You are on page 1of 9

A comparative study of learning curves with

forgetting
Mohamad Y. Jaber and Maurice Bonney

Department of Manufacturing Engineering and Operations Management, University of Nottingham,


University Park, Nottingham, UK

Although there is almost unanimous agreement that the form of the learning curve is as presented by Wright,’
scientists and practitioners have not yet developed a full understanding of the behavior and factors affecting
the forgetting process. Mathematical models of the forgetting process are reviewed, and then three models, the
VRZF, VRVF, and LFCM models are compared and their differences and similarities are discussed. 0 1997
by ELseuier Science Inc.

Keywords: learning, forgetting, minimum break

1. Background point of interruption. They showed that their model was


able to predict the time estimates resulting from a labora-
Industrial learning curves have been studied for almost six
tory study conducted by Globerson et al.” with less than
decades, and the results have been comprehensively sur-
1% error. The Globerson et al9 experiment was per-
veyed by Yelle’ and Belkaoui.3 However a full under-
formed in a computer-oriented manufacturing environ-
standing of the behavior and factors affecting the forget-
ment simulated in a microcomputer laboratory. The 120
ting process has not yet been developed. Hoffman4 and
subjects participated in the experiment in which they
Adler and Nanda’ presented two refined mathematical
performed a data-entry task. Each subject participated in
techniques for incorporating the effects of production
breaks into planning and control models. Elmaghraby6 two sessions, separated by a break. In each session the
participant had to key in 16 forms, containing information
presented the variable regression to invariant forgetting
(VRIF) model in which he assumes that both learning and on 16 jobs. The break length varied from 1 to 82 days (a
forgetting are functions of time.* The theory assumes that typical data set is presented in Figure I of Globerson et
a1.9). In a later study Shtub et al.” used the forgetting
the longer the period of production, the more the produc-
tivity increases, whereas the longer the stoppage, the model developed by Globerson et al.9 (simulated data-en-
greater the forgetting. Elmaghraby hypothesised that there try setting), as well as the data, to validate Bailey’s”
is a unique forgetting function that intercepts the axis (assembly/disassembly setting) hypothesis that forgetting
representing the time to produce a unit at y (later called is a function of the amount of learning prior to the
T’). The fixed value for y is either given a priori or is interruption and the elapsed time of the interruption, and
derived on the basis of the first lot produced and a to identify a power forgetting model and estimate its
“doubling factor,” F, which together define the forgetting parameters. Shtub et al.” used stepwise regression, i.e.,
function. This differs from the assumption made by the the same method as Bailey,” to identify the most influen-
Carlson and Rowe model,’ referred to by Elmaghraby6 as tial independent variables. Fourteen variables were used.
the variable regression to variable forgetting (VRVF) model, Bailey’s hypothesis was confirmed, since the two variables,
which assumed that y varied with each new interruption. the amount of learning and the length of the break,
The Jaber and Bonney’ learn-forget curve model (LFCM) entered the regression equation while the learning rate
used a forgetting slope (rate) that is dependent on: (1) the and initial performance time did not. The power forget-
minimum production break, t,, over which total forgetting ting model developed by Shtub et al.,” which was based
is assumed to occur; (2) the learning slope (rate), 1; and on Wright’s power function, required the estimation of
(3) the amount of equivalent units accumulated, u, by the

*In this paper we will represent both learning and forgetting as


Address reprint requests to Dr. M. Y. Jaber at the American Univer- functions of the output quantity.
sity of Beirut, Faculty of Engineering and Architecture, 830 Third
Avenue, 18th Floor, New York, NY 10022-6297. Received 18 August 1995; revised 7 May 1997; accepted 23 May 1997

