Professional Documents
Culture Documents
of Change Orders
Awad S. Hanna, P.E., M.ASCE1; Wafik B. Lotfallah2; and Min-Jae Lee, S.M.ASCE3
Abstract: This paper presents a hybrid approach to quantify the impact of change orders on construction projects using statistical
regression and fuzzy logic. There are many qualitative variables affecting the impact of change orders on labor productivity; statistical
analysis falls short of addressing the fuzziness of those variables. Because of their complementary nature, fuzzy logic and regression
analysis can be integrated; regression analysis is used to determine the membership functions of the input linguistic values. In this paper,
each input variable is statistically treated before entering a general rule relating its space to the space of loss in labor productivity. The
relative weight of each input variable is determined by its coefficient of determination (R 2 ) value. The expected loss of labor productivity
and its standard deviation are then determined from the output fuzzy membership function. The proposed methodology is general and can
be applied in areas of system analysis and decision making when a complex input-output function is to be predicted in the presence of
some fuzzy knowledge and a large number of real input-output data.
DOI: 10.1061/共ASCE兲0887-3801共2002兲16:4共252兲
CE Database keywords: Productivity; Fuzzy sets; Change order; Construction.
Hybrid Statistical-fuzzy Method The owner’s budget or schedule might change and force a reduc-
tion of scope. Unforeseen natural events might occur. All of these
This paper presents a hybrid system that encapsulates both statis- reasons for change frequently cause disruption in the planned
tical analysis and fuzzy logic to study the effect of the quantita- work schedule, and result in increased costs through rework and
tive and qualitative input variables on enhancing or reducing the decreased efficiency of the base contract work. Some examples of
impact of change orders on labor productivity. Regression analy- causes of inefficiency due to change orders are 共Fig. 1兲
sis is used here to determine the membership functions of the
• Increased frequency of planning and replanning,
input linguistic values as well as the forms of the if-then rules.
• Loss of efficiency due to interruption, interference, and lack of
Equivalently, each input variable is statistically treated before en-
availability of tools, labor, and materials to meet the require-
tering a general rule relating its space to the space of loss in labor
productivity. Also, the relative weight of each input variable is ments of the changes,
determined by its coefficient of determination (R 2 ) value. When • Increased project management and supervision involvement,
linguistic input values are determined, the general rules are sepa- • Loss of efficiency due to a ripple impact that is a direct result
rately fired and the resulting fuzzy values are aggregated with of change orders, such as stacking of trades, schedule com-
their relative weights. The expected loss of labor productivity as pression, and out-of-sequence work, and
well as its standard deviation are then determined from the output • Difficulty in determining equitable adjustment compensation
fuzzy membership function. for the parties involved.
The novel approach in this paper is the use of statistical data to If x is L, then y is P
choose the fuzzy membership functions and to form the fuzzy The values of H, M, and L are determined in a similar way.
if-then rules. The following example illustrates this method. Our method can be applied even if the regression function y
Suppose that a fuzzy if-then rule is to be formed between the ⫽ f (x) is not a linear function. Another way to view this method
universe of an input variable x, and an output variable y. Assume is that we used the function f (x) to map the input variable x to an
further that x lives in the unit interval 关0, 1兴 and y lives in the intermediate variable u⫽ f (x) such that the relationship between
interval 关⫺1, 1兴. For the output variable y, we fix the linguistic u and the output variable y can be represented by an identity
values P 共for positive兲, N 共for negative兲, and Z 共for zero兲, whose fuzzy functional; i.e., we have
fuzzy membership functions are given by
再
If u is N, then y is N; If u is Z, then y is Z;
y for 0⭐y⭐1
P共 y 兲⫽ ; If u is P, then y is P
0 otherwise
再
The above functional can be generalized to the following state-
⫺y for ⫺1⭐y⭐0 ment:
N共 y 兲⫽ ;and
0 otherwise
If u is A, then y is A
Z共 y 兲⫽ 再 1⫺y
1⫹y
for 0⭐y⭐1
for ⫺1⭐y⬍0
where A⫽any fuzzy set.
