0 évaluation0% ont trouvé ce document utile (0 vote)

511 vues7 pagesgood

© Attribution Non-Commercial (BY-NC)

PDF, TXT ou lisez en ligne sur Scribd

good

Attribution Non-Commercial (BY-NC)

0 évaluation0% ont trouvé ce document utile (0 vote)

511 vues7 pagesgood

Attribution Non-Commercial (BY-NC)

Vous êtes sur la page 1sur 7

with regression analysis and neural networks

Rifat Sonmez

Abstract: Conceptual cost estimates play a crucial role in initial project decisions, although scope is not finalized and

very limited design information is available during early project stages. In this paper, the advantages and disadvantages

of the current conceptual cost estimation methods are discussed and the use of regression, neural network, and range

estimation techniques for conceptual cost estimation of building projects are presented. Historical cost data of continu-

ing care retirement community projects were compiled to develop regression and neural network models. Three linear

regression models were considered to identify the significant variables affecting project cost. Two neural network mod-

els were developed to examine the possible need for nonlinear or interaction terms in the regression model. Prediction

intervals were constructed for the regression model to quantify the level of uncertainty for the estimates. Advantages of

simultaneous use of regression analysis, neural networks, and range estimation for conceptual cost estimating are dis-

cussed.

Key words: conceptual cost estimation, regression analysis, neural networks, range estimation.

Résumé : Les devis estimatifs de projets de construction jouent un rôle important dans les prises de décision initiales

du projet bien que la portée ne soit pas finalisée et que très peu d’information sur la conception soit disponible au

cours des premières étapes du projet. Cet article aborde les avantages et les désavantages des méthodes d’estimation

des projets de construction et discute de l’utilisation de la régression, des réseaux neuronaux et des techniques

d’évaluation de la fourchette de coûts dans l’estimation des coûts des projets de construction. Les données historiques

de coût des projets de villages-retraite de soins de longue durée ont été compilées afin de développer des modèles de

régression et de réseaux neuronaux. Trois modèles de régression linéaire ont été étudiés dans le but d’identifier les va-

riables significatives affectant le coût du projet. Deux modèles de réseaux neuronaux ont été développés afin d’examiner

le besoin possible de termes non linéaires ou de l’interaction dans le modèle de régression. Des intervalles de prévision

ont été élaborés pour le modèle de régression afin de quantifier le niveau d’incertitude des estimations. Les avantages

de l’utilisation simultanée de l’analyse de régression, des réseaux neuronaux et de l’évaluation de la fourchette de coûts

dans l’élaboration de devis estimatifs de projets de construction sont présentés.

Mots clés : devis estimatif de projet de construction, analyse de régression, réseaux neuronaux, évaluation de la four-

chette des coûts.

[Traduit par la Rédaction] Sonmez 683

Introduction since project scope is not finalized and very limited design

information is available during the predesign stages of a pro-

Conceptual cost estimates, also referred to as predesign ject. A quick, inexpensive, and reasonably accurate estimate

cost estimates, are prepared at the very early stages of a pro- is needed, however, based on the available information.

ject, generally before the construction drawings and specifi-

Several techniques have been suggested for conceptual

cations are available. At this stage, cost estimates are needed

cost estimation. Regression analysis, simulation, and neural

by the owner, contractor, designer, or lending organization

networks are among these cost estimation techniques that are

for several purposes, including determination of the feasibil-

used during the early project stages. In regression analysis

ity of a project, financial evaluation of a number of alter-

the project cost is estimated with a regression model includ-

native projects, or establishment of an initial budget.

ing a number of independent variables. Kouskoulas and Koehn

Conceptual cost estimates are not expected to be precise,

(1974) used location, construction year, building type, num-

ber of floors, quality, and building technology as the inde-

pendent variables to explain the variations in the project

Received 26 August 2003. Revision accepted 22 March 2004.

