Construction project control using artificial neural networks · to-date project plans at regular intervals during project execution. 2. ARTIFICIAL NEURAL NETWORKS Artificial intelligence

Artificial Intelligence for Engineering Design, Analysis and Manufacturing (1997), / / , 45-57. Printed in the USA.Copyright © 1997 Cambridge University Press 0890-0604/97 $11.00 + .10

Construction project control using artificialneural networks

H. AL-TABTABAI,1 N. KARTAM,2 I. FLOOD,3 AND A.P. ALEX4

'Civil Engineering Department, College of Engineering and Petroleum, Kuwait University,P.O. Box 5969, Safat 13060, Kuwait.

2Civil Engineering Department, College of Engineering and Petroleum, Kuwait University,P.O. Box 5969, Safat 13060, Kuwait.

3ME Rinker Sr. School of Building Construction, Gainesville, Florida 32611, USA""Civil Engineering Department, College of Engineering and Petroleum, Kuwait University,

P.O. Box 5969, Safat 13060, Kuwait.

(RECEIVED June 17, 1996; ACCEPTED December 2, 1996)

Abstract

Artificial neural networks are finding wide application to a variety of problems in civil engineering. This paper de-scribes how artificial neural networks can be applied in the area of construction project control. A project controlsystem capable of predicting and monitoring project performance (e.g., cost variance and schedule variance) based onobservations made from the project environment is described. This project control system has five neural networkmodules that allow a project manager to automatically generate revised project plans at regular intervals during theprogress of the project. These five modules are similar in design and implementation. Therefore, this paper will presentthe main issues involved in the development of one of these five neural network modules, that is, the module foridentifying schedule variance. A description of a graphical user interface integrating the neural network modules de-veloped with project management software, and a discussion on the power and limitations of the overall system con-clude the paper.

Keywords: Construction Project Control; Artificial Neural Networks; Schedule Control; Cost Control

1. INTRODUCTION

The traditional planning and controlling methods practicedin the construction industry demand the project manager tobase the estimate of various control parameters (e.g., costand schedule variances) on status reports that become avail-able from time to time. Project managers evaluate thesestatus reports to predict the variations in these control pa-rameters over the duration of the project. These methodsare satisfactory, but when hundreds of tasks have to be pre-cisely choreographed, these predictions become difficult tomake. Whether such an unaided subjective estimate by theproject manager is completely reliable is an important ques-tion. For effective control, project managers have to corn-

Reprint requests to: Prof. Hashem Al-Tabtabai, Department of CivilEngineering, College of Engineering and Petroleum, Kuwait University,P.O. Box 5969, Safat 13060, Kuwait. E-mail: [email protected].

pare the performance of future work against the originalbaseline estimate to identify likely problems and possiblesolutions. The efficiency of the project manager in makingan intuitive estimate about the future, and how far he/shecould effectively integrate this information with the currentplan, has a strong bearing on the success of a constructionproject. The approach presented here is based on the devel-opment of an artificial neural network tool that will aid theproject manager in this task. Artificial neural networks(ANNs) would seem to offer a potentially powerful tool forestimating project control parameters from current projectconditions. The prime reason for the selection of an ANNparadigm from other available artificial intelligence tools istheir ability to learn and adapt a solution to a problem fromexperience.

This paper describes how a neural network can be de-signed and trained to assist project managers in their deci-sion-making process. The ANN model presented in this paper

45

Downloaded from https://www.cambridge.org/core. 29 Jun 2020 at 06:26:00, subject to the Cambridge Core terms of use.

https://www.cambridge.org/core

46 H. Al-Tabtabai et al.

can identify the deviations in project control parameters fromthe baseline plan, and then generate meaningful and up-to-date project plans at regular intervals during projectexecution.

2. ARTIFICIAL NEURAL NETWORKS

Artificial intelligence techniques such as expert systems andartificial neural networks are becoming widely imple-mented in construction with the main objective of assistingproject personnel in decision-making. In the case of expertsystems, the intent is to make an expert-like device by storinglarge amounts of domain-specific knowledge in condition-action form, with an inference engine that simulates humanreasoning. However, the serial architecture and reasoningprocess of expert systems and the limitations in knowledgestructure, along with the difficulty in knowledge gathering,have been observed as major causes of bottlenecks in thedevelopment of expert systems. Another argument againstexpert systems is that they are performance systems (Si-mon, 1991), that is, they can execute only what they areprogrammed to perform, are unable to learn by themselves,cannot process incomplete or partial attributes, and can-not process previously unseen problems. Artificial neuralnetworks have advantages over expert systems in these re-spects. They allow self-learning, self-organization, and par-allel processing, and are well suited for problems involvingmatching of input patterns to a set of output patterns wheredeep reasoning is not required. A good introduction to thegeneral field of ANNs can be found in Lippmann (1987) orCaudill (1989). Flood and Kartam (1994) provide a de-tailed description about the fundamentals of neural net-works along with their potential applications in civil andconstruction engineering. Another source for understand-ing the scope of ANNs is a recent ASCE monograph en-titled Artificial Neural Networks for Civil Engineers: Fun-damentals and Applications (ASCE 1997). Moselhi et al.(1991) present an in-depth review of neural networks as tools

in construction, along with a comparison of commonly usedneural paradigms.

