these bad genes are identified, the repair method attempts to fix these genes with the most cost-effective treat-
ment. This process is repeated several times until the solution is feasible or the repair is impossible. If the solu-
tion is not feasible, the penalty method penalizes the fitness according to the degree of the constraint
violations.

3.3. Stochastic simulation-based GA procedure

The simulation-based GA procedure for solving the pavement maintenance program is displayed in Fig. 2
and is summarized as follows:
P. Chootinan et al. / Transportation Research Part A 40 (2006) 725–743 733

Pavement Inventory
1
Initial pavement condition
Initial traffic loads
Growth Rate

Generate Initial Maintenance


Solutions
Update Maintenance Strategies
by GA Operators
Reproduction
Crossover
Mutation 3
2 N=1

S =1 N = N +1

Generate Random Traffic S = S+1


Loading

Evaluate Maintenance Strategies


Objective function
Constraints No

Collect Statistical Inferences S > S max

Yes

No Stochastic
Constraints met?

Yes

Constraint No
N > N_ pop
Handling

Yes

Solution and Probabilistic 5 Yes Stopping No


Summary 4
criteria met?

Fig. 2. Flow diagram for a stochastic simulation-based GA for PMS.

1. Initialize a set of maintenance solutions (Npop solutions) according to the pavement inventory data and
pavement performance model.
2. Set N = 1 (the counter for number of solutions) and S = 1 (the counter for number of sample points). Each
maintenance solution will be evaluated based on the objective function, and checked for the feasibility Smax
times by the stochastic simulation. If the maintenance solution violates the stochastic constraints, con-
straint-handling procedure is applied.
3. Rank the maintenance solutions based on their fitness to the problem. Update solutions (maintenance
plans) by GA operators, reproduction, crossover and mutation in order to obtain a new set of maintenance
solutions.
734 P. Chootinan et al. / Transportation Research Part A 40 (2006) 725–743

4. Repeat steps 2 and 3 until stopping criterion is met (e.g., the maximum number of generations).
5. Report the solution and its probabilistic summary.

4. Case study

4.1. General information

To illustrate the feasibility of the proposed method, a case study was conducted. The study network consists
of two routes of flexible pavement. Both routes consist of a 31.450 lane-miles (50.621 lane-kilometers) of flex-
ible pavement with three control sections, one control section on the first route and the remaining two control
sections on the second route. This study aggregates the pavement inventory data into approximately 3281-feet
(1000-m) intervals across both roadways. As a result, there are 53 pavement segments for the case study (see
pavement inventory data in Chootinan, 2001). In this case study, age of pavement segments ranges from 3
years to 9 years with an average pavement age of 4.75 years. Initial PSR values range from 2.49 to 3.48 with
a network average PSR of 3.16.
A 10-year maintenance plan (year 2000–2009) is programmed in order to maintain the pavement condition
above a PSR value of 2.5 with an assumed annual budget of US$ 61,410. Present worth calculations are con-
ducted using an arbitrary discount rate of 12%. The available treatment options, as shown in Table 2, are: 0—
no treatment, 1—routine maintenance, 2—surface treatment, 3—minor rehabilitation, 4—major rehabilita-
tion. The prediction of future traffic loadings utilizes an annual growth rate of 5%. Prediction error of the
future traffic loadings is simulated using a normal distribution and is assumed to increase proportionally with
time. In addition, a confidence level of 90% (ast = 0.90) is set to ensure that stochastic constraints are fulfilled
to at least 90%.

4.2. Selection of simulation-based GA parameters

Although a large population pool in GA has a higher probability of obtaining a better solution, it might
be impractical in a simulation-based framework since the number of function evaluations per generation is
not simply the number of chromosomes in the population pool. The number of function evaluations is
essentially the number of chromosomes times the sample size used in the stochastic simulation. Conse-
quently, it is important to size the population pool such that it is reasonably small yet allows for reason-
ably good solutions. The following settings of GA are used in this study: the population size is 32.
Crossover probability is 0.50 (uniform crossover) while mutation probability is 0.01 (random mutation).
The readers may refer to Goldberg (1989) for the detail description of GA operators. The initial popula-
tion is randomly generated. The roulette wheel (proportional) selection and the half-replacement strategy
are used for reproduction. Chromosomes in each generation are ranked based on their fitness (penalized
objective value) and divided into two parts. Only the chromosomes in the top half (i.e., better solutions)
are eligible for reproduction. Chromosomes in the bottom half will be replaced by the offspring generated
by crossover and mutation. The genetic search will be performed until the maximum number of
generations (50,000 generations) is reached, In the case that the infeasible solution cannot success-
fully be repaired; a very high penalty will be applied. The sample size of the stochastic simulation is
100.