Published on the NRC Research Press Web site at cost. Karshenas (1984) argued that because the type of

http://cjce.nrc.ca on 19 August 2004. building has an important effect on the project cost, different

types of buildings should be studied separately. Karshenas

R. Sonmez. Department of Civil Engineering, Middle East focused on conceptual cost modeling for multistory steel-

Technical University, Ankara 06531, Turkey (e-mail: framed office buildings and used published data for model-

rsonmez@metu.edu.tr).

ing. The model developed by Karshenas included nonlinear

Written discussion of this article is welcomed and will be terms to express project cost in terms of typical floor area

received by the Editor until 31 December 2004. and building height. Karshenas, along with Kouskoulas and

Can. J. Civ. Eng. 31: 677–683 (2004) doi: 10.1139/L04-029 © 2004 NRC Canada

678 Can. J. Civ. Eng. Vol. 31, 2004

Koehn, performed a residual analysis to check model predic- tion, and an historical cost index (time index) was used to

tions. Neither of the models was validated by a technique quantify the variability of project cost due to inflation

such as cross validation, however. (Waier et al. 1996).

Touran (1993) suggested using the Monte Carlo simula- Total building area for a CCRC was determined by adding

tion technique with subjective correlations for probabilistic the gross building areas of residential, health center, com-

cost estimating. Published data on low-rise building projects mons, and structured parking facilities. The percent area of

were used to show the application of the suggested proce- health center and commons was calculated by dividing the

dure. The simulated data were used against actual data to sum of health center and commons area by the total building

evaluate the accuracy of the technique. Isidore and Back area. Similarly, percent parking area was the ratio of struc-

(2002) discussed the importance of quantifying the risk of tured parking area to total building area. The number of

an estimate by using probabilistic range estimating. Isidore units is the total of all residential units including studios,

and Back developed an integrated range estimation and one-, two-, and three-bedroom apartment units, and du-

probabilistic scheduling technique called MSAT and used a plexes. The area per unit was determined by dividing the to-

construction project to demonstrate the technique. tal building area by the number of units.

Hegazy and Ayed (1998) utilized neural networks for con- Seven variables were used to model total project cost. De-

ceptual cost estimation of highway projects. Hegazy and tailed cost estimates for the projects were available and were

Ayed used 14 projects for training and four projects for test- used as the dependent variable in the conceptual cost mod-

ing the neural network models. Three neural networks mod- els. The detailed cost estimates for the 30 projects were very

els with 10 input variables were developed by different close to the actual project costs, and the majority of devia-

methods. tions were due to scope changes.

The techniques involving regression analysis, simulation,

and neural networks each have certain advantages and disad-

vantages for conceptual cost estimation. The advantage of Regression analysis

regression models lies in their generally parsimonious use of In this study, parsimonious models were considered for

parameters compared to neural networks. The technique of both regression and neural network modeling. A parsimoni-

regression analysis, however, requires the user to make deci- ous model can be defined as a model that fits the data ade-

sions about the use of interaction terms and also about the quately without using any unnecessary parameters. The

class of relationships (linear, quadratic, etc.). One of the principle of parsimony is important because, in practice, par-

main advantages of probabilistic estimating techniques is simonious models generally produce better forecasts

their capability to develop a range of estimates for quantify- (Pankratz 1983). In a study by Sonmez and Rowings (1998)

ing the uncertainty of the estimate. In simulation, however, a it was observed that models including only a few significant

probability distribution function needs to be selected to parameters forecasted labor productivity more accurately

model the data. On the other hand, in the common use of than the models including insignificant parameters. To

neural network models apart from the choice of neural net- achieve parsimonious models, a backward elimination

work architecture, the user does not need to exert too much method was used in which all of the independent variables

effort to decide on the class of relations or the probability were considered in the initial regression model and variables

distribution of the variables. As each technique has certain that were not contributing to the model were eliminated one

advantages, a pragmatic approach is to use a mix of tools at a time. Two regression statistics, significance level (P

drawn from regression analysis, probabilistic estimating, and value) and coefficient of determination (R2), were used for

neural networks, which is the focus of this study. determination of variables to be eliminated. The P value

gives an indication of the significance of the variables in-

Description of the data cluded in the model, whereas R2 gives a measure of the vari-

ability explained by the model.