Appl. Math. Modelling 1997, 21523-531, August


0 1997 by Elsevier Science Inc. 0307-904x/97/$17.00
655 Avenue of the Americas, New York, NY 10010 PII S0307-904X(97f00055-3
Learning curves compared with forgetting: M. Y. Jaber and M. Bonney

two parameters, the time to produce the first unit after an of a power form:
interruption and the learning slope, whereas the Glober-
son et a1.9 model required the estimation of one parame- 9x= f,xf (2)
ter, the time to produce the first unit after an interrup-
tion. Learning followed the same curve, that of Wright,’ as illustrated in Figure I where fX is the time for the xth
as before the break. Badiru” presented an approach to unit of lost experience of the forget curve, x is the
developing multivariate (multifactor) models of learning amount of output units that would accumulate if interrup-
curves that account for alternate periods of learning and tion did not occur, T, is the equivalent time for the first
forgetting. To demonstrate the usefulness of a multifactor unit (intercept) of the forget curve, and f is the forgetting
learning curve Badiru ‘* added a production break variable slope.
into the univariate model given by Wright.’ The resulting Assume that an interruption occurs immediately after
bivariate model was of a similar form to that developed by producing the qth unit and that, in intermittent produc-
Globerson et a1.9 Dar-El et a1.13 considered long-cycle tion runs, the gap is of sufficient length that some of the
tasks composed of a sequential set of short nonrepetitive learning accumulated when producing the previous lots is
cycle tasks. They addressed forgetting as a consequence of not retained. When forgetting is accounted for the effort
a specific sub-task, reappearing in the next cycle after a required to produce the first unit after the interruption is
whole cycle time of other activities is completed. Dar-El et higher than the effort required to produce the last unit in
aLI used Wright’s power model to depict the learning the previous cycle. The increase in time to produce the
phenomenon. In their authors experience Wright’s power first unit in the next production run depends on the length
model is the only model used in practice and possibly is of the interruption and the time to produce the 4 th unit.
the only learning model covered in textbooks on produc- When the next production run starts, learning recom-
tion and operations management. Dar-El et a1.13 con- mences.
cluded that the main factors involved in the learning Three of the models mentioned above, namely, VRIF,
behavior of long-cycle tasks are the task length, the learn- VRVF, and LFCM, are now examined in greater detail.
ing slope, the time for executing the first cycle, and the Elmaghraby6 in his VRIF model assumes that there is a
length of the break over which forgetting is assumed to unique forgetting function with a single intercept point.
occur. The Jaber and Bonney’ assumptions are consistent The rationale for this hypothesis is that the intercept, T,,
with these. and the forgetting slope, f, are system-dependent parame-
This paper examines the forgetting models by ters similar to Tl and I of the learning function given in
Elmaghraby6 (VRIF), Carlson and Rowe’ (VRVF), and equation (1). The VRVF model assumes that the intercept
Jaber and Bonney’ (LFCM), each of which uses the of the forgetting function, T,, varies for each lot. The
learning curve phenomenon reported by Wright.’ The rest VRIF and the VRVF models each adopt a fixed forgetting
of the paper is organised as follows. Section 2 provides an slope. The Jaber and Bonney’ LFCM modifies the VRVF
introduction to the theory of the learning and forgetting model by using a forgetting slope that is not fixed. The
curves. Section 3 presents the mathematics required for intercept of the forgetting function, as in Carlson and
the analysis of the VRIF, VRVF, and LFCM forgetting Rowe,’ also varies. A detailed discussion of each model is
models. Section 4 illustrates numerically the differences now presented. A study of both the VRVF and the VRIF
and similarities among the three forgetting models pre- models is discussed in Elmaghraby.6
sented in Section 3. Section 5 investigates the accuracy of
the LFCM model. Section 6 offers a summary and conclu-
sions.
3. The mathematics of the forgetting curves
This section presents the mathematics required for the
2. Introduction VRIF, VRVF, and LFCM models to estimate the future
performance of an intermittent production operation.
The learning curve phenomenon reported by Wright’
implies that the unit production time decreases by a
constant percentage (e.g., 90%, 80%, etc.) each time the 3.1 The KWF model
cumulative quantity produced doubles. Wright’s learning
curve expresses an exponential relationship between di- Elmaghraby’ hypothesised that there is a tmique forget-
rect man-hour input and cumulative production in the ting function with a single origin point T, (previously
form: called y), which either can be given a priori or can be
derived on the basis of the first lot produced. This to-
gether with a “doubling factor” (forgetting slope), f, de-
7; = T, j-’ (1) fines the forgetting function presented in equation (2).
Consider now that there are several production runs.
where 5 is the time to produce the jth unit, j is the F&re 2 assumes that an interruption to manufacturing
cumulative production count, Tl is the theoretical time to occurs at the point where a nominal qi units have been
produce the first unit, and 1 is the learning slope. Simi- produced, where i is the number of the production run
larly for each interruption some forgetting occurs. The (i= 1,2,3,... ) and qi (qi = ui + ui) represents the equiva-
forgetting curve relationship is commonly assumed to be lent units of continuous production experience accumu-