4.3. Numerical results

Multi-year pavement maintenance plans were developed using both the deterministic and stochastic formu-
lations. All three objectives, which are the performance maximization, maintenance cost minimization, as well
as the bi-objective model, are considered in both formulations. Comparison of the maintenance plans devel-
oped by these two formulations is used to investigate the effects of uncertainty on pavement deterioration
under different maintenance objectives.
P. Chootinan et al. / Transportation Research Part A 40 (2006) 725–743 735

39.0

Deterministic Approach (37.23)


37.5

36.0
Stochastic Approach (36.53)
Cumulative PSR

34.5

33.0

31.5

30.0

28.5

27.0
0 5,000 10,000 15,000 20,000 25,000 30,000 35,000 40,000 45,000 50,000

Number of Generations

Fig. 3. Convergence curve of pavement performance maximization.

4.3.1. Pavement performance maximization


Under the pavement performance maximization objective, the developed maintenance programs attempt to
fully utilize the available maintenance budget such that the serviceability level of the lowest (minimum) con-
dition among all pavement sections will be as high as possible. Fig. 3 displays the convergence curves mea-
sured in terms of the summation of minimum pavement segment PSR values over the 10-year planning
period for both the deterministic and stochastic formulations. As can be seen from the figure, the deterministic
and stochastic approaches converge to the solutions (i.e., maintenance plans) with the objective values of 37.23
and 36.53, respectively.
In order to evaluate the performance of the maintenance plans developed under the deterministic and sto-
chastic formulations, the pavement performance curve and pavement deterioration curve of both maintenance
plans were displayed in Fig. 4. The pavement performance curve displays the network average PSR values
before treatment and after treatment for each year of the 10-year maintenance period. To illustrate the sto-
chastic nature of pavement deterioration, a 90% confidence level is also included for both pavement-perfor-
mance and pavement-deterioration curves. The evaluation of performance curves showed that both
maintenance plans (deterministic and stochastic approaches) steadily improve the average network-perfor-
mance throughout the duration of the 10-year planning period. In nearly every year of the planning period,
the stochastic formulation generates a lower average network-performance than that of the deterministic for-
mulation. This indicates that the uncertainty increases the predicted pavement deterioration and results in a
lower-than-expected pavement performance. This is confirmed by evaluating the allocation of maintenance
budget in Fig. 5 and the maintenance activities summarized in Tables 3 and 4.
From Fig. 5, it can be seen that the stochastic formulation utilizes a slightly larger budget than the deter-
ministic formulation for nearly every year of the planning period. When accounting for the uncertainty of
pavement deterioration, pavement sections are expected to deteriorate at a higher rate. As a result, a higher
level of treatments (e.g., treatment Type 2) is generally required to maintain the required level of pavement
standard. In Tables 3 and 4, it can be seen that a higher percentage of pavement area is being treated by treat-
ment Type 2 in the stochastic formulation compared to the deterministic formulation. The increase in pave-
ment area being treated by treatment Type 2 results in a maintenance cost difference of US$ 6807 and a
network average PSR difference of approximately 0.01. Under this situation, the stochastic formulation sug-
gests that a 1.1% increase in the maintenance cost would be required to sustain a comparable pavement
736 P. Chootinan et al. / Transportation Research Part A 40 (2006) 725–743

Deterioration Lower Deterioration Mean


Deterioration Upper Performance Lower
Performance Mean Performance Upper
Deterministic Performance

4.500

4.000
Network Average PSR

3.500
Stochastic Deterministic
3.000 Performance Curve Performance Curve

2.500

2.000

1.500
1999 2001 2003 2005 2007 2009
Year

Fig. 4. Performance of pavement under performance maximization.