The data for this study were compiled from 30 continuing The first regression model (RM1) included all variables

care retirement community (CCRC) projects built by a con- and was in the following form:

tractor in the United States. A CCRC can be defined as an

organization established to provide housing and services, in- [1] C = β0 + β1T + β2L + β3 A + β4H + β5U + β6 F + β7S

cluding health care, to people of retirement age. A CCRC is

generally a mix of residential, health center, and commons in which C is the total project cost estimate (in US$); T is

buildings, and some CCRCs may also have structured park- the time index for the project year; L is the location index

ing. for the project location; A is the total building area (in m2);

The cost of CCRC projects is affected by numerous vari- H is the percent health center and commons area; U is the

ables. Construction year (time) and location, total building total area per unit (in m2); F is the number of floors; S is the

area, combined percent area of health center and commons, percent structured parking area; and β0, β1, …, β7 are the re-

area per unit, number of floors, and percent area of struc- gression coefficients. Data from the 30 projects were used to

tured parking were the variables included in this study. The determinate the regression coefficients and statistics.

projects used in this study were built in 14 different states Regression statistics for RM1 are given in Table 1, includ-

during the time frame 1975–1995. The number of floors for ing R2 and the variable corresponding to the coefficient with

the residential, health center, and commons buildings was the highest P value. The P value 0.663 for the regression co-

between one and 12. A city cost index (location index) was efficient corresponding to the percent parking (P) indicate

used to quantify the variability in project cost due to loca- that this variable probably will not have a significant contri-

Sonmez 679

Independent Variable corresponding to the P value of the

Model variables R2 coefficient with the highest P value coefficient

RM1 T, L, A, H, U, F, S 0.951 S 0.663

RM2 T, L, A, H, U, F 0.950 F 0.383

RM3 T, L, A, H, U 0.949 L 0.110

bution to the model and therefore was dropped from the re- Table 2. P values for regression model RM3.

gression model to achieve a parsimonious model.

Independent variable P value of the coefficient

The second regression model (RM2), which included six

independent variables, was similar to RM1 but did not in- T 0.000

clude the variable percent parking. The R2 value for RM2 L 0.110

has only dropped to 0.950 from the value 0.951 for RM1 A 0.000

with the elimination of the variable percent parking (S). In H 0.049

model RM2 the coefficient for the number of floors (F) had U 0.095

a P value of 0.383, and this variable was dropped from the

model to obtain the third regression model RM3. In RM3,

the coefficient for the variable location index (L) had the

highest P value; however, this variable was kept in the Fig. 1. Neural network model.

model because the contribution of this variable with a P

value of 0.11 was considered to be significant. The P values

for the regression coefficients given in Table 2 indicate that

the contribution of the variables time index (T) and total

building area (A) was more significant than the contribution

of the variables location index (L), percent health center and

commons area (H), and the area per unit (U). Coefficients of

all five variables were significant at the 0.11 level, however,

and were included in the final regression model.

In the third regression model RM3, although an R2 value

of 0.949 indicates that the majority of variability in cost is

explained by the five independent variables, there may be

certain significant nonlinear or interaction relations that had

not been included in the model. A neural network model was

developed to examine whether the current linear model RM3

could be improved further.

Two feed-forward neural network models, as shown in The final regression model (RM3) and neural network

Fig. 1, were developed as an alternative to the regression models (NM1 and NM2) were compared in terms of close-

cost model RM3. The input variables of the neural networks ness of fit and prediction performance. Two error measures

developed for conceptual cost estimating included time in- were used for the comparison: mean squared error (MSE)

dex (T), location index (L), total building area (A), percent and mean absolute percent error (MAPE). MSE and MAPE

health center and commons area (H), and total area per unit were calculated as follows:

(U), which were also the independent variables of the final n

∑ (actuali − predictedi)2

1

regression model. The output variable of the neural networks [2] MSE =

consisted of the dependent variable total project cost esti- n i =1

mate (C). n actual − predicted

∑

Two neural networks with a different number of hidden 1 i i

[3] MAPE = × 100

units were trained using the same variables with a back- n i =1 predictedi

propagation algorithm incorporating a sigmoid transfer func-

tion (Rumelhart et al. 1986). The purpose of training two in which i is the project number. The measures for closeness

neural networks was to seek a number of hidden units that of fit were calculated by using data from all 30 projects to

would result in an adequate prediction performance with a develop the models and then by comparing the model pre-

reasonable closeness of fit. The first neural network model dictions with the actual data.