524 Appl. Math. Modelling, 1997, Vol. 21, August


Learning curves compared with forgetting: M. Y. Jaber and M. Bonney

Re-commencement after
complete forgetting

Time to
produce
a unit

(4) (4 + R) Units
k%Gi Time
P-t----t~~
Learn Forget Learn
Figure 1. Time to produce a unit versus units of output for an interrupted operation: t, is the time in production to produce Q units; t,
is the minimum time for total forgetting; R is the potential additional quantity that would be produced if no interruption occurred;
(q),(q+R) are the number of units produced or which potentially can be produced in times tp and r,+t,, respectively; and x is the
number of units produced or which potentially can be produced in times t, and t,ft,, where x Q g+R and t, d ta.

lated by the point of interruption, where ui is the number expressed as:


of units remembered at the beginning of the cycle and ui
is the quantity produced in cycle i.
It is at this point that the forgetting function is defined
fl,Lfl = T&4j+, + 1>-’

by equating the time required to produce the qith unit on Similarly for cycle i, fI,i is the time to produce the first
the learning curve to the time required on the forgetting
unit in cycle i, after the break in cycle i - 1 and is given
curve. This @ done by equating equations (1) and (2) and
as:
solving for T,, which gives:
fir = T,(u; + 1)~’ (6)
ii, = T,qjP(‘+f) where i = 1 (3)
In the VRIF model the parameters f1 and f are held
Assume that after making qi units the process is inter-
constant for all cycles i.
rupted for a period of length tbi, during which, if there
had been no interruption, an additional si units would
have been produced. In any cycle i, si i R, when tbi G t,. 3.2 The WWF model
Then, from equation (2), the time required to produce
unit number qi + si by the end of the break period on the Carlson and Rowe7 assumed that the intercept of the
forget curve is expressed as: forget curve varies after each interruption based on the
number of units produced in that cycle. Following the
same method as in Section 3.1, equations (3), (4), and (5)
(4) can be modified to, respectively:

As denoted in Fipre 2 u, + 1 is the amount of equivalent


units remembered at the beginning of cycle i + 1 after
interruption in cycle i. From equations (4) and (11, u~+~ is
(8)
expressed as:

(9)
(5)

Elmaghraby6 argued that the value of fql+R does not


Denote F, I+, as the time required to produce the first converge to T1 when total forgetting is assuked in the
unit in cycle’ i + 1, after the break in cycle i, and it is VRVF model. Denote t, as the minimum interruption to

Appl. Math. Modelling, 1997, Vol. 21, August 525


Learning curves compared with forgetting: M. Y. Jaber and M. Bonney

Tl

Forgetting curve

Time
to produce
r Unit i

Learning curve

L I

%+I hii) (4i + si) Units

Time

Figure 2. The learning-forgetting function at the end of cycle i: fPi is the time in production to produce 9i units in cycle i; tbi is the
length of the interruption period in cycle i, tbi =gtB; si is the potential additional quantity that would be produced if no interruption
occurred, si G Ri when rbi d ra. Note that t, is assumed to be fixed for all cycles i ii= 1,2,3,. . . ); (9i), (9;+s,) are the number of units
produced or which potentially can be produced in times tpi and t,,St,;, respectively; and u,+, is the experience remembered at the
beginning of production cycle i+ 1, knowing that the process was interrupted for a period of length tbi after accumulating 9i units in
cycle i, i.e., the period of interruption t,; means that at the start of cycle iS 1 the experience remembered reverts to u;+,.

