Computer-Aided Civil and Infrastructure Engineering 16 (2001) 143–157

REVIEW ARTICLE

Artificial Intelligence−Based Decision Support

Technologies in Pavement Management
Sten Sundin* & Corinne Braban-Ledoux
School of Transportation and Society, Dalarna University, S-781 88 Borlänge, Sweden

Abstract: The field of road pavement engineering has because of the complexity and magnitude of operations,
seen an explosion of artificial intelligence–based applica- automation, and the chain reaction that an error can cause
tions since the late 1980s. Such applications are found in many parts of an organization. Also, the decisions must
at key stages of the decision process involved in pave- be made quickly, and the information necessary to make
ment management. In the analysis phase, they perform decisions may be difficult to access. Therefore, it is dif-
pavement diagnosis and deterioration modeling tasks. In ficult to rely on the trial-and-error approach to pavement
the design phase, they enable rehabilitation needs to be management. To support effective decision making, man-
assessed and contribute to the identification and selec- agers now must learn to use new tools and techniques that
tion of maintenance actions. Finally, in the choice phase, can assist them directly in their most important task: mak-
they are applied for priority programming of rehabilita- ing decisions.
tion and maintenance. They either represent alternative In the field of pavement engineering, a great deal of the
approaches to existing systems or collaborate to make the decision making depends on the judgment and experience
overall system more efficient. The purpose of this article of technologists and engineers. Since many experienced
is to summarize the findings of up-to-date research articles people are nearing retirement, it becomes fundamentally
concerning the application of artificial intelligence to pave- necessary to capture as much as possible of the knowledge
ment management and to illustrate the potential such tools and experience currently existing for the benefit of inexpe-
can offer to pavement engineers. Artificial intelligence tech- rienced people entering the area, contractors, road authori-
niques include expert systems, artificial neural networks, ties, and consultants.
fuzzy logic, genetic algorithms, and hybrid systems. New technologies applicable to pavement management
and maintenance are continually evolving. It is crucial
for people active in this field to continually survey and
1 INTRODUCTION
study areas of new technology for possible application.
Expert systems are the most common technology used to
An important function of pavement management is plan-
substitute for human expertise. They are of great interest
ning that involves a series of decisions. What should be
to organizations because of their ability to enhance pro-
done? When? How? Where? Decision making is the pro-
ductivity and to augment workforces in many specialist
cess of choosing among alternative courses of action for
areas where human experts are becoming increasingly dif-
the purpose of attaining a goal or a set of goals. For years,
ficult to find and retain. Artificial neural networks represent
managers have considered decision making a pure art—a
another example of such promising new technology that
talent acquired over a long period of time through experi-
has attracted attention in many engineering disciplines. This
ence (learning by trial and error). Decision making today
information-processing technology originates from a desire
is more difficult and complicated than in the past for sev-
to simulate how the human brain processes information.
eral reasons. The number of available alternatives is much
The neural networks can learn from experience, generalize
larger, and the cost of making errors also can be very large
from previous examples to new ones, and abstract essential
* To whom correspondence should be addressed, E-mail: characteristics from inputs containing relevant data. They
[email protected]. offer many advantages over traditional methods of model-

© 2001 Computer-Aided Civil and Infrastructure Engineering. Published by Blackwell Publishers, 350 Main Street, Malden, MA 02148, USA,
and 108 Cowley Road, Oxford OX4 1JF, UK.
144 Sundin & Braban-Ledoux

ing in situations where the process to be modeled is com- every computerized system used to support decision mak-
plex to the extent that it cannot easily be represented explic- ing in an organization.
itly in mathematical or statistical terms. There are, how-
ever, several other technologies in the artificial intelligence 2.2 Primary decision support technologies
field, such as fuzzy logic and genetic algorithms, that can Decision support can be provided by one or more deci-
be used to support special situations. sion support technologies (tools). According to Turban,45
The motivation of this study is to review and analyze the the major decision technologies are
potential of using artificial intelligence (AI) in pavement
management and maintenance and to help researchers iden- • Group support systems (GSSs), including group DSS
tify the opportunities where this new technology is appli- (GDSS)
cable. What problems have been attacked? What kinds of • Executive information systems (EISs)
artificial technologies have been used? What are the main • Expert systems (ESs)
achievements to date? What are the obstacles to further • Artificial neural networks (ANNs)
progress? • Hybrid support systems (HSSs)
This article is divided into five major parts: Section 2 A GDSS is an interactive computer-based system that
introduces the different technologies in decision support. facilitates the solution of unstructured problems by a group
Section 3 consists of a short introduction to some of the of decision makers working together. An EIS is a com-
AI technologies that can provide decision support for man- puterized system that is designed to support executive
agerial decision makers. Section 4 sets the background work by presenting key performance indicators. ESs, now
of the data analysis part of a pavement management sys- often called knowledge-based systems, and ANNs are more
tem and provides a brief understanding of the decision explained in Section 3. In an HSS, several techniques work
process involved in pavement management. The remainder either in a collaborative manner or in a competitive way to
of the article reviews the most recent and relevant appli- produce better results.
cations involving AI techniques in pavement management, The decision process involves three major phases:
including expert systems, artificial neural networks, fuzzy analysis, design, and choice. The decision technologies are
logic, genetic algorithms, and a combination of any of related to different phases in the decision process. In the
these techniques, namely, hybrid systems. Finally, Section 6 analysis phase, reality is examined, and the problem that
concludes the article. calls for decisions is identified and defined. In this phase,
EISs and ANNs are helpful by continuously monitoring
both external and internal information, looking for early
2 DECISION SUPPORT TECHNOLOGIES symptoms on problems. The design phase involves gener-
ating, developing, and analyzing possible courses of action.
2.1 The concept of decision support systems A model of the problem is constructed, tested, and vali-
Decisions can be made on different types of problems dated, and criteria are set for the evaluation of possible
ranging from structured (sometimes called programmed) courses of action. The consequences of using various alter-
to highly unstructured (nonprogrammed). In a structured natives are forecasted. The generation of alternatives for
problem, the decisions can be supported completely by complex problems requires expertise that could be provided
algorithms. In an unstructured problem, human intuition by a human expert or an ES. The choice phase includes
often serves as the basis for decision making. There search, evaluation, and recommending an appropriate solu-
are also semistructured problems having some structured tion to the model. Here an ES can be used to assess the
features and some unstructured characteristics. Solving desirability of certain solutions as well as to recommend
them involves a combination of both standard solution the final solution. If a group makes decisions, a GDSS can
procedures and human judgment. Structured and some help in this choice phase and for brainstorming activities
semistructured decisions have been supported by computers in the design phase. Once the proposed solution seems rea-
since the 1960s. Less structured problems can be handled sonable, it is finally ready for implementation.
by a decision support system (DSS) that includes modeling
2.3 Models for decision support
capabilities.
There is no universally accepted definition of DSS, but DSSs rely on models. The basic idea is to perform the anal-
a classic definition15 refers to “a system that couples the ysis on a model of reality rather than on the reality itself.
intellectual resources of individuals with the capabilities Also, experimentation by manipulating the model is easier
of the computer to improve the quality of decisions.” It than manipulating the real system. According to Bemmel,46
is a computer-based support system for management deci- decision support models can be grouped into different cat-
sion makers who deal with semistructured problems. DSS egories (Figure 1). The main categories are the quantitative
is sometimes used as an umbrella term to describe any or and the qualitative decision support models.
 Artificial intelligence–based decision support technologies in pavement management 145

concerned with studying the thought processes of humans

and deals with representing those processes via machines,
as described in the following sections. AI provides the
scientific foundation for several growing commercial tech-
nologies, and AI-based decision support is provided by a
variety of applied AI tools such as expert systems, artificial
neural networks, fuzzy logic, and genetic algorithms.
Fig. 1. Decision-support models can be grouped into different
categories.
3.1 Expert systems