Thus, by statistically treating the input x, we managed to simplify
We also have some statistical data represented by the points the fuzzy rules used to the general rule
(x i ,y i ), for i⫽1,...,n, from which we get the regression line If f 共 x 兲 is A, then y is A
y⫽ f 共 x 兲 ⫽ax⫹b (1)
where the values of a and b are picked to minimize the sum of the Fuzzy Set Concepts „Aggregating Rules…
squares of the errors
n Starting with k input variables x 1 ,x 2 ,...,x k affecting an output
E⫽ 兺
i⫽1
共 y i ⫺ax i ⫺b 兲 2 variable y, we use the method of the previous section to get a
separate regression function u i ⫽ f i (x i ) for each input variable x i
In traditional regression analysis x and y are correlated and a with the generalized fuzzy rule
⫽0. We then prefer to write Eq. 共1兲 in the form Ri If f i 共 x i 兲 is A, then y is A
y⫽ f 共 x 兲 ⫽a 共 x⫺c 兲 Now suppose that for each input variable x i we have the fuzzy
Let us assume for now that a⬎0. We then seek some if-then rules statement
of the form f i 共 x i 兲 is A i
If x is H, then y is P; If x is M , then y is Z; Since our belief in rule Ri can be measured by the coefficient
If x is L, then, y is N of determination R 2i of the correlation between x i and y, we can
deduce that our belief that ‘‘y is A i ’’ is also measured by R 2 . We
where the values H 共for high兲, M 共for medium兲, and L 共for low兲 can then aggregate the fuzzy membership A i by the formula
are yet to be determined.
According to the above rules, it is plausible to say that our 兺 i⫽1
k
R i2 •A i 共 y 兲
belief-strength that some particular value of x is high can be iden- A共 y 兲⫽ (2)
兺 i⫽1
k
R i2
tified with our belief-strength that the corresponding predicted
value of y is positive. This suggests that we can define the fuzzy Calling
membership function H by
R i2
H 共 x 兲 ⫽ P 关 f 共 x 兲兴 w i⫽
兺 i⫽1
k
R i2
and similarly Eq. 共2兲 can be written in the form
M 共 x 兲 ⫽Z 关 f 共 x 兲兴 and L 共 x 兲 ⫽N 关 f 共 x 兲兴 k
L共 x 兲⫽ 再 a 共 c⫺x 兲
0
for x⭐c
otherwise
; and y *⫽
兰 y•A 共 y 兲 dy
兰 A 共 y 兲 dy
representing the expected value of the output y, and the standard since the R 2 value of %Additions is 0.000, this variable (x 5 ) will
deviation with not appear in our statistical-fuzzy model.
兰 共 y⫺y * 兲 2 •A 共 y 兲 dy
2⫽
兰 A 共 y 兲 dy Case Study
representing the accuracy of our estimation. In practice we may
want to use discrete spaces for the input and the output variables, To illustrate our method, we show the calculations for the follow-
and replace integration by summation. ing case study, where each input variable takes either a given
crisp value or a fuzzy value determined by the available informa-
tion:
Application: Impact of Change Orders on 1. Not an industrial project (x 1 ⫽0).
Productivity 2. Estimated/actual peak work hours is 0.71 (x 2 ⫽0.71).
3. Change orders initiated by the owner is 90% (x 3 ⫽0.9).
Using regression analysis, Hanna et al. 共1999a, b兲 developed a 4. Change orders hours approved by the owner is 81.3% (x 4
linear model that quantifies the impact of change orders on labor ⫽0.813).
productivity. In their model, the output dependent variable was 5. Additions or deletions of change order hours is 75% (x 5
the percent loss of labor productivity 共Delta兲. They pointed out ⫽0.75).
some significant independent variables affecting Delta. Table 1 6. The contractor’s productivity tracking system is not
gives the definition and the range of possible values for each sophisticated, but adequate. In this case, we assume that
independent variable. The regression model found was x6 has the following fuzzy value: x6
⫽ 关 0.0,0.2,0.4,0.6,0.8,1.0,0.8,0.6,0.4,0.2,0.0兴 .
%Delta⫽0.3495 7. Absenteeism was not recorded, but the contractor identified
his strength of belief about the absenteeism ratio as follows:
⫹0.139 Industrial 0–5% with strength 70%; 6 –10% with strength 30%. Thus,
as a fuzzy subset of 兵1,2,3,4其, x 7 ⫽ 关 0.7,0.3,0.0,0.0兴 .
⫺0.0984 EA – P
Next we determine the fuzzy sets A i representing f i (x i ). As
Delta ranges between ⫹50 and ⫺50%, we take its discrete uni-
⫺0.0368%OwnInitiatedCO
verse to be
⫺0.190%OwnInitiatedCO* Industrial D⫽ 兵 ⫺0.5,⫺0.4,⫺0.3,⫺0.2,⫺0.1,0.0,0.1,0.2,0.3,0.4,0.5其 .