which total forgetting is assumed. If the production pro- ratio of t,, the minimum time for total forgetting, to t(qi),
cess is interrupted for t, units of time, then the time to the amount of time required to produce qi units. t(qi) is
produce the first unit after the interruption is T1 (i.e., determined by integrating equation (1) over the proper
tbi = t, and si ,= Ri). In the case of Elmaghraby,6 who limits as:
assumed that Ti and f are constant for all cycles i, we
have: T,q/ -’
t(qi) = iq’TIj-‘dj = -
(1 - 1)
$,+R , = f&ji + R,)f (10)
Therefore Ci is represented as:
while in the case of Carlson and Rowe,’ who assumed fii
-1
is calculated for each cycle i while f is fixed for all cycles,
&q;-l
we have: ci = t,
[ I (13)

fq,+,, = ~,i(qi + Ri>f (11) At the point of interruption in cycle i in Fipre 2, equa-
tion (1) is set equal to equation (2), T4, = Tql, then the
intercept of the forget curve is determined from equation
(14), which is derived by modifying equation (7), to adopt
3.3 The LFCM model the changing value of the forgetting slope, to:
Jaber and Bonney’ expressed the forgetting rate as:
fii = T1q;‘i+f~’ (14)
1(1 - Olog qi
fi = where i= 1,2,3,... (12) The coordinates (qi +~~,f~,+~,) on the forgetting curve
log(C, + 1) have equivalent coordinates on the learning curve
(ui+ 1, Tu,,,), where ui+ 1 is the number of units remem-
where fi, which varies in every cycle, is the forgetting bered at the beginning of cycle i + 1 and T, + , = Tp +s,.
slope after interruption in cycle i and Ci = tB/t(qi) is the Equating (1) to (2) after substituting ui+ 1 = j kr equation

526 Appl. Math. Modelling, 1997, Vol. 21, August


Learning curves compared with forgetting: M. Y. Jaber and M. Bonney

(1) and qi + si = x in equation (2) and solving for ui+, The intercept of the forgetting function, f,, is determined
gives (Jaber and Bonney’, equation [ll]): from equation (3) as:

f, = Tlqp”+f,’ = 0.2 x 1~~-(o.l"2+o.Ixl)


ui+, = qj’+fl)“(q, + S;) -f,/l (15)
= 0.0432 days
By taking the limit of equation (2) as si + R, (tbi + te), as If the production run was not interrupted at unit number
illustrated in Figures 1 and 2, the value of fq,+ R con- 100, but rather continued over the break period, a total of
verges to T, when total forgetting is assumed, that is, sr units would have been accumulated given by:
when tbi = t,, and it is elaborated as:

1
l/Cl-I)
l-l
1-I
s,= ~ lb, +9, -91
fq,+R,
= &(qi + R;)’ = T, (16) [ T,

1 - 0.152
= x 10 + 1~~l-".l52
- 100
Similarly, ui+ , = 0 when si + Ri in equation (15
0.2 I
= 107 units

Therefore the amount of equivalent units remembered at


4. Numerical example
the beginning of cycle 2 for the LFCM is determined from
This section illustrates the different solution methods of equation (15) as:
the three mathematical models. Suppose that there exists
a production situation where 100 units are produced in u2 I.FCM= ~~~~(~l.lS2+ll lXl)/l~.lS2 x 207-11 IXl/O.lSZ

each cycle. The elapsed time between ceasing production = 42 units


at unit Qi <Qi = CL=, un; e.g., unit number 100, 200,
300,. . , etc.) and recommencing it at Qi + 1 (e.g., unit The amount of equivalent units remembered at the begin-
number 101, 201, 301,. . . , etc.) is assumed to be 10 days. ning of cycle 2 for the VRIF is determined from equation
Assume that the time required to produce the first unit (5) as:
with no previous experience is 0.2 days, and the process

1
l/ll.lS2
has a 90% learning curve (E = -log(O.9)/log2 = 0.152). 0.0432
Also assume that if the break time extends to 300 days, U 2,VRlF =
o2 x 207”.“’ = 42 units
[ .
then total forgetting occurs,
Elmaghrabyh as well as Carlson and Rowe’ did not Following the same procedure the value of u2 for the
state how they selected the doubling factor, f (the forget- VRVF is 42 units. In the first cycle the learning and
ting slope), in their VRIF and VRVF models, and it was forgetting curves are the same for all models. Therefore
assumed fixed for all cycles. To compare these models the time required to produce the first unit in the second
with the LFCM model the forgetting slope, f, is calculated cycle after the first interruption (e.g., unit number 101) is
based on the interruption in the first cycle using the given as:
LFCM approach. This value of f will then be fixed
throughout all cycles for the VRIF and the VRVF forget-
ting models. It will, however, change cycle by cycle in the %, = f~2,LFC%l = flZ,“RiF = %.VRVF

