Technological Forecasting Based On Segmented Rate of Change

Portland State University
PDXScholar
Dissertations and Theses
Dissertations and Theses
Winter 3-16-2015
Technological Forecasting Based on Segmented

Rate of Change
Dong-Joon Lim
Let us know how access to this document benefits you.

Follow this and additional works at: http://pdxscholar.library.pdx.edu/open_access_etds
Recommended Citation
Lim, Dong-Joon, "Technological Forecasting Based on Segmented Rate of Change" (2015). Dissertations and Theses. Paper 2220.
This Dissertation is brought to you for free and open access. It has been accepted for inclusion in Dissertations and Theses by an authorized
administrator of PDXScholar. For more information, please contact [email protected].
Technological Forecasting Based on Segmented Rate of Change
by
Dong-Joon Lim
A dissertation submitted in partial fulfillment of the

requirements for the degree of
Doctor of Philosophy
in
Technology Management
Dissertation Committee:
Timothy R. Anderson, Chair
Tugrul U. Daim
Antonie J. Jetter
Wayne W. Wakeland

2015
2015 Dong-Joon Lim
Abstract
Consider the following questions in the early stage of new product development.
What should be the target market for proposed design concepts? Who will be the
competitors and how fast are they moving forward in terms of performance
improvements? Ultimately, is the current design concept and targeted launch date feasible
and competitive?
To answer these questions, there is a need to integrate the product benchmarking with
the assessment of performance improvement so that analysts can have a risk measure for
their R&D target setting practices. Consequently, this study presents how time series
benchmarking analysis can be used to assist scheduling new product releases.
Specifically, the proposed model attempts to estimate the auspicious time by which
proposed design concepts will be available as competitive products by taking into
account the rate of performance improvement expected in a target segment.
The empirical illustration of commercial airplane development has shown that this
new method provides valuable information such as dominating designs, distinct
segments, and the potential rate of performance improvement, which can be utilized in
the early stage of new product development. In particular, six dominant airplanes are
identified with corresponding local RoCs and, inter alia, technological advancement
toward long-range and wide-body airplanes represents very competitive segments of the
market with rapid changes. The resulting individualized RoCs are able to estimate the
arrivals of four different design concepts, which is consistent with what has happened
since 2007 in commercial airplane industry.
i
In addition, the case study of the Exascale supercomputer development is presented to

demonstrate the predictive use of the new method. The results indicate that the current
development target of 2020 might entail technical risks considering the rate of change
emphasizing power efficiency observed in the past. It is forecasted that either a Cray-built
hybrid system using Intel processors or an IBM-built Blue Gene architecture system
using PowerPC processors will likely achieve the goal between early 2021 and late 2022.
This indicates that the challenge to improve the power efficiency by a factor of 23 would
require the maximum delay of 4 years to reach the Exascale supercomputer compared to
the existing performance curve.
ii
Table of Contents
Abstract............................................................................................................................... i
List of Tables ................................................................................................................... vii
List of Figures................................................................................................................. viii
Glossary ............................................................................................................................ ix
List of Symbols ................................................................................................................. xi
I. Motivation ...................................................................................................................1
1.1 Introduction ...........................................................................................................1
1.2 Problem Statement ................................................................................................5
1.3 Research Objective ...............................................................................................8
1.4 Overview of Dissertation ....................................................................................10
II. Background ...............................................................................................................12

2.1 Technological Innovation Theories ....................................................................12
2.1.1 S-curve Models ..........................................................................................12
2.1.2 Disruptive Innovation ................................................................................19
2.1.3 New Product Target Setting .......................................................................26
2.2 Technology Assessment Methods.......................................................................30
2.2.1 Qualitative Models .....................................................................................30
A. Delphi and Expert Opinion ...................................................................30
B. Scenario Analysis ..................................................................................30
C. Risk Assessment and Environmental Analysis .....................................31
D. Relevance Trees ....................................................................................32
E. Morphology ...........................................................................................33
F. Analogy..................................................................................................34
G. Causal Model ........................................................................................34
2.2.2 Quantitative Models ...................................................................................35
A. Decision Analysis .................................................................................35
iii
B. Economic (Cost-Benefit)Analysis.........................................................36
C. Modeling (Simulation and System Dynamics) .....................................36
D. Extrapolation .........................................................................................37
2.3 Focused Review on Multi-Attribute Extrapolation Methods..............................39
2.3.1 Intuitive Models .........................................................................................39
A. Scoring Model .......................................................................................39
B. Technology Development Envelope .....................................................40
2.3.2 Parametric Frontier Models .......................................................................41
A. Planar Frontier Model (Hyper-plane) ...................................................41
B. Corrected Ordinary Least Squares ........................................................43
C. Stochastic Frontier Analysis .................................................................44
D. Ellipsoid Frontier Model .......................................................................45
E. Multi-Dimensional Growth Model ........................................................47
2.3.3 Non-parametric Frontier Models ...............................................................49
A. Data Envelopment Analysis ..................................................................49
B. Stochastic (Chance-constrained) Data Envelopment Analysis .............54
2.4 Summary of Critical Review ..............................................................................56
III. Model Development ..................................................................................................58

3.1 Conceptual Framework .......................................................................................58
3.1.1 Segmented Rate of Change ........................................................................58
3.1.2 Infeasible Forecasting Target .....................................................................64
3.2 Formulation .........................................................................................................67
3.3 Illustrative Example ............................................................................................76
IV. Validating through Case Studies.............................................................................80

4.1 Ex Post Analysis: Revisit Earlier Studies ...........................................................81
4.1.1 Risk Analysis: Commercial Airplane Development ..................................81
A. Research Framework.............................................................................81
B. Analysis of Current State-of-the-art ......................................................83
iv
C. Risk Analysis.........................................................................................85
D. Proof of Concept ...................................................................................88
4.1.2 Validation using Past Datasets ...................................................................90
A. Forecasting Accuracy Evaluation Techniques ......................................90
B. Test Results from Earlier Studies ..........................................................92
4.2 Ex Ante Analysis: Focused Application..............................................................95
4.2.1 Exascale Supercomputer Development .....................................................95
A. Background ...........................................................................................95
B. Analysis .................................................................................................99
a) Dataset ...............................................................................................99
b) Model building ................................................................................101
c) Model validation .............................................................................108
d) Forecasting ......................................................................................111
C. Discussion ...........................................................................................114
D. Summary of Findings ..........................................................................118
V. Conclusion ...............................................................................................................120
5.1 Contributions to Application Area ....................................................................120
5.1.1 Exascale Supercomputer Development ...................................................120
5.2 Contributions to Managerial Insight .................................................................122
5.2.1 Risk Analysis ...........................................................................................122
5.2.2 New Product Target Setting .....................................................................123
5.3 Contributions to Methodological Development................................................124
5.3.1 Identification of Local Rate of Change ....................................................124
5.3.2 Identification of Individualized Rate of Change ......................................124
5.3.3 A Finite Forecast for an Infeasible Target ...............................................125
5.4 Limitations ........................................................................................................126
5.5 Future Work Directions ....................................................................................127
5.5.1 Innovative Measure ..................................................................................127
5.5.2 Alternate Efficiency Measures ................................................................127
v
5.5.3 Weight Restrictions ..................................................................................128

5.5.4 Time Varying Rate of Change .................................................................129
5.5.5 Direction of Technological Progress .......................................................129
5.5.6 Stochastic Frontier ...................................................................................129
References .......................................................................................................................131
Appendix. Model Building Guideline...........................................................................158
vi
List of Tables
Table 1. Research gap 1 .......................................................................................................6

Table 4. Research objective .................................................................................................8
Table 5. Research question 1 ...............................................................................................8
Table 8. Commonly used growth models ..........................................................................17
Table 9. Case studies on incumbents success ...................................................................21
Table 10. Summary of literatures on the technology trajectory mapping..........................25
Table 11. Summary of literatures on the new product target setting .................................29
Table 12. Summary of multi-attribute technology assessment models .............................57
Table 13. Forecast results for infeasible targets.................................................................77
Table 14. Commercial airplane dataset ..............................................................................82
Table 15. Local RoC of SOA airplanes at the frontier year of 2007 .................................82
Table 16. Four airplane concepts in 2007 ..........................................................................85
Table 17. Results summary ................................................................................................89
Table 18. Forecast accuracy comparisons .........................................................................92
Table 19. TOP500 dataset from 1993 to 2013 .................................................................100
Table 20. TFDEA model parameters ...............................................................................102
Table 21. Specifications of 13 SOA supercomputers ......................................................103
Table 22. Model validation using a rolling origin hold-out sample tests ........................109
Table 23. Exascale computer as a forecasting target .......................................................111
Table 24. Forecast results of Exascale supercomputer ....................................................113
Table 25. Research question 1 .........................................................................................124
Table 28. Alternate non-radial efficiency measures ........................................................128
vii
List of Figures
Figure 1. Complexity of technology assessment .................................................................4

Figure 2. Overview of dissertation.....................................................................................11
Figure 3. Ambiguity of using time as a proxy for engineering effort ................................15
Figure 4. Technology assessment methods ........................................................................38
Figure 5. Frontier models in input-output space ................................................................44
Figure 6. Two-dimensional illustration of ellipsoid frontier .............................................46
Figure 7. Parametric frontier by MDGM ...........................................................................48
Figure 8. Non-parametric frontier by TFDEA ...................................................................52
Figure 9. Evolution of the SOA frontier ............................................................................59
Figure 10. Illustration of segmented rate of change ..........................................................62
Figure 11. Regions of infeasible super-efficiency under VRS ..........................................65
Figure 12. Forecast deviation distributions ........................................................................78
Figure 13. 2007 state-of-the-art frontier with regard to four design concepts ...................87
Figure 14. Various forecast accuracy tests ........................................................................91
Figure 15. Relative comparison of segmented RoC with constant RoC............................94
Figure 16. System tradeoffs to be considered for the future HPC trend ...........................97
Figure 17. 13 State of the art supercomputers considering system tradeoffs ..................102
Figure 18. Performance trajectories of different processor families................................105
Figure 19. Individualized RoCs with respect to the local RoCs ......................................112
viii
Glossary
Abbreviation
Description
AHP
Analytical Hierarchy Process
AMD
Advanced Micro Devices
C3
Cryogenic Computer Complexity
CAPM
Capital Asset Pricing Model
CI
Confidence Interval
CM
Customer Matrix
COLS
Corrected Ordinary Least Square
CPU
Central Processing Unit
CRA
Comparative Risk Assessment
CRS
Constant Returns to Scale
DEA
Data Envelopment Analysis
DMU
Decision Making Unit
DOE
Department of Energy (U.S.)
DRS
Decreasing Returns to Scale
DVD
Digital Video Disk
EIA
Environmental Impact Analysis
EIS
Entry Into Service
FCM
Fuzzy Cognitive Map
FLOPS
FLoating-point Operations Per Second
FMEA
Failure Mode and Effects Analysis
FS
Frontier Shift
GDF
Geometric Distance Function Model
GPU
Graphics Processing Unit
HDM
Hierarchical Decision Model
HPC
High Performance Computing
IARPA
Intelligence Advanced Research Projects Activity
IO
Input-Oriented
IRS
Increasing Returns to Scale
LCC
Life Cycle Cost
LCD
Liquid Crystal Displays

ix
MCDA(M)
Multi-Criteria Decision Analysis(Method)
MDGM
Multi-Dimensional Growth Model
MPI
Malmquist Productivity Index
MW
Mega-Watt
NDRS
Non-Decreasing Returns to Scale
NEC
Nippon Electric Company
NFC
Near Field Communication
NIRS
Non-Increasing Returns to Scale
NNSA
National Nuclear Security Administration (U.S.)
NSF
National Science Foundation
NUDT
National University of Defense Technology
OEM
Original Equipment Manufacturer
OLED
Organic Light Emitting Diodes
OO
Output-Oriented
PFE
Passenger Fuel Efficiency
PPS
Production Possibility Set
P-SBM
Proportional Slack-Based Model
RAICS
RIKEN Advanced Institute for Computational Science
RAM
Range-Adjusted Model
RCA
Revealed Comparative Advantage
Rmax
Maximal LINPACK performance achieved in Gigaflops
RMSE
Root Mean Square Error
RoC
Rate of Change
RTS
Returns To Scale
SBM
Slack-Based Model
SDEA
Stochastic Data Envelopment Analysis
SFA
Stochastic Frontier Analysis
SOA
State Of the Art
TDE
Technology Development Envelope
TEC
Technical Efficiency Change
TFDEA
Technology Forecasting using Data Envelopment Analysis
UHD
Ultra-High Definition
VRS
Variable Returns to Scale
W-LAN
Wireless Local Area Network

x
List of Symbols
Symbol
j(k)
Each product/technology being compared to (being evaluated)
Input variable index
Output variable index
The number of input variables
The number of output variables
Time of current frontier
Input i of product j
Output r of product j
Output-oriented efficiency of product k at the time of release
Time of release of product j
Output-oriented efficiency of product k at current frontier

Input-oriented efficiency of for product k at the time of release
Input-oriented efficiency of product k at current frontier
Output-oriented intensity vector of product k citing product j at the time of
release
Output-oriented intensity vector of product k citing product j at current
frontier
Input-oriented intensity vector of product k citing product j at the time of
release
Input-oriented intensity vector of product k citing product j at current
frontier
Output-oriented rate of change of product k at current frontier
Output-oriented local rate of change of product j at current frontier
Input-oriented rate of change of product k at current frontier
Output-oriented radial super-efficiency of product k
Output-oriented radial extremity of product k
Input-oriented radial super-efficiency of product k
Input-oriented radial extremity of product k
Description
Input-oriented local rate of change of product j at current frontier
Output-oriented forecasted time of arrival of product k

Input-oriented forecasted time of arrival of product k
xi
I. MOTIVATION
1.1 INTRODUCTION
If the future unfolded as foretold, individuals and governments would get a great
benefit from their actions taken in advance. In addition to handsome payoffs from Wall
Street, semiconductor companies could perfectly meet their market demands with newly
built fabs, and sportswear companies could get the maximum advertising effects by
grasping rising sports stars with long-term contracts. However, in practice, black swan
events often render our plans just plain useless or ineffective [1], [2]. Hence, the choice
between alternative pathways under estimated future states may significantly alter
competitive performance. Note that decision making is inseparable from how the future is
expressed. Indeed, we explicitly or implicitly pay attention to the trends and ideas that are
shaping the future which form the basis of our everyday decisions.
Future research community theoretically differentiates prediction from forecasting. A
prediction is concerned with the future that is preordained and no amount of action in the
present can influence the outcomes. Therefore, it is an apodictic, i.e., non-probabilistic,
statement on an absolute confidence level about the future [3]. Clearly, the goodness of a
prediction lies in whether it eventually comes true. A forecast, on the other hand, is a
probabilistic statement on a relatively broad confidence level about the future.
Fundamentally, it aims to affect the decision making process by investigating possible
signals related to the future events using systematic logic that forecasters must be able to
articulate and defend [4]. Thus, except for particular purposes (e.g., benchmark study), a
good forecast is determined not by whether it eventually came true but by whether it
1
could provide reasonable grounds to support an organizations actions to anticipate

identified uncertainties, thereby resulting in a better future than what was most likely
without specific actions. This is consistent with a popular saying in Oriental philosophy:
The ultimate goal of forecasting is to make that forecast wrong.
Just as in other areas of forecasting, technological forecasting intends to improve the
quality of decisions by providing specific pieces of information focused on technologies.
Lenz identified six major roles of technological forecasting that can be summarized as
follows [5]:
1. To identify limits of current technology systems
2. To establish rates of progress
3. To describe technology alternatives
4. To indicate the feasibilities of technology alternatives
5. To provide a reference standard for the new product development plan
6. To furnish warning signals
Technological forecasting methods can be classified as either exploratory or
normative by whether they extend present trends (exploratory) or look backward from a
desired future to determine the developments needed to achieve it (normative) [6]. The
correct assessment of the future environment and of the corresponding goals,
requirements, and human desires can be better made when exploratory and normative
components are joined in an iterative feedback cycle [3]. Here, it is crucial to have an
accurate understanding of the technological inertia we have today so that exploratory
methods extend the current rate of progress, while normative methods determine how
2
much the speed of such progress might need to be adjusted. However, as technology
systems become more sophisticated, the rate of change varies more dramatically due to
the maturity levels of component technologies [7]. This structural complexity makes
todays forecasting even more challenging, which leads to the question: What is the best
way to combine growth patterns of the various attributes used to describe multi-objective
technology systems?
To answer the above question, two things must be considered: multi-attribute
evaluation and technology segmentation. Multi-attribute evaluation strives to define the
goodness of technology systems that consist of different levels of subsystems. Figure 1
illustrates the difficulty of doing this. Technology B seems to have made a disruption in a
high-end market, while technology A is overshooting the market in terms of technical
capacity 1. However, technology 2 might have been superior and recently challenged by
technology 1 on a different technical dimension. Possibly, different dynamics are taking
place in other dimensions as well where the levels of market demands also vary. This
implies that a single performance measure may be no longer capable of capturing
advancement in a new direction, which makes the holistic assessment of technology
systems difficult. Therefore, it is critical to examine not only which performance
measures are playing a major role in current technological progress but also which
alternate technologies show disruptive potential with respect to emerging performance
measures.
Technology segmentation is related to the identification of homogeneous technology
clusters. Technologies belonging to the same cluster may have a similar mix of technical
capabilities whereby they satisfy the similar target markets. From the technological
3
forecasting point of view, each technology cluster is expected to involve a particular

progress pattern to which similar types of future technologies are subjected. Thus, it is
imperative for a forecasting model to be able to identify distinct technological segments,
if they exist, and treat them separately in the process of capturing the rate of change and
estimating the future performance specifications and features.
This dissertation aims to develop a new frontier analysis method for technological
forecasting that can deal with foregoing two issues: multi-attribute evaluation and
technology segmentation.
Figure 1 Complexity of technology assessment
1.2 PROBLEM STATEMENT

Technology growth curve (or life cycle) theory asserts that the accumulated
performance of a technology forms an S-shaped curve when it is plotted over time, and
the diminishing returns to effort is successively overcome by following growth curves
that possess higher performance limits [8][11]. Although S-shaped growth patterns have
been observed in numerous studies [12][14], a fact well known is that fitting some
portion of cumulative data to a predefined S-curve function, i.e., predictive use of the
curve, is susceptible to several technical assumptions such as engineering effort, is
constant over time, and an upper limit is known in advance.
Disruptive innovation theory integrates the technological growth patterns with
different market acceptance levels to explain the process of new value network creation
[15]. Even though disruptive innovation has contributed to the understanding of industry
dynamics, the practical implications remain debatable [16][18]. In particular, the theory
is criticized that it lacks the mechanism for the predictive use [19]. Christensen suggested
a diagram that jointly portrays trajectories over time of performance demanded by
different market segments and of performance provided by alternative technologies [20].
However, this comes down to the difficulties of employing a growth curve model as an ex
ante analysis.
The recent development of product categorization methods can be viewed as another
stream of literature with regard to the topic of risk analysis for new product development
[21][25]. However, although these approaches can shed light on the new product target
setting practices, there remains a need to integrate the product benchmarking with the
assessment of performance improvement so that analysts can have a risk measure for
their product launch strategy.
The above mentioned problems can be summarized as below.
Table 1 Research gap 1
GAP
#1
Current technological innovation and new product development theories

do not provide a quantitative framework to facilitate the predictive use of
the theory.
Frontier analysis models attempt to form a surface that can represent the same level
of technology systems at given point in time. The evolution of surfaces is then monitored
to capture the rate of change by which future technological possibilities can be estimated.
In the case of parametric frontier methods, an iso-time frontier is constructed as a
functional combination from individual growth curves. Specifically, actual observations
are fitted to an a priori defined functional form, and those growth patterns are combined
together to constitute an iso-time frontier. Therefore, it is difficult to identify distinct
technological segments from the resulting frontier.
Non-parametric frontier methods, on the other hand, have an advantage with regard
to the technology segmentation since the frontier is directly constructed by dominating
technologies that are located on the frontier. This enables the model not only to
characterize each frontier segment but also to identify proper segments that dominated
technologies belong to. However, current non-parametric frontier models dont
incorporate this property into the forecasting process. Instead, they simply aggregate rate
of changes captured from the surpassed technologies to indicate the technological
progress as a whole.
6
The above mentioned issues can be summarized as below.

