1994 INTERNATIONAL SYSTEM DYNAMICS CONFERENCE _
System Dynamics and Learning Curves
M M Naim and D R Towill
Cardiff School of Engineering, University of Wales College of Cardiff,
Queen's Buildings, P.O. Box 917, Cardiff, CF2 1XH, UK
Tel: +44 222 874421 ext 5944
Fax: +44 222 874420
Abstract
Learning curves are met in a wide variety of industrial situations, They have become particularly
important in modern business strategy because product life cycles are constantly reducing as
companies seek to gain competitive advantage via a rapid response to customer demands. The paper
describes a family of System Dynamics models that have been found particularly appropriate in
modelling and forecasting the performance of business organisations; including the performance of
manufacturing systems and the penetration of new products into the market place.
The System Dynamics learning curve models have a servomechanism analogue that yields valuable
insight into the parameter estimation problem. The models are required under two quite different
circumstances. The first is based on historical information where the model is to be added to a
company or consultancy data base. The second is for on-line forecasting and control of a business
activity. An enhanced stability least squared error predictor is described which covers both
qui The paper with industrial applications of the System Dynamics models.
Business Decision-Making, page 164
1994 INTERNATIONAL SYSTEM DYNAMICS CONFERENCE
System Dynamics and Learning Curves
INTRODUCTION
The concept of learning is generally accepted in a wide ranging set of disciplines and professions.
Learning, in the dictionary sense, means the possession of knowledge via study. This conjures up
the thought of individual human learning. The individual's output in such circumstances is difficult
to quantify, although conceptualised as being time dependent. As an individual learns more with
time, his level of knowledge grows. The time dependent output may be modelled via a learning
curve following a classic 'S'-shape. It is this hypothetical model that is often cited by most
authors on learning curve theory especially psychologists. The curve may model an individuals
level of knowledge throughout his life time, or may just represent his performance for a single
task. The learning curve may be made up of a combination of single tasks, either taken
sequentially or simultaneously.
Such a human learning concept may be transferred over to "group" learning, where individuals,
learning particular tasks, integrate them in order to obtain a group output. Therefore the learning
of the group is dependent on the i ion of the individual learning P and the
complexity of the overall task performed. The i i h may be via lised
work procedures or via a group leader.
For an individual or group then the concept of system learning arises intuitively. When talking of
"soft" systems, such as people, organisations or any group of individuals, this concept is easy to
understand. But when talking of "hard" systems this becomes a bit more difficult. After all it is
not easy to imagine an inanimate object such as a machine tool learning. Clearly there is no such
thing as a perfect hard system that requires no human interaction, so that some softness, and
thereby a component of human learning, will undoubtedly exist. Nevertheless, even an
hypothetically ideal machine tool would go through an improvement phase prior to steady state
operation. The existence of failures would typically occur in a “bath tub" fashion. Such a
reliability curve suggests that, when inverted, the performance of the system would follow an 'S'-
shaped pattern. So hard/soft systems (such as a Flexible Manufacturing System) may be
accommodated by the learning, or improvement, concept.
LEARNING CURVE MODELS
The ability to mathematically model the learning curve has found favour in a wide ranging set of
(industrial) applications. These include human operator training and learning performance [Bevis,
1970, Baran, 1986], the diffusion of innovations and product life cycles [Mahajan & Peterson,
1985, Tchijov & Norov, 1989, Yelle, 1980, Edquist & Jacobsson, 1988, Butler, 1988], the
di of ing technologies (AMT) [Towill et al, 1989], the reliability
aspects of a nuclear power station [Sabri & Husseiny, 1979] and the thermal efficiency of power
stations [Sharp & Price, 1990, Naim & Towill, 1990]. This in itself tends to underline the
versatility of learning curves.
Modelling the learning curve allows insight into the system under study. It provides a simplistic
representation of the system's time varying performance so that causal relationships may be
determined, comparisons made with similar systems and forms the basis of decision making
policies. In conjunction with modelling, many applications require a forecasting capability. Given
a limited amount of performance data downstream from saturation, a model may predict, both in
the long-term and short-term, likely system performance in the future.