Budget
0.06 US$ 61,410
Annual Maintenance Cost ($M)

0.05

0.04

Deterministic
0.03 Stochastic

0.02

0.01

0.00
2000 2001 2002 2003 2004 2005 2006 2007 2008 2009
Year

Fig. 5. Maintenance cost allocation under pavement performance maximization.

condition to the deterministic formulation. In general, the following observations can be made about the per-
formance maximization results:

1. The maintenance plans recommended by both deterministic and stochastic formulations are composed pri-
marily of treatment types ‘‘1’’ and ‘‘2’’. However, the results in Tables 3 and 4 indicate that, by accounting
for the uncertainty in the stochastic formulation, a larger number of pavement sections are required to
undergo a higher level of treatment (treatment Type 2).
2. Higher maintenance cost required for the stochastic formulation results in an increase of US$ 216 per lane-
mile of roadway compared to the deterministic formulation. This indicates that the deterioration of pave-
ment is underestimated in the deterministic case, which results in an underestimation of the maintenance
cost.
P. Chootinan et al. / Transportation Research Part A 40 (2006) 725–743 737

Table 3
Summary of maintenance activities under performance maximization (deterministic case)
Treatment option Area of treatments (m2)
2000 2001 2002 2003 2004 2005 2006 2007 2008 2009
0 (Do-nothing) 130,648 157,493 107,994 110,390 141,694 95,567 162,499 128,662 115,591 50,176
1 (Routine) 204,534 166,510 233,313 229,901 191,619 251,746 161,033 205,967 226,696 312,131
2 (Surface treatment) 27,125 38,304 21,000 22,016 28,994 14,994 38,775 27,678 20,020 0
3 (Minor rehabilitation) 0 0 0 0 0 0 0 0 0 0
4 (Major rehabilitation) 0 0 0 0 0 0 0 0 0 0
% of total area treated 63.94 56.53 70.19 69.53 60.89 73.62 55.15 64.49 68.10 86.15
Maintenance cost (US$) 60,182 60,948 61,324 61,402 59,022 60,520 60,219 60,871 59,302 61,337
Network ave. PSR before 3.16 3.33 3.51 3.68 3.84 3.95 4.08 4.14 4.14 4.14
treatment
Network ave. PSR after 3.52 3.70 3.86 4.02 4.13 4.26 4.31 4.32 4.31 4.32
treatment

Table 4
Summary of maintenance activities under performance maximization (stochastic case)
Treatment option Area of treatments (m2)
2000 2001 2002 2003 2004 2005 2006 2007 2008 2009
0 (Do-nothing) 194,347 142,606 107,675 149,676 124,521 108,424 146,648 127,477 121,968 109,541
1 (Routine) 115,871 186,396 234,332 176,329 210,941 232,988 180,659 207,082 214,448 231,577
2 (Surface treatment) 52,089 33,305 20,300 36,302 26,845 20,895 35,000 27,748 25,891 21,189
3 (Minor rehabilitation) 0 0 0 0 0 0 0 0 0 0
4 (Major rehabilitation) 0 0 0 0 0 0 0 0 0 0
% of total area treated 46.36 60.64 70.28 58.69 65.63 70.07 59.52 64.82 66.34 69.77
Maintenance cost (US$) 61,155 61,172 61,008 61,402 61,235 61,183 61,294 61,142 61,221 61,122
Network ave. PSR before 3.16 3.31 3.48 3.66 3.82 3.98 4.11 4.15 4.14 4.13
treatment
Network ave. PSR after 3.50 3.67 3.84 4.00 4.16 4.29 4.32 4.31 4.30 4.29
treatment
Standard deviation of PSR 0.00 0.020 0.024 0.027 0.029 0.028 0.024 0.026 0.029 0.031

4.3.2. Maintenance cost minimization


For the cost minimization objective, the maintenance plans developed under both deterministic and sto-
chastic formulations aim to minimize the present worth of the total maintenance cost spent throughout the
planning period. The deterministic and stochastic formulations recommended the maintenance plans with
the total maintenance costs (present value) of US$ 138,946 and US$ 159,615, respectively (see Fig. 6). As
expected, the stochastic formulation requires a higher maintenance cost compared to that of the deterministic
formulation in order to maintain the same level of pavement standard. To gain more insight on the increase of
maintenance cost, the budget allocation and average network-performance curves displayed respectively in
Figs. 7 and 8 were investigated.
Similar to the results shown in the previous section, a larger amount of maintenance cost is required for the
stochastic formulation. Evaluation of the performance and deterioration curves reveals that the stochastic for-
mulation maintains a higher average network PSR. Although this is counterintuitive, the treatment allocation
scheme recommended by the stochastic formulation in Table 6 creates a higher level of pavement perfor-
mance. Because the pavement conditions in the stochastic formulation deteriorate at a higher rate than the
deterministic formulation, a larger pavement area requires treatment, and, in some cases, a higher level of
treatment. This results in a 13.6% increase in maintenance cost in the stochastic formulation to maintain
the same level of pavement standard. Investigation of the treatment scheme summarized in Tables 5 and 6
indicates that the deterministic formulation treated a larger area of pavement while the stochastic formulation
738 P. Chootinan et al. / Transportation Research Part A 40 (2006) 725–743