LFCM model. In the LFCM, t, is assumed constant and = 0.2 x (1 + 42)-“““* = 0.1129 days
can be determined from the organisation’s performance
records, e.g., as discussed in Jaber and Bonney.N Interruption occurs again at the end of the second cycle,
The value of C for cycle 1, is determined from equa- i.e., after unit number 200 has been produced. At this
tion (13) as: stage the equivalent number of units of experience accu-
mulated is 142 units. As before the forget slope, f2, and
the number of potential units that would have been accu-
mulated on the second break period, s2, are 0.213 and 111
units, respectively. The equivalent number of units of
experience remembered at the beginning of cycle 3 for the
0 .
2 x 1()()-".I52 -1
VRIF, U3,VaIF, VRVF, U3,VRVFj and the LFCM, $LFCMj
= 300 x I = 300/l 1.712 are 48 units, 70 units, and 63 units, respectively. Th_erefore
1 - 0.152
the time required to produce unit_ number 201, _Tzol, for
the three models, i.e., T,, VRIF, T12 VRVF, and T,, LFCM,
The forgetting slope is then determined from equation are 0.1108 days, 0.1043 days, and 0.1663 days, respectively.
(12) as: Table I shows the results obtained from repeating the
same procedure for all three models over five consecutive
0.152 x (1 - 0.152) x log 100 production cycles. The learning curve was plotted against
fi = log(1 + 300/11.712)
= 0.181 the VRIF, VRVF, and LFCM forgetting curves in Figures
3, 4, and 5, respectively. Figure 3 shows that the learning

Appl. Math. Modelling, 1997, Vol. 21, August 527


Learning curves compared with forgetting: M. Y. Jaber and M. Bonney

Table 1. Results for the three mathematical models

VRIF VRVF LFCM


(Elmaghraby’) (Carlson and Rowe’) (Jaber and Bonney’)

Cycle i ui f,, f,, r, “i f,, ?I, f, “, f,; 71, f,

1 0 0.2000 0.0432 0.181 0 0.2000 0.0432 0.181 0 0.2000 0.0432 0.181


2 42 0.1129 0.0432 0.181 42 0.1129 0.0384 0.181 42 0.1129 0.0327 0.213
3 48 0.1108 0.0432 0.181 71 0.1043 0.0361 0.181 63 0.1063 0.0289 0.228
4 48 0.1106 0.0432 0.181 93 0.1002 0.0347 0.181 74 0.1037 0.0272 0.235
5 48 0.1105 0.0432 0.181 111 0.0977 0.0337 0.181 80 0.1026 0.0263 0.238