But perhaps due to the very different applications there has resulted from different researchers no
recognisable "standard" curve fitting practice. In fact, the type of curves fitted, the methods of
fitting and forecasting are just as wide ranging as their applications. Even the names given to the
learning curve differ, such as 'S'-shaped curves, saturation curves, diffusion curves and experience
curves, although all have similar mathematical properties.
Business Decision-Making, page 165
1994 INTERNATIONAL SYSTEM DYNAMICS CONFERENCE
Various models have been advocated to describe the growth of dynamic systems, some of the
most common as indicated in Table I which includes both trend line and smoothing models. The
merits, or otherwise, of such models in describing growth, and subsequently in forecasting (both
short- and long-term), have had id p Although a certain model may be adopted
as it provides the "best fit" for a particular application of many of these models have one or
more of the following snags associated with them;
¢ inability to predict a final saturation level despite the fact that is what happens in the real
world
¢ holding of meaningless parameter values which have no interpretation in the real world
© require data near the saturation level prior to obtaining realistic predictions
* gives only short range forecasts despite requiring much data
Model Equation
li =a+bt
Parabolic Y=atbttct”
Power Y=ar’
Exponential Y¥ =ae"
Logistic y= NK . a)
Gompertz Yeatbhe" +a?
von Bertalanffy Y= a(1 —be™* y
ARMA Y = ARMA(p,q)
le I: mmonly u: it
Besides this, such models are used in an ad hoc manner with no recognised methodology by which
a particular model may be chosen. Various models may be attempted until the "best fit" model is
found.
This paper highlights the hypothesis put forward by Towill [1982] that a family of generic
System Dynamics models, as indicated in Figure 1, may be used to represent, and control, the
dynamics of "soft"/"soft-hard" systems, just as they are used in the engineering of "hard"
systems. In particular, with care, the disad Bes pI ly d of the models in Table I
may be avoided.
THE SYSTEM DYNAMICS LEARNING CURVE MODELS
The analogue between the learning curves of industrial systems and the responses of "soft"
systems has been noted in the literature. For example;
"It appears that the stages of economic growth... are analogous to the growth of
knowledge in a given field... We will further elaborate on the subject by showing that
there exists analogies between the stages of OR/MS knowledge growth and the kinetics of
chemical reactions." [Reismann & Xu, 1992]
While such analogies are beneficial, industrial systems learning curves are probably better
compared with "hard" engineering systems. As Towill has pointed out, in particular with
reference to production start-up, many engineers become plant on-line managers as well as
attaining position where an app ion of system d: is required. This is especially true
for those graduates "who have mastered control engineering .... [with an] ability to model existing
Business Decision-Making, page 166
1994 INTERNATIONAL SYSTEM DYNAMICS CONFERENCE
8 Ripple model
gf oe
5 Time <
&
Can fs
8
eA
E
2| |
aie Time delay model
Pe ee
iT ese OE mee Time
All pass model
om ae
Time
Figure 1; Family of transfer function
RIPPLE
TIME CONSTANT
~ | ey
fhe He ts 4 | |
_ i ee .
A- 4- 4 it | | i | i
pas yo af
| : i
i. ~ _ = an od fae
Figure 2: block f
‘system dynamics learning curve models
Business Decision-Making, page 167
1994 INTERNATIONAL SYSTEM DYNAMICS CONFERENCE
systems from scanty experimental data together with an appreciation of problems associated with
real time control." [Towill, 1991]
The suite of System Dynamics learning curve models shown in Figure 1 are derived from the unit
step responses of simple transfer function models. The block diagram model representation of
those transients in the family that adequately replicate learning behaviour, namely the Time
Constant, 'S' and Ripple curves, are given in Figure 2. The hierarchical nature of the family lends
itself well to a structured methodology by which an analyst is able to select the most appropriate
transient by which to model a particular application. Thus, if the Time Constant model proves to
be inadequate for a particular application then the analyst has a logical line of progression along
the family of System Dynamics models based on increased model complexity.