The quantitative or mathematical category is often based Expert systems (ESs) are computerized advisory programs
on well-defined statistical methods or artificial neural net- that imitate the reasoning process and decision-making
works and makes use of training sets of data. Features skills of human experts when solving specific problems in
are selected primarily by using statistical methods and by some narrow domain. They are now in operation and more
assessing their discriminatory powers. Statistical methods used than other applied AI technologies that are still at the
are often used to test the probability of the occurrence research level.
of some event. In decision making based on statistics or An ES, also called a knowledge-based system, includes a
neural networks, all features (signs, symptoms, measure- knowledge base and an inference mechanism. The knowl-
ments, etc.) generally are used simultaneously. It should be edge base contains permanent facts (as well as tempo-
noted that statistical methods, too, still assume an underly- rary facts depending on the implementation) and rules for
ing qualitative model of features and classifications. making inferences from these facts and other types of
The qualitative methods may be inspired or are the per- knowledge (that are usually used by an expert to derive
ceptions of human reasoning and use features that generally a conclusion) about the specific problem. The inference
are proposed by expert systems. Expert systems are used engine, often called the rule interpreter or reasoning struc-
to represent the knowledge and emulate the reasoning pro- ture, is a program that executes the inference-making pro-
cess of a human expert in a specific problem domain. They cess and draws conclusions from the knowledge base. An
can be used to solve problems that do not have algorith- ES has a software interface through which the user formu-
mic solutions. The decision support methods in the quali- lates queries and supplies other needed information. The
tative category generally use symbolic reasoning methods, system also can explain the reasoning process employed to
such as logical deduction, which may be best understood arrive at a specific conclusion.
in terms of Boolean logic (also called combinatorial logic
or symbolic logic). 3.2 Artificial neural networks
Artificial neural networks (ANNs) are a type of AI-based
3 DECISION SUPPORT SYSTEMS AND THE modeling technique inspired by studies of the brain and
ARTIFICIAL INTELLIGENCE FIELD the nervous system. ANNs work by processing information
between a large number of connected processing elements,
Decision makers rely primarily on their knowledge when the computer equivalent of neurons. Typically, ANNs are
making decisions. Acquiring this knowledge may take used to approximate some complex input-output relation
many years. As the amount of knowledge and complexity and have proved to be powerful tools for function approxi-
increase, the accessibility of knowledge becomes more dif- mation. They are usually fed with a large amount of data to
ficult. A knowledge-based system can enhance the capabil- approximate the underlying relationship. The organization
ities of decision makers by providing them with knowledge of the neurons and weights of the connections determine
together with expertise in data management and model- the output in response to an external input stimulus.
ing. Expert systems are the most common technology used There are multitudes of different types of ANNs. Some
to substitute for human expertise. The purpose is not to of the more popular include the multilayer perceptron,
replace the experts but to make their knowledge and expe- which is generally trained with the backpropagation of error
rience more available. However, several other technologies algorithm; the radial basis function, and the Hopfield and
presented here also can be used to support decision sit- Kohonen models, to name a few.
uations. All these are considered applications of artificial ANNs are particularly effective for pattern matching
intelligence (AI). when the networks have a large database of prior examples
Artificial intelligence is a term that encompasses to draw on. Strictly speaking, a neural network implies a
many definitions. Basically, AI is a unique approach to nondigital computer, but neural networks can be simulated
programming computers. Most experts agree that AI is on digital computers.
146 Sundin & Braban-Ledoux

The ANN approach has proved to be useful in certain overall budget constraints. This level can be further divided
areas that involve recognizing complex patterns, such as into a project-selection level and a program level. The
robotics, data analysis, diagnosing, forecasting, image pro- project-selection level focuses on a particular location and
cessing, pattern recognition, and voice recognition. involves prioritization to identify which projects should be
carried out in each year of the program period. At the pro-
3.3 Fuzzy logic gram level, budgets are established and general allocations
made over the entire network. The project-level work then
Fuzzy logic extends the notions of logic beyond simple
follows in time at the appropriate time in the schedule and
true and false values to allow for partial (or even continu-
represents the actual physical implementation of network
ous) truths. Inexact knowledge and imprecise reasoning are
decisions.
important aspects of expertise in applying common sense
In the late 1960s and early 1970s, the term pavement-
to decision-making situations. Fuzzy logic is based on the
management system (PMS) began being used to describe
concept of the fuzzy set, where degrees of membership in
the entire range of activities involved in providing pave-
fuzzy sets are expressed as a continuum of probability val-
ments. According to Haas,18 a PMS consists of a set of
ues ranging from 0 to 1. Fuzzy logic expresses the extent tools or methods that assist decision makers in finding opti-
to which an individual pertains to a set. For example, under mal strategies for providing and maintaining pavements in a
classic logic, a theoretical tree x is a member of the set of serviceable condition over a given period of time. The func-
tall trees; in contrast, under fuzzy logic, x pertains partly tion of a PMS is to improve the efficiency of decision mak-
to the set of tall trees and can be described as fairly tall. ing, provide feedback on the consequences of decisions,
Fuzzy logic has proved to be particularly useful in ESs and facilitate the coordination of activities within the agency,
other AI applications. and ensure the consistency of decisions made at differ-
ent management levels within the same organization. This
3.4 Genetic algorithms involves comparing alternatives, making decisions, and see-
Genetic algorithms (GAs) are intelligent heuristic search ing that they are implemented in an efficient and economi-
methods. In the computer they use some of the known cal manner.
evolution mechanisms as key elements in their design and A PMS can be broken down into three basic components:
implementation. For a specific problem, the solution is rep- database (which includes inventory and condition data),
resented as a chromosome, which often contains 0s and 1s, data analysis, and system output. A critical component of
indicating the values of a vector of decision variables that a PMS is the data analysis part of the system. Most pave-
describes the alternative courses of action. For this string ment management software provides the following analyt-
of values, the objective value can be computed. A generic ical capabilities: pavement performance modeling, rehabil-
method starts with a randomly generated population of itation needs modeling, models to determine rehabilitation
solutions and randomly combines portions of chromosomes and maintenance alternatives, and priority models.
together to form new solutions with an occasional mutation.
The new solutions are tested for feasibility with the help of 4.1 Pavement performance modeling
some fitness function The best fitted ones from the previ- Most PMSs provide software for determination of cur-
ous and current generations are selected (natural selection) rent condition levels or prediction of future condition. The
to survive to reproduce. After several combinations or iter- most interesting condition of the pavement is the perfor-
ations, the best solution is usually found near an optimal mance, which is evaluated by studying the functional or
solution to the decision-making problem. GAs have been structural behavior of the pavement. Performance modeling
applied to many large-scale combinatorial (difficult) math- is one of the key elements of pavement management. In
ematical programming problems, for example, large-scale these models, the dependent variable of observed or mea-
scheduling problems. sured structural or functional deterioration is related to a set
of independent variables. The performance can be evalu-
4 DECISION SUPPORT IN PAVEMENT ated by periodic observations or measurements of the pave-
MANAGEMENT ment ride quality coupled with historical records of traffic.
Performance models also can be used to analyze pavement
Conventionally, the decision process in pavement manage- life-cycle costs and estimate the type and timing of pave-
ment has two basic levels: the network level and the project ment maintenance and rehabilitation needs.
level. Network-level management involves the evaluation of According to a classification suggested by Broten,5 two
all pavements under an agency’s jurisdiction. The primary broad categories of models can be considered:
purpose of the network-level management is the develop- • Deterministic models, which estimate the average value
ment of a priority program and schedule of work within of the dependent variable (such as the remaining life
 Artificial intelligence–based decision support technologies in pavement management 147