GAP
#2
Current technological forecasting methods do not take into account

technology segmentation.
Non-parametric frontier methods reflect the distinct characteristics of the technology

systems by directly adapting to observed data without relying on arbitrary functional
assumptions [26]. However, this characteristic often leads to an infeasibility problem
since it doesnt generate the frontier facet for non-existent production possibilities in the
past [27]. This becomes a critical problem when the model estimates the distance from
the current frontier to forecasting targets. In other words, the model may fail to measure
the goodness of future technologies if it has an unprecedented mix of technical
capabilities. This issue can be summarized as below.
GAP
#3
Current non-parametric frontier analysis methods often suffer from

concerns regarding infeasibility.
1.3 RESEARCH OBJECTIVE

Based on the aforementioned research gaps, the objective of this research is specified
as below.
Table 4 Research objective
Research
Objective
Develop a new frontier analysis method for technological forecasting

that is capable of multi-attribute evaluation by considering technology
segmentations with the corresponding rate of changes.
The technology segmentation process can be understood as two procedural stages.

First, it is required for the model to identify distinct facets on the frontier in which
local rates of change can be obtained from the surpassed technologies. By doing so, each
state-of-the-art technology will have a local rate of change which indicates how much
progress has been observed by each technological segment represented by those
technologies. This corresponds to the first research question as below.
Table 5 Research question 1
Research
Question
#1
How do we capture the local rate of change from past technologies?
Once the local rates of change are ascertained with respect to each frontier segment,
they can be utilized to obtain the individualized rate of change for each forecasting target.
This procedure makes it possible for the model to apply the customized progress rate
suitable for each forecasting target, thereby reflecting the characteristics of identified
segments into the forecast. This leads to the second research question as below.
8

Research
Question
#2
How do we determine the individualized rate of change for future

technologies?
As previously discussed, non-parametric features of identifying the technology

segmentations come at a cost to the infeasibility problem. Therefore, a new model should
be able to provide an alternative forecasting mechanism in a consistent manner, i.e.,
without violating a radial distance measure, for the case when the infeasibility problem
occurs. This requires an assumption that the superiority of unprecedented types of
technologies can be estimated by the consideration of how far the target technology is
away from the closest technology segment as well as how good the target technology is
from a perspective of the closest technological segment. The formulation of this
procedure is subject to the following research question.
Research
Question
#3
How do we deal with infeasible targets?
1.4 OVERVIEW OF DISSERTATION

This dissertation is organized as follows. In Chapter II, critical literature reviews on
technological innovation theories and technological forecasting methodologies are
provided to address the objective of this dissertation. In Chapter III, the notion of
segmented rate of change is illustrated using a numerical example as well as
mathematical formulation to supply insight into the problem being discussed. In Chapter
IV, a set of applications is provided to demonstrate and validate the proposed model.
Finally, Chapter V summarizes contributions and suggests possible future research
directions.
10
Figure 2 Overview of dissertation

11
II. BACKGROUND
2.1 TECHNOLOGICAL INNOVATION THEORIES

2.1.1 S-CURVE MODELS
Beginning with Kondratievs early observation in 1925, technological innovation has
been believed to exhibit a wave-like phenomenon that consists of life cycles of individual
technologies [8][11]. According to this view, the accumulated performance of a
technology forms an S-shaped curve when it is plotted over time, and the diminishing
returns to effort is successively overcome by following growth curves that possess higher
performance limits. In this belief, Cesare Marchetti once claimed that Anything that
begins and ends an existence will fit a logistic (also known as a Pearl curve) during the
conversation with Theodore Modis, who holds the Guinness Book of Records for carrying
out the greatest number of logistic fits [28].
Simply speaking, an S-curve is a natural outcome which is a cumulative function of
growth levels from typical evolution stages: introduction stage (slow acceleration),
growth stage (speed up), and maturity stage (deceleration). Therefore, it may be worth
understanding the reason why the rates of progress in each stage are different and
especially what hinders the growth during introduction and maturity stages.
In the introduction (or embryonic) stage, a new technological platform makes slow
progress, mostly due to the insufficient interests of the research community [29]. A
consensus between alternate ways of overcoming the bottlenecks is called for; however,
there may be no measures for assessing the new technology properly, which makes the
circumstances of the decision be based less on problem solving ability than on future
12
promise [30]. Therefore, until the new approach has gained established legitimacy as a
worthwhile endeavor, great effort is often spent exploring different paths to identify
meaningful and feasible drivers of advancement. For example, OLED (organic light
emitting diodes) technology has been recently introduced as a new alternative to LCD
(liquid crystal displays) in the flat panel industry. However, it requires a sufficient
amount of time and effort to identify the direction of incremental innovation.
In the growth stage, a new technological platform finally crosses a threshold with
continuous engineering effort, which allows rapid progress [13]. The emergence of a
dominant design, in particular, plays a key role not only to attract researchers to
participate in its development but also to coalesce product characteristics and consumer
preferences [12]. The cumulative efforts reap the greatest improvement per unit of effort,
which creates a virtuous cycle by stimulating more attention devoted to the current
technological platform.
In the maturity stage, the progress slowly and asymptotically reaches a ceiling [10].
Utterback suggested that as a market ages, the focus of innovation shifts from product to
process innovation [31]. Sahal also argued that technology has inherent limits of scale
and/or complexity which restrict the steady growth of performance improvement [14]. As
such, a marginal performance increase requires more cost and engineering efforts, which
eventually deglamorize the current technological platform. As the current technological
platform loses its luster, the research community searches for alternative paths and
rapidly loses cohesion, which reduces the switching costs to the upcoming technology.
Although the S-shaped growth pattern has been observed in a number of studies that
conducted retrospective analysis on industry dynamics, it is well known that fitting some
13
portion of cumulative data into a predefined S-curve function is susceptible to several

technical assumptions, which renders the predictive use of the curve ambiguous [32].
The first assumption is that the engineering effort is constant over time.
Fundamentally, the S-curve reflects the level of technological capability relative to the
cumulative effort invested in developing the technology [33]. However, engineering
effort is rarely used as an abscissa in practical applications due to the difficulty in
tracking the cumulative effort over an entire technology cycle. Instead, there are two
alternative parameters that are often used in place of engineering effort: research and
development investment or time. However, the problem arises when these proxy
measures are not proportional to actual engineering efforts. Fig. 3 shows that the resulting
curve can obscure the true relationship, which could appear to atten out much more
quickly, or not to atten out at all when this assumption is not supported [34].
14
Figure 3 Ambiguity of using time as a proxy for engineering effort (modified from [32])
The second assumption is that the upper limit of a growth curve, i.e., L, is given.
However, it is rare that the true limit of a technology is known in advance, and there is
often considerable disagreement about what the limits of a technology are. A well-known
case of misperception can be found in the disk drive industry [35]. In 1979, IBM had
reached what it perceived as a density limit of ferrite-oxide-based disk drives; therefore,
the company moved to developing thin-lm technology that had a greater potential for
increasing density. Hitachi and Fujitsu, however, continued to ride the ferrite-oxide S-
15
curve, and ultimately achieved densities that were eight times greater than the density that
IBM had perceived to be a limit.
Due to the lack of information, researchers often have no choice but to employ
regression-based calculation to estimate the upper limits of technology growth curves [6].
However, this approach is controversial in the literature [36], [37]. Danners simulation
showed that the accuracy of the resulting limit is highly sensitive to any error present in
the segment of the available data [32]. Martino also argued that the productivity of early
technology development is only minimally influenced by the upper limit because
historical data from the early stages of development contain little information as to the
location of the upper limit [38]. In this sense, he claimed that even a small error in the
upper limit estimation can result in a fairly significant error in the forecast.
The third assumption is that the appropriate growth model is predefined. However,
similar to the estimation of upper limits, a growth curve should not be selected based on
goodness of fit from historic data but on matching the behavior of the selected growth
curve to the underlying dynamics of technology growth [39], [40]. In fact, there are
various equations that represent S-shaped curves which can be categorized into two main
groups: absolute and relative models. The former quantifies the technical capability, ,
as a function of the independent parameter time, t, whereas the latter quantifies the rate of
change in technical capability, , as a function of the most recently achieved level of
technical capability, 1 (see Table 8) [32], [41].
Youngs study showed that relative models were more accurate than absolute models,
and in particular both the Bass and Harvey growth models performed well under most
circumstances [37]. Danner claims that this may be because the inherent characteristic of
16
the relative model that each new data point is anchored to the previous data point. That is,
changes in the relative model are proportional to both the progress to date and the
distance to the upper limit, whereas changes in the absolute model are only proportional
to the distance from the upper limit [32].
Table 8 Commonly used growth models [32], [37]

Type
Name
Equation
Absolute
Logistic (Pearl) [42]
Gompertz [43]
Linear Gompertz [41]

Mansfield-Blackman [44]
Weibull [45]
Relative
1 +
ln ( ln ( )) = 0 + 1
) = 0 + 1
ln (

|) = 0 + 1 ln
ln(ln |

Von Bertalanffy [43]
= (1 )3
S-curve [43]
= (/)
Bass [46]
= 0 + 1 1 + 2 1 2
Non-symmetric Responding
Logistic [47]
ln = 0 + 1 ln(1 ) + 2 ln( 1 )
Harvey [48]
ln = 0 + 1 + 2 ln(1 )
Extended Riccati [49]
1
) + 3 ln(1 )
= 0 + 1 1 + 2 (
1
1
However, it should be noted that the underlying assumption of the relative model
that future advancement is facilitated by technical capability already achieved may not
always be true. In a similar context, Youngs finding, based on 46 historical data sets,
17
might not guarantee the most appropriate selection for the predictive problem at hand.
Therefore, model selection should consider how the models behavior matches the
process that generated the data rather than simply fitting the historical data [50].
The shape of a technologys growth curve is neither set in stone nor given to the
analyst. The limitation of the current architecture can be overcome by technological
innovation which affects the growth rate and possibly allows a higher performance to be
achieved than what had been perceived to be a limit. On the contrary, the lifecycle of a
given technology could be terminated by the unexpected adoption of alternative
technologies even before it passes the inflection point of the curve [51]. Therefore, fitting
a portion of data into an a priori defined growth function should be accompanied by a
deep understanding of the dynamics of the industry being investigated.
18
2.1.2 DISRUPTIVE INNOVATION 1

Christensen asserted his theory of disruptive innovation by integrating technological
growth patterns with different market acceptance levels to explain the process of new
value network creation [15], [52]. The core premises of his theory can be summarized as
follows [17], [18]:
a) A new technology consisting of off-the-shelf components put together in product
architecture provides different values from mainstream technologies and initially
underperforms the dominant one;
b) Products based on disruptive technologies could, therefore, only serve niche
segments without attracting attention of the leading firms;
c) Disruptive technologies steadily improve in performance until it meets the
standards of performance demanded by the mainstream market;
d) Further development could raise the disruptive technologys performance to a level
sufficient to satisfy mainstream customers, while the established technology could
have exceeded the demand of mainstream customers, resulting in performance
overshoot.
Although Christensens work on disruptive innovation has contributed to the
understanding of industry dynamics, the mechanism and practical implications of his
theory remain controversial.
First of all, the theory lacks a clear-cut definition to determine whether or not a given
technology is considered to be disruptive. Christensen explained that disruptiveness
depends on whether it is consistent with their business model. For example, the Internet is
1
This section is adapted from a paper accepted in R&D Management [312]
19
sustaining to catalog retailers and discount brokers, but it is disruptive to department

stores and full-service brokers [20]. Further he described the characteristics of disruptive
innovation as being typically simpler, cheaper, and more reliable and convenient than the
established one [53]. However, Danneels rebutted that these characteristics may be
typical but not necessary characteristics of disruptive technology. He gave examples of
Amazon.com, digital cameras, and digital video disk (DVD) that have had higher
performance when they were introduced in the mainstream market [16]. Chesbrough also
noted that Christensens studies used inconsistent terminology; in other words, they
lacked common criteria to classify different types of technologies [54]. This ambiguous
definition also raised a question: What determines whether incumbents fail or succeed in
the face of disruptive technology? (see Table 9).
In addition, the predictive use of disruptive innovation theory has been a controversial
issue in managerial practice. Cohan contended in his book The Dilemma of the
Innovators Dilemma that Christensen tried to support his theory using cherry-picked
examples, i.e., only case studies of disruptive technologies that succeeded [19]. He urged
that retrospective analysis is subject to bias, and the real challenge to any theory,
especially if it is to be useful managerially, is how it performs predictively. Christensen
responded to this by suggesting a diagram that jointly portrays trajectories over time of
performance demanded by different market segments and of performance provided by
alternative technologies [20]. For ex post case studies, using trajectory maps is fairly
straightforward since the key performance dimension that resulted in a disruption has
been identified and that data on performance demanded and supplied are available.
However, ex ante analysis requires predicting what performance the market will demand
20
along various dimensions and what performance levels technologies will be able to
supply [16].
Table 9 Case studies on incumbents success

Author (Year)
Application area
Cooper and Schendel (1976) [55]

Foster (1986) [10]
Tushman and Anderson (1986) [56]
Henderson and Clark (1990) [57]
Henderson (1993) [58]
Utterback (1994) [11]
Christensen and Bower (1996) [52]
Chesbrough (1999) [59]
McKendrick et al. (2000) [60]
Tripsas and Gavetti (2000) [61]
Helfat and Lieberman (2002) [62]
Tushman and Anderson (1986) [56]
Chesbrough (1999) [59]
Cohan (2010) [19]
McKendrick et al. (2000) [60]
King and Tucci (2002) [63]
Danneels (2004) [16]
Klepper and Simons (2000) [64]
Chandy and Tellis (2000) [65]
Tripsas (1997) [66]
Rothaermel (2001) [67]
Darby and Zucker (2001) [68]
Videotape, Teleconferencing, etc.

Video game console
Competence destroying discontinuities
Copy machine, Radio receiver, etc.
Photolithographic alignment equipment
Electric lamp, QWERTY, etc.
Seagate, Control Data
Hard disk drive (in U.S)
Disk drive (transition to 5.25-inch)
Polaroid
Canon and Nikon
Competence enhancing discontinuities
Hard disk drive (in Japan)
Charles Schwab, Kodak and Fuji
Disk drive (transition to 8 & 3.5-inch)
Disk drive
Non-fossil fuel powered automobile
Television sets
Office products and consumer durables
Mergenthaler Linotype
Pharmaceutical industry
Chemistry to biotechnology
Did incumbents
succeed?
No
No
No
No
No
No
No
No
No
No
No
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Trajectory mapping has been employed in a wide range of applications. The most
famous application of a trajectory map may be the hard disk drive case from
Christensens original work [35]. He used disk capacity as a performance axis and
interpreted the dynamics of industry that smaller disks have replaced bigger ones,
improving their capacities over time. Schmidt later extended Christensens work by
classifying the disk drive case as a low-end encroachment that eventually diffused
21
upward to the high-end [69]. Martinelli conducted patent analysis in the

telecommunication switching industry to discover seven generations of technological
advances from the different paradigmatic trajectories [70]. Kassicieh and Rahal also
adopted patent publication as a performance measure in search of potential disruptive
technologies in therapeutics [71]. Phaal et al. proposed a framework that has been tested
by developing more than 25 diverse emergence maps, analogous to a trajectory map, of
historical industrial evolution, building condence that the framework might be
applicable to current and future emergence [72]. Keller and Hsig analyzed Googles
web-based office application to see if it can pose a disruptive threat to incumbent
technologies, namely Microsofts desktop office application [73]. Barber-Toms and
Consoli tried to identify potential disruptive innovations in the medical industry,
especially for artificial discs, by counting the number of granted patents over time [74].
Husig, et al. (2005), conducted one of the rare ex ante analyses that mapped out
trajectories of both the incumbent technology and a potential disruptive technology [75].
They made a forecast based on a trajectory map that wireless local area network (W-LAN)
technologies would not be disruptive for incumbent mobile communications network
operators in Germany. This is because the average growth rate of the bandwidth supplied
by W-LAN had been overshooting the average growth rate of the bandwidth
requirements of all customer groups.
There are a few studies that used composite performance measures to draw the
technology trajectories. Adamson plotted R2 values from the multiple regression analysis
on the trajectory map to investigate the fuel cell vehicle industry [76]. The results showed
that the subcompact vehicles R2 values were increasing over time while compact
22
vehicles were decreasing. The author used this information and identified key utility
attributes that could command a significant premium before the product reaches the mass
market. This study has significant implications for identifying key drivers of technology
progress using the trajectory map. Letchumanan and Kodama mapped out the correlation
between Revealed Comparative Advantage (RCA), which is generally used to measure
the export competitiveness of a product from a particular country in terms of world
market share, and R&D intensity to examine who was making the most disruptive
advancement at a national level [77]. Even though Koh and Magee didnt utilize any
function to develop composite performance measures, their research has a significance as
they took different trade-offs into consideration to draw a trajectory map [78]. Their
results suggested that some new information transformation embodiment such as a
quantum or optical computing might continue the trends given the fact that information
transformation technologies have shown a steady progress.
Few researchers have proposed the predictive approach of the disruptive innovation
theory considering multidimensional aspects of technology systems. Schmidt suggested
using part-worth curves in search of low-end encroachment [69]. Paap and Katz provided
general guidance for ex ante identification of future disruption drivers [79]. Several
authors have suggested using extant methods for technological forecasting to assess
potential disruptive technologies [16], [17]. Govindarajan and Kopalle argued that
capturing a firms willingness to cannibalize could be a sign of ex ante prediction of
disruptive innovation [80]. Doering and Parayre presented a technology assessment
procedure that iterates among searching, scoping, evaluating, and committing [81].
23
Table 10 summarizes 40 studies that have employed the trajectory map to identify
disruptive alternatives (technology, product, service, etc.). The majority of the studies
adopted a single performance measure to draw the trajectory map. A trajectory map,
however, should be able to take multiple perspectives into account not to miss potential
disruptive indications. Many ex post case studies have shown that disruptions have
occurred from an entirely new type of performance measure that hadnt been considered.
Furthermore, it was often observed that the new technology started below the prior one in
performance on the primary dimension but was superior on a secondary one [18]. This
implies that the current performance measure may no longer be capable of capturing
advancement in a new direction. Therefore, it is crucial to examine not only which
performance measures are playing a major role in current progress but also which
alternate technologies show disruptive potential with respect to the emerging
performance measures.
24
Table 10 Summary of the literature on the technology trajectory mapping

