Professional Documents
Culture Documents
Abstract: This paper presents a modeling framework for planning pavement resurfacing activities on highway networks in the case of
continuous pavement state, discrete time, and infinite horizon. Optimal resurfacing policies that minimize discounted life-cycle costs are
Downloaded from ascelibrary.org by University Of Ulster AT on 06/15/15. Copyright ASCE. For personal use only; all rights reserved.
obtained by solving a multidimensional dynamic program, where travelers’ route choices and the agency’s resource allocation decisions
are considered simultaneously. To reduce computational difficulty, policy iteration is used together with a parametric function approxi-
mation technique. Numerical examples show that the proposed approach solves the planning problem efficiently. The effect of various
factors 共travelers’ route choice, agency budget, etc.兲 on the optimal policies is also analyzed.
DOI: 10.1061/共ASCE兲1076-0342共2007兲13:1共65兲
CE Database subject headings: Pavement management; Highways; Transportation networks; Traffic assignment; Life cycle cost.
formulated as
再冕 冎
algorithms have been proposed; see Sheffi 共1992兲 for a review.
⬁ n +1 The agency costs for pavement resurfacing consist of a fixed
min 兺 兺
=0 i=0
Ci关si共t兲,v共t兲兴e−rtdt + M i关wi共兲兴e−r 共1a兲 setup component 共e.g., for machine rental, operation, and labor兲,
and a variable component that is proportional to the intensity
共e.g., material expenditures兲. Therefore, the undiscounted agency
dsi共t兲 cost for a resurfacing of intensity wi may be represented by the
Subject to = Fi关si共t兲, vi共t兲兴, t 苸 共, + 1兲, ∀ i,
dt following function:
共1c兲
M i共wi兲 = 再 mi0 + mi1 · wi,
0,
if wi ⬎ 0
otherwise
∀i 共3兲
再冕 冎
where di0⫽free flow travel time, and parameter ␣i decreases with
t
the traffic flow capacity of facility i. For simplicity, we assume
that the vehicle operating costs are approximately proportional to si共t兲 = si共+兲 · exp 关bi0 + bi1vi共t兲兴dt , t 苸 共, + 1兲, ∀ i,
roughness 共Li and Madanat 2002; Ouyang and Madanat 2004,
2006兲, such that the user cost per unit time per vehicle is in the Fig. 1 illustrates some of these definitions.
Problem formulation 共1a兲–共1f兲 is a deterministic infinite-horizon rithm we propose in this paper is a combination of parametric
dynamic program with multidimensional continuous state vari- approximation and policy iterations, which is a slight modifica-
ables s and v, while v is directly determined by s from Eq. 共1d兲. tion of the parametric policy iteration 共PPI兲 algorithm in Benitez-
We now write the multidimensional control variable w as a policy Silva et al. 共2000兲 and Howitt et al. 共2002兲. The proposed PPI
w共s兲, and let J共s兲 be the value function in an infinite horizon. The algorithm for our pavement management problem involves the
optimal policy can be found by solving the deterministic Bellman following major steps.
再
equation • Parametric approximation: We approximate the value function
冉冕 冊
n
J共s兲 with a finite-linear combination of basis functions, k共s兲,
1
k = 1 , . . . , K, 共e.g., a complete set of ordinary or Chebychev
J共s兲 = min 兺 Ci关si共t兲,v共t兲兴e−rtdt + M i共wi兲e−r
polynomials in each si兲; i.e.
冎
w共s兲 i=0 0
K
+ e−r · J共s⬘兲 , ∀s 共8a兲 J共s兲 ⬇ 关1共s兲,2共s兲, . . . ,K共s兲兴 · = 兺
k=1
kk共s兲 共10兲
where
where, = 关1 , 2 , . . . , K兴T is a column vector of real param-
v共t兲 = H关s共t兲,D共t兲兴, ∀ t 苸 共0,1兲 共8b兲 eters.
• State-space sampling: To estimate , we take a finite sample
si⬘ = si · exp 冉冕 1
关bi0 + bi1vi共t兲兴dt 冊 共e.g., N grid points兲 of states from the state space, s1 , . . . , sN,
where N ⬎ K. The number N does not have to be very large.