of a pavement or its level of distress). Most determin- After selecting the triggering criteria, the next step is to
istic models used in pavement management are based set up limits that will trigger specific repairs. The limits
on regression analysis. Many ANNs belong to this cate- defined by a road agency are set according to different fac-
gory, since they can be regarded as nonlinear regression tors such as the traffic level, functional classification, and
models. the available resources and policies of the agency. A deci-
• Probabilistic models, which estimate a range (or distribu- sion tree or a matrix is often used to set these limits, which
tion) of values for a dependent variable. Most probabilis- are then incorporated into the PMS.
tic models used in pavement management are based on
Markovian theory. This theory is founded on the assump- 4.3 Maintenance and rehabilitation strategies
tion that the probability that something will change from
one condition state to another only depends on its cur- The process used to select alternatives considered by an
rent state. For each given condition states, estimates are agency for rehabilitation and for maintenance, both pre-
developed to predict what percentage of the pavement ventive and corrective, usually represents current practice.
sections in that state will (1) stay in the same condition All the available rehabilitation and maintenance alterna-
or (2) move to another condition state. tives are identified by the agency. A procedure exists for
determining which ones are feasible for given situations
Performance models also can be classified as mechanis-
so that a priority programming process can be carried out.
tic, empirical or mechanistic-empirical depending on their
These strategies are designed by a relatively small number
formulation and whether mechanistic variables, such as
of engineering experts that use their knowledge, judgment,
stress, strain, or deflection, are used in the model. Empir-
and experience from which design and investment decisions
ical models are based on results of experiments or expe-
are derived. There is a real need to formalize this expertise
rience. Empirical-mechanistic models incorporate elements
and to make it available to a larger number of agencies.
of both approaches.
The selection of feasible rehabilitation alternatives from
The major difficulty in developing performance mod-
a set of available alternatives can range from a simple judg-
els and attempting to make model errors at a reason-
able model is in capturing the large array of explanatory ment to a selection defined in the pavement management
variables and their interactions. Most deterministic per- software. The PMS normally uses models to determine
formance models used in pavement management to date which alternatives would be appropriate to use. These
are based on classic regression techniques, generally using models are usually set up to determine and select appropri-
field data. These databases seldom meet the assumptions of ate repair alternatives based on factors such as pavement
least-squares regression, however. Limitations/complexities condition in the repair year, surface type, and traffic level.
of stochastic and correlated errors in the measurements The most common ways to accomplish these selections are
of explanatory variables are encountered when developing using strategy matrices or decision trees. Here, the driving
regression models of pavement performance.13 These can factors for selecting feasible repair types can be the type of
lead to ambiguous results or underestimated errors when distress and the overall condition index.
regression is performed. Some agencies may have many repair alternatives avail-
The effectiveness of regression analysis also depends on able, whereas others may have only a few alternatives that
the engineer’s ability to comprehend the equation form they use. When more than one repair alternative has been
a priori. What is desired is a technique adaptable to solving identified as feasible for repairing a given pavement seg-
the functional relationships that often feature nonlinearities ment, the PMS software also uses models to select the
for which many of the regression methods are not very well recommended alternative. At their most basic, these mod-
suited. The ANN is a nonlinear technique that provides an els can be defined to mimic past practice by selecting the
alternate approach for this purpose. treatment alternative that has been used historically by the
agency. However, a more sophisticated approach is to base
4.2 Rehabilitation needs modeling the selection on some type of benefit-cost comparison.

Analytical routines are used in the pavement-management

4.4 Priority programming of rehabilitation and
process to identify when a segment should be repaired.
maintenance
These models can be set up to trigger rehabilitation based
on any measure of condition an agency wants, such as One of the key components of pavement management is
roughness, condition index, structural adequacy, and sur- to compare investment alternatives at both network and
face friction. Within a PMS, multiple criteria are often used project levels within some funding or budget constraints.
(e.g., roughness, condition index, and structural adequacy Priority models are used to determine the priority in which
could all be used together) to identify when a pavement pavement segments will get funded when there are insuffi-
segment will require repair. cient funds available for all the triggered projects.
148 Sundin & Braban-Ledoux

These priority assessment methods may vary from simple

subjective ranking to complex optimization models. Every
priority programming method should be directed to answer
the questions of which sections should be rehabilitated or
maintained, how they can be rehabilitated or maintained,
and when they should be rehabilitated or maintained. If
the investment is to be truly optimized, all possible com-
binations of these three questions must be evaluated. Con-
sequently, a model that can deal with all these questions
simultaneously will not be simple, and therefore, the mod-
els of most agencies only deal with one or two of these
questions.
The result of the comparison should be a priority pro-
gram of pavement maintenance and rehabilitation actions.
The most cost-effective alternative for each project or main-
tenance section within a program is identified. The length
Fig. 2. Relating AI-based decision-support technologies to the
of a program period can range from a single year-by-year PMS decision-making process.
program to multiyear programs. A practical approach is
to develop a 5- or 10-year program, but considering the
uncertainties in the future years, only the first 2 or 3 years approach and (2) to provide the local engineering user with
are fixed, and then annual or biannual updates are carried an interactive analysis and design tool for development of
out. However, the priority programming does not make the pavement rehabilitation strategies.
decisions. It only presents a useful guide for the engineer DESIGNER is a knowledge-based expert system for
to follow. assisting highway engineers in selecting recommended
highway pavement design features.23 The knowledge base
is built from the Catalog of Recommended Pavement
5 APPLICATIONS OF ARTIFICIAL
Design Features, which collates good-practice recommen-
INTELLIGENCE TO PAVEMENT
dations for design features of flexible and rigid pavements
MANAGEMENT
for highway engineers and pavement researchers. The core
part of the prototype is programmed using the CLIPS
A total of 37 most recent and relevant publications avail-
expert system shell (C Language Integrated Production
able in the highway pavement literature have been reviewed
System). Based on input information on traffic loading,
and presented hereafter. Articles are classified with respect
subgrade support, and climate provided to the system by
to the AI techniques applied (knowledge-based expert sys-
the designer, design feature recommendations, including
tems, artificial neural networks, fuzzy logic, genetic algo-
pavement cross section, structural design, and materials
rithms, and a combination of any of these techniques,
required to meet minimum performance requirements for
namely, hybrid systems) and the specific problems they
both flexible and rigid pavements, are made available to the
tackle. Reviewed articles are presented according to the
designer. The system also includes a set of tools for evalu-
natural order in which the issues they address appear in
ating different pavement design alternatives.
the pavement-management decision process, as illustrated
OVERDRIVE is a rule-based expert system that helps
in Figure 2. The review first starts with applications involv-
local engineers design the structural thickness of flexi-
ing knowledge-based expert systems, since they have been
ble pavement overlays.38 The major tasks addressed by
the first AI technique to be applied to pavement manage-
OVERDRIVE include determining the effective thick-
ment. After this follows a description of applications devel-
ness of the existing pavement structure, determining a
oped on the basis of fuzzy logic and integrating artificial
new full-depth asphalt-concrete construction thickness, and
neural networks and genetic algorithms. Finally, a review
assessing the consequent need for an overlay. The knowl-
of hybrid systems concludes the article.
edge base of OVERDRIVE is the result of knowledge
engineering efforts with a pavement specialist combined
5.1 Knowledge-based expert systems
with a synthesis of state-of-the-art and other reports, papers,
Historically, knowledge-based expert systems have been the and manuals related to the Asphalt Institute overlay design
first AI technique to be applied to pavement management. method for asphalt-concrete overlays on flexible pave-
Expert systems are powerful tools able (1) to embody the ments. It contains more than 100 rules, expressed using the
knowledge, experience, and judgment of expert pavement knowledge engineering shell EXSYS. The rules are fired
engineers by first identifying a main feasible rehabilitation according to the forward-chaining inference method.
 Artificial intelligence–based decision support technologies in pavement management 149