0.30

0.28

Total Maintenance Cost ($M)


0.26

0.24

0.22

0.20
Stochastic Approach (US$ 159,615)
0.18

0.16

0.14

0.12 Deterministic Approach (US$ 138,946)


0.10
0 5,000 10,000 15,000 20,000 25,000 30,000 35,000 40,000 45,000 50,000
Number of Generations

Fig. 6. Convergence curve of maintenance cost minimization.

Budget
0.06 US$ 61,410
Annual Maintenance Cost ($M)

0.05

0.04

Deterministic
0.03 Stochastic

0.02

0.01

0.00
2000 2001 2002 2003 2004 2005 2006 2007 2008 2009
Year

Fig. 7. Budget allocation under maintenance cost minimization.

treated less area of pavement with a higher level of treatment. A budget increase of US$ 28,206 from the deter-
ministic approach to the stochastic approach results in a PSR value increase of 0.08.
In general, the following observations can be made about the cost minimization results:

1. The deterministic formulation underestimates the pavement deterioration resulting in the reduction of
maintenance cost throughout the planning period. This underestimation of maintenance cost translates
to a cost difference of US$ 897 per lane-mile of roadway.
2. Maintenance activities recommended by the stochastic formulation require a larger amount of maintenance
cost. This indicates that the deterministic formulation underestimates the pavement deterioration, which
leads to the underestimation of the required maintenance cost.
P. Chootinan et al. / Transportation Research Part A 40 (2006) 725–743 739

Deterioration Lower
Deterioration Mean
Deterioration Upper
Performance Lower
Performance Mean
Performance Upper
Deterministic Performance
3.500
3.300

3.100
Network Average PSR

2.900
2.700

2.500
2.300
2.100
1.900

1.700
1.500
1999 2001 2003 2005 2007 2009
Year

Fig. 8. Performance of pavement under maintenance cost minimization.

Table 5
Summary of maintenance activities under cost minimization (deterministic case)
Treatment option Area of treatments (m2)
2000 2001 2002 2003 2004 2005 2006 2007 2008 2009
0 (Do-nothing) 263,607 284,626 249,396 233,894 246,981 233,830 288,324 304,584 294,435 318,346
1 (Routine) 98,700 60,865 105,911 121,413 108,242 118,777 73,983 57,723 61,852 43,961
2 (Surface treatment) 0 16,816 7000 7000 7084 9700 0 0 6020 0
3 (Minor rehabilitation) 0 0 0 0 0 0 0 0 0 0
4 (Major rehabilitation) 0 0 0 0 0 0 0 0 0 0
% of total area treated 27.24 21.44 31.16 35.44 31.83 35.46 20.42 15.93 18.73 12.13
Maintenance cost (US$) 19,396 24,353 25,971 29,018 26,491 30,489 14,539 11,343 16,591 8639
Network ave. PSR before 3.16 3.10 3.05 3.02 3.03 3.01 3.01 2.92 2.78 2.66
treatment
Network ave. PSR after 3.29 3.24 3.20 3.21 3.18 3.18 3.09 2.95 2.83 2.69
treatment

3. On average, the stochastic formulation is likely to recommend a higher level of treatment in order to main-
tain the minimum pavement condition at the PSR value of 2.5. The uncertainty of pavement deterioration
considered in the stochastic formulation led to 96% more of Type 2 treatment being applied when com-
pared to the deterministic formulation.