再 冊冎
0 Each sample state, si, represents a possible initial pavement
− G wi,si · exp 冉冕 1
0
关bi0 + bi1vi共t兲兴dt , ∀i 共8c兲
condition at time t = 0.
• Initialization: Specify an initial parameter vector and initial
policies w共si兲 = w0 for all i.
• Parameter update: With the parametric approximation, Eq. 共9兲
n
is transformed into the following:
兺 M i共wi兲 ⱕ B 共8d兲
0 ⱕ wi ⱕ g2si · exp
i=0
再冕 1
冎
关bi0 + bi1vi共t兲兴dt + g3, ∀ i 共8e兲
兺
K
k=1
kk共s兲 = 兺
i=0
n
再冕 0
1
Ci关si共t兲,v共t兲兴e−rtdt + M i关wi共s兲兴e−r 冎
K
兺
0
+e · −r
kk共s⬘兲
Note that we only need to focus on t 苸 关0 , 1兴 from now on. The k=1
optimal policy w共s兲 and value function J共s兲 should satisfy the
following equation: and thus
J共s兲 = 兺
n
i=0
再冕 1
0
Ci关si共t兲,v共t兲兴e−rtdt + M i共wi兲e−r + e−r · J共s⬘兲, 冎 兺
n
i=0
再冕 0
1
Ci关si共t兲,v共t兲兴e−rtdt + M i关wi共s兲兴e−r 冎
共9兲
K
subject to Eqs. 共8b兲–共8e兲.
For problems with finite states, Eq. 共9兲 can be solved numeri-
= 兺
k=1
关k共s兲 − e−r · k共s⬘兲兴 · k . 共11兲
cally via either policy iteration or value iteration. For our case,
both state and control variables are continuous. There is consid- Within each iteration, we numerically calculate from Eqs. 共8b兲
erable difficulty regarding efficient algorithms to solve this type and 共8c兲 the deterioration of pavement conditions over time,
of high-dimensional dynamic program. A straightforward method s共t兲, and the corresponding traffic flow distribution,
is to discretize the continuous state space into a very large number v共t兲 = H关s共t兲 , D共t兲兴 , ∀ t 苸 共0 , 1兲, for each sample point 关i.e.,
of grid points and then solve the resulting finite-state dynamic s共0兲 = si, i = 1 , 2 , . . . , N兴. This calculation also yields the final
program. This method obviously involves a lot of variables and is pavement conditions s⬘ at time t = 1+, and the following vari-
computationally expensive. ables for each sample point:
再冕 冎
Fig. 3. Traffic demand
n 1
yª 兺 Ci关si共t兲,v共t兲兴e−rtdt + M i关wi共s兲兴e−r ,
i=0 0 updating step is similar to a simpler mixed-integer linear
program.
xk ª k共s兲 − e · k共s⬘兲, ∀ k
−r
The PPI algorithm repeats the parameter update and policy
Regressing y on x ª 关x1 , x2 , . . . , xK兴 over the N observations update steps iteratively at the N sample points until the parameter
Downloaded from ascelibrary.org by University Of Ulster AT on 06/15/15. Copyright ASCE. For personal use only; all rights reserved.
gives a new estimation of the parameter vector new. We up- vector converges. Each iteration step involves N numerical in-
date by ← 共1 − ␣兲 + ␣new, where 0 ⬍ ␣ ⬍ 1. tegrations, one regression, and N optimizations. At convergence,
Normally, small ␣ improves asymptotic convergence, but too J共s兲 can be calculated from Eq. 共10兲 and the optimal policy w共s兲
small ␣ would slow down convergence speed. An alternative can be computed for any state s with one more round of policy
choice is to decrease ␣ gradually, i.e., use ␣ = ␣m in iteration m, update.
such that ␣m → 0 and 兺␣m → ⬁. It should be noted that the parametric policy iteration algo-
• Policy update: Within each iteration, we also update the opti- rithm does not always guarantee convergence 共Benitez-Silva et al.