Completed and ongoing research at the University of from the database, the decision as to whether to rehabili-
California at Irvine has resulted in the development of a tate the section can be made. The user is only asked about
microcomputer-based integrated set of interacting expert judgmental factors not available from the database, such as
systems and algorithmic models known as PARADIGM38 geometric design requirements, administrative information,
(Pavement Rehabilitation Analysis and Design Mentor). pavement history, etc. The final recommendation is saved
PARADIGM consists of two main component systems: in the pavement database to be used later by the PMS. The
OVERDRIVE and SCEPTRE. SCEPTRE (Surface Condi- system can further provide a detailed description of how
tion Expert for Pavement Rehabilitation) evaluates project- to best apply the recommended rehabilitation and mainte-
level pavement surface distress and other user inputs and nance strategy for the particular pavement section.
assists highway engineers in planning and developing cost- An expert system developed to assist local highway
effective pavement rehabilitation strategies. SCEPTRE 1.4, agencies in developing countries in managing pavement
described by Ritchie and colleagues,39 addresses state- maintenance is described by Kotb and colleagues.28 The
maintained flexible pavements and reflects practices and kernel of the system comprises an interactive algorithmic
conditions found primarily in Washington State. The program and an expert system. The algorithmic program
knowledge base has been constructed from the combined simulates the evaluation phase and creates a complete diag-
expertise of two pavement specialists with extensive experi- nostic report of the pavement condition using the pave-
ence in pavement rehabilitation in the states of Washington ment condition index (PCI) procedure. The rule-based
and Texas. It contains about 140 production rules expressed expert system determines and ranks different maintenance
using the knowledge engineering shell EXSYS. The infer- alternatives. The system differentiates between four major
ence mechanism uses the backward-chaining inference maintenance and rehabilitation strategies (M&R), namely,
method. Ten basic rehabilitation strategies are considered: do nothing, recurrent maintenance (10 different activities),
do nothing, fill cracks, fog seal, friction course, chip seal, periodic maintenance (thin or thick overlay), and strength-
double chip seal, thin asphalt concrete overlay, medium ening (reconstruction). The system also includes a cost-
asphalt concrete overlay, thick asphalt concrete overlay, estimation module that calculates the total cost of the
mill, and replace. They can be modified to include a rec- assigned M&R alternative. This component simulates a
ommendation to first “prelevel” or “prelevel or mill” the detailed industrial cost-estimation procedure to determine
pavement as part of the rehabilitation strategy, resulting in total required budget and a long-term economic analysis of
a total of 24 rehabilitation strategies. The selection of fea- the M&R investments. The whole system runs as follows:
sible strategies is keyed to the following six basic factors: The user supplies distress data from a condition survey
type of surface distress, amount of surface distress, severity sheet, and the program computes the PCI. The expert sub-
of surface distress, existing pavement performance, traffic system then inquires about highway characteristics (traffic
levels, and climate. The specific types of surface distress volume, traveling speed, etc.) and condition history (pre-
considered comprise alligator cracking in wheel paths, rut- vious PCI evaluation, rate of previous maintenance, any
ting, longitudinal cracking in wheel paths, and transverse previous strengthening, and the age since strengthening
cracking. SCEPTRE is coupled with an external module or reconstruction). The report from this diagnostic phase
built for cost-effectiveness analysis of feasible rehabilita- includes the following information: traffic volume, section’s
tion strategies. Underlying the analysis for each strategy PCI, short- and long-term rates of deterioration, level of
is a pavement-performance curve, which relates the age of previous maintenance, skid resistance, list of applicable
the pavement to an overall pavement condition rating. This M&R activities based on local circumstances, unit cost and
functionality enables a visual display of expected pavement projected service life for each activity, and existing distress
deterioration for each rehabilitation strategy and serves as types/severity. Based on previous information, the expert
the basis for cost-effectiveness analysis. subsystem determines the appropriate M&R activities.
Lee and colleagues32 describe an expert system being The project PADMA29 (Pavement Deterioration and
developed for the Huntington Town Highway Depart- Maintenance) has been developed very recently in India to
ment in New York. The system supports engineers in the evaluate the feasibility of ES application in pavement man-
decision-making process involved at the project-selection agement. The rule-based system is used for estimating the
level. Given the present pavement conditions, the most causes for the deterioration together with the most appro-
appropriate rehabilitation or maintenance strategy is recom- priate treatment to be carried out. It also provides infor-
mended by the system after all other subjective factors have mation on the appropriate material type and thickness of
been considered. The interface program searches through overlays required for rehabilitation actions considering the
the pavement database to locate the pavement section life-cycle costs of different overlays. Twenty types of symp-
requested by the user and returns the condition survey data toms of defects in bituminous surfacing are grouped under
to the knowledge base. Using distress information obtained four categories: surface defects, cracks, deformation, and
150 Sundin & Braban-Ledoux

disintegration according to the practice followed for main- an alternative way for expert systems to capitalize on the
tenance of roads in India. To specify a cause of a failure or expertise from pavement managers. They are used to serve
distress, the user has to input 30 different facts about the the same purpose, i.e., to offer decision support to decision
present surface condition, material characteristics, and con- makers, but as will be detailed further in the remainder of
struction practices to the system. The information required this article, their approach is different from the expert sys-
for finding the cause of deterioration is acquired from the tems. ANNs are data-driven models, whereas expert sys-
PMS database or from the knowledge of experts. The pro- tems are rule-driven models. ANNs are used for estimating
gram has been developed using the expert system shell, current pavement condition, predicting their future deterio-
DEKBASE, elaborated at the Department of Civil Engi- ration, and finally assisting engineers in selecting “optimal”
neering at the Indian Institute of Technology, in Madras, maintenance or rehabilitation actions.
India. The expert system model has been calibrated using
field data from 10 pavement sections. The authors conclude 5.3.1 Estimating current pavement condition. Pavement
the article on the effectiveness of the system to identify the condition rating is a key process in the estimation of cur-
type of distress and the type of treatment to be carried out. rent pavement condition. It consists of assigning relative
weights to various levels of pavement distresses in order to
5.2 Fuzzy set theory obtain a combined score that indicates the current condi-
Grivas and colleagues16 investigate the combination of tion of a roadway. Pavement condition indices describe the
fuzzy theory and knowledge graphs as a decision sup- pavement’s functional and structural performance expressed
port tool to help pavement engineers handle uncertain in terms of ride quality, safety, surface distress, and struc-
information properly in order to select the most appropriate tural adequacy. Mathematical analytical models tradition-
maintenance/rehabilitation strategies. Knowledge graphs ally are used to determine pavement condition indices.
are used to represent expert knowledge on the relation- ANNs are applied either to output an estimation of such
ships between symptoms of bad road condition and pos- aggregated indices 2,8,47 or to produce detailed information
sible treatments. The weights of the connections between necessary to the computation of the indices.25 This sec-
the symptom nodes and the treatment nodes of the graph tion, focusing on estimating current pavement condition,
express the strength of belief for the possible treatments. also reports comparative studies that have been carried out
The relationships established in the knowledge graphs for to evaluate the performances of ANNs against other more
condition diagnosis and treatment identification can be conventional techniques.3,13
expressed as a conditional statement involving fuzzy rela- Eldin and colleagues8 present a feed-forward ANN used
tion. Fuzzy set theory helps establish these relationships. for condition rating of flexible pavements. The purpose is to
Potentials of the proposed method are shown on a practical build an ANN-based pavement condition-rating model that
problem. The load-carrying capacity of a highway pave- provides the same outputs as the condition-rating scheme
ment structure has been affected mainly by the occurrence established by the Oregon State Department of Transporta-
of cracking, rutting, and spalling. The purpose of the case tion (ODOT). The pavement condition-rating model devel-
study is to identify the treatment options based on rut depth oped at ODOT consists of a long, stepwise calculating
data collected from a roughness survey and based on engi- scheme based on cracking and rutting indices on alliga-
neering judgment about the preservation of pavement struc- tor, transverse cracks, block cracks, patching, bleeding, and
tures. Three knowledge graphs are drawn. The simplest one rutting distresses. ANNs are then regarded as a simpler
represents a direct symptom-treatment relationship. The alternative way to perform pavement condition rating. Input
second one corresponds to a symptom-condition-treatment information to the neural network consists of the severity
relation. It is used to represent relationships between fac- of each aforementioned distress types. A set of 744 and
tors occurring as a consequence of the symptoms. The third 1736 randomly generated patterns is used for training and
graph is a complex symptom-condition-condition-treatment validating the network. The performances of the neural net-
relationship. For each graph, the degree of truth of a spe- work are tested against different levels of noise introduced
cific action is derived from the computation of the weights to the output unit. Neural network performances are fur-
of all connections. ther assessed by comparing the neural network estimations
with expert opinion and outputs from the ODOT’s model.
5.3 Artificial neural networks Results reported by ODOT indicate that on the same por-
The literature review has revealed the great deal of attention tions of two interstate highways, the neural network outper-
that has been paid on artificial network networks (ANNs) forms the ODOT’s model in estimating subjective condition
and their application to pavement management systems ratings provided by the maintenance personnel.
(PMSs) since the beginning of the 1990s. ANNs cover most The index used by Van der Gryp and colleagues47 to esti-
of the tasks involved in PMSs. They can be considered mate the overall pavement condition is the visual condition
 Artificial intelligence–based decision support technologies in pavement management 151

