Probabilistic Approach to Modeling Pavement Performance Using IRI Data

Juan Diego Porras-Alvarado, MS.

Graduate Research Assistant
Email: [email protected]
The University of Texas at Austin
Department of Civil, Architectural & Environmental Engineering (CAEE)
Austin, TX 78712

Zhanmin Zhang, Ph.D.

Associate Professor and Fellow of the Clyde E. Lee Endowed Professorship in Transportation
Engineering
Email: [email protected]
The University of Texas at Austin
Department of Civil, Architectural & Environmental Engineering (CAEE)
Austin, TX 78712

Luis Guillermo Loría Salazar, Ph.D.

General Coordinator
E-mail: [email protected]
Transportation Infrastructure Program LanammeUCR

Word Count:
Text 3802
Figures 4 x 250 = 1000
Tables 3 x 250 = 750
Total 5552

Abstract = 187 words

Paper Submitted for Presentation and Publication at the 93rd Annual Meeting
of the Transportation Research Board

TRB 2014 Annual Meeting Original paper submittal - not revised by author
 J.D. Porras, Z. Zhang & L. Loria 1

1 ABSTRACT
2
3 Accurately predicting pavement performance is an essential element in road infrastructure
4 management. Pavement performance prediction methods can be either deterministic or
5 probabilistic, depending on the method employed to simulate the deterioration or aging process.
6 Deterministic models predict the condition on the basis of mathematical functions of observed or
7 measured deterioration without taking uncertainties associated with the deterioration process into
8 consideration. Probabilistic models, on the other hand, take uncertainties into consideration and
9 predict the condition as the probability of occurrence in a range of possible outcomes. To
10 overcome this shortcoming of deterministic models, probabilistic approaches have been
11 investigated by various researchers. Probabilistic methods can be summarized into three
12 categories: econometric models, Markov Chain Process (PMC) models, and reliability analysis.
13 This paper presents a Markov Chain-based methodological framework to characterize pavement
14 performance in support of pavement management decision makings. The International
15 Roughness Index (IRI) data from the National DOT in the Costa Rican was used for the
16 numerical case study to illustrate the application of the developed methodological framework.
17 The findings from this study show that the proposed methodological framework is a viable
18 approach to modeling pavement deterioration process.
19

TRB 2014 Annual Meeting Original paper submittal - not revised by author
 J.D. Porras, Z. Zhang & L. Loria 2

20 INTRODUCTION
21
22 Pavement performance prediction is necessary for rational budget planning and resource
23 allocation. At the network level, pavement performance prediction is needed for programming
24 maintenance and rehabilitation (M&R) activities, while at the project level it is needed for
25 determining the most appropriate M&R actions to be taken for a specific project, such as
26 preventive maintenance, rehabilitation, or reconstruction (1,2,3,4,5).
27 Research shows it is less expensive to maintain a road than to repair it once it has
28 significantly deteriorated (2,3,4,5,6). This is why pavement management systems (PMSs)
29 prioritize corrective actions depending on pavement condition and available funds. If the system
30 is in good condition, PMS would prioritize preventive maintenance rather than the reconstruction
31 of roads in poor condition. Prioritizing the repair of bad sections of roadway over preventative
32 maintenance for good roads is not the optimal use of funding resources. However, if current
33 conditions of the network are fair or poor then preventive maintenance are not a PMS priority,
34 and rehabilitation and reconstruction become the best alternative. In terms of life-cycle cost and
35 long-term pavement conditions, following the aforementioned strategy results in better system
36 performance (1,3,4,5).
37 Accurate prediction of pavement performance requires reliable pavement deterioration
38 models. Pavement performance prediction methods can be either deterministic or probabilistic,
39 depending on the method employed to simulate the deterioration or aging process. Deterministic
40 models predict the condition on the basis of mathematical functions of observed or measured
41 deterioration without taking uncertainties associated with the deterioration process into
42 consideration. Probabilistic models, on the other hand, take uncertainties into consideration and
43 predict the condition as the probability of occurrence in a range of possible outcomes
44 (2,4,6,7,8,9,10).
45 Most deterministic models are based on establishing regression relationships between
46 performance indicators and independent variables related to pavement performance, such as
47 applied traffic loadings, material characteristics, and environmental conditions. Though
48 deterministic models provide reasonably good prediction results, their deterministic nature does
49 not allow them to be used to capture the inherent uncertainty in the process of pavement
50 deterioration. In other words, these models are constrained by the fact that they cannot take the
51 stochastic nature associated with the pavement performance into consideration (2,4,5,9).
52 To overcome this shortcoming of deterministic models, probabilistic approaches have
53 been investigated by various researchers. Probabilistic methods can be summarized into three
54 categories: econometric models, Markov Chain Process (MCP) models, and reliability analysis.
55 Econometric models are widely used to correlate pavement distress with explanatory variables,
56 such as structural number (SN), thickness of the surface layer, and number of wheel passes per
57 unit strength of pavement (2). MCP is used to determine the transition from one state condition
58 to another of a pavement section or network, using the Transition Probability Matrix (4,6,8). The
59 limitation associated with the traditional TPMs is that TPM cannot directly account for the
60 impact of pavement types, environmental factors, traffic loading, and other relevant factors on
61 the deterioration process. The improved econometric methods such as ordered probit model,
62 Poisson model, and random-effects probit models were therefore proposed to connect the
63 relevant explanatory variables with the transition probabilities. Time-based models, belonging to
64 the reliability model category, are considered alternatives to Markov Chain models as they focus