mal resurfacing intensity at each sample point. The updated 2000兲, especially for small K and N. However, it is obvious that
parameter vector is used to obtain function J共s⬘兲 on the the approximated function 共10兲 should asymptotically converge to
right-hand side of Eq. 共8a兲, which becomes the following the true continuous value function J共s兲 if 共1兲 K goes to infinity;
minimization problem: and 共2兲 the basis functions, k共s兲, k = 1 , . . . , K, are properly cho-
兺 再冕 冎 冋兺 册
n 1 n sen. Following this observation and the discussions in Bellman
min Ci关si共t兲,v共t兲兴e−rtdt + e−r M i共wi兲 + J共s⬘兲 共1961兲, Howitt et al. 共2002兲 hinted that the solution from the PPI
w i=0 0 i=0 algorithm should asymptotically converge to the true optimum
under Conditions 共1兲 and 共2兲.
Note that s⬘ is defined by Eqs. 共8b兲 and 共8c兲, and w is con-
strained within a convex polytope W共s兲 defined by Eqs. 共8d兲
and 共8e兲. Given s, the first summation term is independent of Numerical Example
再兺 冎
w; thus, the optimization simplifies to
n For the convenience of illustration, we analyze a simplest net-
min z共w兲 = M i共wi兲 + J共s⬘兲 共12兲 work with n = 2 pavement facilities connecting a residential area
w苸W共s兲 i=0 to a business center 共see Fig. 2兲. Travelers commute from the
residential area to the business center. The travel demand D共t兲 is
At the interior of W共s兲, we find, by substituting Eq. 共8c兲 into
a scalar that alternates between a high value of 10,000 vehicles
Eq. 共12兲, that the gradients of z,
per hour and a low value of 1,000 vehicles per hour, each lasting
z J s⬘ for a period of T / 2 共e.g., half a day, such that T = 1 / 365 year兲 共see
wi
= mi1 + 兺j s⬘j wji Fig. 3兲.
We extract from past literature 共e.g., Tsunokawa and
J si⬘ Schofer 1994; Ouyang and Madanat 2006兲 reasonable values for
= mi1 + the parameters in the performance models:  = 4, c0 = 0,
si⬘ wi
c1 = 2 ⫻ 10−4 $ / QI/ vehicle, c2 = 0.5 $ / min, g1 = 0.66, g2 = 0.55,
J and g3 = 18.3. Other parameters are listed in Table 1. Facility 1 has
= mi1 −
si⬘ a larger traffic capacity and larger resurfacing costs; it is assumed
to be “major.”
冢 冒再 冊冎 冣
g1
⫻ , ∀i
g2 + g3 si exp 冉冕 0
1
关bi0 + bi1vi共t兲兴dt
Scenario 1: No Budget Constraint
We first consider a benchmark case with B = ⬁. Constraint 共1e兲 is
essentially nonbinding. The pavement facilities are related only
共13兲
through the travelers’ route choices.
is independent of w. Thus, an optimal solution always exists at We approximate the value function by a complete set of ordi-
an extreme point of W共s兲. Thus, the complexity of the policy nary polynomials of total degree 5 in s1 and s2; i.e.
Fig. 4. Optimal solution with no budget constraint: 共a兲 estimated J共s兲; 共b兲 optimal w1共s兲; and 共c兲 optimal w2共s兲
兵sk11sk22: ∀ k1,k2,0 ⱕ k1 + k2 ⱕ 5其 共s1 ⬎ 80 QI兲 关see Fig. 4共b兲兴. A similar pattern exists in Fig. 4共c兲.
The reason is obvious. Since travelers tend to choose a “better”
The total number of parameters K = 21. We then discretize the pavement facility, the traffic flow on one facility is smaller if
state space with N = 400 grid points 共20 each in the s1 and s2 other facilities are in better conditions. The user costs on this
dimensions, from 7.5 QI, 15 QI, . . . , to 150 QI兲. Then we apply facility will slightly decrease. This makes “do nothing” a more
the iterative procedure with the following initial parameters: appealing option.
k = 1, ∀ k = 1,2, . . . ,K, and
Scenario 2: Budget Constraint
w1共si兲 = w2共si兲 = 0, ∀ i = 1,2, . . . ,400
Benitez-Silva et al. 共2000兲 reported that the parametric policy We then impose a finite budget B = 6 ⫻ 106 $ to see how limited
iteration algorithm does not necessarily guarantee convergence, resource affects the optimal policies. In this case, the pavement
but with our choice of initial parameter values and constant ␣ facilities are related not only through travelers’ route choices but
= 0.5, convergence is achieved within 50 iterations. also through constraint 共1e兲.