index (VCI). Although VCI is a well-defined analytical cri- the basis for pavement design practice during the last 30
terion, it contains some weighting factors whose values are years and is based on a factorial test of pavement dura-
determined according to the subjective appreciation of an bility performed from 1958 through 1960 known as the
expert panel based on the importance of one distress type AASHO road test. It consists of a parametric model that
compared with the other. A one-hidden-layer feed-forward relates the surface thickness, base thickness, subbase thick-
network is trained to estimate VCI values, based on the ness, axle load, and accumulated axle load to the PSI. The
severity and extent of various types of distress, includ- AASHO road test offers an opportunity to compare the
ing failure, surface cracks, longitudinal cracks, transverse ANN models with well-established practice-tested results.
cracks, rutting, patching, potholes, bleeding, and pumping. In the reported study, the AASHO road test is reexamined
Reported simulations make it difficult to conclude on the using a Monte Carlo hierarchical adaptive random parti-
effectiveness of the ANN. tioning (MC-HARP) algorithm. This algorithm can be clas-
Attoh-Okine2 evaluates the capabilities of ANNs for esti- sified as an ANN model that combines the best features of
mating the International Roughness Index (IRI) progression neural networks with traditional statistics tools. The com-
in flexible pavements based on structural deformations, parative study stresses the better performance of the ANN.
surface defects, and environmental and non-traffic-related It also shows that the MC-HARP approach, which is a local
mechanisms. Synthetic data are generated for training approximation model, results in better models of service-
a feed-forward neural network with the backpropagation ability trends over the entire input domain than a global
algorithm. A fairly good relationship between actual and approximation provided by the AASHO formula.
estimated roughness is reported despite problems with pre- Evaluating pavement surface condition is still largely a
processing input data and selecting an appropriate set of manual process involving visual inspection by field per-
learning rules and transfer functions. Additional work is sonnel. This is labor-intensive, tedious, and often danger-
needed to identify which pavement condition variables ous, as well as a subjective process resulting in (1) a
should be used to characterize roughness progression. high degree of variability and systematic error among per-
Although the author states that he uses ANNs for prediction sonnel, (2) difficulties in transferring expertise, and (3)
purposes, he actually performs some estimation because the inconsistent evaluation. To overcome these problems, more
time factor is not involved in the model. automated processes are needed. The following reviewed
Significant improvement in modeling pavement perfor- articles present some results on the feasibility of new
mance by use of ANNs compared with classic regression ANN-based approaches to the automated detection, classi-
is attained in a study to estimate the pavement condition fication, and quantification of pavement surface cracking
rating (PCR) index.13 The PCR index rates the pavement based on image-processing techniques. The methodology
condition on an ordinary scale of 0 (worst condition) to 100 proposed by Kaseko and Ritchie24−27 is purely image-
(best condition) according to the type, severity, and extent processing-oriented. The pavement images used in the
of distresses as well as the roughness of the pavement. All research are a sample of the images acquired by the firm
pavement data, including the PCR of pavement sections PASCO USA, Inc., for the U.S Strategic Highway Research
and factors affecting PCR (pavement structure, pavement Program. PASCO collected the pavement images using the
age, and traffic volumes), are extracted from the Mississippi ROADRECON instrumentation vehicle. Images extracted
Department of Transportation PMS database. Over 300 from the whole database are digitized to 512 × 464 pixel
field data observations are made available for three differ- digital images with an 8-bit gray scale. Multilayer percep-
ent families of pavements (flexible, composite, and jointed tron (MLP) neural networks, trained with the backpropa-
concrete pavements). Potentials of classic regression mod- gation algorithm, are used for segmenting the images and
els, Bayesian regression models, and ANNs for estimating for classifying the images. Image segmentation consists
the PCR attribute are investigated. Reported results show of processing the images to increase the contrast between
that the ANN outperforms the other two models in terms the objects of interest and the background, extracting the
of higher correlation coefficient. By using Bayesian regres- objects, and then converting the gray-scale image into a
sion, a definitive improvement is made over classic regres- binary image. As far as classification of images is con-
sion models. In the Bayesian approach, the field data are cerned, images are classified into four categories according
supplemented with judgment and expertise collected from to the nature of cracks: transverse, longitudinal, alligator,
knowledgeable experts. and block cracking. An overall performance of 92 percent
Banan and colleagues3 report the results from a is reached on the test set. A comparison of traditional
comparative study on the effectiveness of ANNs and and neural classifiers for the classification of subimages
the American Association of State Highway Officials is also carried out. Three different models are considered:
(AASHO) formula to estimate the pavement durability of a Bayesian classifier, a k-nearest-neighbor classifier, and an
road sections expressed in terms of the present service- MLP. Results24 demonstrate the better performance of the
ability index (PSI). The AASHO formula has served as neural network.
152 Sundin & Braban-Ledoux

On the same original pavement images as used by the AC back-calculated elastic modulus value, the unbound-
Kaseko and colleagues, Nallamothu and colleagues33 layer thickness, the unbound-layer back-calculated elastic
apply radial basis function neural networks for classify- modulus value, the subgrade back-calculated elastic mod-
ing pavement distress into four types: transverse cracks, ulus value, the annual number of days with a temperature
longitudinal cracks, raveling, and no crack. Each feature higher than 90◦ F, the freeze index, the annual precipita-
vector consists of an image that represents a distress of tion level, the average annual equivalent axle loads, the
compressed-size 24 × 9 pixels. On the recall of the test set, age of the pavement at the first IRI observation, and the
85 percent of patterns are classified correctly. Because of corresponding IRI measurement. The output layer contains
the differences in the preprocessing and feature-extraction four neurons; the first one represents the predicted value of
techniques applied, a direct comparison with the perfor- the IRI criterion 1 year into the future, the second repre-
mances reached by Kaseko and colleagues is not possible. sents the predicted value of the IRI criterion 2 years into
the future, etc. Data from 50 road sections are used for
5.3.2 Predicting future pavement condition. Being able to training and validating the ANNs. The comparison with
predict pavement performance and/or key pavement char- the predicted and actual IRI values is considered satisfac-
acteristics such as roughness from the smallest set of actual tory with a root mean squared error of 0.11 mm/m over
measured characteristics is a crucial issue for pavement the 144 observations used. Long-term-prediction issues are
management because it can yield large savings in data col- also addressed. To allow the ANN to estimate the IRI over
lection. The value of having a model that can predict future a time horizon longer than 4 years, the last predicted IRI
value is taken as the “new current IRI” and the “new age
deterioration accurately is also high in the optimization
of the pavement” is increased by another 4 years. Overall
of maintenance and capital improvement strategies. This
short- and long-term performances on the training set look
naturally raises the issue of the accuracy of the predic-
promising. Nevertheless, it needs to be further investigated
tions, since planning optimal rehabilitation actions strongly
on the validation set.
depends on forecasted road deterioration. Results from the
The PMS used by KDOT includes a project-level opti-
following literature review reveal that ANNs outperform
mization system that requires models for estimating the
multiple linear regression techniques, which are used tra-
probability that a given level of distress will occur at
ditionally in predicting pavement deterioration. ANNs pre-
some future time. Huang and colleagues22 use multiple lin-
dict either future values of IRI22,31,41 or the progression of
ear regression and two ANN models to predict roughness
some specific type of distress such as the rut depth.42
distress level probability for bituminous pavements. The
Roberts and colleagues41 investigate the use of a models are based on historical pavement condition data and
quadratic function ANN to predict IRI. The quadratic specific project-level data concerning pavement structural
function ANN is a generalized adaptive feed-forward neu- characteristics, traffic, and climatic conditions. Two sets
ral network that combines supervised and self-organizing of data representing 121 and 146 road sections are used.
learning. It develops its own network architecture using an The output variables are binary: one if the pavement exists
evolutionary mechanism. Both a conventional MLP trained within a given roughness level and zero if the pavement is
with the backpropagation algorithm and the quadratic in any other distress level. A backpropagation neural net-
function ANN predict IRI for three types of bituminous work with one hidden layer is employed, and 17 indepen-
pavements (composite, full-depth bituminous, and partial- dent variables are selected. The success rates for testing
design bituminous). They are built from 105 patterns drawn data range from 73 to 93 percent. The ANN models are
from the Kansas Department of Transportation’s (KDOT) found to be generally much better predictors of roughness
PMS. Each pattern consists of 11 variables of 6 types: distress level probability than traditional multiple regres-
rutting, fatigue cracking, transverse cracking, block crack- sions. However, the binary nature of the response variable
ing, equivalent axle loads, and the IRI. Fatigue cracking creates inherently a difficulty for regression analysis that
is further divided into 4 variables, and transverse cracking gives the comparison little relevance.
is divided into 3 variables. The self-organizing quadratic ANNs are also applied to predict the rut-depth change
function ANN performs significantly better than the MLP. in road pavements.42 A feed-forward network is built
Because of this improvement, the authors recommend based on 1776 observations collected from the Swedish
devising methods to investigate different combinations of road network. According to the Swedish National Road
genetic and other types of self-organizing algorithms on Administration, the key data to consider as inputs to pre-
ANN models. dict the rut-depth change are the type of pavement for the
La Torre and colleagues31 apply an MLP to predict the upper layer, the previous value of the rut depth, the pre-
IRI of a flexible pavement section for the following 4 years vious change in rut depth, climate condition, maintenance
from the year the analysis is carried out. Each input pat- information such as the type of last pavement action, the
tern consists of the asphalt concrete (AC) layer thickness, number of years since last maintenance action, and traffic
 Artificial intelligence–based decision support technologies in pavement management 153