The equations defining the transients of the latter three models are given by;
Time Constant Model
Ym (t) = Ye +¥,(i-
+y
imto= Yer, 14 uf . | 2
2 am Tage
Limiting 'S'-curve (1) =1))
-1
Ymt=¥e+¥ft-(1+ Ze | 3
Ripple model
va
Ym()=Ye+¥/(I-6 )+Asin(o.r+ oe’ 4 wd
Ymi(t) = model output at time, t
Yo = model output at time, t= 0
Yor Ye = model output at time, t = c
& time constant, a measure of the rate of learning
4,72 time constants, the ratio of which describes the 'S'-
curve characteristics
A = amplitude of cyclical component
o = frequency of cyclical component
o = phase shift
td = damping coefficient
SYSTEM DYNAMICS LEARNING CURVE APPLICATIONS
Since the Time Constant model was first used in 1970 [Bevis, 1970] it has been utilised in many
industrial applications. These include; various manual and assembly tasks [Hitchings & Towill,
1975], the modelling of nuclear power station operator reliability [Sabri & Husseiny, 1979],
determining the average thermal efficiency of nuclear power stations [Sharp & Price, 1990, Naim
& Towill, 1990], capacity assessment of process plants [Brennan & Stephens, 1985], and the
start-up of FMS [Towill et al, 1989].
Business Decision-Making, page 168
1994 INTERNATIONAL SYSTEM DYNAMICS CONFERENCE
The widespread application of the Time Constant model has been possible through the
development of a number of sophisticated curve fitting techniques. These include a Taylor series
expansion method for least squared error (LSE) estimation of i LSE search
methods, Kalman filters and maximum likelihood. Of these the LSE search method may be the
simplest to apply although it is rather t and time ing in terms of
properties especially if the solution space is too wide or localised optima are reached. Kalman
filters have been found to be effective but require a iderable number of data points and an
appreciation of the statistics of the data prior to curve fitting.
The Taylor series LSE expansion method is that favoured by the authors as it is simple to apply
and is ly easy for to und d as it is based on simple regression and as with all
methods the aim is to minimise the error between the model and the raw data. This method has
subsequently been used, and adapted, in the development of curve fit algorithms for the 'S'-shaped
and Ripple learning curve models. Appendix I gives the curve fit sequence for the 'S'-curve that
utilises parallel adaptation in order to improve the robustness of the algorithms [Naim, 1993].
The aim of modelling the learning curve is to glean information from the model itself and from
the residuals scattered around the model. Costs may then be Pp d. In
order for the analysis to be of any use in instigating action about tthe process it must be a
component in an overall management information and control system (MICS). Towill [1985]
P a simple for a p based MICS for on-line manufacturing system start-
up. This has been developed to take into account the learning curve associated with the diffusion
of new products and is shown in Figure 3. Such a model therefore takes into account the systems
view of a business in which di Pp to develop mutually fulfilling
strategies.
The model not only accounts for distinct and separate manufacturing or marketing changes but
also the knock-on effect of one on the other as well as change occurring in parallel. An
important point to note is that the target learning curves have to be realistic. Imposed learning
“has an adverse affect when the imposed learning pace is slightly faster than the unpaced
[natural] learning rate [due to the individual attempting to keep up with the imposed rate and
thereby leading to stress and poor performance], and has no effect when the imposed pace seem
to be too difficult", the individual reverting back to his natural pace. [Globerson & Seidmann,
1988]. This can also be extended to group and organisational learning.
The structure and interactions between the functions are obviously more complex in practice but
the system boundary under consideration is as shown in the causal loop diagram of Figure 4. This
shows the open-loop cascade effect of efficient project management; leading to a rapid, smooth
start-up learning curve and therefore a steep cost function curve that ultimately results in an
effective product diffusion learning curve [Yelle, 1980].