Figs. 4共a–c兲 show the value function J共s兲 and the optimal We use exactly the same PPI procedure to obtain the optimal
policy w1共s兲, w2共s兲 at convergence. Obviously, J共s兲 is monotoni- solution to this problem. The value function J共s兲 and optimal
cally and concavely increasing with both s1 and s2. In this case, policies w1共s兲, w2共s兲 are plotted in Figs. 5共a–c兲. Again, J共s兲 is
the optimal policy for each pavement facility roughly has a monotonically and concavely increasing with s1 and s2, in a shape
threshold structure. Figs. 4共b and c兲 show that it is not optimal to similar to the benchmark case in Fig. 4共a兲. Fig. 6 plots the change
resurface a facility unless its roughness reaches a certain value of J共s兲 due to the imposition of the finite-budget constraint. As
共i.e., about 70 QI for Facility 1 and about 130 QI for Facility 2兲. expected, the expected cost J共s兲 increases for all s, and the in-
This is similar to the optimal policies for a single pavement fa- crease is most significant for large s1 共i.e., when Facility 1 is in
cility 共Ouyang and Madanat 2006兲. However, due to travelers’ bad condition兲. This is intuitive. When the “major” facility is in
route choices, the threshold for one facility is dependent of the bad condition, it is more costly to fully resurface it, and then the
conditions of other facilities. For example, when Facility 2 is in a budget constraint becomes most restrictive.
very good condition 共s2 ⬍ 30 QI兲, we tend to refrain from resur- As in Figs. 4共b and c兲, Facilities 1 and 2 shall not be resur-
facing Facility 1 unless s1 deteriorates to a larger value faced unless their conditions reach certain thresholds. When a
Fig. 5. Optimal solution with a finite budget: 共a兲 estimated J共s兲; 共b兲 optimal w1共s兲; and 共c兲 optimal w2共s兲
finite-budget constraint is imposed, facilities tend to have lower infinite budget. Facility 2 shall never be resurfaced if that poses a
resurfacing thresholds. For example, the threshold for Facility 1 competition of resource against Facility 1; i.e., when Facility 1 is
reduces to about 60 QI in Fig. 5共b兲. When s1 ⬇ 50 QI, similarly, in a moderate to bad condition 共s1 ⬎ 60 QI兲. This seemingly coun-
Facility 2 will be resurfaced as long as s2 reaches 120 QI 关see Fig. terintuitive finding can be related to the concept of economy of
5共c兲兴. scale—when resources are limited, investing all or nothing on
Also clear from Figs. 5共b and c兲, major facilities are given alternative facilities may be the most cost-effective choice. In
resurfacing priority if resource 共budget兲 is limited. In our case, the practice, however, highway agencies may pose additional con-
optimal policy for Facility 2 is very different from the one with straints to prevent certain pavement facilities from deteriorating
infinitely. Such constraints are not considered in this model.
The change of optimal policies can be explained as follows.
We know from Eq. 共13兲 that the gradient of z共w兲 satisfies
z / wi = mi1 + 共J / si⬘兲共si⬘ / wi兲 , ∀ i, while si⬘ / wi is always
negative. Observe that when the budget constraint is imposed, not
only J共s兲 but also J / si⬘ increases for most of s⬘. Therefore,
when budget constraint is imposed, z / wi will have a smaller
value during every policy update step even for the same s and s⬘.
This reduction in gradient will make a larger wi, ∀i, more favor-
able. In our case, Facility 1 has priority for budget allocation
because J / s1⬘ increases the most significantly. This makes in-
creasing w1 the most favorable.