volume. The prediction performances of the feed-forward of pavement maintenance treatments. Altogether, there are
neural network are compared with those of multiple regres- about 40 different variables and factors influencing R&S
sion models. It is concluded that the application of ANNs decisions. Past solutions to this problem have relied on
improves the quality of the predictions. rule-based systems, but the implementation of such systems
requires considerable development and programming effort.
5.3.3 Assessing maintenance needs and selecting mainte- ANNs are regarded as alternative solutions that require sub-
nance actions. A neural network has been used as a part stantially less time and efforts for development. For both
of an automatic procedure for preliminary screening and the ANN and the rule-based system, the recommending
recommending roadway sections for pavement preservation R&S treatment problem is formulated as the selection from
at the Arizona Department of Transportation10 (ADOT). a set of numbers 0, 1, 2, . . . , 10 that indicate the desir-
Automation of the selection process is expected to reduce ability of R&S. Definite rejection of R&S treatment is
subjectivity and minimize the probability of missing sec- indicated by 0, whereas 10 means that it is highly desir-
tions that should be programmed. Maintenance needs are able. The feed-forward neural network inputs the 40 dif-
identified according to the deterioration of the road sec- ferent variables and factors influencing R&S decisions and
tion (cracking and roughness severity over the past 3 years, is trained with the backpropagation algorithm. The rule-
rutting, and patching), the skid resistance, the structural based system is ROSE.18 Results show that for lower desir-
number, flushing, the maintenance costs, the daily traffic ability, between 0 and 5, there are substantial differences
volumes, and the “rate” index. The rate index is used by the between the two techniques, whereas for higher desirability
PMS and district engineers for preselecting the road sec- (between 6 and 10), the results are very similar. ANN tech-
tions for maintenance. It is computed as the weighted sum nology is seen as a powerful and efficient alternative tech-
of the cracking severity expressed in percentage, the rough- nique to rule-based systems. The ANN technique is easier
ness, the rutting, and the average maintenance cost for the and faster to implement, and only the ranking of sections
last 3 years. The feed-forward MLP is fed with additional with high desirabilities is important, because only these
input variables such as the road classification and the region sections are actually considered for R&S treatment and
the road section belongs to. The input vector to the MLP is only the highest desirabilities receive the treatment. After
coded or stratified by classifying each component in up to analyzing in-depth the advantages and drawbacks of each
a maximum of five ranges corresponding to the following technique with regard to the R&S problem, the authors con-
categories: low, medium, high, very high, and extremely clude on the benefits to actually combine both techniques.
high. The output layer of the ANN consists of one neu- Recently, genetic algorithms have emerged as a search
ron representing the decision to recommend the road sec- algorithm to design ANN models. Taha and colleagues43
tion for maintenance. If the output value is larger than 0.5, use a hybrid evolutionary-learning system using gradient-
then the section is recommended for maintenance; if it is descent learning as well as a genetic algorithm (GA) to
less than 0.5, the section is not recommended. The ANN is determine the network connection weights in a system
trained and tested on data from all projects programmed at for selecting the optimal maintenance strategy for flexible
ADOT from 1990 to 1996. The optimal number of hidden pavements. The factors affecting the maintenance strategy
units is set to 15. Reported results indicate that the ANN selection are identified as distress type, density of distress,
learns 100 percent of the training patterns and 76 percent of riding comfort index (RCI), traffic volume, climate, crack
the test patterns. These performances are considered accept- type, and severity of distress. The different pavement main-
able, especially if one considers that not all factors influenc- tenance strategies available are do nothing, crack seal coat-
ing the selection of the preservation projects can be traced ing, route and seal, cold-mix patching, hot-mix patching,
back to the PMS database. The next stage entails integrat- hot-mix recycled patching, and reconstruction. The ANN
ing the ANN into the PMS designed at ADOT.9 A simi- resulting from the combination of GAs and the backprop-
lar approach is applied by Algsugair and colleagues1 for agation misclassifies only 6 of 100 unseen cases with an
selecting maintenance actions on the Riyadh road network error rate of 0.024. The work of Taha and Hanna stresses
in Saudi Arabia. that a critical issue to the effectiveness of ANNs is the opti-
Hajek and colleagues18 present the results from an inter- mization of the global error function and also that ANN
esting comparison on the performances of a rule-based sys- performances can be improved by applying GAs.
tem and ANNs for selecting and recommending routing and
sealing (R&S) treatments. Recommendation of R&S treat-
5.4 Genetic algorithms
ments is a complex problem involving many input param-
eters. In addition to the amount and width of cracks, R&S Kwasi and colleagues30 explain in a tutorial way the gen-
decisions also depend on crack type, pavement serviceabil- eral concept of genetic algorithms (GAs) and describe
ity, pavement structure and age, presence of other pavement how they can be applied for estimating roughness progres-
distress (raveling, flushing, rutting, etc.), and the existence sion in flexible pavement. Available information consists
154 Sundin & Braban-Ledoux

of roughness index, equivalent standard axles, the age of Maintenance actions are ranked according to their urgency
the pavement system, the rut depth and the thickness of level (from the most to the least urgent action) and the num-
the asphalt layer, and the structural number of the over- ber of working days they require. Two interesting conclu-
all pavement system. Effects of crossover probabilities on sions can be drawn from the numerical simulations. First,
the estimation capability of the GA model are investigated. although each GA run produces a different set of solutions,
It is shown that high crossover probabilities around 0.95 because of the random initialization of the population, the
or higher (but less than 1) give the best fit for this spe- objective function values computed at the end of each run
cific problem. Despite promising results, the article puts are very close. It implies that the GA, for this specific prob-
too much emphasis on the fundamentals of GAs and not lem, produces near-optimal solutions. Second, none of the
enough on applying the technique itself to solve the rough- GA solutions reaches 100 percent of the available resources
ness estimation problem. expressed in terms of budget, personnel, and equipment.
The remaining reviewed articles apply GAs for finding It means that resource allocation is overestimated and that
optimal solutions to the complex problem of programming GAs are powerful tools for their optimization.
maintenance activities. The models described support pave- Fwa and colleagues11 go deeper into the analysis of
ment engineers in their decision on when to activate main- the fundamentals of GAs. The numerical examples used
tenance works and which type of maintenance work would to illustrate the applications of GA models program-
produce the maximum overall system effectiveness and how ming maintenance activities are discussed in-depth in early
long the work is scheduled for. publications.
Fwa and colleagues12 present in a pedagogical way the The problem tackled by Yuge and colleagues48 is to mini-
formulation of a GA model developed for programming mize the costs of pavement maintenance over the following
pavement maintenance and rehabilitation at the network 20 years. Available information includes rut depth, surface
level over a planning horizon of 20 years. The first half of smoothness, and cracking ratio aggregated into the mainte-
the article describes in-depth the nature and complexity of nance control index (MCI) and cumulative traffic flow vol-
the problem tackled, whereas the second half is dedicated ume. The article suffers from a lack of clarity and techni-
to building the GA model itself. General and fundamental cal details, and this prevents the reader from understanding
issues such as coding the problem parameters, generating how the authors actually build their model.
the initial parent pool of solutions, evaluating the parent
pool by means of an objective function, and generating 5.5 Hybrid systems
the offspring are largely discussed. To prevent the GA Chou and colleagues7 describe a complete image-
model from violating key constraints imposed by the spe- processing system for assessing road pavement condition.
cific problem definition and thus prevent it from generating They combine fuzzy set theory and ANNs. The following
invalid solutions, heuristics are introduced to the general five stages are dealt with: acquisition of pavement video
GA procedure. The proposed algorithm is tested on a sim- images, segmentation of images, feature extraction, classi-
ple hypothetical road network consisting of 30 three-km fication of images, and computation of crack severity and
pavement segments. The article concludes on the effective- extent. The image size for each frame is 482 × 512 pixels,
ness of the GA model. and the gray scale is from 0 to 255. An algorithm based on
Hoque and colleagues20 briefly explain the concept of fuzzy set theory and principles of maximum fuzzy entropy
the penalty method as a constraint-handling technique and is proposed for image segmentation. Feature extraction is
show how the genetic technique combined with the penalty the process of extracting features that provide a description
method solves a practical example problem. The basic of the objects of interest in the images. It is done using
feature of the penalty method lies in the conversion of a Hu moments,21 Bamieh moments,4 and Zernike moments.44
constrained problem to an unconstrained one. Candidate A feed-forward perceptron performs the classification of the
solutions are chosen randomly from the pool, and if any image distress into seven classes: longitudinal crack, trans-
solution violates any constraint, a penalty value is assigned verse crack, combined longitudinal and transverse crack,
to its evaluation value. The problem consists of planning right and left diagonal crack, alligator crack, and nondis-
maintenance activities for four highway types such as urban tress. Input data consist of the 18 moments previously
interstate, urban arterial, rural interstate, and rural primary computed. The hidden layer contains 17 units. A total
over the next 45 working days. The urgent need for mainte- of 126 longitudinal cracks, 162 combined longitudinal and
nance is categorized into three classes (high, medium, and transverse cracks, 78 transverse cracks, 126 right diagonal
low) and estimated according to the distress severity. Four cracks, 126 left diagonal cracks, 117 alligator cracks, and
possible routine maintenance activities (shallow patching, 42 nondistresses are considered for training and validating
deep patching, premix leveling, and sealing longitudinal the network. The competition winner-takes-all approach is
cracks and joints) are considered. For each road section, applied. Performances of 100 percent classification on the
the GA model provides a complete maintenance schedule. recall of the training and validation sets are mentioned.
 Artificial intelligence–based decision support technologies in pavement management 155