Artificial neural networks, also known as connectionistsystems, are a class of modeling tools inspired by the work-ings of biological neural systems. Artificial neural net-works are composed of neurons or processing elements (PE)and connections that are organized in layers: an input layer,middle or hidden layer(s), and an output layer, as illustratedin Figure 1. The signals entering the input layer flow to theoutput layer through the middle layer(s), with the input sig-nals detailing the problem to be solved and the output sig-nals representing the network's solution to that problem. Theconnections between the neurons are associated with nu-merical values called connection weights, which determinethe influence of one neuron on another. The connectionweights modify the output signal on each of the connectionpaths, making some connections stronger and other connec-tions weaker. The neurons in the input layer receive theiractivation from the environment, while the activation levelsof neurons in the middle and output layers are computed asa function of the activation levels of the neurons feedinginto them. Typically, this involves the summation of all in-coming signals (along with a bias associated with the mid-dle layer neuron) followed by the application of a nonlinearfunction termed the activation function or transfer function.

During the training process, the connection weights arelearned by the network as training examples are presentedrepeatedly. The training process can be supervised or unsu-pervised. In supervised training (the approach adopted forthe work presented here), the connection weights are mod-ified continuously until the error between the desired out-put and the actual output is minimized. The neural networkmodifies the weights between layers in successive itera-tions until the network is able to generate the desired outputof the system to within a specified accuracy or until the user-specified number of iterations has been performed. Knowl-edge is effectively learned and stored by the weights on theconnections between the neurons. Once training has been

Variancesin

controlparameters

/Construction1

projectenvironment

variables

[ Input layerj [ Hidden layerj ( Output layer Fig. 1. A simple multilayered neural net-work mapping of the problem.



Construction project control using ANNs 47

completed, it is anticipated that the neural network will beable to generate the required output for example problemsnot considered during training.

3. CONSTRUCTION PROJECT CONTROL

Construction project control consists of verifying that all ofthe activities and resources involved in the project are usedeffectively to ensure that the project is within budget, ontime, and provides satisfactory technical performance. Thecontrol process detects any deviation from the baseline planand the necessity of any corrective action that should betaken accordingly.

Project control in construction includes the following fourelements (Cleland, 1985):

a. The performance standards and plans formulated anddeveloped from the project's objectives, goals, andstrategies.

b. The performance measurement techniques.

c. A comparison of the planned and actual performances.

d. Corrective action that is required to get the project backon track.

In any construction project, a Work Breakdown Structure(WBS) classification (US DOE, 1986) can be adopted toorganize work items systematically. WBS began with thedevelopment of the Gantt chart and eventually evolved intothe Critical Path Method (CPM). Work Package (WP) un-der WBS classification consists of a prescribed amount ofwork with a single responsible authority and a specific bud-get. The use of WBS in planning and control is supportedby various researchers (Neil, 1982; Meuller, 1986; Adrian,1987; Al-Tabtabai & Diekmann, 1990; Keisk & Selby, 1990;Riggs, 1990). Using WP, the deviations can be evaluated bycomparing the earned values with the actual expenditure todetermine the efficiency and productivity. Thus, each WPcan be associated with information such as:

a. descriptive data—code of WP, start date, status, typeof work;

b. cost data—budgeted cost of work scheduled (BCWS),budgeted cost of work performed (BCWP), actual costof work performed (ACWP); and

c. scheduling data—duration, start date, percentage ofwork completed, available float, etc.

Sophisticated project management systems that performproject control functions (e.g., engineering control, purchas-ing inventory management, quantity measurement) are avail-able, but their use has been limited almost exclusively toalgorithmic solutions. However, many construction projectcontrol situations are not amenable to purely algorithmicsolutions. When dealing with real-life situations, it is im-possible or unfeasible to obtain the amount of data nec-

essary for a realistic representation of the system underexamination (Warszawaski, 1985). Traditionally, the projectmanager relies on his/her own intuition and expertise to dealwith these ill-structured problems.

To perform effective project control, a project managershould have expert knowledge gained through education andexperience in the construction environment. This knowl-edge must include technical skills, economic and financialknowledge, social and communication skills, and legal aswell as political knowledge. No existing project manage-ment system can be effective in providing the project man-ager with all of these types of knowledge (Woolery &Crandall, 1983; Levitt & Kunz, 1985). This is because suchsystems are based only on quantitative methods of plan-ning, scheduling, and monitoring (Al-Tabtabai & Diek-mann, 1990; Kartam & Levitt, 1990; Kartam & Levitt, 1991).

When project progress deviates from the baseline plan,the principle problem for the project manager is to selectthe most effective response to forecast performance vari-ance and apply mitigating strategies before actual projectperformance suffers. The ability to use construction knowl-edge and expertise to analyze project progress deviationsand then forecast project performance represents, perhaps,the most important single feature that has not been accom-modated in existing computerized project management sys-tems. However, recent advances in artificial intelligence andcomputing provide new tools to capture and structure hu-man knowledge and expertise in fields where conventionalalgorithmic programming does not apply.

Once the environment variables that affect the project con-trol parameters (variations in schedule, cost, quantity, etc.)are identified and their status is determined for each of theWPs, an expert in project control can make an intuitive judg-ment about the expected influence of these variables on theproject control parameters, based on analogy and withoutthe necessity of deep reasoning. A systematic pattern of thesejudgments can be coded and represented using neuralnetwork-based tools. Each WP can be evaluated using atrained neural network to identify the variations in controlparameters from the planned values. The trained neural net-work mimics the decision process of the expert whose judg-ment is employed for training. Outputs from such systemscan be used to update the project control statistics and togenerate revised plans at various execution stages of the WP.

4. NEURAL NETWORK DESIGN

The neural network approach discussed in this paper to de-velop a project control system follows the steps shown inFigure 2. The following section describes the design param-eters adopted for the current problem.