4.3.3. Bi-objective model


Unlike the single-objective models, the bi-objective model attempts to optimize both pavement perfor-
mance and maintenance cost simultaneously. Although a wide range of weight combinations could be used
to represent different levels of importance put toward these two objectives, equal weights for both objectives
(w1 and w2 are 0.50) are adopted in this study to illustrate the effects of simultaneous optimization of two
objectives to develop a compromise maintenance plan. Fig. 9 displays the convergence curves for both deter-
ministic and stochastic formulations. From Fig. 9, the deterministic and stochastic formulations recom-
mended the maintenance plans with the total maintenance costs of US$ 284,595 and US$ 287,257,
740 P. Chootinan et al. / Transportation Research Part A 40 (2006) 725–743

Table 6
Summary of maintenance activities under cost minimization (stochastic case)
Treatment option Area of treatments (m2)
2000 2001 2002 2003 2004 2005 2006 2007 2008 2009
0 (Do-nothing) 272,699 253,808 256,794 280,085 253,849 240,723 300,700 327,762 305,614 331,192
1 (Routine) 67,908 85,779 95,522 75,222 101,458 98,210 55,223 34,545 49,819 31,115
2 (Surface treatment) 21,700 22,720 9991 7000 7000 23,374 6384 0 6874 0
3 (Minor rehabilitation) 0 0 0 0 0 0 0 0 0 0
4 (Major rehabilitation) 0 0 0 0 0 0 0 0 0 0
% of total area treated 24.73 29.95 29.12 22.69 29.94 33.56 17.00 9.53 15.65 8.59
Maintenance cost (US$) 26,769 26,049 32,543 30,168 29,915 28,932 17,771 18,372 13,609 10,907
Network ave. PSR before 3.16 3.15 3.16 3.13 3.09 3.07 3.10 3.01 2.87 2.75
treatment
Network ave. PSR after 3.33 3.35 3.32 3.27 3.24 3.27 3.19 3.04 2.93 2.78
treatment
Standard deviation of PSR 0.000 0.020 0.027 0.031 0.035 0.040 0.042 0.044 0.047 0.049

0.39 36.0
Pavement Condition (Deterministic) -34.80

34.4
Total Maintenance Cost ($M)...

0.37

Pavement Condition (Stochastic) -34.22


32.8

Cumulative PSR
0.35

0.33 31.2
Maintenance Cost (Stochastic) -US$ 287,257

0.31 29.6

0.29 28.0

Maintenance Cost (Deterministic)- US$ 284,595


0.27 26.4
0 5,000 10,000 15,000 20,000 25,000 30,000 35,000 40,000 45,000 50,000
Number of Generations

Fig. 9. Convergence curve of bi-objective model.

respectively. The network performances delivered by these two maintenance plans are 34.80 and 34.22,
respectively.
In order to evaluate the performances of the bi-objective maintenance plans developed by the deterministic
and stochastic formulations, the pavement performance curve and pavement deterioration curve in Fig. 10
were developed. The pavement performance curve displays the average network PSR value before treatment
and after treatment for each year of the 10-year maintenance period. Due to the stochastic nature of pavement
deterioration, the 90% confidence level is also included for both the pavement performance and pavement
deterioration curves.
From Fig. 10, it can be seen that the bi-objective models exhibit similar traits as the single-objective coun-
terparts. That is, the first several years of the planning period have steadily increasing pavement performance
values. This is similar to the pavement performance curves of the pavement performance maximization objec-
tive. As the bi-objective model also attempts to satisfy the cost minimization objective, after the sixth year,
when the pavement network reaches its maximum condition, the average network performance level starts
to decrease. This is similar to the pavement performance curves of the cost minimization objective. Similar
P. Chootinan et al. / Transportation Research Part A 40 (2006) 725–743 741

Deterioration Lower Deterioration Mean


Deterioration Upper Performance Lower
Performance Mean Performance Upper
Deterministic Performance

4.500
Network Average PSR

4.000

3.500

3.000

2.500

2.000

1.500
1999 2001 2003 2005 2007 2009
Year

Fig. 10. Performance of pavement under bi-objective model.

to the pavement performance curves, the bi-objective budget allocation of Fig. 11 mimics the budget allocation
of the performance maximization objective in the first five years and the cost minimization objective in the last
five years.
Similar to the results of single-objective models, a larger amount of maintenance cost is required for the
stochastic formulation. Investigation of the treatment scheme summarized in Tables 7 and 8 indicates that
the deterministic formulation treated a larger area of pavement, where as the stochastic formulation treated
less area of pavement with a higher level of treatment. A maintenance cost increase of US$ 1308 from the
deterministic formulation to the stochastic formulation is required to sustain comparable PSR values. In gen-
eral, the following observations can be made about the bi-objective results:

Budget
0.06 US$ 61,410
Annual Maintenance Cost ($M)...