(NM1) had six hidden units (the number of units in the input Prediction performances of the models were compared by

buffer plus the number of units in the output layer), and the a procedure based on the cross-validation technique. The

second neural network model (NM2) had three hidden units. procedure can be summarized in the following steps:

The neural networks were trained with the 30 project cases (1) Three projects were selected randomly as the test sam-

that were used to develop the regression models. ple and a new data set was formed. The training set in-

680 Can. J. Civ. Eng. Vol. 31, 2004

Model MSE MAPE Model MSE MAPE

RM3 2.1 × 1012 9.3 RM3 3.3 × 10 12

11.1

NM1 9.6 × 1011 8.5 NM1 3.6 × 1012 12.3

NM2 1.3 × 1012 8.6 NM2 3.8 × 1012 11.7

cluded data from the remaining 27 projects but not the Fig. 2. Impact of building area on cost per unit area.

data from the three projects selected as the test sample.

(2) Model parameters for the regression and neural network

models were determined with the training data set.

(3) The regression and neural network models with the new

parameters were used to predict costs of the projects,

which were selected as the test sample.

(4) All of the projects were selected as the test sample in

groups of three, and steps 1–3 were repeated for each of

the test samples.

(5) Mean squared error (MSE) and mean absolute percent

error (MAPE) were calculated for the regression and

neural network models.

The MSE and MAPE values of the final regression model

(RM3) and neural network models (NM1 and NM2) for

closeness of fit are given in Table 3. The MSE and MAPE

values of the neural network models NM1 (9.6 × 1011 and

8.5, respectively) and NM2 (1.3 × 1012 and 8.6) were Fig. 3. Impact of time index on cost per unit area.

smaller than those of the regression model RM3 (2.1 × 1012

and 9.3), indicating that the neural network models provided

better fit to the data. On the other hand, however, the MSE

and MAPE values of the regression model RM3 (3.3 × 1012

and 11.1, respectively) for prediction performance given in

Table 4 were smaller than those of the neural network mod-

els NM1 (3.6 × 1012 and 12.3) and NM2 (3.8 × 1012 and

11.7). Although neural network models provided better fit to

the data, the regression model had reasonably good predic-

tion performance; therefore; the regression model RM3

could be considered as an adequate model, and there is no

need to add nonlinear or interaction terms to the model. Al-

though neural network model NM1 had a better fit to data,

both neural network models had similar prediction perfor-

mance and NM2 was used for sensitivity analysis because it

was more parsimonious.

For 29 of the 30 cost estimates predicted by the regression

model the accuracy was within ±24%, and for one project ues, and time index (T) was kept constant at the reference

the accuracy was within ±40%. The accuracy range ±24% is year value to obtain Fig. 2. Figure 2 demonstrates the effect

acceptable and is within the range of –30% to +50% sug- of variable building area (A) on the project cost per unit area

gested by AbouRizk et al. (2002) for conceptual cost estima- for the regression (RM3) and neural network (NM2) models.

tion of building projects. The average absolute accuracy Both the regression and neural network models indicate that

11.1% is also acceptable and is within the suggested range cost per unit area decreases as the total building area in-

of –15% to +25% for the average accuracy (AbouRizk et al. creases. This is mainly due to decreasing indirect cost per

2002). unit area as a result of an increase in the project size and de-

creasing contribution of the exterior walls due to an increase

Model results in the building area. The sensitivity analysis of neural net-

work model NM2 indicates that cost per unit area increases

Sensitivity analysis was used to demonstrate the effect of to some extent from 15 000 to 25 000 m2, probably due to

independent variables on the project cost for the regression characteristics of project data captured by the neural net-

(RM3) and neural network (NM2) models. To perform sensi- work. Figure 3 demonstrates the impact of time index (T) on

tivity analysis, the values of one variable were varied while the project cost per unit area. Both the regression and neural

those of the remaining variables were kept constant. The val- network models suggest similar relations between time in-

ues of the variable total building area (A) were varied, loca- dex and project cost. Similarly, Fig. 4 demonstrates the im-

tion index (L), percent health center and commons area (H), pact of location index (L) on the project cost per unit area.

and area per unit (U) were kept constant at their mean val- The magnitude of the impact of location index on the project

Sonmez 681

Fig. 4. Impact of location index on cost per unit area. Fig. 6. Impact of area per unit on cost per unit area.

Fig. 5. Impact of percent health center and commons area on the project data. There was only one project with 12 floors,

cost per unit area. with the other projects having between one and six floors.