In this research, a continuous function mapping networkhas been adopted since the values of the output variable(schedule variance) and the input variables are continuous.Continuous mapping functions can represent both continu-




Divide the project control domain into specific simple control modules. Such aclassification helps in the efficient development of a neural network for each of

the individual modules.

Identify the factors that affect each of the control parameters to define theinputs and outputs for each of the control modules.

Collect input-output data for each of the control modules from experts andformat it into training examples

| Select a portion of the training examples and train the neural network. \

Test the performance of the developed neural network using test examples notused to train the network. Refine the network by adjusting the network

parameters if required.

Integrate each of the network modules using a graphical user interface (GUI)that is capable of prompting the user with present environment variables and

displaying the forecast of the control parameters.Fig. 2. Development steps of the projectcontrol system.

ous and discrete data, whereas discrete mapping functionsare limited in their ability to represent continuous data. How-ever, certain discrete sets of values (attribute labels) wereused to provide meaningful anchors for the scale values inorder to ease the judgment process of experts during knowl-edge acquisition.

The back-propagation algorithm is the most widely usedtraining technique for continuous function mapping, hasbeen shown to be theoretically sound (Rumelhart et al.,1986), performs well in modeling nonlinear functions, andis simple to code. Back-propagation algorithm training de-velops the input to output mapping by minimizing a sumsquared error cost function measured over a set of train-ing examples. The mean squared error (MSE) is expressed

Transfer Function (x) =

as:

MSE =(D - A)2

where A is the actual output generated by the neural net-work, D is the desired output value, and n is the number oftraining example. Twomey and Smith (1997) provide a re-view of other various error matrices available for neural net-work validation.

The transfer function computes the activation level of aneuron from the sum of input values. The function adoptedfor the current problem was a sigmoidal curve, given by:

1 + e'

where x is the sum of the inputs to the neuron. The sigmoi-dal function generates output values between 0 and 1.

Another important network design variable is the learn-ing rate coefficient (T)), which represents the degree by whichthe weights are changed when two neurons are excited. Eachtime a pattern is presented to the network, the weights lead-ing to a neuron are modified slightly during learning in thedirection required to produce a smaller error at the outputsthe next time the same pattern is presented. The amount ofweight modification is proportional to the learning rate. Thevalue of t) ranges between 0.0 and 1.0, where a value closerto 1 indicates significant modification in weight while a valueclose to 0 indicates little modification. A small learning rateof 0.15 was arbitrarily chosen for the current problem, sincelarger learning rates have often been found to lead to os-cillations in weight changes resulting in a never-endinglearning process. One way to allow faster learning withoutoscillations is to make the weight change, in part, a func-tion of the previous weight change. A momentum coeffi-cient represents this portion of the weight change. In thisstudy, a coefficient of 0.7 was found to perform well.

The number of hidden layer(s) and the number of hiddenneurons in the hidden layers provide the power of internalrepresentation in capturing the nonlinear relationship be-




tween the input and output vectors. Hence, a larger numberof hidden layers and hidden neurons provides the potentialfor developing a more effective network. However, the ad-dition of more hidden neurons increases the number of un-determined parameters (weights and biases) associated withthe network. A large number of training examples is thenneeded to solve these parameters and to get a good approx-imation of the problem domain. When too few training ex-amples are provided, the network will try to memorize,resulting in poor generalization. Carpenter and Hoffman(1995) determined that to provide a good approximation overthe problem domain, one should have overdetermined ap-proximations, that is, the number of training pairs should begreater than the number of undetermined parameters. Un-determined parameters in neural network approximations arethe weights and biases associated with the network. To cap-ture and represent the features within a set of data thereshould usually be one to two hidden layers. One hidden layerwith 14 hidden neurons constitutes the present network ar-rangement.

The training of a neural network is stopped when the er-ror falls below a user-specified level, or when the user-defined number of training iterations has been reached. Inthis case, 20,000 iterations were planned for the final train-ing process, as this was found adequate in a series of testruns.

5. KNOWLEDGE REPRESENTATION

The proposed system has five different control modules rep-resenting two of the main control domains, namely, sched-ule control and cost control, as shown in Figure 3. Variancein schedules is controlled by the schedule variance module(SV-module), while variances in cost are controlled by thelabor productivity module (LP-module), the labor wage ratevariance module (LWV-module), the material price rate vari-ance module (MPV-module), and the quantity variance mod-ule (QV-module). The environment variables that may affecteach of these control domains will act as the input variablesfor the network module, while the expected variance rating,

Workpackagerevised

statistics

Fig. 3. Neural network modules of the project control system.

represented in a suitable form, will be the correspondingoutput variables. The following section will briefly identifythe knowledge representation schema, adapted for each ofthe modules from Al-Tabtabai and Diekmann (1992). Theinput attributes were based on personal interviews with 30project experts in the construction industry in the UnitedStates and Kuwait. All of the experts consulted were engi-neers with more than 20 years of experience in top construc-tion management positions. The authors have adopted everypossible step to choose high-quality experts for attribute iden-tification and, hence, believe that any chance of insignifi-cant additions or significant omissions are very remote.

5.1. Schedule control module

Schedule control is the process of tracking the variations inthe planned schedules and taking corrective actions to fin-ish the work as planned. However, as the project enters itsexecution phase, variations from previous expectations canoccur. These variances related to schedules are representedby the parameter schedule variance, which can be quanti-fied as the variation between the budgeted cost of work per-formed (BCWP) and the budgeted cost of work scheduled(BCWS).

5.7.7. Schedule variance module

The following attributes that affect the estimation of ex-pected schedule variance for a construction WP are selectedas the input pattern for the SV-module.