⫺0.100%COHrsApproved Since the input variables x 1 ,...,x 5 take crisp values, their re-
gression outputs f 1 (x 1 ),..., f 5 (x 5 ) take the following crisp values:
⫹0.0627%Additions
A 1 ⫽ f 1 共 0 兲 ⫽0.077⬇0.1
⫺0.0593 Productivity-Track
A 2 ⫽ f 2 共 0.71兲 ⫽0.115⬇0.1
⫹0.0544 Absenteeism (3) A 3 ⫽ f 3 共 0.9兲 ⫽0.035⬇0.0
From their statistical findings, we get a separate regression A 4 ⫽ f 4 共 0.813兲 ⫽0.080⬇0.1
linear function for Delta against each input variable. Table 2
shows those functions together with their R 2 value. Note that A 5 ⫽ f 5 共 0.75兲 ⫽0.0912⬇0.1
%Own %COHrs
Industrial EA – P Initiated CO Approved %Additions Productivity-Track Absenteeism Project Statistical Statistical
Case project (x 1 ) (x 2 ) (x 3 ) (x 4 ) (x 5 ) (x 6 ) (x 7 ) actual model fuzzy model
Example No 共0兲 0.71 0.9 0.813 0.75 Not sophisticated, 0–5% with strength 70%; 10.80 23.70 7.20
project but adequate 共1兲 6 –10%
with strength 30%
M3132 Yes 共1兲 1.21 1 0.81 0.62 No 共0兲 0–5% 7.24 15.49 9.65
M3182 No 共0兲 0.43 0.4 0.8 0.75 Yes 共1兲 0–5% 7.90 25.46 9.72
M3362 Yes 共1兲 3.96 0.75 0.75 0 No 共0兲 11.20% 6.52 1.70 5.20
E042-2 Yes 共1兲 0.45 0.923 0.846 0.4 Yes 共1兲 0–5% with strength 13.27 17.05 9.54
70%; 6 –10%
with strength 30%
E131 No 共0兲 0.82 1 0.94 0.65 Yes 共1兲 0–5% with strength 10.61 17.39 8.51
70%;
6 –10%
with strength
30%
E262 Yes 共1兲 0.28 0.95 0.5 0.65 Yes 共1兲 Greater than 20% 10 39.45 25.31
with strength 70%;
12–20%
with strength 30%
Since x 6 takes a fuzzy value, the outputs f 6 (x 6 ) take the follow- Thus, as a subset of D
ing fuzzy value:
A 7 ⬇ 关 0,0,0,0,0,0.7,0.5,0.3,0,0,0兴
A 6 ⫽ f 6 共关 0/0⫹0.2/0.1⫹0.4/0.2⫹0.6/0.3⫹0.8/0.4⫹1.0/0.5
where we added the entry 0.5⫽(0.7⫹0.3)/2 to maintain the con-
⫹0.8/0.6⫹0.6/0.7⫹0.4/0.8⫹0.2/0.9⫹0/1.0 兴 兲 vexity of A 7 .
Now we use A(y)⫽ 兺 i⫽1 k
w i A i (y) to find the output fuzzy set A
⬇ 关 0/0.13⫹0.4/0.12⫹0.6/0.11⫹0.8/0.10⫹1.0/0.09⫹0.6/0.08 representing the relative loss in labor productivity, as follows:
⫹0.4/0.07⫹0.2/0.06兴 .
A⫽0.011关 0,0,0,0,0,0,1,0,0,0,0 兴
To get that value, we applied f 6 of Table 2 on the elements of the
universe 共denominators兲, rounded up the values to two decimal ⫹0.268关 0,0,0,0,0,0,1,0,0,0,0,兴
digits, and added the membership values 共numerators兲 corre-
⫹0.203关 0,0,0,0,0,1,0,0,0,0,0 兴
sponding to the same denominator.
If we further try to represent A 6 as a subset of D, we have to ⫹0.153关 0,0,0,0,0,0,1,0,0,0,0 兴
round up the denominators even more to one decimal digit. How-
ever, as the regression function f 6 has a very small slope, A 6 gets ⫹0.000关 0,0,0,0,0,0,1,0,0,0,0 兴
approximated by the crisp number
⫹0.069关 0,0,0,0,0,0,1,0,0,0,0 兴
A 6 ⫽0.1
⫹0.296关 0,0,0,0,0,0.7,0.5,0.3,0,0,0兴
indicating that our discrete model cannot capture the fuzziness of
the variable x 6 . Also, x 7 takes a fuzzy value and the correspond- ⫽ 关 0,0,0,0,0,0.410,0.649,0.089,0,0,0兴
ing f 7 (x 7 ) takes the fuzzy value
Thus, the expected loss in labor productivity 共centroid兲 is
A 7 ⫽ f 7 共关 0.7/1⫹0.3/2⫹0.0/3⫹0.0/4 兴 兲 ⫽ 关 0.7/0.039⫹0.3/0.160
兺 y•A 共 y 兲 0 * 0.410⫹0.1* 0.649⫹0.2* 0.089 0.0827
y *⫽ ⫽ ⫽
⫹0.0/0.281⫹0.0/0.402兴 ⬇ 关 0.7/0.0⫹0.3/0.2⫹0.0/0.3 兺A共 y 兲 0.410⫹0.649⫹0.089 1.148
⫹0.0/0.4兴 ⫽0.072⬇7.2%