curve and the VRIF forgetting curve did not conform to In Table 1 the VRIF gav_e the highest values of ‘?,i for
the hypothesised relationship that the curves intercept at all cycles. The values of TIi obtained from the VRVF
the same point on all cycles. Unlike the VRIF curve the were slightly lower than those of the LFCM. The LFCM
VRVF and LFCM curves did confirm the relationship model combined the characteristics of the VRIF and the
presented in Figzue 2. This is due to the adjustment in VRVF models by giving close estimates of the values of
both the forgetting rate and the intercept of the forgetting the VRVF and the convergence desired by the VRIF.
curve after each interruption.
Assuming that the production break increases from 10
days to 300 days (tbi = t, = 300 days), then tota! forgetting
occurs and si = Ri. To test the convergence (Tq,+R, + T,)
5. How accurate is the LFCM model?
of all three models, as given in equations (lo), (ll), and
(16), the question is then, what if a break of 300 days is As indicated earlier the Jaber and Bonneys learn-forget
experienced in the first cycle? In the second cycle? In the curve model (LFCM) presented in Section 3.3 produced
third cy$e? . . . etc. Table 2 shows that the calculated val- results that are consistent with the results of the Glober-
ues Of Tq +R LFCM converge to a value of 0.2 (T,) for all son et aL9 experimentally derived learning-forgetting
five cycle;, Ghile fq, + R,,vR,F converges to a value slightly model. The Globerson et aL9 experiment was performed
higher than T, (0.2001 in cycle 1, then 0.2!07 in all in a computer-oriented manufacturing environment simu-
remaining cycles). Conversely the value of Tq,+R,,VRVF lated in a microcomputer laboratory. The 120 subjects
diverged away from T, for each consecutive cycle (see divided into six groups participated in the experiment in
column 5, Table 2). Figzue 6 illustrates the behavior of the which they performed a data-entry task. Each subject
three forget curves in the second cycle, where the LFCM participated in two sessions, of 16 repetitions each, sepa-
started asymptotically with the VRVF and ends up asymp- rated by a break. The average lengths of the break time
totically with the VRIF. Both the VRIF and the LFCM between the two consecutive sessions were 1.7 (l-2 days),
curves conform to the hypothesised relationship in Figure 6.1 (3-7 days), 11.5 (8-16 days), 19.1 (16-25 days), 40.8
1, which shows the convergence to a unique value of T,, (26-53 days), and 65.1 days (54-82 days) in Groups 1
while the VRVF did not. The above discussion shows that through 6, respectively. Globerson et al.” use-d seven
mathematical models to estimate the value of T,,,, the
the LFCM was consistent with the hypothesised learning-
calculated performance time for the first unit after the
forgetting relationships of Figures 1 and 2, while the
break if interruption occurs, as a function of Tq+ ,, the
VRIF and the VRVF have some inconsistencies.

0.2 - Forgetting CUNI?


- ForgettIng curve
0.2 f .= 0.18 - - - Learning curve
- - - Learning curve
= 0.16
“, 0.14

1 100 200 300 400 500 1 100 200 300 400 500

Potential units Potential units


Figure 3. The behavior of the VRIF forgetting curve and Figure 4. The behavior of the VRVF forgetting curve and
Wright’s learning curve over five production runs with interrup- Wright’s learning curve over five production runs with interrup-
tions occurring at 100, 200, 300, 400, and 500 potential units. tions occurring at 100, 200, 300, 400. and 500 potential units.

528 Appl. Math. Modelling, 1997, Vol. 21, August


Learning curves compared with forgetting: M. Y. Jaber and M. Bonney

Table 2. The convergence of the time required to produce the


- Forgetting curve
0.2 last unit on the forget curve
.Z 0.18
VRIF VRVF LFCM
= 0.18
; 0.14 Cycle i 9;fR; Tq,+R, 9;fR; fq,+~, 9;+ R; f,,+ R,
2 0.12
B 0.i 1 4793 0.2001 4793 0.2000 4793 0.2000
h 0.08 2 4866 0.2007 4866 0.1784 4866 0.2000
2 0.06 3 4876 0.2007 4916 0.1680 4902 0.2000
: 0.04 4 4877 0.2007 4952 0.1615 4920 0.2006
i= 0.02 5 4877 0.2007 4980 0.1571 4930 0.2000
04 I I, I, 1 I, 1 ,b

1 100 200 300 400 600


for 1 <i < 16, where q is the performance time of the jth
units
Potential unit after the break, and u2 represents knowledge remain-
Figure 5. The behavior of the LFCM forgetting curve and ing from previous learning (prior to the break), and its
Wright’s learning curve over five production runs with interrup-
value was estimated using the following equation:
tions occurring at 100, 200, 300. 400, and 500 potential units.

-l/I
calculated performance time for the q + 1 units after the 1.87 x 7-“.9
q+l x t”.“9
h
u* = -1
break if no interruption occurs, and fb, the break time. Tl
The model that best describes the forgetting phenomenon
was found to be:
For this study the average values of Wright’s model in
equation (1) were T, = 544 seconds, 1 = 0.325. The time to
t/+1= 1.87 x T”.”
q+l x to.“”
I1 (18) produce the 17th repetition is determined from equation
(1) as:
The model presented
in equation (18) has an adjusted
R* = 0.75 (R = 0.86). Assuming that equation (18) appro- T17 = 544 x 17P”.325 = 217 set
priately describes the forgetting phenomenon, Globerson
et al.’ integrated it into the learning model to improve its The calculated performance time for the first unit after an
predictability power. The learning curve model for the average break of 65.1 days is determined from equation
second session was of the form: (18):