1. Performance of the management: This encompassesthe performance of the contractor/consultant in termsof his efficiency and experience in planning and con-trol, the general performance of the subcontractors, thedecision-making ability of supervisors and their abil-ity and experience to cope with technical problems,etc.

2. Cash flow situation: Constraints on the financial frontdue to delay in payments for work executed and block-age of funds can cause unnecessary work stoppage.

3. Material and equipment availability: The unavailabil-ity of materials specified at the time of execution, dif-ficulties in the procurement of the right equipment,import restrictions, delays caused during transit andclearance of materials, the nonavailability of alterna-tive materials, difficulties in the approval of alternatematerials, and lack of available plant and equipmentfacilities for the work demand, and so on, can causeschedule shifts.

4. Labor availability and productivity: Unavailability ofthe appropriate labor for the work, and productivitythat is below minimum expected levels, can lead toproject delay.

5. Weather and other environment influences: Inclementweather and environment influences (such as change




in temperature, rainfall, rise in ground water levels,etc.) can cause major delays in construction opera-tions. Remote sites and poor access cause difficulty inthe transportation of materials, personnel, and equip-ment, thus leading to variations in the schedule fromthe baseline plan.

6. Amount of rework, extra work, and work difficulty:The expected difficulty of a job may be elevated dueto technological change, changes in the planned ex-tent of work, and, in extreme cases, changes in the over-all scope of the activity. The difficulty of work alongwith any possibility of extra work and added difficul-ties at later stages of the project can prolong the du-ration of a WP.

7. Percentage of work completed: This factor representsthe current progress of the WP.

8. Trend in schedule variance: Trends can be used to in-dicate future outcomes from past performance. Sched-uling experts usually compare their present projectperformance with experiences from similar past proj-ects with a similar schedule variance pattern to developan impression of how their project is progressing.

The input variables used in the SV-module are inherentlycontinuous in nature, and hence any of an infinite numberof values could be input. There was a need, however, to se-lect some scale that a user of the system could relate to.Human experts provided their response for the input vari-ables using seven labels with a scale ranging from 1 to 9.Attribute 7 has been scaled/normalized so that a value of 1corresponds to 10% and a value of 9 corresponds to 90%—values that a user can readily relate to. A rating in the formof a percentage change in schedule variance from the cur-rent level schedule variance forms the output vector for thismodule.

The neural network is an empirically derived model, asopposed to a theoretically derived model. A quantificationof each input variable (into one of seven values) is not dif-ficult for an expert in the field. For example, attribute 1 (per-formance of management) can have a value anywherebetween 1 (highly favorable) and 9 (highly unfavorable).When an attribute is labeled as "highly favorable," it refersto the best condition that an expert can think of regardingthat attribute; and when an attribute is labeled as "highlyunfavorable," it refers to the maximum worst condition thatan expert can think of regarding that attribute. Once the ex-pert identifies these extremes from his experience and skill,he can then easily quantify other intermediate scale valuesbetween these extremes. Such experts make such judg-ments on a day-to-day basis in evaluating the performanceof a project, albeit in a less formal manner. The choice ofwords to express each level for a variable (where appropri-ate) facilitates this task by using terms that are familiar tothe expert. Figure 4 represents the typical neural network

representation of input to output vector mapping for the SV-module.

5.2. Cost control modules

Cost control starts with the development of estimates forlabor, material, and equipment. The estimated costs arederived from historical cost databases that are suitably mod-ified with the current cost indices. These costs, once iden-tified, are assigned to work items and loaded in the networkto generate a cost estimate for the entire project. However,the project cost estimate is again susceptible to variationsdue to the effects of various environmental variables. Aschanges occur, the project manager needs to modify the orig-inal estimate. The following modules will help the projectmanager identify the deviation of respective cost elementsand incorporate the effect of these variances in the cost es-timate. The breakup of cost elements as labor cost and ma-terial cost will be helpful in the control process.

a. Labor costs: Estimating labor-related costs representsa major and often ill-defined part of the project. Thebasic approach in estimating labor costs associated witha WP is to divide labor costs into components, de-velop them separately, and use them in the followingequation:

Labor Cost = Quantity of Work (units)X Productivity (hr/unit) X Wage Rate ($/hr),

where Quantity of Work is the volume of work to be per-formed associated with the operation of the WP, and Pro-ductivity is the amount of work that a crew involved in aWP can complete in a defined period of time. Standard ratesof labor productivity are available from historical recordsor established sources. However, the difficulty comes whenthe estimator needs to account for and estimate many of thefactors that can influence labor productivity. These factorsare highly qualitative in nature, and a great deal of experi-ence and intuition is needed to develop the type of informa-tion that is required. Wage Rate is the money associated withhourly wages, which are generally made up of some or allof the direct wages, taxes, holiday pay, sick leave, vacationpay, overtime premiums, insurance premiums, and trainingcosts.

b. Material costs: Material costs can be represented as,

Material Cost = (Quantity of Work)X (Material Price Rate),

where Material Price Rates are usually established by directquotations from manufacturers and suppliers.

The neural network project control system describedin this paper includes the following four modules for costvariance. The attributes affecting each module are self-explanatory, and thus it suffices to list them.




Attr.-

Attr-

Attr.-

Attr.-

Attr.-

Attr.-

Attr.-

AttT.-8TrenchnS.V

Final Schedule Variance will...

Worsen

100% from the current level75%50%25%No improvement/Worsening

25% from the current level50%

100% "150% "

f Weight Array -w j

H.FFS.FA.ES.UFU.FH.UFH.II.S.IN.I/W

Legend= Highly favorable= Favorable= Slightly favorable= As expected= Slightly unfavorable= Unfavorable= Highly Unfavorable= Highly improving= Improving= Slightly improving= No improve/worsen

Fig. 4. A feedforward multilayer neural network representation of input-output vectors for the schedule variance module.