Author (year)
Application area
Performance measure
Plotting method
Walsh (2004)
Keller & Hsig (2009)
Martinelli (2012)
Phaal et al. (2011)
Padgett & Mulvey (2007)
X. Huang & Soi (2010)
Kaslow (2004)
Kassicieh & Rahal (2007)
Christensen (1997)
Schmidt (2011)
Rao et al. (2006)
Bradley (2009)
Lucas & Goh (2009)
Madjdi & Hsig (2011)
Husig et al. (2005)
Walsh et al. (2005)
Figueiredo (2010)
Caulkins et al. (2011)
Adamson (2005)
Belis-Bergouignan et al.
(2004)
Ho (2011)
Werfel & Jaffe (2012)
Microsystems
Office application
Telecommunication
S&T based industry
Brokerage market
General industry
Vaccine
Therapeutics
Disk drive
Disk drive
P2P and VoIP
Medical operation
Photography
W-LAN
W-LAN
Silicon industry
Forestry industry
General industry
Fuel cell vehicle
Organic compound
Critical dimension
Number of operations
Patent citation
Sales
Service integration level
Capacity & Price
Efficacy
Patent publication
Capacity
Part-worth
Data transfer
Noninvasiveness
Price, convenience, etc.
Active Hotspot ratio
Data rates
Number of firms
Novelty & complexity
Market connection
Utility coefficient values
Environmental
performance
Technology sources
Patent
Growth curve
Data accumulation
Data accumulation
Data accumulation
Data accumulation
Data accumulation
Data accumulation
Patent mapping
Data accumulation
Data accumulation
Data accumulation
Data accumulation
Data accumulation
Data accumulation
Data accumulation
Data accumulation
Data accumulation
Skiba curve
Data accumulation
Data accumulation
No & Park (2010)

Letchumanan & Kodama
(2000)
Spanos & Voudouris
(2009)
Frenken & Leydesdorff
(2000)
Watanabe et al. (2009)
Hobo et al. (2006)
Watanabe et al. (2005)

S.-H. Chen et al. (2012)
Epicoco (2012)
Funk (2005)
Raven (2006)
Castellacci (2008)
Kash & Rycoft (2000)
Arqu-Castells (2012)
W.-J. Kim et al. (2005)
C.-Y. Lee et al. (2008)
Koh & Magee (2006)
Barber-Toms & Consoli
(2012)
General industry
Smoking cessation
products
Nano-biotechnology
General industry
(High-tech)
Manufacturing SMEs
Patent
Correlation between
Exports and R&D intensity
AMT
Civil aircraft
Printers
Service oriented
manufacturing
industry
Electrical machinery
Smart grid
Semiconductor
Mobile phone
Renewable energy
Manufacturing and
service industries
Radiation therapy
General industry
DRAM
Home networking
IT
Artificial disc
25
Data accumulation
Reduced form
model
Data accumulation
Data accumulation
Data accumulation
Diffusion rate
(Entropy statistics)
Sales and price
Sales, income, employees,
and productivity
Data accumulation
Marginal productivity
Average age
Devices per chip
Mobile subscribers
Energy Production
Labor productivity
Data accumulation
Data accumulation
Data accumulation
Data accumulation
Data accumulation
Data accumulation
Capability
Patent
DRAM shipment and
Memory density
New household/year
Megabits
Patent
Growth curve
Poisson model
Data accumulation
Price function
Data accumulation
Data accumulation
Data accumulation
Data accumulation
2.1.3 NEW PRODUCT TARGET SETTING

Product target setting is one of the most essential practices in the early stage of new
product development to ensure that the firm pursues the right markets and products from
a strategic viewpoint [111]. This involves decisions about the target market, product mix,
project prioritization, resource allocation, and technology selection [112]. It should be
noted here that, unlike the traditional approach stipulating that the product concept be
frozen before detailed product design commences, it may be optimal to pursue multiple
product concepts and select the best design through an iterative process [113][115].
The body of literature relevant to this topic can be divided into four groups: marketfocused approach, organization-focused approach, operations management-focused
approach, and engineering design-focused approach.
The market-focused approach views customer utility as a function of product
attributes, hence the emphasis is placed on collecting customers value propositions for
product positioning and pricing by adopting various market research methods [116]
[119]. Those classic approaches include brainstorming and Delphi [120], morphology (or
morphological analysis) [121], and lead users analysis [122]. In addition, recent attempts
such as the voice of the customer [123], probe and learn [124], empathic design [125],
fuzzy cognitive map [126], and crowdsourcing [127], [128] have been used to derive
promising product concepts from consumers perception as well as underlying behaviors.
The key difference between the market-focused approach and engineering design-focused
approach is that the former places more emphasis on product concept generation, whereas
the latter is more concerned with product concept selection with the determination of
specific attribute levels [129].
26
The organization-focused approach is focused at a relatively aggregate level on the

determinants of project success from the perspective that a product is an artifact resulting
from an organizational process [111]. Consequently, typical subjects include
development team organization (functional, project, or matrix) [130], [131], team staffing
[132], [133], team performance measurement [134][136], arrangement of the
development team [137][139], infrastructure and training [140][142], and development
stage-gate [143][145]. Recently, the critical roles of leadership and of communication
and conict management training are receiving extensive attention as strategies for
overcoming the challenges to team effectiveness in new product development [146],
[147].
The operations management-focused approach can be viewed as a stream of literature
with regard to the topic of financial and business environment analysis for new product
development [148], [149]. This approach has a viewpoint that a product is a sequence of
development and/or production process steps with the goal being to achieve high
efficiency across the steps [111]. The topics, therefore, mainly focus on capacity
utilization [150], process performance [151], development sequence and schedule [152],
supplier and material selection [153], etc. Specifically, one of the most popular models is
capital asset pricing model (CAPM), which attempts to optimize the level of return at the
lowest possible level of risk within a product portfolio [154]. Failure mode and effects
analysis (FMEA), similarly, provides a framework to identify actions that could eliminate
or reduce the likelihood of the potential failure of a product or process [155], [156]. More
recently, Markeset and Kumars study proposed the integrated model of RAMS, i.e.,
reliability, availability, maintainability and supportability employing the life cycle cost
27
(LCC) analysis within the stage gate model for project and work process management
[157].
Much of the engineering design-focused approach is at a more detailed level of
abstraction with the focus being the individual product or specific market [111]. The
perspective of this approach is that a product is a complex assembly of interacting
components [134]. Perhaps the most known method is conjoint analysis, which attempts
to identify the ideal combination of product size, shape, configuration, function, and
dimensions [158]. Furthermore, recent attention to the product categorization has been
enhanced by benchmarking studies in an attempt to identify distinct combinations of
product attributes. An initial work related to this product-focused approach may be found
in Doyle and Greens study which used a widely known benchmarking technique, data
envelopment analysis (DEA), to identify homogeneous product groups, i.e., competitors,
as well as market niches [21]. Specifically, they applied DEA to classify printers by
ordering them from broad to niche based on the number of times each printer appears in
others reference sets. In a similar vein, Seiford and Zhu developed measures for products
attractiveness and progress by separating context-dependent frontiers [22]. Further, Po et
al. showed how this product feature-based clustering can be used for decision makers to
know the changes required in product design so the product can be classified into a
desired cluster [23]. Amirteimoori and Kordrostami later extended this approach to take
the size of products into account, thereby comparing products grouped by scale [24]. In
addition, Amin et al. clarified the role of alternative optimal solutions in the clustering of
multidimensional observations from the DEA approach [159]. Most recently, Dai and
28
Kuosmanen proposed a new approach that can take into account the cluster-specific
efficiency rankings as well as stochastic noises [25].
Table 11 Summary of literatures on the new product target setting

Classification
Perspective on product
Methods (or main topics of discourse)
Marketfocused
A product is a bundle
of attributes
Brainstorming
Delphi
Morphology
Lead user analysis
Organization- A product is an artifact

focused
resulting from an
organizational process
Development team
organization
Team staffing
Performance
measurement
Operations
managementfocused
A product is a sequence
of development and/or
production process
steps
Capacity utilization
Process performance
Development
sequence and schedule
Supplier and material
selection
Engineering
designfocused
A product is a complex
assembly of interacting
components
Conjoint analysis
Data envelopment
analysis (DEA)
29
Voice of customer
Probe and learn
Empathic design
Fuzzy cognitive map
Crowdsourcing
Team arrangement
Infrastructure and
training
Development stategate
Leadership and
communication
Capital asset pricing
model (CAPM)
Failure mode and
effects analysis
(FMEA)
Reliability,
availability,
maintainability and
supportability
(RAMS)
Life cycle cost
(LCC)
2.2 TECHNOLOGY ASSESSMENT METHODS

2.2.1 QUALITATIVE MODELS
A. DELPHI AND EXPERT OPINION
The Delphi technique was developed during the 1950s by workers at the RAND
Corporation and became publicized by Dalkey and Helmer in 1963 [120]. Delphis
consensus building process is derived by a series of intensive questionnaires interspersed
with controlled opinion feedback. It is believed that as a result of contacting the panelists
by letter (or electronic correspondence), the disturbing factors of group discussion such
as bandwagon effect produced by the majority opinion can be minimized [3]. Delphi may
ask experts about direct forecasts of technological parameters or likelihoods of future
events; however, it was originally intended for use in judgment and forecasting situations
in which pure model-based statistical methods are not practical or possible [160]. Thus,
Delphi is applied to the most important task of setting up goals on higher levels such as
social, national, and corporate goals. Recent applications include emerging infectious
animal diseases [161], health and social care [162], Basque university systems [163], and
essential drugs needed for quality care of the dying [164].
B. SCENARIO ANALYSIS
Scenario analysis postulates a set of imaginative descriptions that can encompass the
plausible range of future aspects [6]. This technique particularly attempts to set up a
synoptic view of as many developments as can be grasped and as may appear relevant to
an experimental simulation of a possible reality. Kahn asserted the importance of
scenario mapping by saying that a specific estimate, conjecture, or context, even if it is
30
later shown to have serious defects, is often better than a deliberate blank which tends to
stop thought and research [165].
This technique is usually integrated with other forecasting models not only to identify
a firm basis of possibilities but also to investigate the impact of technology interactions
under the various conditions. Recent hybrid applications of scenario analysis include
Nowack et al.s Delphi-based model [166], Winebrake and Creswicks analytical
hierarchy process (AHP)-based model [167], Kok et al.s participative back casting-based
model [168], and Jetter and Schweinforts fuzzy cognitive map (FCM)-based model
[169].
C. RISK ASSESSMENT AND ENVIRONMENTAL IMPACT ANALYSIS
Risk analysis pays particular attention to the negative impact of technologies on
social institutions and critical infrastructure [170]. Linkov et al. surveyed the comparative
risk assessment (CRA), multi-criteria decision analysis (MCDA), and adaptive
management methods applicable to environmental remediation and restoration projects
and asserted that it is required to shift from optimization-based management to an
adaptive management paradigm for the conservation of the ecosystem [171]. Recently, as
an attempt to distinguish and categorize the potential risk in advance, more attention has
been paid to developing the predictive models. As an example, Kolar and Lodge
developed an ecological risk assessment model to evaluate the risk of alien species for
nonagricultural systems [172].
In a similar vein, environmental impact analysis (EIA) has become an important and
often obligatory part of todays technology assessment [173]. Ramanathans study
applied a multi-criteria model to capture the perceptions of stakeholders on the relative
31
severity of different socio-economic impacts, which will help the authorities in

prioritizing their environmental management plan [174]. Van der Werf and Hayo
compared 12 indicator-based approaches to assess the environmental impact at the farm
level and provided a set of guidelines for the proper application of EIA [175].
D. RELEVANCE TREES
The notion of relevance trees was first proposed by Churman et al. in an attempt to
aid decision making in general industrial contexts [3], [176]. The qualitative relevance
trees are frequently employed in conjunction with scenario analysis to estimate the
significance of criteria. The process can be simplified as follows. First, the relevance
trees, namely hierarchical decision models, are constructed to assess missions, objectives,
goals, strategies, activities, etc. From the scenario analysis, a number of criteria are
derived for each level of the relevance trees. Then, the weights and significance numbers
are estimated on the basis of the identified scenarios. Finally, the decision matrices are
computed to provide a set of alternative actions for each scenario.
The extended applications using relevance tree can be found in recent studies,
including Ghiculescu et al.s model integrating the decision structure with customer
matrix (CM) [177] and Manuel and Pretoriuss model using criteria derived from
relevance trees as inputs of neural network [178].
32
E. MORPHOLOGY
The morphology (or morphological method) was developed by Zwicky in 1962 [121]
in an attempt to deduce all of the solutions of a given definite problem. The method
proceeds as follows:
1. The statement of the problem, i.e., the object of the morphological device, is
made.
2. The precise definitions of the class of devices are elaborated.
3. Related parameters with sub-elements are grouped as matrices and listed for
connection.
4. The alternative solutions are obtained as chains of selected elements from each
matrix.
5. Determine the performance values of all of the derived solutions.
6. Select the particularly desirable solutions and their realization.
This process provides a framework for thinking in basic principles and parameters, which
is growing in importance, even if practiced in a disordered or ad hoc fashion [3].
Recent developments of morphology technique tend to be in conjunction with data
mining approaches with the current heavy interests on network analysis. Examples can be
found in Feng and Fuhais study [179] and Jun et al.s study [180] that used patent-based
morphological mapping, and Yoon et al.s study [181] that developed the text mining
morphology analysis.
33
F. ANALOGY
As a research technique, analogies have been mainly applied in the social sciences
[3]. Nonetheless, it may improve the anticipatory insight, especially when quantitative
methods suffer from the absence of sufficient data but there exist analogous events in
history. The classic application can be found in Bruces study of The Railroad and the
Space Program-An Exploration in Historical Analogy in 1965 [182]. The study sought
to test the feasibility of using railroad development in a systematic way to forecast space
program development. The historical analogy method, however, tends to neglect the
political, social, and philosophical impact, thereby often providing unsatisfactory
forecasts. A recent application of Goodwin et al. [183], which conducted a comparative
analysis of four different forecasting methods using analogous time series data for a sales
forecast of a new product, also concluded that using an analogy led to higher errors than
the parameters estimated from small but actual data.
G. CAUSAL MODEL
A causal model considers the explicit cause-and-effect relationships that affect the
growth of technology systems [6]. Therefore, this technique relies on the assumption that
the relevant variables and their linkages are known and can be described in a structural
model. However, due to the lack of information, the use of causal modeling is limited to
forecasting adoption or diffusion of innovations where the related parameters can be
measured [50], [184][188].
34
2.2.2 QUANTITATIVE MODELS

A. DECISION ANALYSIS
Multi-criteria decision making (MCDM) methods have been widely applied to the
technology assessment to integrate qualitative values into quantitative factors of
technologies. The traditional applications of decision analysis may start from the check
lists [189], expected value model [190], and stochastic success evaluation model [191].
Recent developments of decision analysis model on technology assessment are facilitated
by more advanced multi-criteria decision theories. Ondrus and Pigneur investigated the
potential of near field communication (NFC) as an upcoming technology for mobile
payments [192]. Their study conducted comparative analysis on an expert panel and
showed that the Swiss industry was enthusiastic about adoption of this technology. Daim
et al. used an applied hierarchical decision model to identify the optimal design
characteristics for the U.S. Northwest off-shore wind turbine [193]. As an application in
the service industry, Tang and Tzeng applied a hierarchy fuzzy MCDM to examine
critical environmental factors relevant to Internet commerce in changes in the
international marketing environment from the perspective of business activity, socioeconomics, and information management [194].
35
B. ECONOMIC ANALYSIS (COST-BENEFIT ANALYSIS )

Economic analysis (or cost-benefit analysis), which is not confined to the technology
assessment method, is frequently used to translate estimates resulting from technological
forecasting into economic terms with the purpose of evaluating the impact of a
technological innovation. This implies that it should be possible to quantify the forecasts
in monetary terms, although this may be a challenge where qualitative aspects such as
social goals or environmental impacts are mainly concerned.
The application of economic analysis to technology assessment stems from the crude
forms of a return on investment or a discounted cash flow to a gain factor considering
technical feasibility or a project acceptability index considering past experience [195].
Moslehi and Kumar delineated the optimistic vision of the smart grid not only from its
technical promises but also from the expected benefits that significantly outweigh the
estimated costs [196]. Friedewald and Oliver analyzed the economic aspect of the
ubiquitous computing and claimed that the cost barrier of RFID transponders and system
integration should be overcome [197].
C. SYSTEM DYNAMICS
System dynamics is defined as a simplified representation of the structure and
dynamics of part of the real world [6]. As an attempt to apply dynamic modeling to
technology assessment, Forrester proposed the concept of system dynamics for complex
systems [198]. Therefore, system dynamics has been particularly popular for shaping the
dynamic ecosystem of the future technologies and their diffusion patterns by taking
feedback, i.e., non-linear behaviors among entities, into account.
36
Daim et al. applied system dynamics to the fuel cell industry and found that the
adoption rate would be increased as a consequence of government policies and
supply/demand relations [199]. Maier developed a new product diffusion model using
system dynamics to incorporate competition and to map the process of substitution
among successive product generations [200]. Suryani et al. constructed a system
dynamics model to forecast air passenger demand and to evaluate some policy scenarios
related with runway and passenger terminal capacity expansion to meet the future
demand [201].
D. EXTRAPOLATION
Extrapolation models employ mathematical and statistical techniques to extend time
series data into the future under the assumption that the past conditions and trends will
continue in the future more or less unchanged [6]. Since estimation is a data-based
forecast, it requires a sufficient amount of good data to be effective. The next section
provides a focused review on the frequently used extrapolation models that can deal with
multi-attribute assessment.
Figure 4 classifies various technology assessment methods in two dimensional plots.
While the qualitative approaches tend to focus more on eliciting multiple perspectives
from the knowledge sources (e.g., expert panels, history, etc.), the quantitative
approaches place more emphasis on drawing meaningful findings by analyzing the
numerical data. It is not surprising that technology focused approaches tend to be
quantitative, whereas society focused approaches are mostly qualitative.
37
Figure 4 Technology assessment methods
38
2.3 FOCUSED REVIEW ON MULTI-ATTRIBUTE EXTRAPOLATION METHODS