iy = T,(j + u2)-’ (19) f,, = 1.87 X 217°.9 X 65.1°.09 = 344.50 z 345 set

A
0.2 -

0.18-

0.16-

*g 0.14-

: 0.12-
2
'0
9 0.1-
-vRlF
; 0.08- - LFCM
_ _ _ VRVF

E 0.06-
i=

0.04-

0.02-
.
O,,,,,,,,,,,,,,,,,,,,~
1 500 1008 1500 2000 2500 3000 '3500 4000 4500

Potential units
Figure 6. The convergence of the three forgetting curves in the second cycle when total forgetting is assumed to occur (t,,=t,).

Appl. Math. Modelling, 1997, Vol. 21, August 529


Learning curves compared with forgetting: M. Y. Jaber and M. Bonney

Similarly, for average breaks of 1.7, 6.1, 11.5, 19.1, and The VRIF, VRVF, and the LFCM models hypothe-
40.8 days, the values of T,, are 248, 278, 295, 309, and 330 sised two relationships in defining the learning-forgetting
set, respectively. Following the same procedure as in relationship. The first hypothesis indicates that when total
Jaber and Bonney,s the level of units remembered at the forgetting occurs, the performance time on the forgetting
beginning of the second run, c~rrc., is given from equa- curve reverts to a unique value equivalent to the time
tion (15) and has values of 7.54, 5.59, 4.83, 4.30, 3.62, and required to produce the first unit with no prior experi-
3.24 units for average breaks of 1.7, 6.1, 11.5, 19.1, 40.8, ence. The second hypothesis is that the performance time
and 65.1 days, respectively. The time required to produce on the learning curve equals that on the forgetting curve
the first unit after an average break period of 65.1 days is: at the point of interruption. The VRIF mode1 was consis-
tent with the first hypothesis but was inconsistent with the
~17. LFCM = 544 x (3.24 + 1 )P”‘325 = 340 set second, whereas the VRVF mode1 was consistent with the
second but was inconsistent with the first. The LFCM
which differs from the result obtained in Globerson et a1.9 mode1 was consistent with both hypotheses.
by - 1.45%, and it is calculated as: The accuracy of the LFCM mode1 was tested, and the
results were shown to be consistent with those of the

%Dev =
f17,LFCM
- ?I’ x lOO=
340 - 345
x 100
Globerson et a1.9 experimentally derived learning-forget-
ting mode1 with a negligible percentage of deviation for
f,, 345
production breaks ranging between 16 and 82 days.
= - 1.45% Unlike earlier works5-’ it is believed that the work in
this paper is the first that has used t, as a parameter. The
Similarly the values of f,, are 271, 295, 307, 316, and 331
primary assumption in the development of the learn-for-
set, respectively, and the percentage deviations are 12.1,
get curve mode1 (LFCM) is that t, is of a fixed value.
5.36, 4.07, 2.27, and 0.3%, respectively. The value of u2 is
Intuitively using a fixed value does not seem unreason-
determined from equation (19). The value of u2 is then
able, provided one starts the production after a period of
incorporated in equation (20) to determine the value of
moderate training. This limiting property is characterised
T,,. By treating u2 as an integer value, for average breaks
by the assumption of t, occurring over the interval [O,tB],
of 1.7, 6.1, 11.5, 19.1, 40.8, and 65.1 days the values of T,,
disregarding any possibility of occurrence over the interval
are 250, 289, 304, 322, 347, and 347 set, respectively. Also
[tB, ~1. Another limitation to the work is that no field data
by treating the values of u2,LFCM as integers the values of
was available to aid the authors in justifying the assump-
‘1, LFCM are 277, 304, 322, 322, 347, and 347 set, respec- tions made to develop the learn-forget curve model
tively and the percentage of deviations from the results of (LFCM). Finally the above discussion suggests that the
Globerson et a1.9 are 10.8, 5.2, 5.98, 0, 0, and O%, respec- LFCM model should be tested with field data in several
tively. The results indicate that the LFCM mode1 pre- industrial settings. If the LFCM mode1 proves to be unsat-
sented in Section 3.3 produces results that are consistent isfactory to represent reality, a new modified mode1 should
with the results of the experimentally derived learning-for- be developed to represent reality more faithfully.
getting curve by Globerson et aL9 for a break length
between 16 and 82 days.