TRB 2014 Annual Meeting Original paper submittal - not revised by author
 J.D. Porras, Z. Zhang & L. Loria 3

65 on estimating the probability distributions of time taken to transit from one condition state to
66 another using duration models (2,4,11).
67 The objective of study is develop a Markov Chain-based methodological framework to
68 characterize pavement performance in support of pavement management decision makings. The
69 International Roughness Index (IRI) data from the National DOT in the Costa Rican was used
70 for the numerical case study to illustrate the application of the developed methodological
71 framework. The rest of the paper is organized in four parts. First, a brief overview of the MCP is
72 presented. Then, the proposed methodology is thoroughly explained. Next, a case of study using
73 the primary road network in Costa Rica is discussed to demonstrate the use of the proposed
74 methodology. Finally, the summary and findings are presented.
75
76 MARKOV CHAIN PROCESS (MCP)
77
78 In this study, MCP is used to determine the TPMs. The Markov prediction model is a stochastic
79 process that: is discrete in time, has a countable or finite state space, and satisfies the Markov
80 property. The Markov property is satisfied if the future state of the process depends on its present
81 state, but not on its past states. In the pavements field, the Markov property is satisfied if the
82 future condition of the network is dependent on the present condition of the network and not on
83 its past condition (8,13). In other words, the Markov Chain model has no memory of the past.
84 Road condition can be modeled by two types of Markov processes, homogeneous or non-
85 homogeneous. In the case of the homogeneous process, the road network will always deteriorate
86 following the transition probabilities of one single transition matrix. If the pattern of
87 deterioration of a particular road network is likely to change at a certain point in time, t, the
88 deterioration process may be modeled by a non-homogeneous process. This implies the use of a
89 different transition matrix before and after t. In this case, the vector of the condition at t will
90 become the starting vector for the second chain, operating with a different transition matrix. This
91 type of arrangement may be performed as many times as required (2).
92 MCP models prove an effective method to predict performance deterioration of
93 infrastructure facilities because of their ability to capture uncertainty of the deterioration process.
94 Additionally, these models show an important applicability because of their relatively simple
95 analytical procedure, becoming a very attractive alternative for DOTs in the U.S. and other
96 highway agencies around the world, especially for network-level analysis (4,6,10,11,12,14).
97 However, to analyze results correctly, model limitations should be considered. The memory-less
98 property of MCP becomes one of the most important limitations because it using different TPMs
99 at different times of analysis horizon implies that past conditions does affect future conditions.
100 In addition, this approach does not provide a mechanism for physical factors important to the
101 deterioration process to be incorporated in the modeling process. Physical factors can be
102 introduced by econometric methods such as ordered probit model, Poisson model, and random-
103 effects probit models. These models have been used successfully by various studies to calibrated
104 TPMs for highway agencies. Despite modeling shortcomings the MCP provides a powerful
105 relative easy methodology to develop pavement performance models at the network level.
106
107 METHODOLOGY
108
109 One of the important goals for highway agencies is to keep the highway infrastructure network in
110 good condition with available funds. For this reason, pavement performance prediction becomes