5.2.1. Labor productivity variance module

The following attributes affecting the estimation of ex-pected labor productivity variance for a construction WPare selected as the input variables for the LPV-module:

• present WP labor productivity variance,

• WP percent complete,

• past trend of labor productivity,

• expected changes in the difficulty of work,

• expected impacts from changes,

• expected changes in labor skill and motivation,

• expected changes in supervisor/management quality,

• expected changes in field support quality,

• expected changes in overtime allocation.

5.2.2. Labor wage rate variance module

The following attributes affecting the estimation of ex-pected labor wage rate variance for a construction WP areselected as the input variables for the LWV-module:

• present WP wage rate variance,


• past trend of wage rate variance,

• expected changes in journeyman/apprentice ratio,

• expected changes in craft mix,

• expected changes in planned overtime.

5.2.3. Material price rate variance module

The following attributes affecting the estimation of ex-pected material price rate variance for a construction WPare selected as the input variables for the MPV-module:




• present material price rate variance,


• past trend of material price rate variance,

• expected storage warehouse expense,

• expected price level changes (interest/inflation),

• expected changes in availability and market stability,

• expected changes caused by scope or quality changes,

• expected changes in transportation costs.

5.2.4. Quantity variance module

The following attributes affecting the estimation of ex-pected quantity variance for a construction WP are selectedas the input variables for the QV-module:

• present WP quantity variance,


• past trend of quantity variance,

• effect of anticipated scope changes,

• anticipated changes in the volume of rework,

• anticipated changes in the volume of waste and scrap,

• quality of quantity estimates.

6. DEVELOPMENT OF NEURAL NETWORKMODULES

Having identified the factors that influence the control pa-rameters, it is now necessary to establish the training sam-ples for each of the control domains. The following is anoutline of the development of the SV-module.

6.1. Knowledge acquisition and formulation

Case profiles with random values for each of the input vari-ables were provided to a number of experts to ascertain theirdecisions. Seven project managers from the local construc-tion industry (Kuwait) with more than 20 years of projectmanagement experience exercised their judgments on 75 WPcase profiles. Each case represented a unique situation of aproject. The experts analyzed the state of a set of inputsdescribing the case and exercised their judgment on the ex-pected change of schedule variance for the described WP.The judgments of each expert on these case profiles werefurther given as feedback to other participants for modifi-cation and revision of judgments. The revised judgmentswere averaged to reach a smoothed output value for each ofthe case profiles. These hypothetical case profiles with theextracted output values were coded to generate a set of train-ing data. The total number of possible cases is effectivelyinfinite, since the input variables are inherently continuous.The fact that the authors chose to limit the actual values ofeach input to a finite set does not alter this point. As far asthe neural network is concerned, it is fitting a continuous

function to the training data. The number of training pointsthat are needed to develop an accurate continuous functionmodel depends on other factors, most notably: the complex-ity of the solution surface being modeled (i.e., how manyhills and valleys it contains); the stochastic content of thedata (i.e., one needs enough training points to prevent biasdue to random fluctuations); and the number of input vari-ables (though not the number of values considered for eachvariable where they are inherently continuous). The valid-ity of the model, as measured in Section 6.3, supports theidea that sufficient patterns were used. Further, the capabil-ity of neural networks to generalize on a limited number oftraining data combined with the randomness in the trainingset will alleviate any chance of bias due to the limited num-ber of training cases.

The eight factors (attributes) that are considered for theprediction of schedule variance become the input variablesfor the SV-module. The industry experts exercised their judg-ments based on these attribute variables to an output scalethat represents the percentage change in schedule varianceexpected from the current level. Figure 4 provides the inputto output mapping of the SV-module.

6.2. Training

The generated input-output data pairs (75 cases) were eachdivided into a training set and a test set. The test set wasderived from the data set by selecting 10-20% of data pairsrandomly. NEUROSHELL™, a commercially available neu-ral network development software from Ward System Group,Inc., was employed to train and develop the neural networkmodule. Eight input neurons and one output neuron with 14hidden neurons constituted the neural network arrangementfor the schedule control module. Network development wasperformed on an IBM-compatible Pentium class machine(100 MHz, 16 MB RAM). Training took 30 minutes, andthe least average error of the network, which was accu-mulated over all of the training sample cases, reached0.0000054.

6.3. Validation of results

6.3.1. Expert validation

The data set used for training the neural network was usedfurther to develop a linear multiple regression analysis(MRA) model with the factors as independent variables andthe judgment as the depended variable. Table 1 provides acomparison between the actual output (expert judgment),the MRA output, and the ANN-generated output for the train-ing set. The mean absolute percentage errors for the MRAand ANN models when applied to the training set were 9and 7%, respectively. It can be observed that the ANN modelcould capture the decision process better than the MRAmodel. A detailed description of the development of the MRA




Table 1. Comparison ofMRA and ANN modelgenerated output vs. desired output (expertjudgment) for training cases

Caseprofile #

Desiredoutput

MRAoutput

ANNoutput

1345678910121415161718192122232426272930313334363738394344454647484950515354555657585961626364656667686971727475

3.27.73.844.97.66.37.14.483.74.94.66.13.87.25.16.13.5482.76.55.37.62.35.92.45.15.46.76.75.13.84.33.83.82.86.86.54.74.97.65.97.65.15.75.45.34.37.74.44.45.66.34.75.44.32.84.6

Mean absolute percentage error

3.87.54.54.8576.26.74.594.14.83.55.92.75.956.22.33.88.12.46.34.56.82.8635.34.66.36.45.33.65.03.63.93.86.86.44.85.67.2684.65.85.64.94.37.643.55.76.55.46.24.43.14

MRA9%

3.87.74.24.54.97.66.36.94.47.93.44.83.463655.72.43.77.62.46.44.472.45.62.55.04.96.26.55.13.64.43.34.03.17.06.34.85.57.55.87.94.95.35.65.04.37.44.24.05.26.44.75.642.64.1

ANN7%

model and its comparison with the ANN model is presentedin Al-Tabtabai et al. (1997).