As technology becomes even more sophisticated, there are few technologies that
truly possess only a single technical capability. The rate of change also varies over time,
being affected by the maturity levels of component technologies. This structural
complexity makes todays forecasting even more challenging, which leads to the question:
how to combine growth patterns of each attribute to describe the multi-objective
technology systems? There are three different approaches to tackle this problem: intuitive
modeling, parametric frontier modeling, and non-parametric frontier modeling.
2.3.1 INTUITIVE MODELS

The intuitive models are used to combine multiple technology attributes into a single
technology measure, often where no physical basis exists to do so [38]. Therefore, these
models are often technology specific, non-unique and highly subjective.
A. SCORING MODEL
Martino suggested distinguishing overriding variables, tradable variables, and
optional variables to construct an appropriate scoring model [202]. His example of the
fighter jet scoring model2 is shown below:
2 2
1 +
(1)
Once the scoring model is obtained, it becomes possible to estimate the overall score
of the future technologies by extending the historical trend. However, while the scoring
The weights and the tradeoff coefficients were determined by an Air Force officers subjective judgment.
39
model provides a composite measure so that each technology system can be put on a
common basis, it is not capable of capturing the necessary information to simultaneously
evaluate each system attribute relative to the remaining attributes.
B. TECHNOLOGY DEVELOPMENT ENVELOPE
Technology development envelope (TDE) was originally developed by Gerdsri to
identify an optimum technology development path as a roadmapping method [203]. The
procedure consists of six steps:
Step 1: Technology forecasting to identify emerging alternatives.
Step 2: Technology characterization to establish evaluation criteria.
Step 3: Technology assessment on identified alternatives based on criteria.
Step 4: Hierarchical modeling to determine the relative importance of criteria.
Step 5: Technology assessment to determine the relative value of alternatives.
Step 6: Formation of TDE to establish an optimum development path.
Within this process, TDE constructs a hierarchical decision model (HDM) to
determine the relative importance of emerging technologies aligned with the
organizations objectives. Technologies having the highest value in each time period
represent the most preferred technology alternatives. In this sense, the path connecting
those technologies from one period to another becomes an optimal technology
development roadmap.
However, since the technology assessment process in TDE is predicated upon HDM,
multiple perspectives on technology attributes are to be averaged within the process of
obtaining a single ranking of technology alternatives for each period. That is,
40
combinatorial values derived from different levels of technology attributes are supposed
to be represented by a single weighting scheme aggregated from a panel of experts
opinions. This becomes a critical issue to identify the better technology when the
market segments exist as having particular customers with differing value propositions on
technology systems.
2.3.2 PARAMETRIC FRONTIER MODELS

The parametric frontier approach is characterized by being defined a priori with
several assumptions on random noise and efficiency distribution in an attempt to
approximate the ideal relationship between inputs and outputs. Since actual observations
are to be compared with generalized production possibilities, the measurement is defined
to be functionally relative.
A. PLANAR FRONTIER MODEL (HYPER-PLANE )
The planar frontier model (or hyper-plane method) was first introduced by
Alexander and Nelson [204]. They assumed that the movement of the tradeoff surface is
describable by a smooth and monotonic function. The function to be estimated by a
multiple regression is given as:
= (1 , , )
(2)
where is the introduction date of a system, and denotes the technical capabilities.
Specification of the functional form and determination of the coefficients of the
equation provide a measure of average technological trend over time. For example, Lim
41
et al. applied a planar frontier model to develop the wireless protocol forecasting model.
The resulting equation is shown as follows [205]:
= 1984.411 2.532 ()
(3)
+ 6.651 ()
A major advantage of this approach is a simple implementation based on multiple

regression analysis, which allows straightforward interpretation of the results. Also,
regression diagnostics can be used to select significant parameters in explaining the
technological progress. However, the linearity assumption underpinning this approach
can be a disadvantage at the same time. In particular, the planar frontier assumes a
constant rate of change without considering acceleration from increased engineering
efforts or deceleration as the system approaches the physical limit. Therefore, this
approach may result in an inaccurate forecast, particularly when the technological
systems experience architectural transition. Furthermore, this approach may not be
applicable to large systems, as there could be difficulties with the use of technical
parameters arising from the tradeoffs between different component systems [206].
Presupposition of treating an ordinal index as a cardinal measure, i.e., using release date
as a proxy for the level of technology, should also be justified to employ this approach
[207].
42
B. CORRECTED ORDINARY LEAST SQUARES

Aigner and Chu introduced a corrected ordinary least squares (COLS) method to
construct a deterministic and parametric production function by extending the ordinary
least squares method [208]. This process first estimates unknown parameters using the
ordinary least squares method:
min ( ( ; ))2
(4)
where ( , ) denotes actual k observations and is a production function. Then, the

model finds the smallest possible correction by introducing an additional coefficient 00
to ensure that all observations are placed below the production frontier with the
maximum error term [209]:
00 = max { ( ; )| = 1, , }
(5)
Once the production function is estimated, the efficiency measurement for each
observed production can be made by comparing them to the maximum (minimum)
possible output (input) for a given input (output) along with a desired directional distance
function.
Since COLS has its root in statistical principles, i.e., the maximum likelihood, the
frontier is constituted to represent the general pattern of actual observations without
taking noise into account [26]. That is, any variation in the dataset, including possible
noise, is considered to contain significant information about the efficiency. Therefore,
this method may not be appropriate when there is a need to identify the underlying
pattern of production possibilities without the impact of the random noise.
43
C. STOCHASTIC FRONTIER ANALYSIS

As another variation of the ordinary least squares method, stochastic frontier analysis
(SFA) was introduced by Aigner et al. in an attempt to construct the stochastic and
parametric production function [210]. That is, the SFA model includes both a stochastic
error term and a parametric inefficiency term defined by a frontier curve. The basic
model, simplified by Bogetoft [26], can be shown as follows:
= ( ; ) + ,
~(0, 2 ),
~+ (0, 2 ),
= 1, ,
(6)
where denotes the noise, the inefficiency, and + a half-normal distribution.

Compared to COLS, SFA distinguishes noise with inefficiency (see Fig. 5).
Therefore, efficiency scores from SFA tend to be higher than COLS or other
deterministic frontier models. This characteristic would be appropriate when the dataset
suffers from random variations, and therefore outliers need to be detected in the process
of frontier formation. On the other hand, this might restrict the perspective of identifying
various tradeoffs that can represent the distinct production possibilities.
Figure 5 Frontier models in input-output space (modified from [26])

44
D. ELLIPSOID FRONTIER MODEL

As a non-linear frontier model, Dodson proposed an ellipsoid frontier model to
quantify the technological advance in relation to the state of the art (SOA) surface [211],
[212]. This model attempts to fit the technology frontier into an ellipsoid functional form
from which tradeoffs among attributes can be explained (see Fig.6). Dodsons measure of
technological advance is defined as ( 2 1), where
Martino later extended Dodsons model to allow use of it in any order [213].
Martinos generalized ellipsoid model is given as follows:

( ) = 1
(7)
where n is the order of the ellipsoid, the value of the ith technical capability, and the
intercept of the ellipsoid on the ith axis.
Martino also suggested using the mean absolute deviation rather than the mean
squared deviation for the fitting procedure to reduce the effect of extreme values. This
allows the fitted frontier surface to be located closer to the median of the observations
instead of the mean.
Although this approach can provide a measure to investigate the SOA formation
process, the fundamental question remains to be resolved: why the technology tradeoff
surface should be following the ellipsoid form? In detail, the ellipsoid frontier model
presupposes that the tradeoff of one technical capability being relinquished for the
45
advancement of the others can be explained by adopting a predefined functional form

rather than by adapting to the data. Dodsons choice of an ellipsoid shape is analytically
sound for the representation of a strictly convex surface but may not always be
representative.
In addition, the ellipsoid function is limited to explaining the tradeoff between
outputs. This requires an assumption that the tradeoff surface is only applicable to the
technical capabilities consuming the same amount of inputs such as engineering
resources.
Figure 6 Two-dimensional illustration of ellipsoid frontier (modified from [212])
46
E. MULTI-DIMENSIONAL GROWTH MODEL

To overcome the limitation of traditional S-curves, being able to deal with only a
single attribute for any subject, Danner developed the multi-dimensional growth model
(MDGM) to generate an iso-time frontier by combining individual technology growth
curves [32]. This approach first decomposes technology systems into individual attribute
levels to obtain a proper curve for each subsystem (see Fig. 7.) The iso-time frontier, at
which the same levels of technology systems are to be located, is then formulated by
combining identified individual growth curves. The resulting model makes it possible to
either forecast technical capabilities at a certain point in time or estimate the time by
which desired levels of technical capabilities will be operational.
Similar to the planar frontier model, a major limitation to the utility of MDGM is the
requirement that all dimensions of technical capability integrated must be statistically
independent. This presupposes that the individual time spans required to advance each
attribute towards corresponding upper limits can be linearly combined to explain the
technology systems growth rate. However, the higher the complexity of technology
systems under evaluation is, the more individual growth rates are likely to be interrelated,
hence the generated iso-time frontier without consideration of concurrent advancement
would not provide an accurate picture of the feasible combinations of technical
capabilities.
Interdependence between technology attributes might also require the modification
of established upper limits. It has been frequently observed in highly complex systems
that individual upper limits become more challenging to achieve as they tend to restrict
47
one another [214], [215]. Therefore, the progress of the iso-time frontier should be guided
by adjusted upper limits with the consideration of the architectural complexity involved.
Figure 7 Parametric frontier by MDGM (modified from [32])
48
2.3.3 NON-PARAMETRIC FRONTIER MODELS

The non-parametric frontier approach forms the best practice frontier without
relying on arbitrary functional assumptions. Instead, it maximizes the flexibility to
capture various production possibilities observed from the actual data. Since this
approach doesnt construct an averaged target with which data points are to be compared,
the measurement is defined to be empirically relative.
A. DATA ENVELOPMENT ANALYSIS
The original data envelopment analysis (DEA) was proposed by Charnes et al. [216].
As the name of decision making units (DMUs) implies, the efficiency measure in DEA is
defined as the ratio of the weighted sum of outputs to the weighted sum of inputs using a
freely chosen weighting scheme for each DMU, and as such, the efficiency measure will
show them in the best possible light. The ratio form of the dual (multiplier) input-oriented
variable returns to scale DEA model can be presented as below:
r ur yr 0 w
max h0
i vi xi 0
u y w
1, j,
v x
r
rj
i ij
u r , vi , w is free
(8)
where 0 denotes the input-oriented efficiency of DMU being assessed, the weight
assigned to output r, the weight assigned to input i, the ith input variable of DMU
j, the rth output variable of DMU j, and w the returns to scale (RTS) parameter.
The above input-oriented multiplier model can be readily translated to the primal
(envelopment) model, which is shown below as a single-stage theoretical formulation:
49
min 0 sr si
i
r
y s y , r 1,..., s
x s x , i 1,..., m
1
,
s
,
s
,
rj
ro
ij
0 io
(9)
where 0 denotes the technical input-oriented efficiency, the loading factor attached to
DMU j, + and the slacks equal the reduced cost of and respectively. Note that if
the optimal value of 0 is less than 1, then 0 is inefficient in that the model (9) will
have identified another production possibility that secures at least the output vector but
using no more than the reduced input vector . Thus, is a measure of the radial
input efficiency of 0 in that it reflects the proportion to which all of its observed
inputs can be reduced pro rata, without detriment to its output levels [209].
DEA studies have often examined the changing performance of units over time
[217][221]. A shorthand notation for a DEA model can be defined as 0 (0 , 0 ) as the
efficiency of the DMU o in time period t with input and output characteristics (0 , 0 ),
being measured against the frontier of peers also in time period t. A peer compared
against units from the following period would then be +1

0 (0 , 0 ) .
If the value
of +1
0 (0 , 0 ) is less than 1.0, then the unit in period t is inefficient relative to units from
period t+1. If the value of +1

0 (0 , 0 ) is greater than 1.0, then the unit in period t
outperforms units from period t+1 in some manner and is efficient.

It would be expected that a particular units efficiency scores such as 0 (0 , 0 ) and
+1 +1
+1
0 (0 , 0 ) would vary over time, but separating an effect from improved operation of
a unit from different conditions affecting all units cannot be determined simply from the
50
efficiency scores. Fre et al. introduced a DEA-based Malmquist productivity index (MPI)
to measure the technical efficiency change (TEC) and the frontier shift (FS) over time as
an extension of the original concept introduced by Malmquist [217], [222]. The inputoriented MPI can be defined as
1
0 (0 , 0 )
0+1 (0+1 , 0+1 ) 0+1 (0 , 0 ) 2
= 0 0 = +1 +1 +1 [ +1 +1
]
0 (0 , 0 )
0 (0 , 0 ) 0 (0 , 0 )
(9)
where 0 denotes DEA efficiency score and 0 , 0 input and output levels at given point
in time t. Therefore, TEC indicates technical efficiency change between period t and t+1:
improves (<1), remains (=1), and declines (>1.) In a similar sense, FS measures the
amount of frontier shift: regress (>1), no shift (=1), and progress (<1).
To extend the time-series application of DEA into technological forecasting, Inman
developed a measure to quantify the rate of frontier expansion by which the arrival of the
following DMUs can be estimated [223]. Specifically, his method, technology
forecasting using data envelopment analysis (TFDEA), establishes the envelopment, i.e.,
SOA technology frontier, using the data points identified as relatively efficient from DEA
(see Fig. 8). Note that the frontier is a set of convex combinations formed by SOA
technologies, hence its not a curved surface but a piecewise linear combination. The
tradeoffs between technical capabilities can be considered as a radial improvement within
this frontier space. The TFDEA iterates the frontier formation process over time to track
the rate of frontier shift. This momentum of progress is then used to make a forecast for
the future technologies (DMUs.)
Unlike the iso-time frontier from MDGM, the frontier constructed by TFDEA
typically consists of multiple vintages of SOA technologies. This allows the model to
51
specify the individual timing, i.e., effective time, of any points on the frontier according
to the corresponding tradeoff surface. This enables TFDEA to identify the starting point
of each forecasting target from which their best forecast can be made. Lim et al.
examined how this approach could improve the forecasting accuracy compared to a
planar frontier model in which the constant baseline, i.e., the regression constant, is
assumed for all forecasts [205].
Figure 8 Non-parametric frontier by TFDEA
In spite of additional benchmarking information provided by DEA within TFDEA

process, traditional TFDEA doesnt utilize this information when it comes to a
consolidation of rate of changes (RoCs) captured from the past technologies. In fact, it
simply employs an average value to make a forecast for future technologies regardless of
their unique characteristics. That being said, RoCs captured from the surpassed
technologies are simply aggregated to represent the technological progress as a whole.
52
This might overlook the unique growth patterns captured from different tradeoff surfaces.
Consequently, it was shown at times in previous applications that forecasting based on a
single aggregated RoC did not consider the unique growth patterns of each technology
segment, which resulted in a conservative or aggressive forecast [224], [225]. This issue,
in particular, becomes more visible when the application area contains distinct progress
patterns identified from multiple technology segments. Therefore, it is necessary to
incorporate the notion of segmented RoC into the forecasting procedure so that each
forecasting target can be subject to the individualized RoC that best reflects the potential
growth rate of analogous technologies.
It has been occasionally observed from the past applications that TFDEA may suffer
from instances of infeasible super-efficiencies when variable returns to scale (VRS) was
assumed. In theory, this is also a problem for the input-oriented decreasing returns to
scale (DRS) model and output-oriented increasing returns to scale (IRS) model [27]. This
problem results in failure to make a forecast for the target technology since the model is
unable to measure the superiority of corresponding technology compared to current SOA
technologies. Note that the constant returns to scale (CRS) model is also susceptible to
this problem when zero data is included in any input variables [27]. However, this is rare
in actual applications since it indicates heterogeneous DMUs or technologies [209].
The problem of infeasibilities in the super-efficiency model can be attributed to the
inherent characteristics of a non-parametric frontier since this approach identifies the
production possibilities without spanning unobserved regions. Especially under VRS,
DEA constitutes the frontier purely based on observed DMUs and, therefore, tradeoffs in
uncharted regions remain unknown. This renders forecasting targets subject to those
53
unknown regions impossible to be projected from the current SOA frontier in a radial
manner.
B. STOCHASTIC (CHANCE-CONSTRAINED ) DATA ENVELOPMENT ANALYSIS
Land et al. proposed a data envelopment analysis model that can deal with stochastic
variability in inputs and outputs, which evolved from the earlier technique called chanceconstrained programming developed by Land et al. [226], [227]. The standard inputoriented model is presented below:
min ( + + )
. . ( 0 ) 1 (1 )0 + = 0,
= 1, ,
+ = 0 ,
= 1, ,
(10)
+ , , 0, , ,

where denotes radial input contraction factor, + and slack variables, E the
mathematical expectation, F the distribution function of the standard normal distribution,
the standard deviation of best practice output minus observed output, i.e., s.d. (
0 ), loading factor, the ith input variable of DMU j, the rth output variable
of DMU j, K threshold fraction allowing hyper-efficiency [228].

Formulation (10) minimizes the contraction factor , subject to two sets of
constraints.
54
First, the observed outputs must not exceed best practice outputs more often than
probability of K. For example, in the case of K = 0.01, only 1% or less of DMUs will do
better than the DMU being assessed. That is, K indicates the fraction of DMUs being
located above the frontier in Fig. 5. This constraint can be simplified as below [26]:
( 0 ) ,
= 1, ,
(11)
where P denotes the mathematical probability.