0.05

0.04

Deterministic
0.03 Stochastic

0.02

0.01

0.00
2000 2001 2002 2003 2004 2005 2006 2007 2008 2009
Year

Fig. 11. Budget allocation under bi-objective model.


742 P. Chootinan et al. / Transportation Research Part A 40 (2006) 725–743

Table 7
Summary of maintenance activities under bi-objective model (deterministic case)
Treatment option Area of treatments (m2)
2000 2001 2002 2003 2004 2005 2006 2007 2008 2009
0 (Do-nothing) 170,240 162,132 188,336 186,209 126,764 154,737 193,889 259,813 309,660 362,307
1 (Routine) 151,222 171,706 143,596 142,238 214,634 193,696 161,418 94,500 52,647 0
2 (Surface treatment) 40,845 28,469 30,375 33,860 20,909 13,874 7000 7994 0 0
3 (Minor rehabilitation) 0 0 0 0 0 0 0 0 0 0
4 (Major rehabilitation) 0 0 0 0 0 0 0 0 0 0
% of total area treated 53.01 55.25 48.02 48.60 65.01 57.29 46.48 28.29 14.53 0.00
Maintenance cost (US$) 59,816 54,722 50,602 52,903 57,586 48,288 36,879 24,461 10,346 0
Network ave. PSR before 3.16 3.32 3.46 3.55 3.66 3.81 3.89 3.91 3.82 3.68
treatment
Network ave. PSR after 3.50 3.64 3.73 3.84 3.99 4.07 4.08 3.99 3.85 3.68
treatment

Table 8
Summary of maintenance activities under bi-objective model (stochastic case)
Treatment option Area of treatments (m2)
2000 2001 2002 2003 2004 2005 2006 2007 2008 2009
0 (Do-nothing) 194,584 172,340 181,524 148,083 151,963 152,529 227,774 246,176 315,411 362,307
1 (Routine) 118,429 154,764 156,010 186,343 173,540 183,948 134,533 116,131 46,896 0
2 (Surface treatment) 49,294 35,203 24,773 27,881 36,804 25,830 0 0 0 0
3 (Minor rehabilitation) 0 0 0 0 0 0 0 0 0 0
4 (Major rehabilitation) 0 0 0 0 0 0 0 0 0 0
% of total area treated 46.29 52.43 49.90 59.13 58.06 57.90 37.13 32.05 12.94 0.00
Maintenance cost (US$) 59,598 56,355 48,913 57,165 61,224 55,183 26,437 22,821 9,216 0
Network ave. PSR before 3.16 3.28 3.42 3.55 3.68 3.84 3.96 3.92 3.83 3.68
treatment
Network ave. PSR after 3.47 3.61 3.73 3.86 4.02 4.13 4.10 4.00 3.85 3.68
treatment
Standard deviation of PSR 0.000 0.019 0.026 0.027 0.029 0.032 0.037 0.039 0.043 0.045

1. The deterministic formulation underestimates the pavement deterioration resulting in a reduced mainte-
nance cost throughout the planning period. This underestimation of maintenance cost translates to a cost
difference of US$ 42 per lane-mile of roadway.
2. Maintenance activities of the stochastic formulation require a larger amount of maintenance cost. This indi-
cates that the deterministic formulation underestimates the pavement deterioration; thus, underestimates
the required maintenance cost.
3. On average, the stochastic formulation applies a higher level of treatment in order to maintain comparable
pavement performance values. This led to 9.0% more of Type 2 treatment being applied in the stochastic
formulation when compared to the deterministic formulation.