Table 2 provides the comparison between the actual out-put and the neural network-generated output for the testingset. The mean square error (MSE) for the neural networkmodel when applied to the test cases was found to be 0.481,and the percentage of average operational error for the testcases was found to be only 3.86%. These statistics show theability of the network to predict S V values with moderate tohigh precision. Figure 5 plots the expert judgment (solidline) and the predicted values by the neural network (thinline) for the complete data set of the SV-module.

6.3.2. Statistical test

A statistical test was performed on the test results to com-pare the probability distribution of the expert judgment andthe neural network-generated decision. A nonparametric sta-tistical test was used because of the limited number of cases.A nonparametric test that utilizes information on both thesigns and the magnitude of differences is the WilcoxonSigned-Rank Test (Scheaffer & McClave, 1990). The Wil-coxon Signed-Rank Test for paired differences analyzes thedifference between the expert judgment and the neuralnetwork-generated decisions by calculating the ranks of theabsolute values of the differences between them. The tiedabsolute differences are assigned the average of the ranksthey would receive if they were unequal but successive mea-surements. After the absolute differences are ranked, the sumof the ranks of the positive differences T+ and the sum ofthe ranks of the negative differences T_ are computed. Thetest statistic for the paired difference is the smaller of pos-

Table 2. Comparison of SV-module output (neuralnetwork generated) vs. desired output(expert judgment) on unseen cases

Caseprofile #

Desired output(a)

SV-module output(b)

2

11

13

20

25

28

32

35

40

41

42

52

60

70

73

3.92.34.96.35.07.84.53.33.75.94.92.74.25.44.6

4.52.44.75.25.16.55.73.52.76.94.42.84.05.84.7

Mean square error = 0.481

Average Operational Error = ~ 2n OL

100 = 3 .




10 20 5030 40

Case profile #

Fig. 5. Expert decision vs. neural network-generated decision.

60 70 75

itive (T+) and negative rank sums (71). To test the null hy-pothesis that the two sets of data (expert decision and theneural network-generated decision) are from continuoussymmetric populations, the test hypothesis is formulated asfollows:

HQ = the probability that the distribution of the expertdecision and the neural network decision are iden-tical;

//, = the probability that the distribution of the expertdecision and the neural network decision are dif-ferent.

Test statistics (T)= the smaller of the positive and negative rank sums

T+ and 71 .

The smaller the value of T, the greater the evidence thatthe two probability distributions differ in location. The re-jection region for T can be determined by consulting thetable of critical values of To for the Wilcoxon Paired-Difference Signed-Rank Test. For n = 15 pairs of case pro-files, the value of To for a two-tailed test with a significancelevel of a = 0.01 is 16 (Scheaffer & McClave, 1990, Ap-pendix, Table 10). Therefore, the rejection region for thetest is 7 < 16 for a = 0.01.

Table 3 shows the computations for a Wilcoxon Signed-Rank Test for paired differences. Since the smaller rank sum,71 , does not fall within the rejection region, the alternatehypothesis (//,) cannot be concluded at a = 0.01. Thus theexperiment has not provided sufficient evidence to indicatethat the actual decisions and neural network decisions arenot identical.

Neural network modules computing the material price ratevariance, labor wage rate variance, and quantity variancealso were developed using the same procedure.

7. DISCUSSION

The major advantages of adopting ANNs for the presentproblem are (1) ability to model the complex nonlinear map-

Table 3. Wilcoxon Paired-Difference Signed-Rank Test

Caseprofile

#

211

13202528323540414252607073

7"-statistics

Desiredoutput

3.92.34.96.357.84.53.33.75.94.92.74.25.44.6

= T_ =

T+ =

SV-moduleoutput

4.52.44.75.25.16.55.73.52.76.94.42.84.05.84.7

59.5 T = 59.560.5

Difference

-0 .6-0.1

0.21.1

-0.11.3

-1.2-0 .2

1.0-1 .0

0.5-0 .1

0.2-0 .4-0.1

Absolutedifference

0.60.10.21.10.11.31.20.21.01.00.50.10.20.40.1

Rank

10.02.56.0

13.02.5

15.014.06.0

11.511.59.02.56.08.02.5




ping required in the decision-making process, (2) fault tol-erance capability in smoothing out the noise in the datacollected for training, and (3) ability to generalize on in-complete information. However, the following factors needto be addressed during the development of the system:

a. The accuracy of the control system mainly dependson the soundness of the underlying expert decisions.In other words, the quality of the generated predic-tions by the control system is directly affected by thelegitimacy of judgments used for training.

b. The training cases generated from the judgment of ex-perts are not free of bias because of the intuitive andsubjective nature of the judgment process of individ-uals.

c. Any variables that have a bearing on the WP perfor-mance, if omitted during the knowledge representa-tion stage, will alter the predictions.