Ritchie and colleagues40 aim at developing an innova- clean and refill cracks, patch, reseal, full reconstruction, do
tive contactless intelligent system using a novel artificial nothing). The input layer contains six input neurons, repre-
intelligence–based approach for automatic processing of senting the type and extent of distress, and the classification
flexible pavement images acquired in real time. The system of road type. The effectiveness of the ANN is not clearly
integrates neural network and knowledge-based expert sys- demonstrated. This application is somehow disappointing.
tems in addition to conventional algorithmic and modeling It does not describe an active cooperation between the ANN
techniques. The ANN is used to determine the type, sever- and the knowledge-based expert system. The knowledge-
ity, and extent of distresses from digitized video image rep- based expert system acts more as a front-end interface for
resentations of the pavement surface. The knowledge-based the user rather than an AI system.
expert system takes inputs from the ANN, as well as other Prechaverakul36 describes only the theoretical back-
pertinent characteristics, and accesses historical records ground for using a knowledge-based expert system and
such as rates of deterioration, design details, and climatic fuzzy logic for minor rehabilitation projects in Ohio.
factors from a facility data base, to determine the structural A rule-based expert system is built to model the pavement
adequacy, remaining life, safety of the structure, and fea- engineer’s reasoning process when selecting proper treat-
sible rehabilitation and repair strategies. The knowledge- ments for rehabilitation of deteriorated road section. The
based expert system already exists in prototype form.37, 38 fuzzy logic module is elaborated for solving the multiob-
A three-layer perceptron is trained to classify 32 × 32 pixel jective decision-making problem that applies when decision
pavement images by the type of cracking: no distress, trans- makers must select only one rehabilitation scheme from a
verse cracking, longitudinal cracking, and a combination set of alternatives subjected to a set of constraints to be
distress. The 32 × 32 pixel images are extracted by simple satisfied.
thresholding from the original 512 × 512 pixel images pro-
vided by Roadman-PCES instrumented vehicle. The allo- 6 SUMMARY AND CONCLUSION
cation of these subimages to their actual class is performed
by manual inspection. A set of features is extracted from Many of the decision problems in pavement management
each subimage to form an input pattern to the ANN. The have a complex nature. They are often highly nonlinear
feature vector comprises the mean number of distressed with numerous and complex data sources. By introducing
pixels per row or column of the subimage, the variance of AI methods, the goal is to improve the quality and expand
the number of distressed pixels per row, and the variance of the boundaries of decision support in many different areas
the number of distressed pixels per column. A total of 256 in the field of pavement management. This article has pre-
and 4864 subimages are used, respectively, for training sented the fundamentals of different AI methods and iden-
and testing. The performances reported are rather impres- tified where this new technology is applicable.
sive: 100 percent correct classification for the training set, AI systems are found at key stages of the decision pro-
and 99 percent for nondistressed subimages, 93 percent cess involved in pavement management. In the analysis
for transversal cracking, and 96 percent for longitudinal phase, they perform pavement diagnosis and deterioration
cracking on the recall of the test set. For the combination modeling tasks. In the design phase, they enable rehabilita-
distress class, only 60 percent of the patterns are classi- tion needs to be assessed and contribute to the identification
fied correctly. Three explanations are provided: presence of and selection of maintenance actions. Finally, in the choice
noise in the subimages that makes it difficult to classify phase, they are applied for priority programming of rehabil-
some subimages, small number of combined patterns in the itation and maintenance. Either AI systems represent alter-
training set, and the fact that the combination-distress type native approaches to existing systems, or they collaborate
includes various distress patterns that do not have homoge- to make the overall system more efficient. Figure 2 summa-
neous characteristics. rizes how the different AI technologies in this review have
Goh14 reports an integrated ANN and knowledge-based been used for different problems in the decision-making
expert system for selecting proper rehabilitation schemes of process in pavement management.
deteriorated road pavement. The knowledge-based expert Which of these major technologies to use depends on
system provides the pavement engineer with a list of rec- the nature of the problem and the specific decision support
ommended repair schemes together with their associated system configuration. The key is a successful solution to
costs. It also offers the possibility to rank the rehabilitation the decision problem, not the use of a specific tool or tech-
actions according to the pavement condition-rating index. nique. As nonlinear regression models, artificial neural net-
A feed-forward ANN is coupled with the knowledge-based works are particularly suited for processing numerical data
expert system and feeds it with the list of recommended and retrieving information from numerical historical data.
repair schemes. Nine actions are possible (50-mm asphalt Expert systems are especially efficient in problem solving
overlay, 25-mm asphalt overlay, 25-mm asphalt and partial that involves extensive expert knowledge and human rea-
reconstruction, 50-mm asphalt and partial reconstruction, soning processes that are too complex to be represented
156 Sundin & Braban-Ledoux

and implemented in an analytical way. The characteristics 5. Broten, M. Local Agency Pavement Management Applica-
of the two technologies are so different that, in many cases, tion Guide, Washington State Department of Transportation,
combining expert systems and neural networks results in Olympia, WA, 1997.
powerful systems. 6. Chou, C. J. & Liau, T., Development of automated algorithms
Many of the reported applications show that using AI for pavement condition survey, Transportation Research
Record, 1536 (1996), 103–9.
offers significant advantages over other more conventional
7. Chou, J., O’Neill, W. A. & Cheng, H., Pavement distress eval-
models. Unfortunately, a large number of these AI systems uation using fuzzy logic and moments invariants, Transporta-
have been developed on the basis of synthetic data. The real tion Research Record, 1505 (1995), 39–46.
challenge is to develop an application that performs signifi- 8. Eldin, N. N. & Senouci, A. B., A pavement condition rating
cantly better than the models commonly used by pavement model using backpropagation neural networks, Microcomput-
engineers on the basis of real data collected from the field. ers in Civil Engineering, 10 (6) (1995), 433–41.
AI-based systems are continually evolving, and the cur- 9. Flintsch, G. & Zaniewski, J., Expert project recommendation
rent trend is toward their integration into pavement manage- procedure for Arizona Department of Transportation’s pave-
ment so that they can increase productivity and enhance the ment management system, Transportation Research Record,
1592 (1997), 26–34.
quality of the decision-making process. The system result-
10. Flintsch, G., Zaniewski, J. & Delton, J., Artificial neural net-
ing from the integration of AI techniques into pavement
work for selecting pavement rehabilitation projects, Trans-
management systems is meant to be an adjunct to pavement portation Research Record, 1524 (1996), 177–93.
management engineers and administrators to extend their 11. Fwa, T. F., Chan, W. T. & Hoque, K. Z., Network level
capabilities but not to replace their judgment. It still implies programming for pavement management using genetic algo-
collaboration between the user and the system. Neverthe- rithms, in Fourth International Conference on Managing
less, this integration does not solve all problems pavement Pavements, Durban, South Africa, May 17 to 21, 1998, Vol. 2,
engineers are confronted with, and pavement engineers pp. 815–29.
have to be aware of the limitations of the models so that 12. Fwa, T. F., Chan, W. T. & Tan, C. Y., Genetic algorithm pro-
they use them in their domain of validity. gramming of road maintenance and rehabilitation, Journal of
Transportation Engineering, 122 (3) (1996), 246–53.
13. George, K. P., El-Rahim, A. M. A. & Shekharan, A. R.,
ACKNOWLEDGMENTS Updates of pavement performance modeling, in Third Inter-
national Conference on Road and Airfield Pavement Tech-
nology, Proceedings, Vol. 1, Beijing, China, April 1998,
We express our sincere appreciation for the funding pp. 402–10.
provided by the Swedish National Road Administration 14. Goh, A. T. C., A hybrid neural network based pavement man-
(SNRA), Center of Excellence for Research and Educa- agement system, Road & Transport Research, 6 (4) (1997),
tion in Operation and Maintenance of Infrastructure (CDU), 62–71.
and the Swedish Transport and Communications Research 15. Gorry, G. M. & Scott Morton, M. S., A framework for man-
Board (KFB). This report is part of a research effort agement information systems, Sloan Management Review, 13
to apply artificial neural networks in pavement mainte- (1) (1971), 55–70.
nance. The input and feedback provided by Johan Lang 16. Grivas, D. A. & Shen, Y. C., Use of fuzzy relations to man-
at the Swedish National Road Administration is gratefully age decisions in preserving civil infrastructure, Transporta-
tion Research Record, 1497 (1995), 10–26.
acknowledged.
17. Haas, R., Hudson, W. R. & Zaniewski, J., Modern Pave-
ment Management, Krieger Publishing Company, Malabar,
FL, 1995.
REFERENCES 18. Hajek, J. J., Haas, R. C. G. & Phang, W. A., ROSE:
A knowledge-based expert system for routing and seal-
1. Alsugair, A. M. & Al-Qudrah, A. A., Artificial neural network ing, in Proceedings of the Second North American Pave-
approach for pavement maintenance, Journal of Computing ment Management Conference, Toronto, Canada, Nov. 1987,
in Civil Engineering, 12 (4) (1998), 249–55. pp. 2.301–2.341.
2. Attoh-Okine, N. O., Predicting roughness progression in flex- 19. Hajek, J. J. & Hurdal, B., Comparison of rule-based and
ible pavements using artificial neural networks, in Third Inter- neural network solutions for a structured selection problem,
national Conference on Managing Pavements, San Antonio, Transportation Research Record, 1399 (1993), 1–7.
TX, 1994. 20. Hoque, K. Z., Fwa, T. F. & Chan, W. T., Programming of
3. Banan, M. R., & Hjelmstad, K. D., Neural networks and pavement maintenance based on genetic algorithm, in Third
AASHO road test, Journal of Transportation Engineering, International Conference on Road and Airfield Pavement
122 (5) (1996), 358–66. Technology, Proceedings, Vol. 2, Beijing, China, April 1998,
4. Belkasim, S. O., Shridhar, O. M. & Ahmadi, M., Pattern pp. 1108–16.
recognition with moment invariants: A comparative study and 21. Hu, M. K., Visual pattern recognition by moment invariants,
new results, Pattern Recognition, 24 (12) (1991), 1117–38. IRE Transactions on Information Theory, 8 (1962), 179–87.
 Artificial intelligence–based decision support technologies in pavement management 157