The independent variables included in this study were lim-

ited to seven, since the contractor compiled information only

on these variables. Study of other variables that may influ-

ence project cost related project management, site condi-

tions, and building and project characteristics that could be

identified at the conceptual stages would possibly improve

the accuracy of the models presented. On the other hand,

however, the R2 value 0.949 of regression model RM3 indi-

cates that the majority of variability in cost is explained by

the five independent variables, and also the average absolute

accuracy of 11.1% is acceptable for a conceptual cost

model.

Prediction intervals

The neural network and regression models developed pro-

vide a point estimate for the project cost. During conceptual

cost captured by neural network model NM2 is more signifi- stages of a project, however, since the scope is not finalized

cant than the magnitude of the impact suggested by regres- and very limited design information is available, there are

sion model RM3. Sensitivity analysis of the both the numerous uncertainties that cannot be reflected by a point

regression and neural network models indicates that the cost estimate. Therefore, the use of range estimating is essential

per unit area increases as the percent health center and com- for quantifying the level of uncertainties.

mons area (H) increases (Fig. 5). Cost per unit area is ex- The use of prediction intervals for the regression models

pected to increase as the percent health center and commons makes range estimating possible. Before prediction intervals

area increases because the cost per unit area for the health are constructed, the assumptions of zero expectation, con-

center and commons buildings is usually more expensive stant variance, and independence of residuals are examined

than that of the residential buildings. Lastly, the sensitivity by residual plots for model RM3. One of the important

analysis of the regression and neural network models also assumptions for constructing prediction intervals is the nor-

indicates that cost per unit area decreases as the area per unit mality of the residuals. The normality assumption of residu-

increases (Fig. 6). This is because, for a given total building als can be examined using quantile to quantile plots and the

area, as the number of units decreases, the cost of accesso- correlation coefficient for the quantile to quantile plot (John-

ries per unit area associated with the number of units (such son and Wichern 1992), which was calculated as 0.97. Since

as plumbing and kitchen and bathroom accessories) is ex- this value is larger than the value 0.96 at n = 30 and α =

pected to decrease, resulting in a lower cost per unit area. 0.05 level of significance, the hypothesis for normality was

The magnitude of the impact of area per unit (U) on the pro- not rejected.

ject cost captured by neural network model NM2 is more Prediction intervals for the cost estimate using mean val-

significant than that suggested by regression model RM3. ues of the variables and reference time index were con-

As a result of regression analysis, the effects of the vari- structed for model RM3. Figure 7 illustrates a range of cost

ables percent area of structured parking (S) and number of predictions for different probability values. Estimates repre-

floors (F) on project cost were not found to be significant senting different levels of risk can be selected using the

for inclusion in the models. The insignificance of the vari- probability–cost curve. As an example, a 20–90% range of

able number of floors may be due to the characteristics of probability would be US$18 625 915 to US$22 463 544; in

682 Can. J. Civ. Eng. Vol. 31, 2004

(A) Model inputs

Time Location Residential Health center Commons

Total building area (m2) index index unit count area (m2) area (m2)

25 960 100.0 106.4 147 1795 2871

Lower Expected Higher

Probability of beating the noted estimate (%) 20 50 10

Total programmed project cost (US$) 21 437 388 23 028 230 25 450 391

Fig. 7. Probability–cost (US$) curve. range for the prediction. The cost prediction of the model for

the case project was US$23 028 230. The model also indi-

cated that the cost of the project could be less than

US$21 437 388 with a probability of 20%, and also the cost

could be higher than US$25 450 391 with a probability of

10%. The cost estimate of the case project was calculated as

US$22 600 989 after quantity take-off and detailed estima-

tion was performed. The detailed cost estimate was within

the cost prediction range of the model and very close to the

expected cost.

Conclusions

Conceptual cost models for continuing care retirement

community projects were developed with regression analysis

and neural networks. The models were compared for close-

other words, for this range the probability of the detailed ness of fit and prediction performance. Both of the modeling

cost estimate being less than the lower value (US$18 625 915) techniques provided reasonably accurate cost estimates.

would be 20% and the probability of the detailed cost esti- Once the models were developed, cost estimates could be

mate being more than the upper value (US$22 463 544) made immediately.

would be 10%. A nonsymmetrical range was selected to de- Neural networks can identify the relations between vari-

crease the chance of an underestimate. A point prediction of ables and the project cost. Regression analysis, however,