The provision of cognitive feedback of judgments amongthe experts allows a reduction in the bias in judgments to agreat extent. In addition, the ability of neural networks tooperate with noisy and incomplete data suggests that theabove limitations can be overcome to some degree. Thisability can be enhanced by choosing high-quality, well-experienced professionals in the local construction industryas domain experts for knowledge representation. Increas-ing the number of training cases with a wide representationof various possible situations will enable the network to gen-eralize and learn the problem more accurately. The /-statisticsfrom the MRA suggest that the identified attributes were allsignificant for the SV-module. Further, the validity of themodel, as demonstrated in Section 6.3, suggests that no keyvariables were excluded from the input to the neural net-work.

Training cases also can be obtained from actual projects.In this case, the environmental variables and the effects ofthese variables on the WP are known and can be related toeach of the control domains. Careful observation at regularintervals is necessary to generate such training cases. Thus,a more realistic representation can be obtained for training.The network modules can be trained further to represent anyunique and peculiar characteristics of the current project.New data can be appended to the training set for subsequentupdate of the network.

8. INTEGRATION

A graphical user interface (GUI) that integrates the devel-oped neural network modules with various project manage-ment systems was developed using VisualBasic™ software.The GUI utilizes the dynamic data exchange (DDE) facil-ity available for Window™ applications for the efficienttransfer of data between various systems. The GUI receivesWP data and feeds them into the neural network module,

which, in turn, provides the predicted statistics. The GUIfacilitates the transfer of revised WP statistics back into thesource application. In this case, MS Project™ was selectedas the scheduling component of the integrated system. Fig-ure 6 represents the GUI architecture.

A typical run through this system for schedule controlwould be as follows:

• Import the existing schedule data file from MS Pro-ject™ to the GUI environment.

• WPs that are currently under progress are selected bythe GUI.

• Answer the prompts of the GUI with values that re-flect the current and anticipated state of environmentvariables.

• The GUI will provide the predictions on each of theconcerned WPs after passing the input data through theSV-module and the original schedule plan is modifiedwith revised activity duration and a new activity finishdate.

• The modified schedule data file then is transferred backto the MS Project™ environment. This new data filecan be processed by MS Project™ to generate revisedschedule plans.

The GUI can be used to modify the training set with newtraining cases generated from the actual projects. This willprovide an up-to-date training set that reflects the currentenvironmental conditions and allows development of a morevalid neural network for each of the modules.

9. SUMMARY AND CONCLUSIONS

The paper has demonstrated the potential for applying neu-ral networks to construction project control. Neural net-works are well suited to decision-making in analogy-basedproblems using the intuition and experience of experts. Thisresults from their ability to learn and generalize from expe-rience. The experiments showed that neural networks aresuitable for modeling complex relations between construc-tion environment conditions and performance of the projectas reflected in its WP.

As shown, it is best to break down the project controldomain into easily manageable modules with individual butsystematic neural network representation, and then inte-grate these using an interface to provide the final solutions.The breakup of a problem domain into control modulesreduces the complexity of a network and facilitates theknowledge-gathering process. The results obtained from theschedule variance module were compared with the recom-mendations provided by the experts. The validation testshowed the neural network solutions to be accurate. The testcases were limited to 15, and a large number of test caseswould provide greater confidence in these results. The GUIsupports the integration of the developed project control sys-tem with traditional project management software for sched-




Project Management Tools

Schedule datafiles generatedfrom MSProject™ orPrtmavera™

| Cost data files11 generated from| Precision1 Estimator™

DDE LinkRevised schedule and costinformation produced by theinterface can be poked into thesource application

Neural NetworkModules

QV-Module |

MPV- Module |

LWV-Module |

LPV-Module [

SV-Module |

NEURUSllfcLL™ .

Environment

D D E |Link |

Reports

Updatetreport oscheduland cos

\

f

data

A

Project Control System

Functions of the system:

Read the data files,select the concerned work package,prompt the user for values of attributes,generate the forecast from neural modules,update the data files with changes forecasted by the model,generate reports, etc.

GUI Fig. 6. Graphical user interface (GUI) that links the neural net-work modules with existing project management systems.

uling (e.g., MS Project™ and Primavera™) and estimating(e.g., Precision Estimator™). Such integration facilitatesthe flow of data between a newly developed system and ex-isting software. As a result, the integrated system is bothfeasible and practical for use by consulting engineers per-forming supervision and for construction managers respon-sible for project management and control functions.

REFERENCES

Adrian, J. (1987). Construction Productivity Improvement. Elsevier Sci-ence Publishing Co., Inc., New York.

Al-Tabtabai, H., & Diekmann, J. (1990). PROCON: A knowledge basedapproach to construction project control. Proc. CIB, Sydney, Australia,385-397.

Al-Tabtabai, H., & Diekmann, J. (1992). Judgmental forecasting in con-struction projects. Constr. Mgmt. Econ., 10, 19-30.

Al-Tabtabai, H., Kartam, N., Flood, I., & Alex, A.P. (1997). Expert judg-ment in forecasting construction project completion. In EngineeringConstruction and Architectural Management. Blackwell Science Ltd.,Oxford, UK (Forthcoming).

ASCE (1997). Artificial Neural Networks for Civil Engineers: Fundamen-tals and Applications (Kartam, N., Flood, I., and Garrett, J., Eds.) ASCE,New York.

Carpenter, W.C., & Hoffman, M.E. (1995). Training backprop neural net-works. AI Expert. March, 30-33.

Caudill, M. (1989). Neural networks primer. Al Expert Feb., 61-67.Cleland, D.I. (1985). A strategy for on-going project evaluation. Proj. Mgmt.