Second, there is a deterministic constraint for inputs stating that the benchmark
being compared must not use more inputs than the reduced inputs, 0 .
The efficiency score is determined by the above model for each DMU, where
denotes optimum chance-constrained DEA efficiency score, such that sub-efficient ( <
1 = 1 ), efficient ( = 1 ), and hyperefficient ( > 1) [226].
Similar to the difference between COLS and SFA, this approach is able to separate
noise from inefficiency included in the traditional DEA. However, this procedure should
be supported by statistical requirements such as input and output variables that are known
to be normally distributed. Information about joint probability distributions of the random
variables should also be justified to make the resulting efficiency reliable [229].
55
2.4 SUMMARY OF CRITICAL REVIEW

Table 12 summarizes technology assessment models that can take tradeoffs from
multiple technology attributes into account. Aside from intuitive models, the primary
difference between parametric frontier models and non-parametric frontier models is how
much weight is put on uncertainty versus flexibility. The former approach places a higher
importance on the uncertainty by being defined a priori except for a finite set of
unknown parameters that are estimated from the data [26]. Therefore, a parametric
approach tends to be robust to extreme points by filtering them with a predefined
general pattern. The latter approach, in contrast, purely adapts the SOA frontier to data
without being shaped a priori, which renders the resulting frontier to be a piecewise
linear combination rather than a curved surface. This property makes it possible for the
non-parametric approach to take full advantage of the distribution of the dataset without
relying on statistical corrections.
Among the investigated multi-attribute evaluation models, the non-parametric frontier
approach shows the favorable features that make it possible to take account of
technological segmentation by classifying generated frontier facets. In particular, the
deterministic model may be preferred to the stochastic model in the practical sense that
the flexible nature enables a versatile application by not being restricted to statistical
conditions. It is worth noting here that, by definition, the shape of the SOA frontier
indicates system tradeoffs among technical capabilities. That being said, while the
parametric approach presupposes that the shape of SOA surface, namely system
tradeoffs, would not change over time, the nonparametric approach attempts to elaborate
its changes in each time period with the given data. One can argue that a lack of
56
stochastic estimations results in a mean structure of technological innovation degraded by

random noises, i.e., extreme technologies. However, deterministic frontier analysis,
which takes advantage of superior technologies rather than being averaged by mediocre
or inferior technologies, has shown its usefulness in a wide range of technological
forecasting studies [225], [230][232]. In addition, it is expected that robustness of the
model can be partly compensated by the frontier segmentation approach by treating
distinguished groups of technologies separately. Meanwhile, the infeasibility problem
must be dealt with to guarantee the coherent and complete forecasting process.
The subsequent sections will elaborate how the new approach is formulated with
respect to identified research questions.
Table 12 Summary of multi-attribute technology assessment models

Frontier analysis method
Intuitive method
Pros
High flexibility in
model building
Parametric frontier model

Capable of detecting
outliers
Relies on
Difficult to estimate
subjective opinions
required parameters
No consideration
Sensitive to
of multiple
multicollinearity
tradeoffs
Model Scoring model
Planar model
(Hyper-plane)
TDE (Technology
o COLS (Corrected
Development
Ordinary Least Squares)
Envelope)
o SFA (Stochastic Frontier
Analysis)
Ellipsoid frontier
MDGM (MultiDimensional Growth
Model)
Cons
Non-parametric frontier model

Free from parametric
requirements
Capable of identifying
frontier segments
Sensitive to noises/extremes
Susceptible to infeasibility
TFDEA (Technology
Forecasting using Data
Envelopment Analysis)
o SDEA (Stochastic Data
Envelopment Analysis)
: Technological forecasting models
57
o : Econometric models
III. MODEL DEVELOPMENT
3.1 CONCEPTUAL FRAMEWORK

3.1.1
SEGMENTED RATE OF CHANGE
Since TFDEA has at its core the widely used technique of DEA, TFDEA inherits the
ability to provide many of the same rich results. One of the key results yielded by DEA is
the identification of targets and efficient peers [233]. Specifically, DEA constitutes the
frontier of a production possibility set (PPS) based on best practice DMUs. Within this
framework, relative efficiency is determined by comparing the performance of each unit
against that of a (virtual) target formed by efficient peers. A practical interpretation is that
efficient peers can serve as role models which inefficient DMUs can emulate so that they
may improve their performances. In other words, those benchmarks have a mix of inputoutput levels similar to that of DMUs being compared, which indicates that they are
likely to operate in analogous environments and/or to favor similar operating practices
[209].
The implementation of TFDEA relies on a series of benchmarking processes over
time [223]. This is depicted in Fig. 9, assuming an output-oriented DEA model under
variable returns to scale (VRS) [234]. The frontier year, T, is the point in time at which
the analysis is conducted. Products G, H, and I are identified to be the most competitive
at time T and therefore define the SOA frontier at time T. Products A~F, in contrast, were
themselves SOA when they were first released but were superseded by subsequent
products and hence are located below the frontier. Products J and K are future products,
58
i.e., sets of specifications used as forecasting targets that are placed beyond the current
SOA frontier.
Figure 9 Evolution of the SOA frontier
The TFDEA process can be understood as three procedural stages. First, it iterates
the DEA process to obtain efficiency scores of products both at the time of release and at
the frontier year. Second, it estimates an RoC that represents how fast products have been
replaced by the next generation products. In other words, the RoC indicates a potential
growth rate of the SOA frontier in the future. Finally, the model makes a forecast of
future products based on the average RoC.
The original TFDEA process simply aggregates RoCs from the past products and
uses the average RoC to make a projection without taking technological segmentation
into account. However, as previously discussed, DEA provides pragmatic information
regarding benchmarks, which enables an identification of distinct product clusters. This
59
information can be obtained either by reference sets in the envelopment model or by

weighting schemes in the multiplier model.
For example, two different product clusters are identified in Fig. 9. The first cluster
can be characterized by an optimized weighting scheme, that is, a facet connecting
products G and H. This can be interpreted that inefficient products from this cluster,
namely B and E, may have similar mixes of input-output levels such that a corresponding
weighting scheme will show them in the best possible light. This can also be recognized
as a reference set in the envelopment model since their performances are compared
against virtual targets constituted by efficient peers, namely products G and H.
In the same manner, the second cluster can represent another weighting scheme, that
is, a facet connecting products H and I. Even though the underlying products, A, C, D,
and F, have less efficient input and output amounts, they must have similar ratios of the
input-output levels that require the common weighting scheme to optimize their
operations. The envelopment model, likewise, will constitute virtual products interpolated
by products H and I for these inefficient products.
The idea of segmented RoC arises when there is a need to draw a distinction
between each cluster; hence, the growth potential should be explained by local RoCs
rather than a universal RoC. In our example, one may notice that cluster 2 has observed
faster RoCs than cluster 1. Specifically, products B and E did not show a large
performance gap compared to the current SOA frontier even though the old product B, in
particular, had stayed on the SOA frontier for a long time and only recently became
superseded. This implies that the technological progress within cluster 1 has been neither
fast nor frequent. In contrast, products pertinent to cluster 2 have shown successive
60
replacements with substantial performance advancement over time. This may imply that
more engineering effort has been invested in cluster 2-type products, which results in
more frequent introductions of advanced products over time.
Once distinguishing clusters are identified with varying RoCs, it is readily possible
to make a forecast using those local RoCs. For example, the estimated arrival of future
product J can be determined by measuring how far it is from the current SOA frontier and
then extracting the root of that distance using local RoCs from cluster 1 given the fact
that it is projected to the frontier facet of cluster 1. In the same manner, the arrival of
future product K can be estimated using local RoCs from cluster 2. One may expect that
if both products were achievable with the same amount of engineering advances, the
arrival of product K might be earlier than that of product J since faster progress is
expected from cluster 2-type products. In other words, requiring the same amount of time
to reach the technological level of product J would entail significant development risk.
Figure 10 depicts how the local RoC and individualized RoC can be obtained.
Product L had been located on the SOA frontier in the past but later became obsolete by
the current SOA frontier formed by new competitive products: M, N, and O. As
aforementioned, the fact that product L is compared to its virtual target, i.e., L,
constituted by its peers: M, N, and O indicates that product L may have a similar mix of
input-output levels with those peers although the absolute level of attributes may vary,
which can classify them as homogeneous products. Hence, the technological
advancement, namely the performance gap between L and L during a given time period,
can be represented by the peers as a form of local RoC. Consequently, each local RoC
61
indicates a growth potential for adjacent frontier facets based on the technological
advancement observed from the related past products.
Figure 10 Illustration of segmented rate of change
Once the local RoC of current SOA products are obtained, it is straightforward to
compute the individualized RoC for the new product concepts. Suppose product
developers came up with a product concept Q. Note that by definition, a better product
would be located beyond the current SOA frontier as superseded products are located
below, namely enveloped by, the current SOA frontier. It is seen that the virtual target of
Q, i.e., Q, is subject to the frontier facet constituted by current SOA products N, O, and
P. Thus, the individualized RoC of Q can be obtained by combining local RoCs with the
reference information: how close Q is from N, O, and P respectively. It should be noted
here that technological advancement observed from the product L may have affected the
individualized RoC of Q as SOA product N and O are involved in both sides of the facets
62
by having intermediate technological characteristics. This information can give insight to

the product developers not only about who the major players in a target market are but
also about how competitive the proposed design concept is. In other words, this can
provide a diagnostic if the proposed design concept is aggressive or conservative in terms
of scheduled delivery to the market considering the current rate of technological progress
expected in a target segment. One can also utilize this information to estimate the arrival
of a competitors design targets as a post product launch strategy of their own.
As we increase the dimension, i.e., the number of structural characteristics and/or
functional features of the product being considered, the problem can be better handled by
algebraic formulation than graphical analysis. The underlying formulation is introduced
in section 3.2.
63
3.1.2
INFEASIBLE FORECASTING TARGET3
It has been observed in past super-efficiency DEA applications that the infeasibility
problem occurs when variable returns to scale (VRS) was assumed [27]. In theory, this is
also a problem for an input-oriented decreasing returns to scale (DRS) model and outputoriented increasing returns to scale (IRS) model. This problem results in failure to make a
forecast in TFDEA since the model is unable to measure the super-efficiency,
0 (0> , 0> ) , i.e., superiority of specified technical capabilities from the current
frontier. Note that the constant returns to scale (CRS) DEA model is also susceptible to
this problem when a zero value is included in an input variable [209]. However, this is
rare in TFDEA applications since it indicates heterogeneous technologies.
Figure 11 depicts possible occasions in which infeasible super-efficiency occurs
under VRS. It is readily seen that target E and F are subject to infeasibility from the
current frontier in the input-oriented (IO) model and output-oriented (OO) model
respectively, whereas target D is infeasible in both orientations. Therefore, the arrivals of
those targets in corresponding orientation from the current frontier are unable to be
computed using the traditional TFDEA model.
This section is adapted from a paper accepted in International Transactions in Operational Research [313]
64
Figure 11 Regions of infeasible super-efficiency under VRS DEA model
Alternate measures have been developed to deal with the infeasible super-efficiency
problem. Lovell and Rouse suggested employing a user-defined scaling factor to make
the VRS super-efficiency model feasible [235]. Cook et al. developed a radial measure
of super-efficiency with respect to both input and output direction so that one can derive
the minimum change needed to project a DMU to a non-extreme position, and the other
can reflect the radial distance of that shifted DMU from the frontier formed by the
remaining DMUs [236]. In a similar vein, Lee et al. proposed a slack based superefficiency model that can consider both input savings and output surplus in cases where
infeasibility occurs [237]. Lee and Zhu further extended this model to deal with the
infeasibility problem caused by zero input values [238].
In this study, Cook et al.s alternate super-efficiency measure is adopted for two
main reasons: a) it returns bi-oriented L1 distances for infeasible targets and hence it
65
secures the existing RoC calculation; b) it returns the same radial distance as the standard
super-efficiency measure [239] when the target is feasible.
Cook et al. [236] defined the term extremity to indicate a minimum radial
movement in either direction needed for a DMU to reach a non-extreme position. For
example, in the input-oriented model, target E will have an extremity of 0.75 (=15/20), to
bring it down to the closest feasible point, i.e., E (20, 15). The radial input augmentation
is then applied, i.e., 1.25 (=25/20), from this shifted point E to the peer unit C.
Consequently, the input-oriented super-efficiency of target E from the current frontier can
be defined as 2.583 (=1.25+1/0.75). In a similar sense, the output-oriented superefficiency of target F from the current frontier is 4.5 (=2+1/0.4), and target D has 6.333
(=5+1/0.75) and 12 (=2+1/0.1) from the input-oriented model and output-oriented model
respectively.
Once the super-efficiency score of each forecasting target from the current frontier is
obtained, RoCs can be applied to the estimation of those target technologies arrivals.
Note that targets that contain extremities in their super-efficiency scores require RoCs
from both orientations. That is, the time period for the extremity can be estimated by the
RoC from the opposite orientation model. In the case of target F, for example, the outputoriented TFDEA model should be able to compute how long it will take to reduce the
input from 10 to 5 based on the RoC that one would obtain from the input-oriented model
as well as to augment the output from 2 to 5 based on the output-oriented RoC. This
indicates that performing TFDEA in both orientations is required to deal with the
infeasible forecasting targets.
66
3.2 FORMULATION
I now turn to the TFDEA formulation incorporating the proposed approach under
VRS. The entire process can be divided into four separate stages.
The first stage iterates efficiency measurement in a time series manner so that the
evolution of the SOA frontier can be monitored. As mentioned above, it is required to
obtain RoCs in both orientations to make a forecast for targets containing extremities.
{,}
Therefore, this stage computes radial efficiencies:

{,}
model shown by (12)-(18) and
from the output-oriented
from the input-orientation model shown by (19)-
(25). Note that the model can be formulated as a single large LP, it may also be
formulated and solved as a series of equivalent, smaller LP models for the time of release
(R) and models for the current frontier time (C) depending on the implementation
algorithm. Specifically, represents the th input and represents the th output for
each technology j = 1,, n, and j = k identifies the technology to be evaluated.
The objective functions for each orientation, (12) and (19), incorporate minimizing
effective dates as well to ensure reproducible outcomes from possible alternate optimal
solutions by distinguishing between Pareto-efficient technologies4 [240], [241].
Constraints (15), (16), (22), and (23) limit the reference sets so that two types of
efficiencies, one at the time of release (R) and the other at the current frontier time (C) in
4
Unlike the iso-time frontier from parametric frontier models, the technology frontier constructed by
TFDEA typically consists of multiple vintages of SOA technologies. This allows the model to specify the
individual timing, i.e., effective date, of any points on the frontier according to the corresponding tradeoff
surface. Therefore, the issue of alternate optimal solutions occurs either due to weakly efficient technology
or to an efficient but not an extreme technology, namely F type or E type in Charnes et al.s classification
[314]. Both cases can be dealt with by introducing the secondary objective to choose the reference
technology presenting either in the farthest time horizon, i.e., maximum sum of effective date, or in the
closest time horizon, i.e., minimum sum of effective date. Note that depending on the application area,
slack maximization may be preferred to prevent weakly efficient technologies from setting the effective
date. Further discussion can be found in [240], [315].
67
which the forecast is conducted, are obtained. That is, and each measure the
amount by which technology k is surpassed by the technologies available at the time of
release since constraint (15) and (22) allow the reference set of technology k to only
include technologies that had been released up to . Similarly, and can be
interpreted as how superior technology is compared to the current SOA frontier by
constraint (16) and (23).
Note that the current time is defined as a fixed time T, which can be either the
most recent time in the dataset or a certain point in time as a forecasting origin when the
time series hold-out sampling is performed. The variable

describes how much of
technology is used in setting a target of performance for technology .

Note that in the case of the VRS model, constraint (17) and (24) would allow
replacing the denominator in the second term of (12) and (19) with a 1, making the
objective functions linear. Here, it is imperative that the value of a non-Archimedean
infinitesimal, , not be implemented as a finite approximation to avoid inaccuracies and
erroneous results [241]. Instead, the actual implementation is to use a two-stage
preemptive linear programming to first identify the radial efficiency and then to either
maximize (or minimize) effective dates or to maximize the slacks according to the need.
68
=1

[ (
)]
=1
(12)
=1
. .
,
, , {, }
(13)
, , {, }
(14)
. .
= 0,
(, )| >
(15)
. .
= 0,
(, )| >
(16)
, {, }
(17)
, , {, }
(18)
=1
. .
,
=1
. .
= 1,
=1
. .
0,
=1

[ + (
)]
=1
(19)
=1
. .
,
, , {, }
(20)
, , {, }
(21)
. .
= 0,
(, )| >
(22)
. .
= 0,
(, )| >
(23)
, {, }
(24)
, , {, }
(25)
=1
. .
,
=1
. .
= 1,
=1
. .
0,
69
The non-VRS models such as non-decreasing returns to scale (NDRS), nonincreasing returns to scale (NIRS), or CRS would render the objective function, (12) and
(19), to no longer be linear as the denominator is not constrained to be equal to 1. For
computational purposes, the same general secondary goal of minimizing effective years
can also be approximated by subtracting the sum of reference vectors, =1

and
=1
respectively, in the objective function as seen in (26) and (27). While this
substitution is not technically a numerical approximation, it is generally consistent with

minimizing effective year and has the advantage of remaining linear [240].
[
=1
[
=1
(
=1
)]

(26)
=1
(
=1
)]

(27)
=1
The second stage, shown by (28)-(31), calculates the RoCs from each orientation,
( ), by taking all technologies that were efficient at the time of release, = 1
( = 1), but later superseded by new technologies at the current frontier time, > 1
( < 1). Having calculated RoCs of past technologies in (28) and (30), the idea of
segmented RoC can then be implemented by taking the weighted average of RoC for
each technology on the current SOA frontier. This leads to the calculation of local RoCs
in (29) and (31), where ( ) represents the local RoC driven by technology j at current
time T. Note that technology j has an efficiency score of 1 at the current frontier; in other
70
words, it is one of the SOAs that constitutes the frontier onto which future technologies
are to be projected. The numerator of (29) and (31) indicates the weighted sum of RoCs
from past technologies that have set technology j as a (or one of) benchmark(s). The
denominator indicates the accumulated contribution of technology j to the evolution of
the SOA frontier. Consequently, ( ) represents the local RoC that only counts RoCs
in which SOA technology j has been used as a benchmark5.
=1 ,
=1 ,
= ( )
| = 1, > 1
(28)
=1 ,

=1, >0 ,
| = 1
(29)
=1
=1
1
= ( )
| = 1, < 1
(30)
=1

=1, >0
| = 1
(31)
A special case [316] was observed in which the RoC for one product, i.e., ( ), exceeded 10.0 due to
the short time period between the effective date and the actual release date. An RoC of 10.0 indicates that
the technology is advancing at a rate of 1000% per time period (day, month, year). This greatly exceeds
even the rapidly moving portions of the computer industry such as microprocessors and therefore is
considered an unreliable estimate of RoC. The current implementation assumes the maximum acceptable
RoC to be 10.0 and hence drops those having RoCs greater than this limit from the local RoC calculation.
Exploring this further is a topic for future work.
5
71
The third stage solves super-efficiency models for the forecasting targets of future
products. Since the purpose of this stage is to measure the super-efficiency of each
forecasting target from the current frontier, the reference set is confined to the current
SOA technologies by (35), (36), (41), and (42). M is a user-defined large positive number
to give a preemptive priority to the identification of a minimum radial shift of inputs (or
outputs) to render the model feasible. In the output-oriented model, shown by (32)-(37),
radial output reduction and extremity are obtained as 1 and 1 + , respectively.
Likewise, in the input-oriented model, shown by (38)-(43), radial input augmentation and
extremity are defined as 1 + and 1 , respectively.
( + )
(32)
| >
. .
(1 + ) ,
(33)
(34)
=1
. .
(1 ) ,
=1
. .
= 1,
| = 1
(35)
. .
= 0,
(, )| >
(36)
. . ,
0,
(37)
=1
72
( + )
(38)
| >
. .
(1 + ) ,
(39)
(40)
=1
. .
(1 ) ,
=1
. .
= 1,
| = 1
(41)
. .
= 0,
(, )| >
(42)
. . ,
0,
(43)
=1
The last stage makes a forecast of the arrival of future technologies. The
individualized RoC for each forecasting target k can be computed by combining the local
RoCs of SOA technology j that constitutes the frontier facet onto which technology k is
being projected. That is, in the case of the output-oriented (input-oriented) model, the
estimated elapsed time for the extremity, if any, is computed using the individualized
RoC from the input-oriented (output-oriented) model. For target F in Fig. 11, for example,
the time span for the extremity, namely distance from F to F, 2 (=10/5), from the outputoriented model is estimated by individualized RoC combined with input-oriented local
RoCs of its peers: A and B. In addition, the time span for radial output reduction, namely,
distance from F to A, 0.4 (=2/5), is estimated by individualized RoC from its outputoriented peer, A. Consequently, the forecasted arrival time of F is obtained by the sum of
those estimated elapsed times and the reference time of the current frontier. Likewise, the
forecasted arrival time of D under the output-oriented (input-oriented) model is obtained
73
by the sum of the estimated time span for the distance from itself to a radially shifted
point, i.e., D (D), using input-oriented (output-oriented) local RoC of C (A) and
estimated time span for the distance from the shifted point to its peer, A (C), using
corresponding output-oriented (input-oriented) local RoC.
(
=
)
+
1
1
=1 ,
=1 ,
=1 ,
=1 ,
=1 ,
=1 ,
(1 + )
(1 + )
1
1
=1 ,
=1 ,
| >
(44)
| >
(45)
=1 ,
=1 ,

=1 ,

=1 ,
Equations (32) and (33) yield the same solution, when the original TFDEA model is
feasible, and provide results with consistent interpretation when the original TFDEA
model is infeasible. The following proofs of theorem (32) and (33) guarantee that the
proposed TFDEA extension returns a feasible and a finite solution.
Theorem 1 (44) always yields a finite forecast.