TRB 2014 Annual Meeting Original paper submittal - not revised by author
 J.D. Porras, Z. Zhang & L. Loria 4

111 essential in order for decision makers to allocate funds as efficiently as possible. Figure 1 shows
112 the proposed methodological framework for pavement performance modeling using MCP in the
113 context of the generic infrastructure management process. The methodology consists of three
114 main modules: data analysis, modeling, and results analysis.
115

116
117 FIGURE 1 Methodology
118
119 Data Analysis
120 Data analysis serves as the first module of the proposed methodology. The objective is to
121 determine the available information in the dataset that can be potentially used for pavement
122 prediction modeling. The availability of such information as pavement structure, materials
123 characterization, traffic volumes, climatic conditions, and M&R history should be determined by
124 examining the existing datasets; if the information is available, it should be carefully analyzed in
125 terms of its accuracy, format, completeness, integrity.
126
127 Modeling
128 The second module is centered on modeling pavement deterioration using a probabilistic
129 approach. More specifically, the proposed methodology is based on MCP, but tailored to satisfy
130 the specific needs of this research, as illustrated in figure 1. The modeling module is divided into
131 five steps: time interval determination, condition state formulation, distribution of pavement
132 sections in the condition states, definition of the MRD and TPMs’ computation, and optimization
133 of the TPMs.
134 The first step in this module is to determine the time interval of the data collection, which
135 in turn is used to define the time duration for the transition probability over which the pavement
136 condition state will change from one cycle to the next. Most of the DOTs in the U.S. and certain
137 highway agencies of other countries collect pavement condition data every one or two years.
138 Usually, data collected can be classified in two categories: structural condition and functional
139 condition, measuring the structural integrity and ride quality of pavement, respectively (3,4,5). It
140 should be clarified that pavement condition is defined as a snapshot of the pavement structure in
141 time t, while pavement performance is the trend of pavement condition over a period of time.
142 The main objective of this study is to develop a probabilistic approach to predict pavement
143 performance or the trend of pavement condition over time using available data.
144 After the time interval is defined, pavement condition indices and rating scores should be
145 divided in a finite number of pavement condition states. The condition states should be carefully

TRB 2014 Annual Meeting Original paper submittal - not revised by author
 J.D. Porras, Z. Zhang & L. Loria 5

146 chosen such that they capture the full range of pavement behavioral conditions over the design
147 life. More specifically, bands that bound condition states must be defined over a condition
148 indicator if it is in rational scale, so that the probability of a pavement structure transitioning
149 from one condition state to the next can be determined.
150 The third step is to distribute the network into condition states previously established, i.e.,
151 the condition of the network is depicted with the percentages of the total network in each
152 condition state. Studies have shown that distribution can be presented in various attributes such
153 as the number of pavement management sections, pavement lane miles, percentage of the road
154 network, remaining service life, or distress measurements (4,5,6,13). Since the distribution
155 defines the overall condition of the pavement network in each cycle being analyzed and, in turn,
156 the TPMs, careful attention should be given to the selection of the attribute to be used to present
157 the distribution.
158 The next step is to define the Markov Rate Diagram (MRD) considering the pavement
159 condition states determined in previous steps. MRD is a graphical approach to understanding the
160 transition of the pavement network from one condition state to the next. Additionally, it allows
161 constraints to be incorporated to adjust for structural behavior of a pavement or to analyze

= 0 for > , it implies that the pavement condition cannot be

162 different condition scenarios. For example, by setting the transition probability from condition

= 0 can be used to represent a

163 states i to j to equal to 0, or
164 improved unless an M&R treatment is applied. Similarly,
165 holding condition state whereby where the pavement has reached its worst condition and cannot
166 deteriorate further. Figure 2 shows an example of an MRD and its corresponding mathematical
167 formulation in the form of a TPM. Each node represents the current condition and arrows
168 indicate the transition rate when the pavement deteriorates from one state to the next.
169 Additionally, the circular arrows indicate the probability that the pavement remains in the same
170 condition state. After the MRDs are defined, the initial TPMs for each of the scenarios being
171 considered can be developed.
172

0 0
0 0
=
0 0
0 0 0 1

FIGURE 2 Markov Rate Diagrams Example

173
174 The final step in this module is to optimize the TPMs by minimizing the error between
175 the real road network distribution and the calculated distribution using the initial TPMs . The
176 TPMs are optimized using the generalized reduced gradient nonlinear optimization code
177 incorporated as an add-in to the software Microsoft Excel (14). The objective function used
178 follows the form (6):
179