PRACTICAL EXAMPLES
The _impl ion_of condition monitoring (C-M) Ani. to_a_ continuous casting
iT wv
The start-up tonnage output for a particular CONCAST steel mill [James, 1988] over a 60 month
period is shown in-Figure 5. The time scales relate to the period for which data was available and
not necessarily from the actual start date of production. Sequential models are required in this
particular case with the boundary between phases disti d by the i ion of C-M
tools. The Time Constant model is suitable for the first phase of production, up to Month 29,
forecasting a final saturation level of 39.23. thousand tonnes per month (95% of which will be
reached by Month 75).
The impl ion by the pany of C-M techni is preceded by a in prodi
as may be expected owing to disruptions during the installation of C-M hnology. An
improvement in output follows which may be modelled via the Limiting 'S'-curve which predicts a
final performance output of 47.11 thousand tonnes per month. The final level of attainment is
Business Decision-Making, page 169
1994 INTERNATIONAL SYSTEM DYNAMICS CONFERENCE
Updated a Observed
» [Mata a [Oper ator Lost Manufacturing gS Performance, Og Wee Trend curve
earning eck
——
facion activity, 7 Hogging ‘modeling |
curve ‘
. Woot _, Observed
rketing Sales . ag Sales ~~ Sag5 [Performance] DIS, “Trend | ,_curval
strategy => decisions activity
° oe
~-tasnng 9868088 acon sity Wogging "meeting
4 Time
reproduction __
planning effort es
, v vr
Available Rate of Variance of
capacty start-up startup
bs :
er 4
Delivery costs
performance
r
Pnees
2
Figure 4. The causal and product diffusion
Business Decision-Making. page 170
1994 INTERNATIONAL SYSTEM DYNAMICS CONFERENCE
Sales (thousands)
6 Tonnes (thousands)
“Y= 23.24 + 23.87(1 -(l+ t/7.2)exp(-t/7.2)) |
gpa eOeE |
C-M Implementation
Y¥ = 12.33 + 26.9(1 - exp(-t/28.62))
vassal : an
10 20 30 40 50 60
Time (months)
Time Constant ' Raw Data =~? Limiting ‘S'-curve
os
“AST start-up performance
194y 19S— 19S] 19S2 198319841985 1956 1957 1958 1959 1960 1961
Ye
ear
- Raw data ~— Time Constant model
~ Liming ‘S'-curve + Standard 'S'-curve
Figure 6: Sales of home freezers (1947-61)
Business Decision-Making, page 171
1994 INTERNATIONAL SYSTEM DYNAMICS CONFERENCE
reached by Month 57. Thus the introduction of C-M techniques has led to a predicted 20%
in final production output with an associated 24% reduction in learning rate as
defined byt.
Modelling Sales Data
Figure 6 shows the application of the System Dynamics learning curve models to the diffusion of
products into the market place. For the purposes of comparison the sale of home freezers [Bass,
1973] has been modelled by the generalised “S'-curve, the limiting case of the 'S'-curve and the
Time Constant model. Table II shows the residual mean square (RMS)! value for each of the
models.
Model => Limiting 'S'-curve Generalised 'S Time Constant
curve
Sales suplication RMS value ('000s) | RMS value (000s) | RMS value ('000s)
Home freezers 15 16 18
Table Il: Summary of RMS values for home freezer sales
Using the sum of errors squared as a performance measure there would have been little to
sh between the lised 'S'-curve model and its limiting case. But biasing the measure
towards the simpler models using the RMS value indicates the Limiting 'S'-curve as the better
modeller of the trend. At the same time it can be ‘seen that the Time Constant model also gives a
reasonable fit to the data alth h not ig the infancy stage of the trend.
Utilising the Time Constant.curve as the trend model, corresponding residual time series and
correlogram have shown a cyclical component within the data although the significance is limited
by the lack of sufficient data points. There is therefore a case for utilising the Ripple curve to
model the trend.
CONCLUSION
The paper has highlighted the development of a suite of learning curve models, based on the
servomechanism principles of System Dynamics, that may be utilised as a management decision
support tool for POnOrNe and controlling the -start-up phases of new products and the
fi . The family of transients has been applied to a variety of
industrial systems of which only two have been given as example.