US$20 147 235 for the project cost is limited in terms of requires a decision about the class of relations (linear, qua-

quantifying possible variability of the conceptual cost esti- dratic, etc.) to be used in modeling. On the other hand, the

mate, but the range estimate presented provides crucial in- regression models, most of the time, use fewer model pa-

formation about the uncertainty of the estimate. rameters than the neural networks, which may result in

In this study, linear regression analysis provided adequate better prediction performance if the relations between the

models for conceptual modeling of construction costs for variables are presented adequately. Comparison of the re-

building projects. Depending on the project type and charac- sults from the neural network model with those from the re-

teristics of data used for modeling, however, and also de- gression model may help to identify the need for nonlinear

pending on the independent variables included in the or interaction terms in the regression model. By using re-

models, neural network models can also be selected for pre- gression analysis and neural network techniques simulta-

diction of construction costs at the conceptual project stages. neously, a satisfactory conceptual cost model (which fits the

It is also possible to develop prediction intervals for the neu- data adequately and has a reasonable prediction perfor-

ral network models by applying techniques similar to the re- mance) can be achieved.

gression analysis (Dybowski and Roberts 2001). Prediction intervals were constructed for the regression

model for range estimation. The use of range estimation dur-

Case example ing early project stages would help to quantify the uncertain-

ties involved with the conceptual cost estimate. The writer

The latest project was selected as a case project to illus- believes that the techniques presented in this paper could im-

trate the use of the developed conceptual cost model. The re- prove estimation of costs in the early stages of a project. Im-

maining 29 projects were used to determine the regression provements in conceptual cost estimates would hopefully

model used to predict the detailed cost estimate for the case lead to more realistic expectations and better execution strat-

project. The model can be utilized as a spreadsheet program egies.

so that the user can easily make conceptual cost range esti-

mates (Table 5). Once the user enters the input values related

to the total building area, time and location indices, number

References

of residential units, health center area, and commons area, AbouRizk, S.M., Babey, G.M., and Karumanasseri, G. 2002. Esti-

the model can predict an expected cost and a 20%–90% mating the cost of capital projects: an empirical study of accu-

Sonmez 683

racy levels for municipal government projects. Canadian Journal function for buildings. Journal of the Construction Division,

of Civil Engineering, 29: 653–661. ASCE, 100(CO4): 589–604.

Dybowski, R., and Roberts, S.J. 2001. Confidence intervals and Pankratz, A. 1983. Forecasting with univariate Box–Jenkins mod-

prediction intervals for feed-forward neural networks. In Clini- els. John Wiley & Sons, New York. pp. 81–82.

cal applications of artificial neural networks. Edited by R. Rumelhart, D.E., Hinton, G.E., and Williams, R.J. 1986. Learning

Dybowski and V. Gant. Cambridge University Press, Cam- internal representations by error propagation. In Parallel distrib-

bridge, UK. pp. 298–327. uted processing. Vol. 1. Foundations. Edited by D.E. Rumelhart

Hegazy, T., and Ayed, A. 1998. Neural network model for paramet- and J.L. McClelland. MIT Press, Cambridge, Mass. Chapt. 8.

ric cost estimation of highway projects. Journal of Construction Sonmez, R., and Rowings, J.E. 1998. Construction labor productiv-

Engineering and Management, 124(3): 210–218. ity modeling with neural networks. Journal of Construction En-

Isidore, J.L., and Back, W.E. 2002. Multiple simulation analysis gineering and Management, 124(6): 498–504.

for probabilistic cost and schedule integration. Journal of Con- Touran, A. 1993. Probabilistic cost estimation with subjective cor-

struction Engineering and Management, 128(3): 211–219. relations. Journal of Construction Engineering and Management,

Johnson, R.A., and Wichern, D.W. 1992. Applied multivariate sta- 119(1): 58–71.

tistical analysis. Prentice Hall Inc., Englewood Cliffs, N.J. Waier, P.R., Akins, T.J., Chiang, J.H., Ferguson, J.H., Kaplan,

Karshenas, S. 1984. Predesign cost estimating method for multi- M.H., Jr., Mewis, R.W., Mossman, M.J., Moylan, J.J., Murphy,

story buildings. Journal of Construction Engineering and Man- J.D., and Page, J.R. (Editors). 1996. Building construction cost

agement, 110(1): 79–86. data 1995. 54th ed. R.S. Means Company, Kingston, Mass.

Kouskoulas, V., and Koehn, E. 1974. Predesign cost-estimation