TRB 2014 Annual Meeting Original paper submittal - not revised by author
 J.D. Porras, Z. Zhang & L. Loria 6

= − ′ (1)
180 where:

′
181 = ith term of the distributions obtained from the TPMs.
182 = ith element of the original data distribution obtained from step three
183
184 The objective function minimizes the difference between the distribution of pavements
185 condition from the data set and the distributions obtained using the generated TPMs in the fourth
186 step. From the definition of the objective function, it is obvious that the transition probability
187 values or the elements of the TPMs are optimized through the optimization procedure (6).
188
189 Results Analysis
190 It is important to examine the probabilities values obtained in the TPMs to ensure that they
191 correspond to the MRD. The next step is to use the TPMs to predict pavement performance in
192 the following years to provide an insight on the M&R needs. Additionally, it is important to
193 consider incorporating calibration procedures to adjust the TPMs when new data is collected.
194 The new data obtained from new evaluation surveys will provide a feedback that should be used
195 to further calibrate the TPMs.
196
197 CASE STUDY: COSTA RICAN ROAD NETWORK
198
199 The Costa Rican highway system consists of 22,258 miles (35,820 km), divided among
200 municipal and national roads. A total of 17,398 miles (28,000 km) are part of the municipal
201 roads network, while the other 4,859 miles (7,820 km) constitute the national roads network.
202 Costa Rica has the second highest rate of roads by square kilometers in Latin America and is the
203 number 52nd in the world with 72 km of roads per square kilometer area. From the national roads
204 network, 60 percent are classified as flexible pavement, less than one percent is classified as
205 rigid pavement, and gravel roads make up the remaining 39 percent. However, 98 percent of the
206 budget used for maintenance and rehabilitation is invested in flexible pavement.
207
208 Data collection process
209 LanammeUCR has the ability to collect pavement performance data such as IRI, pavement
210 deflections and pavement surface friction. However, the data set provided by LanammeUCR
211 only comprises data related to the IRI and the FWD. Due to issues with the use of corrections for
212 moisture and temperature, FWD data was discarded. Thus, the study utilizes pavement roughness
213 as the only data to generate a probabilistic model for pavement performance for the Costa Rican
214 primary road network. Moreover, the data collection process includes advanced data verification
215 and quality control procedures. In Costa Rica, pavement roughness measurements are used for
216 quality control and project acceptance. The largest application of IRI in the nation is the National
217 Road Network Evaluation, which started in 2004 and is conducted every two years. The
218 evaluations are performed using a Dynatest Inertial Profiler, Model 5051 Mark III Roadway
219 Surface Profiler (RSP), property of LanammeUCR (15). The RSP computes the longitudinal and
220 transverse profile, measures rutting, and registers the operational speed of the equipment.

TRB 2014 Annual Meeting Original paper submittal - not revised by author
 J.D. Porras, Z. Zhang & L. Loria 7

221
222 Data analysis
223 For the evaluation of national road network, the results of the IRI are given at 100-m intervals,
224 and reported as the average value of the IRI for the left and right sensors of the laser profiler. The
225 average value of right and left IRI collected is known as the Mean Roughness Index (MRI).
226 Additionally, the data is reported in millimeters/meters or meters/kilometers, as standardized by
227 AASHTO R54 “Standard Practice for Accepting Ride Quality When Measured Using Inertial
228 Profiling Systems”. The data obtained is organized in shape files that can be manipulated in
229 geographical information systems (GIS). Information such as the M&R zone, route, IRI from left
230 and right wheel paths, and coordinates was found for each 100 meter section. Currently data is
231 available for the following years: 2004, 2006, 2008, and 2010.
232
233 Pavement condition states
234 Because data obtained is collected in two-year cycles, for this study the time interval was defined
235 as two years. Then the IRI data was discretized with the boundaries determined by using
236 information from the National Road Network Evaluation Program, as shown in Table 1.
237
238 TABLE 1 Pavement Conditions State
Pavement Condition State IRI in in/mile (m/km)
Good 0.00 – 70.87 (0.00 – 3.00)
Fair 70.87 – 106.30 (3.00 – 4.50)
Poor 106.30 – 151.18 (4.50 – 6.40)
Failed >151.18 (>6.40)
239
240