REFERENCES
Baran, R.H., 1986, A modified Ainsworth measure of learning efficiency, Proceedings of the
1986 IEEE International Conference on Systems, Man and Cybernetics, October, Atlanta, GA,
USA.
Bass, F.M., 1973, A new product growth model for consumer durables: in Creating and marketing
new products, Granada Publishing Ltd., Section 4.6, pp 445-463.
Bevis, F.W., 1970, An exploratory study of industrial learning with special reference to work
study standards, M.Sc. Thesis, UWCC, Cardiff, Wales.
Brennan, D.J. and Stephens, G.K., 1985, The extent and rate of capacity increase in process
plants, Process Economics International, Vol. V, Nos. 3 & 4, pp 15-25.
Butler, J.E., 1988, Theories of technological innovation as useful tools for corporate strategy,
Strategic Management Journal, Vol. 9, pp 15-29.
Sum_errors_ squared
TRMs =
No. datapoints ~ No. parameters
‘Business Decision-Making, page 172
1994 INTERNATIONAL SYSTEM DYNAMICS CONFERENCE
Edquist, C. and Jacobsson, S, 1988, Flexible automation: the global diffusion of new technology in
the engineering industry, Blackwell, Oxford.
Globerson, S. and Seidmann, A., 1988, The effects of imposed learning curves on performance
improvements, IIE Transactions, Vol. 20, No. 3, pp 317-324.
Hitchings, B. and Towill, Di! R., 1975, Error analysis of the time constant learning curve model,
1 Journal of P: th, Vol. 13, No. 2, pp 105-135.
James, B., 1988, Process industry - a user's view, I.Prod.E. Seminar, Condition Monitoring in
Manufacturing - The Key to Reliable and Economic Performance, UWCC, Cardiff, U.K., 16 June.
Mahajan, V. and Peterson, R.A., 1985, Models for innovation diffusion, Sage Publications,
Beverley Hills.
Naim, M.M., 1993, Learning curve models for predicting performance of industrial systems, PhD
Thesis, University of Wales College of Cardiff, United Kingdom.
Naim, M.M. and Towill, D.R., 1990, An engineering approach to LSE modelling of experience
curves in the electricity supply industry, International Journal of Forecasting, Vol. 6, No. 4, pp
549-556.
Reismann, A. and Xu, X., 1992, On stages of knowledge growth in the management sciences,
JEEE Transactions on Engineering Management, Vol. 39, No. 2, pp 119-128.
Sabri, Z.A., and Husseiny, A.A., 1979, lytical modelling of nuclear power station operator
reliability, ‘Annals of Nuclear Energy, Vol 6, pp 309-325.
Sharp, J.A. and Price, D.H.R., 1990, Experience curves in the electricity supply industry,
International Journal of Forecasting, Vol. 6, No. 4, pp 549-556.
Tchijov, I. and Norov, E., 1989, Forecasting methods for CIM technologies, International
Journal of Engineering Costs and Production Economics, Vol. 17, pp 323-329.
Towill, D.R., Davies, A. and Naim, M.M., 1989, The dynamics of capacity planning for flexible
manufacturing system startup, International Journal of Engineering Costs and Production
Economics, Vol. 17, pp 55-64.
Towill, D.R., 1982, How complex a learning curve model need we use?, Radio and Electronic
Engineer, Vol. 52, No. 7, pp 331-338.
Towill, D.R., 1991, Engineering change; or is it change engineering? A personal perspective, IEE
Proceedings A, Vol. 138, No. 1, pp 11-21.
Towill, D.R., 1985, M systems ications of learning curves and progress functions,
International Journal of Engineering Costs and Production Economics, Vol. 9, pp 369-383.
Yelle, L.E., 1980, Industrial life cycles and learning curves: interaction of marketing and
production, Industrial Marketing Management, Vol. 9, pp 311-318.
Business Decision-Making, page 173