5. Summary and conclusion

In order to address the growing concerns of pavement management at the network level, accounting for
uncertainty in pavement maintenance programming is essential. Based on the results of this study, it is evident
that the pavement maintenance programming models using a deterministic formulation for the development
of pavement maintenance plan underestimate the level of pavement deterioration in future years. A direct
result of this is the underestimation of the required maintenance cost and the overestimation of the expected
pavement performance. By accounting for uncertainty, in the form of prediction error, multi-year pavement
maintenance plans can more accurately predict the future needs of the pavement network.
P. Chootinan et al. / Transportation Research Part A 40 (2006) 725–743 743

Through the development of the simulation-optimization framework proposed in this paper, PMS can
develop multi-year pavement maintenance programs that can account for uncertainty in the prediction pro-
cess. While accounting for uncertainty, a wide range of PMS objectives can be implemented into the frame-
work while utilizing existing deterioration models. This allows for simple adoption of these methods since
existing PMS procedures do not need to change. In addition to the needs of PMS, decision makers are allowed
to limit the probability or risk that the selected maintenance plan would fail to maintain the required levels of
performance. This allows for the generation of maintenance plan alternatives that account for various degrees
of risk.

Acknowledgement

This research is supported by the Community/University Research Initiative (CURI) grant from the State
of Utah.

References

Al-Suleiman, T.I., Sinha, K.C., Riverson, J.D., 1991. Effects of pavement age and traffic on maintenance effectiveness. Journal of
Transportation Engineering 117 (6), 644–659.
Ang, A.H.S., Tang, W.H., 1984. Probability Concepts in Engineering Planning and Design, Decision, Risk, and Reliability, vol. II. John
Wiley & Sons, New York.
Ben-Akiva, M., Humplick, F., Madanat, S., Ramaswamy, R., 1993. Infrastructure management under uncertainty: latent performance
approach. Journal of Transportation Engineering 119 (1), 43–58.
Chan, W.T., Fwa, T.F., Tan, C.Y., 1994. Road-maintenance planning using Genetic Algorithms—I. Formulation. Journal of
Transportation Engineering 120 (5), 693–709.
Chootinan, P., 2001. Pavement maintenance programming using a stochastic simulation-based genetic algorithm approach. Master Thesis,
Department of Civil and Environmental Engineering. Utah State University, Logan, UT.
Durango, P.L., Madanat, M., 2002. Optimal maintenance and repair policies in infrastructure management under uncertain facility
deterioration rates: an adaptive control approach. Transportation Research 36A, 763–778.
Goldberg, D.E., 1989. Genetic Algorithm in Search, Optimization, and Machine Learning. Addison-Wesley, Massachusetts.
Haas, R.C.G., Hudson, W.R., 1982. Pavements Management Systems. McGraw-Hill, Inc., New York.
Irrgang, F.C., Maze, T.H., 1993. Status of pavement management systems and data analysis models at state highway agencies.
Transportation Research Record 1397, 1–6.
Lee, Y.H., Mohseni, A., Darter, M.I., 1993. Simplified pavement performance models. Transportation Research Record 1397, 7–14.
Li, N., Xie, W., Haas, R., 1996. Reliability-based processing of Markov chains for modeling pavement deterioration. Transportation
Research Record 1524, 203–213.
Li, N., Haas, R., Xie, W., 1997. Investigation of relationship between deterministic and probabilistic prediction models in pavement
management. Transportation Research Record 1592, 70–79.
Liu, B., 1999. Uncertain Programming. John Wiley & Sons, Inc., New York.
Mbwana, J.P., Turnquist, M.A., 1996. Optimization modeling for enhanced network-level pavement management system. Transportation
Research Record 1524, 76–85.
McKay, M.D., 1988. Uncertainty analysis. In: Ronan, Y. (Ed.), Sensitivity and Uncertainty Analysis Using a Statistical Sample of Input
Values. CRC Press, FL, Chapter 4.
Peterson, D.E., 1987. NCHRP synthesis 135: pavement management practice. Transportation Research Board, National Research
Council, Washington, DC.
Robert, W.E., Marshall, A.R., Lin, S.S., Shepard, R.W., Aldayuz, J., 2002. Integration of agency rules with the preservation optimization
model in the Pontis bridge management system. Transportation Research Record 1795, 74–81.
Utah LTAP, 2004. Transportation Asset Management System.
Wang, K.C.P., Zaniewski, J., Way, G., 1994. Probabilistic behavior of pavements. Journal of Transportation Engineering 120 (3), 358–
375.
Zimmerman, K.A., 1995. Pavement Management Methodologies to Select Projects and Recommended Preservation Treatments: NCHRP
Synthesis 222. Transportation Research Board, National Research Council, Washington, DC.

You might also like