TRB 2014 Annual Meeting Original paper submittal - not revised by author
 J.D. Porras, Z. Zhang & L. Loria 8

241 Pavement condition for the years analyzed

242 The distribution of the pavement sections according to its IRI is computed in terms of percentage
243 of the entire network. Each pavement section is classified with the boundaries specified in Table
244 1. Once the pavement sections are classified in each category, the percentages are calculated by
245 dividing the total network’s length of each category by the total length of the network evaluated.
246 Table 2 shows the percentage of the road network for each year distributed in the four condition
247 states previously established.
248 It can be seen from Table 2 that the pavement network experiences degradation in terms
249 of IRI from 2004 to 2010. The “Good” conditions percentage is reduced by 8 percent, while the
250 “Fair” condition increased by approximately 8 percent. There is a reduction in “Poor” condition.
251 However, the “Failed” condition increases by almost 5 percent. Since maintenance and
252 rehabilitation information is not available, all sections were considered in calculating the
253 network distribution.
254
255 TABLE 2 National Road Network states condition
Year Condition (percent) Control Network
Good Fair Poor Failed Sections Length (km)
2004 30 28 28 14 539 3460
2006 26 30 25 19 780 4365
2008 28 32 24 16 780 4365
2010 22 35 23 20 780 4365
256
257
258 MRDs and TPMs
259 The Markov Rate Diagrams (MRDs) were defined by considering pavement conditions states
260 shown in Table 1. In order to analyze different conditions for the road network, various Markov
261 diagrams are generated. As examples, four diagrams are shown in figure 3. After distributing the
262 road network into established categories, the initial TPMs for each case under consideration are
263 computed. Using data from years 2004 and 2006, a preliminary TPM was computed for each of
264 the four rate diagrams cases shown in Figure 3.
265

TRB 2014 Annual Meeting Original paper submittal - not revised by author
 J.D. Porras, Z. Zhang & L. Loria 10

266

forward can occur, meaning that ≤ has to be

For this condition all transitions moving

0
true for all . The transitions backwards are

=
0 0
restrained. The cell with zero values means that
there is no probability of moving from a higher
0 0 0 1
state condition to a lower condition state. The last
assumption considers that a road cannot improve
itself without first receiving treatment.
a)
In this case, a further restriction was added
to the analysis. The control sections were allowed
to deteriorate by no more than one state in each

0 0
cycle. This configuration in pavement

0 0
deterioration modeling is commonly used because

=
it is assumed that degradation follows the natural

0 0
sequence, moving from one state to another rather

0 0 0 1
than moving more than one state at the time
(3,7,8,10). For the two configurations previously

expected since < for all entries

discussed a steady-state condition was not
were
restricted to be equal to zero.
b)

The next two rate diagrams were designed

=
to include repair actions. Information regarding
specific repairs for each section was not available.
Therefore, special treatment analysis was not
carried in this study.
For (c), all transition probabilities are
c) assumed to be possible as shown in the matrix
form (all entries are allowed).
The last rate diagram (d) shows the

0 0
degradation condition explained previously for the

0
second configuration. However, repair actions

=
0
were introduced. In this design, the repair is

0 0
assumed to move any condition state back to the
best condition state, signifying that maintenance
actions will recover the pavement performance in
terms of the IRI to a good condition.
d)
267 FIGURE 3 Markov Chain Rate Diagrams (a) MRD case 1, (b) MRD case 2, (c) MRD
268 case 3 and (d) MRD case 4

TRB 2014 Annual Meeting Original paper submittal - not revised by author
 J.D. Porras, Z. Zhang & L. Loria 11