Proof If the target k is feasible from the original super-efficiency model in [239], this
means that = 0 by theorem 3 in [236]. This reduces (44) to the traditional outputoriented TFDEA model in [223]. If the target k is infeasible from the original superefficiency model in [239], this means that > 0 by theorem 3 in [236]. In both cases,
74
1/(1 ) > 0 by theorem 4 in [236]. Therefore,
is always feasible. This
completes the proof.
Theorem 2 (45) always yields a finite forecast.

Proof If the target k is feasible from the original super-efficiency model in [239], this
means that = 0 by theorem 1 in [236]. This reduces (45) to the traditional inputoriented TFDEA model in [223]. If the target k is infeasible from the original superefficiency model in [239], this means that > 0 by theorem 1 in [236]. In both cases,
1 + > 0 by theorem 2 in [236]. Therefore _ is always feasible. This
completes the proof.
75
3.3 ILLUSTRATIVE EXAMPLE

To illustrate the use of the proposed approach, this section revisits the TFDEA
application to the LCD panel industry [232]. The dataset includes 389 models from 20
manufacturers that have been introduced to the market from 1997 to 2012. For the sake of
showing the comparative results, a hold-out sample test is conducted by dividing the
dataset into two parts, a training set (1997~2007) and testing set (2008~2012), to validate
the forecasting model using two input parameters: bezel size and weight and three output
parameters: screen size, resolution, and contrast ratio.
Table 13 summarizes the forecast results of those products which were infeasible in
the traditional input-oriented VRS TFDEA model. That is, infeasibilities rendered the
traditional TFDEA model unable to make a forecast for the listed 31 targets out of 95
future technologies from the forecasting origin of 2007. This is shown by the value of
extremities in the fourth column indicating the necessity of output reductions to be able to
compute the radial input-oriented super-efficiency scores shown in the fifth column.
As previously discussed, an infeasible target in the input-oriented TFDEA may occur
when the output levels of the target technology are unprecedented in the past. In other
words, there is no way a radial expansion in the inputs of the target technology, keeping
outputs fixed, can cross the current SOA frontier. Therefore, one can interpret this large
portion of infeasible targets as an indicator of active technological advancements,
particularly in terms of output parameters that post-2007 LCD panels achieved. This is
actually in line with the dynamics of the flat panel industry in which the manufacturers
kept investing their engineering efforts toward a higher performance standard such as
ultra-high definition (UHD) with lifelike contrast ratio.
76
Table 13 Forecast results for infeasible targets

DMU
(k)
77
165
166
212
218
248
265
266
268
271
273
274
276
277
282
283
284
293
298
305
321
325
330
331
336
341
344
349
350
353
355
357
LCD
panel name
T520HW01 V0
V562D1-L04
V460H1-LH7
LTA550HF02
V400H1-L08
LK460D3LA63
LTA460HM03
LTA460HQ05
P460HW03 V0
V460H1-L11
V460H1-LH9
LK601R3LA19
LTA550HJ06
P546HW02 V0
P645HW03 V0
P650HVN02.2
R300M1-L01
LTF320HF01
V315H3-L01
T400HW03 V3
V420H2-LE1
V370H4-L01
V400H1-L10
LTA460HN01-W
V500HK1-LS5
BR650D15
LK600D3LB14
LK695D3LA08
LTI700HA01
T706DB01 V0
V546H1-LS1
Actual
year of release
( )
Extremity
(1 )
Radial distance
(1 + )
Individualized
output-oriented RoC
=1 ,

(
)
=1 ,
=1 ,

(
)
=1 ,
2008
2008
2009
2009
2009
2010
2010
2010
2010
2010
2010
2010
2010
2010
2010
2010
2010
2010
2010
2010
2010
2010
2010
2011
2011
2011
2011
2011
2011
2011
2011
0.9351
0.9632
0.8183
0.9151
0.8508
0.9778
0.9778
0.9778
0.9778
0.8183
0.8653
0.6737
0.8159
0.9151
0.9674
0.9674
0.7910
0.9590
0.9590
0.9018
0.8893
0.9211
0.9018
0.9778
0.9489
0.8571
0.6012
0.8268
0.9289
0.9060
0.8159
1.6236
1.0035
1.3059
2.0317
1.2724
2.2312
1.7685
2.0760
2.0760
1.2929
1.4340
1.4558
1.7228
2.3603
1.8781
1.7607
2.7735
1.1092
1.3354
1.2863
1.9215
1.3982
1.4399
1.9895
2.5756
1.3431
1.5420
1.8865
1.6223
2.8878
2.0305
1.1778
1.1876
1.1932
1.1708
1.1984
1.1752
1.1826
1.1824
1.1824
1.1932
1.1900
1.1609
1.1738
1.1415
1.1413
1.1603
1.1156
1.2043
1.2039
1.1955
1.1832
1.1982
1.1952
1.1825
1.1354
1.1650
1.1686
1.1446
1.1510
1.1286
1.1688
1.1972
1.6580
1.1879
1.1986
1.1849
1.1941
1.1941
1.1941
1.1941
1.1879
1.1898
1.4553
1.1940
1.1986
1.2088
1.2088
1.6838
1.1817
1.1817
1.1866
1.1877
1.1849
1.1866
1.1941
1.1962
1.2028
1.1866
1.2050
1.2110
1.2345
1.1940
Individualized
input-oriented RoC
Effective date
=1 ,

(
)
=1 ,
Forecasted
time of release
_
(
)
2006.72
2007.00
2006.86
2006.70
2006.90
2006.77
2006.77
2006.77
2006.77
2006.86
2006.83
2007.00
2006.77
2006.70
2006.56
2006.56
2007.00
2006.95
2006.95
2006.88
2006.86
2006.90
2006.88
2006.77
2006.74
2006.64
2006.88
2006.61
2006.52
2006.56
2006.77
2009.82
2007.16
2009.53
2011.16
2009.21
2011.43
2010.11
2011.02
2011.02
2009.48
2009.73
2010.27
2011.09
2012.11
2010.13
2009.75
2009.52
2007.79
2008.91
2008.92
2011.35
2009.33
2009.58
2010.78
2012.43
2009.24
2012.67
2011.42
2009.56
2012.41
2012.06
The sixth and seventh columns both show the individualized output-oriented RoC
and input-oriented RoC. The time span required for output reduction, i.e., extremity, was
therefore obtained using corresponding individualized output-oriented RoC. Likewise,
the time span required for input augmentation, i.e., radial distance, was computed using
the corresponding individualized input-oriented RoC.
The last column shows the forecasted year of release considering the superiority of
each target technology compared to the 2007 SOA frontier. That is, the forecasted year of
release was obtained by the sum of the optimal starting point of the forecast, i.e.,
effective date shown in the eighth column, and the estimated elapsed times for extremity
and radial distance.
Figure 12 Forecast deviation distributions
78
The accuracy of the proposed model can be readily shown by comparing those
forecasted years with actual years of release. The deviation statistics contain this
information. As seen from Fig. 12, forecast deviation distribution of 31 infeasible targets
has a mean of -0.26 years with 0.41 in a 95% confidence interval (CI). This is more
accurate than a forecast of 64 feasible targets, i.e., mean deviation of +1.19 years (0.53),
which could improve the overall forecasting performance of a mean deviation of +0.72
years (0.40). Note that the proposed model yielded the forecast results equivalent to that
of a traditional TFDEA model for feasible targets. Consequently, it is shown that the
proposed model could make a reasonable forecast for formerly infeasible targets as well
as a consistent forecast for feasible targets.
79
IV. VALIDATING THROUGH CASE STUDIES
In this chapter, the proposed approach is applied to actual case studies. This is
organized in two sections. The first section focuses on ex post analysis, which revisits the
past applications to show how this approach can improve the forecasting accuracy using a
hold-out sample technique. In particular, the case study of the commercial airplane
industry is described in detail to fully explain the use of the proposed approach. To
further validate the utility of the proposed approach, six past datasets are revisited and the
comparative results are provided in comparison to the traditional approach.
The second section focuses on ex ante analysis, which addresses how the proposed
approach can be used to solve the actual forecasting problems in the supercomputer
industry. Specifically, the case study aims to investigate technological progress of
supercomputer development to identify the innovative potential of three leading
technology paths toward Exascale development: hybrid system, multicore system and
manycore system.
80
4.1 EX POST ANALYSIS: REVISIT EARLIER STUDIES

4.1.1 RISK ANALYSIS: COMMERCIAL AIRPLANE DEVELOPMENT
A. RESEARCH FRAMEWORK
To illustrate the use of the proposed method, this section assumes a scenario in
which commercial airplane developers are examining four design concepts in 2007. They
have collected data, including 24 aircrafts introduced to the market in the last 40 years,
and attempt to identify which market segment proposed design concepts are subject to
and when the ideal delivery to the market would be as competitive products considering
the rate of technological advancement observed until 2007.
Note that the performance characteristics used in the earlier study by Lamb, Daim,
and Anderson were adopted [225]. In the original study [242], they attempted to develop
technology assessment models based on a multiple-regression analysis. However, the
resulting model was confined to only two predictors due to the insufficient statistical
significance, which resulted in a high unexplained variability. This study revisits and
updates the dataset not only to incorporate the latest information but also to investigate
the industry dynamics with a consideration of different SOA trends as suggested in the
previous study (see Table 14).
81
Table 14 Commercial airplane dataset

Airplane
EIS
(year)
DC8-55
DC8-62
747-100
747-200
DC10-30
DC10-40
L1011- 500
747-300
767-200ER
767-300ER
747-400
MD-11
A330-300
A340-200
A340-300
MD-11ER
777-200ER
777-300
A330-200
A340-600
A340-500
777-300ER
777-200LR
A380-800
1965
1966
1969
1971
1972
1973
1979
1983
1984
1988
1989
1990
1993
1993
1993
1996
1997
1998
1998
2002
2003
2004
2006
2007
Travel
range
(1,000 km)
9.205
9.620
9.800
12.700
10.010
9.265
10.200
12.400
12.200
11.065
13.450
12.270
10.500
15.000
13.700
13.408
14.305
11.120
12.500
14.600
16.700
14.685
17.370
15.200
Passenger
capacity
(3rd class)
132
159
366
366
250
250
234
412
181
218
416
293
295
261
295
293
301
365
253
380
313
365
301
525
PFE
(passengers*km/L)
13.721
16.646
19.559
23.339
18.199
16.844
19.834
25.652
24.327
26.575
25.803
24.595
31.877
25.252
27.335
24.939
25.155
23.713
22.735
28.323
24.334
29.568
28.841
24.664
Cruising
speed
(km/h)
870
870
893
893
870
870
892
902
849
849
902
870
870
870
870
870
892
892
870
881
881
892
892
902
Maximum
speed
(km/h)
933
965
945
945
934
934
955
945
913
913
977
934
913
913
913
934
945
945
913
913
913
945
945
945
EIS: entry into service, PFE: Passenger fuel efficiency derived from passenger capacity, maximum travel range at
full payload, and fuel capacity
Table 15 Local RoC of SOA airplanes at the frontier year of 2007

SOA Airplane
Local RoC
Dominated airplanes
747-300
1.000949
DC8-55, 747-100/200, L1011-500
747-400
1.001404
DC8-55/62, 747-100/200, L1011-500, A340-200
A330-300
1.002188
767-300ER, A340-300
777-300ER
1.002561
767-300ER, A340-200/300/600
777-200LR
1.004606
A340-200/500
A380-800
1.003989
A340-500/600
82
B. ANALYSIS OF THE CURRENT STATE -OF-THE-ART

The commercial aircraft industry has important niches with segmented levels of
competition from regional jets to jumbo jets. Following the scenario, Table 15 records the
local RoCs of six SOA airplanes from the vantage point of 2007. The third column lists
dominated airplanes, i.e., past airplanes that have appointed the airplane in the first
column as a benchmark. As previously discussed, one can notice that airplanes are
grouped together with similarities in their specifications, which characterizes distinct
segments in 2007. While the frontier is five dimensional in this application, airplanes in
the first column are equivalent to products B, C, D, and E from Fig. 10, and column
threes airplanes are obsolete airplanes such as A.
The Boeing 747 series, as its nickname jumbo jet suggests, has been recognized as
the most successful wide-body commercial aircraft [243]. In particular, despite their large
bodies, the advanced aerodynamic design still allowed the 747-300 and 747-400 to reach
a cruising speed of up to 902 km/h [244]. These characteristics can be identified from the
dominated airplanes that include not only 747 predecessors (747-100 and 747-200) but
also Douglas DC8 series and Lockheed L-1011 that were also known as fast-cruising
airplanes. However, gradual technology advancement is observed from the relatively
slow local RoC of the 747 aircrafts, which is consistent with the fact that they had been a
dominant design for a long time until Airbus created a strong market rival [243].
The Airbus series (A3X0) can be best characterized as long-range airplanes. In fact,
the company has primarily targeted the growing demand for high capacity and
transcontinental flights. In addition, they have focused their effort at enhancing the
structural design using advanced winglets and working on aerodynamic improvements
83
for higher fuel efficiency [245]. For example, recent long-range airplanes, the twinjet
A330 and the four-engine A340, became popular for their efficient wing design [246].
Meanwhile, the Airbus A340-500 has an operating range of 16,700 km, which is the
second longest range of any commercial jet after the Boeing 777-200LR (range of 17,370
km) [247]. Therefore, it is not surprising that the A330-300 has been selected as a
benchmark of not only the same family airplane A340-300 but also the Boeing 767300ER, which is also a relatively long-range (11,065 km) airplane with high passenger
fuel efficiency (26.575 passenger*km/L). Additionally, the Airbus A380-800 became the
worlds largest passenger airplane with a seating capacity of 525 [248]. One can also
relate this feature to the reference set which consists of its predecessors: A340-500 and
A340-600 with relatively higher passenger capacities as well. This long-range, wide-body
airplane has emerged as a fast-growing segment as airlines emphasized transcontinental
aircraft capable of directly connecting any two cities in the world [243]. This, in fact, has
initiated a series of introductions of the A340 family for Airbus to compete with Boeing
[249], which is consistent with the fast local RoCs, indicating a very competitive segment
of the market with rapid improvement.
The Boeing 777 series ranks as one of Boeings best-selling aircraft for their high
fuel efficiency, which enable long-range routes [250]. In particular, the 777-300ER is the
extended range version of the 777-300, which has a maximum range of 14,685 km, made
possible by superior passenger fuel efficiency of 29.568 passenger*km/L. These
exceptional characteristics made not only the preceding 767-300ER but also the Airbus
series that pursued higher fuel efficiency (A340-200/300/600) appoint the 777-300ER as
a benchmark for their performance evaluation. Likewise, the 777-200LR has been
84
selected as a benchmark for long-range airplanes that have relatively smaller passenger
capacities: A340-200 and A340-500. Because of rising fuel costs, airlines have asked for
a fuel-efficient alternative and have increasingly deployed the aircraft on long-haul
transoceanic routes [251]. This has driven engineering efforts more toward energy
efficient aircraft, which is reflected in the fast local RoCs of the Boeing 777 series.
C. RISK ANALYSIS
I now turn to the strategic planning for the proposed airplane concepts (see Table 16).
In particular, the planning team would like to identify the relevant engineering targets for
each design concept as well as the corresponding rate of technological advancement, i.e.,
individualized RoC, so that they can examine the feasibility of proposed design concepts
in terms of their delivery target.
As SOA airplanes at the frontier of 2007 represent different types of past airplanes,
future airplanes, namely design concepts, can be classified by the characteristics of their
reference airplanes identified on the 2007 frontier. This allows the model to compute an
individualized RoC under which each future airplane is expected to be released. Figure
13 summarizes the results.
Table 16 Four airplane concepts in 2007