22. Huang, Y. & Moore, R. K., Roughness level probability 35. Pant, P. D., Zhou, X., Arudi, R. S., Bodocsi, A. &
prediction using artificial neural networks, Transportation Aktan, A. E., Neural-network-based procedure for condition
Research Record, 1592 (1997), 89–97. assessment of utility cuts in flexible pavements, Transporta-
23. Jiang, Y. J., Darter, M. I., Von Quintus, H. & Killingsworth, tion Research Record, 1399 (1993), 8–13.
B. M., Designer: A prototype knowledge-based expert system 36. Prechaverakul, S. & Hadipriono, F. C., Using a knowledge-
for selecting recommended highway pavement design fea- based expert system and fuzzy logic for minor rehabilita-
tures, in Third International Conference on Road and Airfield tion projects in Ohio, Transportation Research Record, 1497
Pavement Technology, Proceedings, Vol. 1, Beijing, China, (1995), 19–24.
April 1998, pp. 157–65. 37. Ritchie, S. G., A knowledge-based approach to pave-
24. Kaseko, M. S., Lo, Z. P. & Ritchie, S. G., Comparison of ment overlay design, Transportation Research Record, 1145
traditional and neural classifiers for pavement-crack detection, (1987), 61–8.
ASCE Journal of Transportation Engineering, 120 (4) (1994), 38. Ritchie, S. G., Expert systems in pavement man-
552–69. agement, Transportation Research, 21A (2) (1987),
25. Kaseko, M. S. & Ritchie, S. G., Pavement image processing 145–52.
using neural networks, in Proceedings of the Second Interna- 39. Ritchie, S. G., Che, I. Y., Mahoney, J. P. & Jackson, N. C.,
tional Conference on Applications of Advanced Technologies Surface condition expert system for pavement rehabilitation
in Transportation Engineering, American Society of Civil planning, ASCE Journal of Transportation Engineering, 113
Engineers, New York, 1991, pp. 238–42. (2) (1987), 155–67.
26. Kaseko, M. S. & Ritchie, S. G., A neural network–based 40. Ritchie, S. G, Kaseko, M. & Bavarian, B., Development of an
methodology for automated distress classification of pave- intelligent system for automated pavement evaluation, Trans-
ment images, in Proceedings of the International Conference portation Research Record, 1311 (1991), 112–19.
on Artificial Intelligence Applications in Transportation Engi- 41. Roberts, C. A. & Attoh-Okine, N. O., A comparative analyses
neering, 1992, pp. 195–214. of two artificial neural networks using pavement perfor-
27. Kaseko, M. S. & Ritchie, S. G., A neural network–based mance prediction, in Conference Proceedings, Neural Net-
methodology for pavement crack detection and classification, work Applications in Highway and Vehicle Engineering,
Transportation Research, Part C: Emerging Technologies, 1C 1996, pp. 59–73.
(4) (1993), 275–91. 42. Sundin, S., Predicting rut depth in pavements using artifi-
28. Kotb, A. S. & Moore, C. J., A decision support expert system cial neural networks, in Nimes ’98 Conference on Complex
for maintenance and rehabilitation needs, in World Transport Systems, Intelligent Systems, & Interfaces, La Lettre de l’IA
Research, Proceedings of the 7th World Conference, Vol. 4: Nimes 98, May 26–28, 1998, pp. 263–6.
Transport Management, 1996, pp. 397–410. 43. Taha, M. A. & Hanna, A. S., Evolutionary neural net-
29. Kuncheria, P. I. & Veeragavana, A., PADMA: Expert system work model for the selection of pavement maintenance
for flexible pavement deterioration and maintenance, Indian strategy, Transportation Research Record, 1497 (1995),
Highways, 24 (9) (1996), 27–47. 70–6.
30. Kwasi, A. A. & Attoh-Okine, N. O., Effects of cross proba- 44. Teague, M. R., Image analysis via the general theory of
bilities in evolutionary algorithm trained to predict pavement moments, Optical Society American Journal, 70 (8) (1980),
performance, presented at the 78th Transportation Research 920–30.
Board Annual Meeting, Washington, January 1999. 45. Turban, E. & Aronson, J. E., Decision Support Systems
31. La Torre, F., Domenichini, L. & Darter, M. I., Roughness pre- and Intelligent Systems, Prentice-Hall, Englewood Cliffs, NJ,
diction model based on the artificial neural network approach, 1998.
in Fourth International Conference on Managing Pavements, 46. van Bemmel, J. H. & Musen, M. A., Handbook of Medical
May 1998, pp. 599–612. Informatics, Springer-Verlag, Berlin, 1997.
32. Lee, H. & Galdiero, V., Enhancing the existing pavement 47. Van der Gryp, A., Bredenhann, S. J., Henderson, M.
management systems by using expert system techniques, in G. & Rohde, G. T., Determining the visual condition
First International Conference on Applications of Advanced index of flexible pavements using artificial neural networks,
Technologies in Transportation Engineering, San Diego, CA, in Fourth International Conference on Managing Pave-
February 1989, pp. 76–81. ments, Durban, South Africa, May 17 to 21, 1998, Vol 2,
33. Nallamothu, S. & Wang, K. C. P., Experimenting with recog- pp. 115–29.
nition accelerator for pavement distress identification, Trans- 48. Yuge, F., Kameyama, S., Ishitani, M., Himeno, K. &
portation Research Record, 1536 (1996), 130–5. Kasahara, A., Optimization of pavement maintenance and
34. Owusu-Ababio, S., Cracking prediction in flexible pavements: rehabilitation strategy considering bearing capacity, in Third
Effect of neural network topology on model performance, in International Conference on Road & Airfield Pavement Tech-
Conference on Neural Network Applications in Highway and nology, Proceedings, Vol. 2, Beijing, China, April 1998,
Vehicle Engineering, April 1996, pp. 43–55. pp. 1047–55.