269 TMP Optimization

270 The TPMs were optimized by minimizing the error between the real road network distribution
271 and the calculated distribution using the initial TPMs. The four estimated TPMs are optimized
272 using the generalized reduced gradient nonlinear optimization code incorporated as an add-in to
273 software Microsoft Excel (14) and the formulation was discussed earlier in the Modeling section.
274
275 RESULTS
276
277 This section divides the results in two categories: calculated TPMs and prediction for the year
278 2020.
279
280 Calculated TPMs
281 TPMs were derived for each of the previously discussed four cases, as shown in Table 3. TPM
282 MRD cases 1 and 2 show the results for the cases without M&R treatments applied, as the
283 pavement sections showing improvements in condition were dropped from the dataset. However,
284 the pavements sections that stayed in the same condition state were included in the dataset
285 regardless if M&R treatments were applied. The lower two matrices illustrate the results when
286 improvements were observed, which implies that M&R treatments were applied to the pavement
287 sections.
288
289 TABLE 3 Calculated TPMs
Good Fair Poor Failed Good Fair Poor Failed
Good 0.850 0.150 0.000 0.000 Good 0.851 0.149 0.000 0.000
Fair 0.000 0.800 0.127 0.073 Fair 0.000 0.860 0.140 0.000
Poor 0.000 0.000 0.773 0.227 Poor 0.000 0.000 0.847 0.153
Failed 0.000 0.000 0.000 1.000 Failed 0.000 0.000 0.000 1.000
TPM MRD case 1 TPM MRD case 2
Good Fair Poor Failed Good Fair Poor Failed
Good 0.712 0.077 0.031 0.179 Good 0.756 0.244 0.000 0.000
Fair 0.127 0.809 0.064 0.000 Fair 0.000 0.863 0.137 0.000
Poor 0.036 0.139 0.737 0.088 Poor 0.134 0.000 0.771 0.095
Failed 0.002 0.163 0.138 0.697 Failed 0.028 0.000 0.000 0.972
TPM MRD case 3 TPM MRD case 4
290
291 As expected, the matrix diagonal and the offset diagonal show the highest values for the
292 MRD cases 1 and 2. As briefly discussed earlier, MRD case 3 is a scenario where pavement
293 sections with M&R treatments are included. The MRD case 4 assumes that M&R would only
294 improve the current condition state “Good” (condition state 1).
295 Since neither pavement management system nor pavement prediction models are
296 currently being used by transportation authorities in Costa Rica, TPMs could provide a decision-
297 support tool to allocate funds using a data-driven approach, instead of an experienced-based
298 subjective approach. Though the developed TPMs need to include more pavement performance
299 indicators; however, they can serve as an initial step towards more data-driven pavement
300 management systems for Costa Rica.
301
302

TRB 2014 Annual Meeting Original paper submittal - not revised by author
 J.D. Porras, Z. Zhang & L. Loria 12

303 Costa Rican Road Network 2020 Predictions

304 Based on the calculated TPMs, the state conditions of the network for year 2020 are predicted.
305 The results are presented in pie plots for a more comprehensive comparison among the different
306 studied conditions. Figure 4 illustrates the predictions obtained.
307

12% 12%
33%
50% 21% 27%

17% 28%

Good Fair Poor Failed

Good Fair Poor Failed

(a) (b)

27% 17%
19% 20%

22% 34%
22%
39%

Good Fair Poor Failed

Good Fair Poor Failed

(c) (d)
308
309 FIGURE 4 2020 Pavement Performance Predictions
310 (a) MRD case 1, (b) MRD case 2, (c) MRD case 3 and (d) MRD case 4
311
312 For the two categories analyzed, with and without repair, it shows that results for the year
313 2020 are similar. The matrices for the no-repair condition are similar in terms of magnitude of
314 each of the probabilities. Consequently, the percentages in each condition state for 2020 show
315 the same distribution. The same was observed when repair is added in the analysis. However,
316 when repair is incorporated in analysis, the failed condition percentage is reduced by almost 10
317 percent. This reduction indicates the importance of considering M&R programs to maintain a
318 certain level within budget constraints.
319
320 CONCLUSIONS
321
322 The overall objective of this paper is to develop a Markov Chain-based methodological
323 framework to characterize pavement performance in support of pavement management decision

TRB 2014 Annual Meeting Original paper submittal - not revised by author
 J.D. Porras, Z. Zhang & L. Loria 13