Design
concept
Travel
range
(1,000 km)
Passenger
capacity
(3rd class)
PFE
(passengers*km/L)
Cruising
speed
(km/h)
Maximum
speed
(km/h)
Delivery
target
(year)
14.816
467
28.950
917
988
2010
15.750
280
34.851
913
954
2010
15.000
315
34.794
903
945
2013
14.800
369
35.008
903
945
2015
85
The first design concept aims for a large commercial aircraft carrying 467
passengers while having the fast cruising speed of 917 km/h. As noted earlier, these
characteristics are also reflected in its reference airplanes: 747-400, 777-300ER, and
A380-800. That is, this design concept would compete with these three airplanes in the
current (2007) market with given specifications. The individualized RoC of this design
concept can therefore be obtained by interpolating local RoCs in conjunction with
reference information. Here, the individualized RoC obtained was 1.002748, which
suggests a more rapid technology development in its category compared to the average
RoC of 1.002149. This is about 28% faster and resulted in an estimated entry into service
(EIS) of the current design concept in 2011.49. Therefore, one may consider the delivery
target of 2010 to be an aggressive goal that might encounter technical challenges by
outpacing the rate of technological advancement of the past.
In a similar manner, characteristics of the second design concepts long range of
15,750 km with outstanding passenger fuel efficiency of 34.851 passenger*km/L are
consistent with the nature of its identified reference airplanes: A330-300 and 777-200LR.
As implied in the local RoCs of 777-200LR (1.004606) with its reference information
(0.67), this concept is subject to one of the fastest advancing technology clusters seeking
a high fuel efficiency. Consequently, it was expected that the very fast individualized
RoC of 1.003793 could achieve this level of specification by 2013.45. Similar to the first
design concept, this indicates that the delivery target of 2010 may involve a significant
technical risk since it requires exceeding the past rate of technological advancement.
86
Figure 13 2007 state-of-the-art frontier with regard to four design concepts6
The third design concept has features similar to the second design concept such that
it also aims to be a long-range and fuel efficient aircraft; however, it pursues a larger
passenger capacity of 315. This feature is reflected in the reference set that additionally
includes 777-300ER, which has a large passenger capacity of 365. The relatively slow
local RoC of the 777-300ER and the A330-300 may imply the difficulty of technological
advancement with respect to the travel range and passenger capacity. As a result, the
individualized RoC for this design concept was found to be 1.003494, giving a forecasted
EIS of 2012.45. Given the delivery target of 2013, the current design concept might be
regarded as a feasible goal; however, on the other hand this possibly entails a modest
market risk of lagging behind in the performance competition.
The last design concept is a variation of the third design concept, aiming for a much
larger airplane but with a shortened travel range. Not surprisingly, this different blend of
6
This figure depicts conceptualized frontier facets relevant to the four design concepts under discussion.
87
the same three peers makes a virtual target of this design concept positioned closer to the
777-300ER than to the 777-200LR and A330-300, which would result in a further
conservative prospect based on the slow rate of performance improvement represented by
the 777-300ER. Consequently, the individualized RoC was found to be 1.002568, giving
a forecasted EIS of 2020.16. This indicates that the delivery target of 2015 may be an
overly optimistic goal which could cause a postponement due to technical risks involved.
D. PROOF OF CONCEPT
I now come back to the present and validate the performance of the presented
method (see Table 17). The first design concept was the Boeing 747-8, which began
deliveries in 2012 [252]. In fact, this airplane faced two years of delays since its original
plan of 2010 due to assembly and design problems followed by contractual issues [253].
The second design concept was another Boeing airplane, 787-9, which made its
maiden flight in 2013, and the delivery began on July 2014 [254], [255]. In line with the
results, the originally targeted EIS of 2010 could not be met because of multiple delays
due to technical problems in addition to a machinists strike [256].
The third design concept was the initial design target of Airbus A350-900, which has
been changed and rescheduled to enter service in the second half of 2014 [257], [258].
The delay was mainly imposed by a strategic redesign of the A350, the so-called XWB
(extra-wide-body) program, that allows for a maximum seating capacity of 440 with a 10abreast high-density seating configuration as well as a reinforced fuselage design [259]. It
is interesting to note that Airbus has made a strategic decision by delaying the A350-
88
900s delivery with improved specifications to compete with the Boeing 777 series in the
jumbo jet segment, which was recognized in the analysis results seven years ago.
Similarly, the last design concept was the Airbus A350-1000, whose EIS has also
been rescheduled to 2018 [260]. This airplane is the largest variation of the A350 family
and designed to compete with the Boeing 777-300ER, as is also seen from the reference
information. Nevertheless, the postponed delivery target of 2018 may still be an
aggressive goal considering the technological advancement observed in this segment.
Table 17 Results summary

Design
concept
1 (747-8)
Planned
EIS
2010
Estimated
EIS
2011.49
Delayed
EIS
2012
A330-300, 777-200LR
2010
2013.45
2014
A330-300, 777-300ER, 777-200LR
2013
2012.45
2014
4 (A350-1000) A330-300, 777-300ER, 777-200LR
2015
2020.16
2018
Reference airplanes (competitors)

747-400, 777-300ER, A380-800
2 (787-9)
*
3 (A350-900 )
89
Initial design
4.1.2 VALIDATION USING PAST DATASETS

A. FORECASTING ACCURACY EVALUATION TECHNIQUES
Fundamentally, the true accuracy of a forecasting model should be judged by the
future events that were not known during the model building process. However, this socalled real time assessment has practical limitations for practitioners, which makes a
holdout sample test that measures how the model is able to reproduce data already known
but not used in construction of the model commonplace in forecasting literature [261].
The resulting forecast deviations, i.e., the difference between estimated data and the held
out data, can therefore provide an accuracy measure (or the goodness of fit) of the
forecast model being considered. This is also useful to compare the performance of
different models on the same data [262].
A holdout sample test requires the division of the historical data series into a fit
period and a test period. The forecasting origin is defined as the point from which the
forecasts are generated. In general, the practitioner can adopt either a single forecasting
origin or multiple forecasting origins.
Forecast statistics relying on a single forecasting origin, i.e., a fixed origin evaluation,
often require a fairly long test period since they yield only one set of forecasts from a
given forecasting origin. This necessarily renders the resulting summary statistics to be a
mlange of near-term and far-term forecast errors [263]. In addition, practitioners must
justify the reason to pick only one set of forecasts, which otherwise could raise a criticism
such that the results are susceptible to corruption by occurrences unique to that origin
[261].
90
To overcome the aforementioned problems of using a single origin, a rolling origin

evaluation successively updates the forecasting origin and accumulates forecasts from
each new origin. This technique makes it possible not only to obtain a sufficient number
of forecasts with the same historical data but also to desensitize the error measures to
special events at any single origin [264]. Furthermore, a rolling origin evaluation
produces multiple forecasts for every lead time, i.e., time period, between the origin and
the time being forecasted, which allows one to assess the forecasting accuracy of an
individual times series at each lead time [265].
Figure 14 Various forecast accuracy tests
Swanson and Whites study discussed that forecasting accuracy may also be affected
by an increase of the fit period when a rolling origin is employed [266]. To avoid this,
they suggested a procedure called a fixed size rolling window to maintain a constant
91
length of the fit period. This technique can clean out old data in an attempt to update the
forecasting model, thereby mitigating the influence of data from the distant past [261].
B. TEST RESULTS FROM EARLIER STUDIES7

To validate the performance of the proposed approach, holdout sample tests using
both constant RoC and segmented RoC on six datasets were conducted. Note that a
rolling origin evaluation was implemented from the origin used in the original studies. It
should be also noted here that the accuracy measure of root mean square error (RMSE)
was adopted to represent forecasting errors since our forecast is the arrival of
technologies, i.e., single scale with non-zero occurrence, estimated from their
performance levels [267]. In addition, deviation distributions were tested to distinguish
their differences from random variations with statistical significance. Table 18
summarizes comparative results of forecasting accuracies.
Table 18 Forecast accuracy comparisons

Application area
Commercial airplane [225]
RMSE
(Root mean square error)
Constant
Segmented
RoC
RoC
Deviation statistics
(95% confidence interval)
Constant
Segmented
RoC
RoC
Paired t test
t-stat
p-value
11.9208
6.3084
-9.06(5.18)
-3.56(3.65) -4.3653
0.0023
Fighter jet [230]
7.8229
7.2524
-7.22(3.38)
-6.32(3.17) -2.1274
0.0454
Battle tank [224]
23.1312
16.7987
-15.57(7.62)
-9.30(6.30) -5.3973
0.0001
LCD [232]
2.3061
2.1508
+0.63(0.27)
+0.35(0.30) 6.7182
0.0000
HEV [268]
3.4176
3.3329
-2.33(1.70)
-2.26(1.67) -3.2221
0.0105
DSLR [269]
2.6333
2.6271
-0.43(0.36)
-0.15(0.33) -3.8553
0.0002
This section is adapted from a paper accepted in proceedings of PICMET14 [317]
92
In all cases, the segmented RoC showed not only smaller forecasting errors, i.e.,
> , but also statistically significant distributions
closer to zero than that of constant RoC, i.e., |

| > | | ( < 0.05).
One may infer that forecasting accuracy improvement would be more significant if
unique segments were identified with a greater local RoC contrast to one another, and
future technologies were subject to those unique segments. This can be shown by
comparing the constant RoC with individualized RoCs.
Figure 15 contains this information. Note that RoCs were normalized to show their
distribution in comparison to constant RoCs that were set to be 100% across the
forecasting origins. It is seen that in the case of commercial airplane and battle tank
applications, individualized RoCs for forecasting targets show skewed distributions from
constant RoCs. That is, most of the forecasting targets were subject to relatively fast
progressing segments such that constant RoCs had to yield fairly conservative forecasts
as seen from the deviation statistics. On the other hand, the segmented RoC approach
could reflect those variations by obtaining fast individualized RoCs from the distribution
of local RoCs, which resulted in considerable accuracy improvements.
In contrast, when the local RoC of a certain segment by which most future
technologies are classified was close to the constant RoC, the impact of segmented RoC
would be marginal even if a wide range of local RoCs was identified. This can be seen
from the case of DSLR application in which a constant RoC could reasonably represent
the variations of individualized RoCs as an average value.
93
Figure 15 Relative comparison of segmented RoC with constant RoC
A special case can occur when the regions or clusters do not contain past products
that have been surpassed. In this case, a product may not have a local RoC. Graphically,
this would occur in Fig. 9 if products B and E were not included, which would then result
in G failing to have a local RoC. In such cases, Gs local RoC could be assumed to be the
average RoC of all SOA products (H and I). Another approach would be to average the
RoC for products that are on the same facet(s) of the efficiency frontier (simply H).
94
4.2 EX ANTE ANALYSIS: FOCUSED APPLICATIONS

In the previous section, the proposed model was tested based on historical data. In
contrast, this section provides a focused case study of supercomputer development to
demonstrate how the model can be used in a predictive manner.
4.2.1 EXASCALE SUPERCOMPUTER DEVELOPMENT 8

A. BACKGROUND
Supercomputers have played a critical role in various fields which require
computationally intensive tasks such as pharmaceutical testing, genomics research,
climate simulation, energy exploration, molecular modeling, astrophysical simulation, etc.
The unquenchable need for faster and higher precision analysis in those fields creates the
demand for even more powerful supercomputers. Furthermore, developing an indigenous
supercomputer industry has become a fierce international competition due to its role as a
strategic asset for a nationwide scientific research and the prestige of being the maker of
the fastest computers [270], [271]. While the vast majority of supercomputers have still
been built using processors from Intel, Advanced Micro Devices (AMD), and Nvidia,
manufacturers are committed to developing their own customized systems, e.g.,
interconnect, operating system and resource management, as system optimization
becomes a crucial factor in todays massively parallel computing paradigm [272].
Advances in supercomputers have come at a steady pace over the past 20 years in
terms of speed, which has been enabled by the continual improvement in computer chip
manufacturing [273]. As of March 2014, the worlds fastest supercomputer is the Tianhe8
This section is adapted from a paper accepted in Omega [318]
95
2 built by Chinas National University of Defense Technology (NUDT) performing at

over 33.86 Petaflops, i.e., 33.86 1015 floating point operations per second. This is about
273,000 times faster than the fastest machine from 20 years ago, the Fujitsu Numerical
Wind Tunnel. On average, progress went from being measured by Gigaflops in 1990 to
Teraflops in about 10 years, and then to Petaflops in another 10 years [274]. In line with
this, the next milestone is to build an Exascale computer, a machine capable of doing a
quintillion operations, i.e., 1018 , per second, which had been projected to see light of day
by 2018 [275]. However, there are significant industry concerns that this incremental
improvement might not continue mainly due to several practical problems.
The biggest challenge to build the Exascale computer is the power consumption
[276]. Tianhe-2, which is currently not only the fastest but also the most power hungry
supercomputer, uses about 18 megawatts (MW) of power. If the current trend of power
use continues, projections for the Exascale computing systems range from 60 to 130 MW,
which would cost up to $150 million annually [277]. Furthermore, few sites in the U.S.
will be able to host such power hungry computing systems due to the limited availability
of facilities with sufficient power and cooling capabilities [278]. Therefore, unlike past
advancement mainly driven by performance improvement [279], power efficiency has
now gone from being a negligible factor to a fundamental design consideration. To cope
with these issues, current efforts are targeting the Exascale machine that uses electrical
power of 20 MW using 100 million processors in the 2020 timeframe [276], [280].
Figure 16 illustrates this challenge. The rate of performance progress has been
constant until a recent date, and this would envision the first Exascale computer in 2018.
96
However, this steady performance improvement has been made possible by meeting the
exponentially growing power demands. Now that the 20MW of power consumption is set
as a feasible limit, the engineering effort has to be focusing more on minimizing power
consumption than on maximizing computational power. This implies that extrapolation
relying on a single performance measure, i.e., computing speed, may overlook required
features of future technology systems and could eventually result in an erroneous forecast.
Specifically, the average power efficiency of todays top 10 systems is about 2.2
Petaflops per megawatt. This indicates that it is required to improve power efficiency by
a factor of 23 to achieve the Exascale goal. It is therefore crucial to incorporate the power
consumption and multicore characteristics that identify the power efficiency into the
measure of technology assessment to have a comprehensive understanding of future high
performance computing (HPC) [281]. This requires a multifaceted approach to
investigate the tradeoffs between system attributes, which can tackle questions such as:
How much performance improvement would be restricted by power and/or core
reduction? What would be the maximum attainable computing performance with certain
levels of power consumption and/or the number of cores?
Figure 16 System tradeoffs to be considered for the future HPC trend

97
There are three leading technology paths representing todays supercomputer

development: hybrid systems, multicore systems, and manycore systems [282]. The
hybrid systems use both central processing units (CPU) and graphics processing units
(GPU) to efficiently leverage the performances [283]. The multicore systems maintain a
number of complex cores, whereas the manycore systems use a large number of less
powerful but power efficient cores within the highly parallel architecture [284].
Manufacturers and researchers are exploring these alternate paths to identify the most
promising, namely energy efficient and performance effective, avenue to face challenges
of deploying and managing Exascale systems [285][287]. The comparative analysis of
these technology paths can, therefore, give insights into the estimation of the future
performance levels as well as the possible disruptive technology changes.
This study employs Technology Forecasting using Data Envelopment Analysis
(TFDEA) to measure the technological progress considering tradeoffs among power
consumption, multicore processors, and maximum performance so that supercomputers
are to be evaluated in terms of both power efficiency and performance effectiveness. The
resulting analysis then provides a forecast of Exascale computer deployment under three
different development alternatives in consideration of the current business environment
as well as emerging technologies.
98
B. ANALYSIS
a) Dataset
The TOP500 list was first created in 1993 to assemble and maintain a list of the 500
most powerful computer systems [288]. Since the list has been compiled twice a year,
datasets from 1993 to 2013 have been combined and cleaned up so that each machine
appears once in the final dataset. The purpose of this study is to consider both power
efficiency and performance effectiveness; therefore, lists up to 2007 were excluded due
to the lack of information on the power consumption (see table 19). Variables selected for
this study are as follows:
Name (text): name of machine
Year (year): year of installation/last major update
Total cores (number): number of processors
Rmax (Gigaflops): maximal LINPACK performance achieved
Power (Kilowatts): power consumption
Processor technology/family (text): processor architecture being used
Interconnect family (text): interconnect being used
In the final dataset, there were a total of 1,199 machines, with the number of cores
ranging from 960 to 3.12 million, power ranging from 19KW to 17.81MW, and Rmax
ranging from 9 Teraflops to 33.86 Petaflops from 2002 to 2013. Note that a logarithmic
transformation was applied to all three variables due to their exponentially increasing
trends.
99
Table 19 TOP500 dataset from 1993 to 2013

Data column
Rank
Site
Manufacturer
Name
Computer
Country
Year
Total Cores
Accelerator/Co-Processor Cores
Rmax
Rpeak
Efficiency (%)
Nmax
Nhalf
Power
Mflops/Watt
Measured Size
Processor Technology/Family
Processor Generation
Processor
Proc. Frequency
Processor Cores
Processor Speed (MHz)
System Family
System Model
Operating System
OS Family
Cores per Socket
Architecture
Accelerator/Co-Processor
Segment
Application Area
Interconnect Family
Interconnect
Region
Continent
1993
~2007
2008
~2009
2010-1*
2010-2*
2011-1*
2011-2*
~2013
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O
O: Available, : Unavailable
-1(2) denotes the first (second) list of corresponding years
100
b) Model building
As discussed earlier, power consumption and the number of cores are key variables
representing the power efficiency and therefore were used as input variables, while the
maximum LINPACK performance (Rmax) was used as the output variable. This allows
the model to identify the better performing supercomputer which has lower power,
fewer cores, and/or higher performance if other factors are held constant.
Orientation can be either input-oriented or output-oriented and can be best thought of
as whether the technological progress is better characterized as input reduction or
output augmentation [232]. While power consumption will be a key concern in the
Exascale computing, the advancement of this industry has been driven primarily by
computing performance, i.e., flops, improvement. In fact, the Exascale computing is a
clearly defined development goal, and therefore an output orientation was selected for
this application. It should be noted here that either orientation can deal with tradeoffs
among input and output variables.
As with many DEA applications, variable returns to scale (VRS) was selected for
appropriate returns to scale assumption since doubling the input(s) doesnt correspond to
doubling the output(s) here.
The main purpose of this study is to make a forecast of the Exascale computer
deployment by examining past rates of progress, thus the frontier year of 2013 was used
so as to cover the entire dataset. Lastly, minimizing the sum of effective dates was added
as a secondary goal into the model to handle the potential issue of multiple optima from
the dynamic frontier year [240]. Table 20 summarizes the model parameters used in this
study.
101
Table 20 TFDEA model parameters

Inputs
Output
Orientation
RTS
Frontier year
Frontier type
Second goal
Power, Cores
Rmax
Output
VRS
2013
Dynamic
Min
Figure 17 shows 13 supercomputers identified as SOAs from the analysis. Intel

provided the processors for the largest share (62%) and, inter alia, GPU/Accelerator
based systems showed impressive performances both in power and core efficiencies,
while IBMs Blue Gene, NNSA/SC and Blue Gene/Q showed comparable power
efficiency as manycore based systems.
Figure 17 13 State-of-the-art supercomputers considering system tradeoffs

102
As also seen from specifications of these supercomputers in Table 21,

supercomputers in different sizes are characterized as being equally competitive in
consideration of system tradeoffs. This enables the model to construct technology
frontiers from which various production possibilities can be identified. This characteristic,
in fact, differentiates the TFDEA process from a single dimensional measure such as the
TOP500 list in which technological efforts to become energy efficient and/or core
efficient are not taken into account.
Table 21 Specifications of 13 SOA supercomputers