Nomenclature
6. Summary and conclusion break time to achieve total forgetting
This paper compared three proposed mathematical mod- forgetting slope in cycle i
els that describe the learning-forgetting relationship. The learning slope
LFCM of Jaber and Bonney,’ calculates the value of the L’i number of actual units produced in cycle i
forgetting slope based on three factors. These factors were u, number of equivalent units of experience with forget-
the equivalent number of units of continuous production ting at the beginning of cycle i
accumulated by the point of interruption, the minimum 'i number of equivalent units of experience with no
break to which the manufacturer assumes total forgetting, forgetting at the beginning of cycle i
and the learning slope. The LFCM estimates of the time 9i number of theoretical units of output accumulated at
required to produce the first unit after an interruption the end of cycle i
period were between the_ estimatesgiven by the VRIF and X number of units of output that would have been
the VRVF (f,i,vRvF < T,i,LFCM G T,i,VR,F). The learn-for- accumulated if interruption did not occur
get curve mode1 (LFCM) allows the forgetting slope to be Ri number of units that would have been produced
estimated, whereas the VRIF (Elmaghraby6) and VRVF during t, in cycle i
(Carlson and Rowe’) models do not state the method by j production count/cumulative number of repetitions
which they derive their forgetting slopes (rates), and so tbi
shorter break than t, experienced in cycle i
the LFCM mode1 was used to estimate these to give ci
the minimum value of the ratio of the break time to
results consistent with the VRIF and VRVF equations. the production time in cycle i that will give total
The equivalent time of qi + Ri on the VRVF curve, forgetting
Tr;+R,.v~~~, failed to converge to a unique value of T, fl,
time for the first unit of the forgetting curve in cycle
w en total forgetting was assumed. i

530 Appl. Math. Modelling, 1997, Vol. 21, August


Learning curves compared with forgetting: M. Y. Jaber and M. Bonney

time for the xth unit of lost experience on the 5. Adler, G. L. and Nanda, R. Effects of learning on optimal lot size
forgetting curve determination-Single product case. AIZE Trans. 1974, 6, 14-20
6. Elmaghraby, S. E. Economic manufacturing quantities under
time to produce ui units in cycle i
conditions of learning and forgetting (EMQ/LaF). Prod. Plan-
time to produce the first unit in cycle i with forget- ning Confrol 1990, 1, 196-208
ting after an interruption in production 7. Carlson, J. G. and Rowe, R. G. How much does forgetting cost?
time to produce the first unit in cycle i with no Indust. Eng. 1976, 8, 40-47
forgetting after an interruption in production 8. Jaber, M. and Bonney, M. C. Production breaks and the learning
curve: The forgetting phenomena. Appl. Math. Modefling 1996,
time to produce the jth unit on the learning curve
20, 162-169
9. Globerson, S., Levin, N. and Shtub, A. The impact of breaks on
forgetting when performing a repetitive task. IIE Trans. 1989, 21,
376-381
References 10. Shtub, A., Levitt, N. and Globerson, S. Learning and forgetting
industrial skills: Experimental model. Int. J. Hum. Factors Manu-
Wright, T. Factors affecting the cost of airplanes. J. Aeronaur.
fact. 1993, 3, 293-305
Sci. 1936, 3, 122-128
Yelle, L. E. The learning curve: Historical review and compre- 11. Bailey, C. D. Forgetting and the learning curve: A laboratory
hensive survey. Decision Sci. 1979, 10, 302-328 study. Manage. Sci. 1989, 35, 346-352
Belkaoui, A. The Learning Curve. Quorum Books, Westport, CT, 12. Badiru, A. B. Multivariate analysis of the effect of learning and
1986 forgetting on product quality. Inf. J. Prod. Rex 1995, 33, 777-794
Hoffman, Thomas R. Effect of prior experience on learning 13. Dar-El, E. M., Ayas, K. and Gilad, I. Predicting performance
curve parameters. J. Indust. Eng. 1968, 19, 412-413 times for long cycle time tasks. IIE Trans. 1995, 27, 272-281

Appl. Math. Modelling, 1997, Vol. 21, August 531

You might also like