J. (Special Summer Issue) Aug., 11-17.Flood, I., & Kartam, N. (1994). Neural networks in civil engineering: Prin-

ciples and understanding. J. Comput. Civ. Engrg. 8(2), 131-148.Kartam, N.A., & Levitt, R.E. (1990). Intelligent planning of construction

projects. J. Comput. Civ. Engrg. 4(2), 155-176.

Kartam, N., & Levitt, R. (1991). An artificial intelligence approach to projectplanning under uncertainty. Proj. Mgmt. J. XXII(2), 7-11.

Keisk, T, & Selby, K. (1990). Automating construction estimating. Proc.Annual Conf. and 1st Biennial Environmental Specialty Conference,CSCE, Hamilton, Ontario, Canada, Vol. II-l, 150-160.

Levitt, R.E., & Kunz, J.C. (1985). Using knowledge of construction andproject management for automated schedule updating. Proj. Mgmt. J.16(5), 57-76.

Lippmann, R.P. (1987). An introduction to computing with neural nets.1EEEASSP Mag. Apr., 4-22.

Meuller, F. (1986). Integrated Cost and Schedule Control for ConstructionProjects. Van Nostrand Reinhold Company, New York.

Moselhi, O., Hegazy, T., & Fazio, P. (1991). Neural networks as tools inconstruction. J. Constr. Engrg. Mgmt. 117(4), 606-625.

Neil, J. (1982). Construction Cost Estimating for Project Control. Prentice-Hall Inc., New York.

Riggs, L. (1990). Project control technique. Proc, CIB, Sydney, Australia,11-25.

Rumelhart, D.E., Hinton, G.E., & Williams, R.J. (1986). Learning internalrepresentation by error propagation. In Parallel Distributed Process-ing (Rumelhart, D.E., McClelland, J.L., and the PDP Research group,Eds.), pp. 318-362. MIT Press, Cambridge, MA.

Scheaffer, R.L., & McClave, J.J. (1990). Probability and Statistics for En-gineers, 3rd Ed. PWS-KENT Publishing Co., Boston, MA.

Simon, H. (1991). Artificial intelligence: Where has it been, and where isit going? IEEE Trans. Knowledge and Data Engrg. 3(2), 128-136.

Twomey, J.M., & Smith, A.E. (1997). Artificial Neural Networks for CivilEngineers: Fundamentals and Applications (Kartam, N., Flood, I., andGarrett, J., Eds.). ASCE Press, New York (Forthcoming).

US DOE (1986). Cost/Schedule System Criteria for Contract Perfor-mance Measurement—Data Analysis Guide, USA.

Warszawaski, A. (1985). Decision models and expert systems in construc-tion management. Build. Environ. 20(4), 201-210.

Woolery, J.C., & Crandall, K.C. (1983). Stochastic network model for plan-ning and scheduling. J. Constr. Engrg. Mgmt. 109(3), 342-354.




Hashem Al-Tabtabai is currently a faculty member of civilengineering department at Kuwait University. He receivedthe B.S. and M.S. degrees in civil engineering in 1980 and1984 respectively from North Carolina State University, andthe Ph.D. degree in 1989 from Colorado University at Boul-der. His research focuses on advanced information technol-ogy in project management, project control, cost estimation,cost management and cost control. He has provided exten-sive consultation in the field of project management to Mil-itary Engineering Department and Ministry of Public Worksof Kuwait during the post-war renovation period and alsoactive in providing training to various private and govern-ment organizations in Kuwait. He was a recipient of the BestTeacher award of Kuwait University in 1996.

Nabil Kartam is an Associate Professor of civil engineer-ing at Kuwait University. He received the M.S. degree inconstruction engineering and management in 1985 from theUniversity of Michigan, the M.S. degree in computer sci-ence in 1988, and the Ph.D. degree in civil engineering in1989 from Stanford University. He worked as an engineer,planner, professor, and consultant for the last twelve years.Dr. Kartam is the sole investigator for more than a half mil-lion dollars of industrial research in the area of microcom-puter applications, construction engineering, projectmanagement, construction safety, construction estimating andscheduling, artificial intelligence and CAD. He is the firstrecipient in Maryland's College of Engineering of the dis-tinguished Lilly Teaching Award in 1991.

Ian Flood is Director of the Fire Research Centre and afaculty member of the M. E. Rinker Sr. School of BuildingConstruction, at the University of Florida, Gainesville, USA.Prior to this, he held academic positions in the Departmentof Civil Engineering at the University of Maryland, Col-lege Park, and in the School of Building and Estate Man-agement at the National University of Singapore. His currentareas of research include the development of intelligentmethods of modeling construction processes (for which hereceived the National Science Foundation Research Initia-tion Award), neural network-based approaches to the weigh-in-motion of trucks crossing steel bridges, and the modelingof uncertainty in construction costs and schedules.

Alex P. Alex is a research assistant of civil engineering atKuwait University. A native of India, he did his undergrad-uate and graduate work in Civil Engineering from Manga-lore University, India, obtaining his Bachelor and Masterdegrees in 1991 and 1993, respectively. In 1993 he joinedthe Kuwait University and since then assisted in researchoriented toward the application of new and emerging arti-ficial intelligence paradigms such as neural networks andgenetic algorithms to construction management problems.His research interests are in modelling and optimization ofcomplex decision systems using computational intelligencetechniques, and he has been sponsored by Kuwait Univer-sity and Kuwait Foundation for Advancement of Sciences,since 1993.



Documents

Construction project control using artificial neural networks · to-date project plans at regular intervals during project execution. 2. ARTIFICIAL NEURAL NETWORKS Artificial intelligence