324 makings. The methodology proposed was successfully utilized for the case of study. Major
325 conclusions drawn from this study include:
326
327 • The dynamic nature of the proposed TPMs can be effectively used for pavement
328 deteriorations modeling. Uncertainties related to pavement performance can be taken into
329 consideration using the proposed TPMs, allowing pavement condition to be predicted as
330 the probability of being in one of the pre-defined condition states.
331 • Optimization can be used to minimize the errors associated with TMPs. By comparing
332 the distribution of pavements condition from the data set and that obtained using the
333 generated TPMs, the errors can be significantly reduced by using the proposed
334 optimization techniques. As illustrated in the example, the TPMs closely follow the data
335 set after the optimization is performed, resulting in more reliable TPMs and, in turn, more
336 accurate prediction of pavement performance.
337 • The proposed framework, as demonstrated by the case study, can be applied to a wide
338 range of conditions by various highway agencies. It provides a relative easy methodology
339 for pavement deterioration modeling that could enhance the decision-making process in
340 highway agencies with limited pavement data, especially at the network level.
341

TRB 2014 Annual Meeting Original paper submittal - not revised by author
 J.D. Porras, Z. Zhang & L. Loria 14

342 REFERENCES
343
344 1. Saba, R. (2011). Pavement Performance Models (NordFoU - PPM). Nordic Cooperation
345 Program. Available at:
346 http://www.nordfou.org/documents/pavement/Summary%20report%20v3.pdf
347
348 2. Zheng, L. (2005). A probabilistic and adaptive approach to modeling performance of
349 pavement infrastructure. Dissertation from the University of Texas Austin. Austin, TX.
350
351 3. Golabi, K. & Pereira, P. (2003). Innovative Pavement Management and Planning System for
352 Road Network of Portugal. ASCE Journal Infrastructure Systems, 9(2), 75–80.
353
354 4. Lytton, R. (1997). Concepts of Pavement Performance Prediction and Modeling. 2nd North
355 American Pavement Management Conference (1987)
356
357 5. Prozzi, J. & Madanat, S. (2004). Development of Pavement Performance Models by
358 Combining Experimental and Field Data. Journal of Infrastructure Systems. 10(1), 9-22.
359 March
360
361 6. Ortiz-Garcia J., Costello S, & Snaith, M. (2005). Derivation of transition probability matrices
362 for pavement deterioration modeling. ASCE Journal of Transportation Engineering, Vol.
363 132, No. 2. .Pg.141-161
364
365 7. Haas, R., Hudson, W.R. & Zaniewski, J. (1994). Modern pavement management systems.
366 Krieger publishing company, USA. 583 p.
367
368 8. Wang, K., Zaniewski, J., & Way, G. (1994). Probabilistic Behavior of Pavements. ASCE
369 Journal of Transportation Engineering., 120(3), Pg. 358–375.
370
371 9. Kim, S & Kim, N. (2006). Development of performance prediction models in flexible
372 pavement using regression analysis method. KSCE Journal of Civil Engineering, Vol. 10, No.
373 2, Pg. 91-96.
374
375 10. Ningyuan, L., Haas, R. & Xie, W. (1997). Investigation of Relationship between
376 deterministic and probabilistic prediction models in Pavement Management. Transportation
377 Research Record 1594, Pg. 70-79.
378
379 11. Madanat, S., Mishalani, R., & Ibrahim, W. (1995). ”Estimation of Infrastructure Transition
380 Probabilities from Condition Rating Data.” Journal Infrastructure Systems, 1(2), 120–125.
381
382 12. Mishalani, R. & Madanat, S. (2002). Computation of Infrastructure Transition Probabilities
383 Using Stochastic Duration Models. Journal Infrastructure Systems, 8(4), 139–148.
384
385 13. Do, M., Kobayashi K., & Han D. (2010). Estimation of Markovian Transition Probabilities
386 for Pavement Deterioration Forecasting. KSCE Journal of Civil Engineering, 2010 14(3).
387 Pg.343-351.

TRB 2014 Annual Meeting Original paper submittal - not revised by author
 J.D. Porras, Z. Zhang & L. Loria 15

388
389 14. Fylstra, D., Lasdon, L., Watson, J., & Warren, A. (1998). Design and use of the Microsoft
390 Excel Solver. Interfaces, Vol. 28, No. 5, Pg. 29-55. Available at:
391 http://www.utexas.edu/courses/lasdon/design3.htm
392
393 15. Badilla, G., Elizondo, F. & Barrantes, R. (2008). Determinación de un procedimiento de
394 ensayo para el cálculo del IRI. Unidad de Investigación. Proyecto N UI-03-08. Available at:
395 http://www.lanamme.ucr.ac.cr/templates/university/images/publicaciones/ui-03-08.pdf
396

TRB 2014 Annual Meeting Original paper submittal - not revised by author