Name
Eurora Eurotech
Aurora HPC 10-20
Tianhe-2 TH-IVBFEP
Year
Cores
Power
Rmax
Technology Family
Processor
Interconnect
2013
2,688
46.00
100,900
Intel
InfiniBand
2013
3,120,000
17,808.00
33,862,700
Intel
Custom
HPCC
2013
10,920
237.00
531,600
Intel
InfiniBand
Titan Cray XK7
2012
560,640
8,209.00
17,590,000
AMD
Cray
2012
9,216
45.11
110,500
Intel
InfiniBand
2012
8,192
41.09
86,346
IBM Power
Custom
iDataPlex DX360M3
2011
3,072
160.00
142,700
Intel
InfiniBand
NNSA/SC Blue
Gene/Q Prototype 2
2011
8,192
40.95
85,880
IBM Power
Custom
DEGIMA Cluster
2011
7,920
31.13
42,830
Intel
InfiniBand
2008
1,260
18.97
9,259
IBM Power
InfiniBand
2008
1,024
42.60
9,669
Intel
InfiniBand
2008
960
153.43
14,669
IBM Power
InfiniBand
2007
960
91.55
9,058
Intel
InfiniBand
Beacon Appro
GreenBlade
GB824M
BlueGene/Q, Power
BQC 16C 1.60GHz
BladeCenter QS22
Cluster
Cluster Platform
3000 BL2x220
Power 575, p6 4.7
GHz, Infiniband
BladeCenter HS21
Cluster
103
Figure 18 illustrates performance trajectories based on 1,199 supercomputers from

three dominant processor families: AMD, IBM Power, and Intel (IA-32/64, Core,
Nehalem, Westmere, and Sandy Bridge). Since the Japanese supercomputer, EarthSimulator, was built in 2002 using a Nippon Electric Company (NEC) chip which was
not adopted by other supercomputer manufacturers thereafter, Fig. 18 is drawn from 2005
to focus on the main vendors of processors for todays systems. The ordinate is the
overall performance score from the DEA model. As such, each line indicates a
performance trajectory of the top performing supercomputers from each year against the
frontier year of 2013. A performance score of 100% indicates that the supercomputer has
a superior performance enough to be on the SOA frontier in 2013. A performance score
higher than 100% denotes super-efficiency from the DEA model, which can show how
much the supercomputer is outperforming other SOA supercomputers.
The trajectory of Many/Multicore systems shows that IBM Power (PC) processor
based machines are outperforming AMD processor based machines. AMD processor
based machines, however, showed surpassing performances over IBM Power (PC) based
machines when they were adopted by Cray to build hybrid systems in 2011 and 2012. In
fact, the successful development of Titan Cray XK7 using AMD Opteron CPUs coupled
with Nvidia coprocessors has made Cray Inc. one of the leading supercomputer vendors
to date. Interestingly, this is also consistent with the fact that Cray Inc. was awarded the
$188M U.S. Blue Waters contract, which is a project funded by the National Science
Foundation (NSF), replacing IBM, which had pulled out of the project prior to
completion in 2011 [289].
104
Figure 18 Performance trajectories of different processor families
It is also interesting to point out that there was a distinct performance gap between
Many/Multicore based machines and GPU/Accelerator based machines using AMD
processors in 2011 and 2012. This can be attributed to the strategic partnership between
Cray and AMD. In fact, Cray has been a staunch supporter of AMD processors since
2007, and their collaboration has delivered continued advancement in HPC [290]. In
particular, Crays recent interconnect technology, Gemini, was customized for the AMD
Opteron CPUs Hyper-Transport links to optimize internal bandwidth [291]. Since
modern supercomputers are deployed as massively centralized parallel systems, the speed
and flexibility of interconnect become important for the overall performance of a
supercomputer. Given that hybrid machines using AMD processors all use Crays
105
interconnect system, one may notice that AMD based supercomputers had a significant
performance contribution from Cray interconnect as well as Nvidia coprocessors.
One may notice that top supercomputers based on Intel processors have switched to
hybrid systems since 2010. This is because combining CPUs and GPUs is advantageous
in data parallelism, which makes it possible to balance the workload distribution as
efficient use of computing resources becomes more important in todays HPC structure
[292]. Hybrid machines using Intel processors have all adopted InfiniBand interconnect
for their cluster architectures regardless of GPUs/Accelerators: Nvidia, ATI Radeon,
Xeon Phi, PowerXCell, etc. InfiniBand, manufactured by Mellanox and Intel, enables
low processing overhead and is designed to carry multiple traffic types such as clustering,
communications, and storage over a single connection [293]. In particular, its GPUDirect technology facilitates faster communication and lower latency of GPU/Accelerator
based systems that can increase computing and accelerator resources, as well as improves
productivity and scalable performance [294]. Intel acquired the InfiniBand business from
Qlogic in 2012 to support innovating on fabric architectures not only for the HPC but
also data centers, cloud, and Web 2.0 market [295].
As another possibility, recent attention is focusing on Intels next generation
supercomputer, which will adopt Crays Aries interconnect with Intel Xeon Phi
accelerator as Intels first non-InfiniBand based hybrid system after its acquisition of
interconnect business from Cray [296]. Interestingly, this transition reflects the strategic
decision of Cray, ending an association with AMD to facilitate an independent
interconnect architecture rather than a processor specific one as AMDs recent
performance and supply stability fell behind competitors [291], [297].
106
Unlike AMD or Intel processor based systems, the top performing supercomputers
using the IBM Power (PC) processor were Many/Multicore systems. IBM initially
developed the multicore architecture which later evolved to manycore systems, known as
Blue Gene technology. The Blue Gene approach is to use a large number of simple
processing cores and to connect them via a low latency, highly scalable custom
interconnect [298]. This has the advantage of achieving a high aggregate memory
bandwidth, whereas GPU clusters require messages to be copied from the GPU to the
main memory and then from main memory to the remote node, whilst maintaining low
power consumption as well as cost and floor space efficiency [299]. Currently,
GPU/Accelerator based systems suggest smaller cluster solutions for the next generation
HPC with its promising performance potential; however, the Blue Gene architecture
demonstrates an alternate direction of massively parallel quantities of independently
operating cores with fewer programming challenges involved [300].
107
c) Model validation
To validate a predictive performance of the constructed model, hold-out sample tests
were conducted. Specifically, a rolling origin was used to determine the forecast accuracy
by collecting deviations from multiple forecasting origins so that the performance of the
model can be tested both in the near-term and far-term. This provides an objective
measure of accuracy without being affected by occurrences unique to a certain fixed
origin [261]. The comparative results with the planar model and random walk 9 are
summarized in Table 22.
Since the first hybrid system, the Blade Center QS22, appeared in 2008 in the dataset,
the hold-out sample test was conducted from the origin of 2009 for hybrid systems. For
example, the mean absolute deviation of 1.58 years was obtained from TFDEA when the
model made a forecast on arrivals of post-2009 hybrid systems based on the rate of
technological progress observed from 2008 to 2009. The overall forecasting error across
the forecasting origins was found to be 1.32 years, which is more accurate than the planar
model and random walk.
Although multicore systems showed successive introductions from 2007 to 2012,
technological progress, i.e., expansion of SOA frontier surface, hasnt been observed
until 2010. This rendered the model able to make a forecast only in 2011. The resulting
forecast error of TFDEA was found to be about a year, which is slightly greater than that
of the planar model albeit still more accurate than the random walk. However, care must
be taken to conclude which one was more accurate than the other since the result was
The random walk model simply predicts that the next period value is the same as the current value, i.e.,
the arrivals of forecasting targets = the forecasting origin [319].
108
obtained only from a single forecasting origin in 2011. Therefore, the forecasting of the
multicore Exascale system will be made using both TFDEA and the planar model in the
following section.
Consecutive introductions of manycore systems with a steady technological progress
made it possible to conduct hold-out sample tests from the origin of 2007 to 2012.
Notwithstanding a bigger average forecasting error of 1.49 years due to the inclusion of
errors from longer forecasting windows than the other two systems, TFDEA showed
outperforming forecast results compared to the planar model and random walk.
Table 22 Model validation using rolling origin hold-out sample tests

Mean absolute deviation (unit: year)
Forecast
Origin
Hybrid systems
Multicore systems
Manycore systems
TFDEA
Planar
model
Random
walk
TFDEA
Planar
model
Random
walk
TFDEA
Planar
model
Random
walk
2007
N/A
N/A
N/A
N/A
N/A
N/A
1.8075
2.8166
2.9127
2008
N/A
N/A
N/A
N/A
N/A
N/A
1.4470
2.5171
2.4949
2009
1.5814
2.7531
2.1852
N/A
N/A
N/A
2.0060
2.3593
2.0509
2010
1.1185
1.9956
1.5610
N/A
N/A
N/A
1.4996
2.0863
1.6016
2011
1.8304
1.5411
1.2778
0.9899
0.7498
1.0000
1.2739
1.8687
1.3720
2012
0.7564
1.2012
1.0000
N/A
N/A
N/A
0.8866
2.2269
1.0000
Average
1.3217
1.8728
1.5060
0.9899
0.7498
1.0000
1.4867
2.3125
1.9053
N/A: insufficient data
Overall, it is shown that the TFDEA model provides a reasonable forecast for three
types of supercomputer systems with the maximum possible deviation of 18 months. In
109
addition, it is interesting to note that forecasts from TFDEA tended to be less sensitive to
the forecasting window than the planar model or random walk.
This implies that the current technological progress of supercomputer technologies
exhibits multifaceted characteristics that can be better explained by various tradeoffs
derived from the frontier analysis. In addition, a single design tradeoff identified from the
planar model was shown to be vulnerable to the forecasting window: it showed a
tendency to be less accurate as the forecasting window gets longer.
110
d) Forecasting
I now turn to the forecasting of the Exascale systems. As previously noted, the
design goal of the Exascale supercomputer is not only to have the Exaflops (1018 flop /
second) computing performance but also 20MW power consumption and 100 million
total cores considering the realistic operating conditions (see Table 23) [276], [280]. This
set of specifications was set as a forecasting target to estimate when this level of system
could be operational given the past technological progress identified from the relevant
segments.
Table 23 Exascale computer as a forecasting target

Cores
Power
Rmax
100 million
20 MW
1 Exaflops
Table 24 summarizes the forecasting results from the three architectural approaches.
Exascale performance was forecasted to be achieved earliest by hybrid systems in
2021.13. Hybrid systems are expected to accomplish this with a relatively high
individualized RoC of 2.22% and having the best current level of performance
represented by Tianhe-2. Figure 19 depicts the identified individualized RoC with respect
to the local RoCs. It is seen that the technology frontier of hybrid systems includes a wide
range of progress patterns in terms of local RoCs, i.e., 0.27%~2.71%, and the Exascale
target is subject to the relatively fast advancing segment.
Considering the possible deviations identified in the previous section (1.32 year),
one could expect the arrival of a hybrid Exascale system within the 2020 timeframe. This
promising future of hybrid systems is, in fact, acknowledged by many industry experts
111
claiming that GPU/Accelerator based systems will be more popular in the TOP500 list
for their outstanding power efficiency, which may spur the Exascale development [282],
[285].
Figure 19 Individualized RoCs with respect to the local RoCs
The forecasted arrival time of the first multicore based Exascale system is far
beyond 2020 due to the slow rate of technological advancement: 1.19% as well as
relatively lower performance of current SOA multicore systems. It is also shown from
Fig. 19 that the technology frontier of multicore systems has relatively narrow ranges of
local RoCs, i.e., 0.48%~1.86%, and, inter alia, the Exascale target is pertinent to the
moderate segment.
Note that projection from the planar model also estimated the arrival of multicore
based Exascale system farther beyond the 2020 timeframe10.
This result implies that innovative engineering efforts are required for multicore
based architecture to be scaled up to the Exaflop performance. Even though the RIKEN
10
The arrival of the first multicore Exascale system was forecasted in 2061.62 from the planar model.
112
embarked on the project to develop the Exascale system continuing the preceding success
of K-computer, IBMs cancellation of the Blue Water contract and recent movement
toward the use of a design house raises questions on the prospect of multicore based
HPCs [301], [302].
The first manycore based system is expected to reach the Exascale target by 2022.28.
This technology path has been mostly led by the progress of the Blue Gene architecture,
and the individualized RoC was found to be 2.34%, which was the fastest of the three. It
is interesting to note from Fig. 19 that this fast progress, however, belongs to the
moderate region of the technology frontier where the local RoCs range from 1.09% to
3.40%.
Although this fast advancement couldnt overcome the current performance gap with
hybrid systems in the Exascale race, the Blue Gene architecture still suggests a promising
pathway toward the Exascale computing by virtue of its stable configurations closer to
the traditional design with fewer programming challenges [299].
Table 24 Forecast results of Exascale supercomputer

Hybrid system
Multicore system
Manycore system
Individualized
Rate of change (RoC)
1.022183
1.011872
1.023437
Forecasted arrival of
Exascale supercomputer
2021.13
(2019.81~2022.45)
2031.74
(2030.75~2032.73)
2022.28
(2020.80~2023.77)
( ): Ranges of forecasts considering the possible deviations
113
C. DISCUSSION
The analysis of technological RoC makes it possible to forecast a date for achieving
Exascale performance from three different approaches; however, it is worthwhile to
examine these forecasts with consideration for the business environment and emerging
technologies to anticipate the actual deployment possibilities of the Exascale systems.
The optimistic forecast is that, as seen from the high performing Tianhe-2 and Titan
Cray XK7 system, there would be an Intel or AMD based system with a Xeon Phi or
Nvidia coprocessor and a custom Cray interconnect system. However, given business
realities it is unlikely that the first Exascale system will use AMD processors. Intel
purchased the Cray interconnect division and is expected to design the next generation
Cray interconnect optimized for Intel processors and Xeon Phi coprocessors [303].
Existing technology trends and the changing business environment would make a
forecast of a hybrid Exascale system with a Cray interconnect, Intel Processors and Xeon
Phi coprocessors.
The 2.22% annual improvement for hybrid systems has come mostly from a
combination of advances in Cray systems, such as their transverse cooling system, Cray
interconnects, AMD processors and Nvidia coprocessors. It is difficult to determine the
contribution of each component; however, it is worth noting that only Cray systems using
AMD processors were SOAs. This implies that Crays improvements are the highest
contributor to the RoC for AMD based hybrid systems. Furthermore, Intels recent
decision to move production of Cray interconnect chips from TSMC to its more advanced
processes will likely result in additional performance improvement. Thus, one may
114
expect that the Cray / Intel collaboration might result in a RoC greater than the 2.22%
and reach the Exascale goal earlier.
Another possibility of achieving Exascale systems is IBMs Blue Gene architecture
using the IBM Power (PC) processor with custom interconnects. This approach has
shown a 2.34% yearly improvement building on the 3rd highest rated Sequoia system.
The Blue Gene architecture, with high bandwidth, low latency interconnects and no
coprocessors to consume bandwidth or complicate programming, is an alternative to the
coprocessor (hybrid) architectures being driven by Intel and AMD. Given IBMs more
stable business structure, it may be more effective moving forward while Intel / Cray
work out their new relationship.
Who has the system experience to build an Exascale system? Cray, IBM and Appro
have built the largest SOA original equipment manufacturer (OEM) systems. In 2012,
Cray purchased Appro, leaving just two major supercomputer manufacturers [304].
Based upon the captured RoCs and the business changes, one can expect that the first
Exascale system will be built by either Cray or IBM.
As supercomputer systems become more complex and expensive, it is worth noting
the funding efforts in a story about the future of HPC. The U.S. Department of Energy
(DOE) recently awarded $425 million in federal funding to IBM, Nvidia, and other
companies that will build two 150 Petaflops systems with an option on one system to
build it out to 300 Petaflops [305]. The plan states that IBM will supply its Power
architecture processors, while Nvidia will supply its Volta GPUs, and Mellanox will
provide interconnected technologies to wire everything together [306]. In addition, the
DOE and the National Nuclear Security Administration (NNSA) have announced $100
115
million to fund the FastForward2 project that will develop technologies needed for future
energy efficient machines in collaboration with AMD, Cray, IBM, Nvidia and Intel [307].
One may notice that U.S. science funding will support both hybrid and manycore systems
for producing the next leap toward the Exascale.
Japan had earlier announced a goal to reach Exascale with the total project cost of
$1.2 billion by 2020 [308]. However, the deputy director at the RIKEN Advanced
Institute for Computational Science (RAICS) recently modified the goal and plans to
build a 200 to 600 Petaflops system by 2020. Nonetheless, given the fact that RIKEN
selected Fujitsu to develop the basic design for the system, there is a keen interest in how
much the multicore system could be scaled up with a relatively low power efficiency of
complex cores.
Data driven forecasting techniques, such as TFDEA, make a forecast of technical
capabilities based upon released products, so emerging technologies that are not yet being
integrated into products are not considered. In the supercomputer academic literature,
there is an ongoing debate about when the currently dominating large core processors
(Intel, AMD) will be displaced by larger numbers of power-efficient, lower performance
small cores such as ARM, much like what happened when microprocessors displaced
vector machines in the 1990s and ARM based mobile computing platforms are affecting
both Intel and AMD desktop and laptop sales [282], [287]. Although there is no ARM
based supercomputer in the TOP500 yet, the European Mont-Blanc project is targeting
getting one on the list by 2017, and Nvidia is developing an ARM based supercomputer
processor for use with its coprocessor chips [309]. Small cores are a potentially disruptive
technology as power efficiency is becoming more important; therefore, further analysis is
116
needed to investigate when it will overcome the challenges of building interconnects to

handle a larger number of smaller cores or when software developers will overcome the
synchronization challenges of effectively using more cores.
Another
new
kind
of
supercomputer
attracting recent
attention
is
the
superconducting supercomputer (as one way to enable the quantum computing). Even
though the exact financial and technical details with a timeframe were not disclosed, the
Intelligence Advanced Research Projects Activity (IARPA) revealed that funding
contracts have been awarded to IBM, Raytheon-BBN and Northrop Grumman
Corporation focusing on the development of the Cryogenic Computer Complexity (C3)
program [310]. Early research suggested that a superconducting supercomputer would be
able to provide around 100 Petaflops of performance while consuming just 200 kilowatts
[311]. If the mission of the C3 program can be achieved and the related technologies can
be successfully transferred to practical usages, the next generation supercomputers could
be far different from the ones of the past, and the Exascale goal could be achieved
without concerns of power and cooling capacities.
Lastly, this study set the Exascale target considering the realistic operating
conditions: 20MW of power consumption and 100 million cores. If this set of
specifications was relaxed at the manufacturers free will, the arrival of an Exascale
computer could come earlier than current forecasts as China is believed to be targeting
the 2018-2020 timeframe for continuing their gigantic design of Tianhe-2.
117
D. SUMMARY OF FINDINGS
The HPC industry is experiencing a radical transition which requires improvement
of power efficiency by a factor of 23 to deploy and/or manage the Exascale systems. This
has created an industry concern that the nave forecast based on the past performance
curve may have to be adjusted. TFDEA is well suited to deal with multiple tradeoffs
between systems attributes. This study examined comparative prospects of three
competing technology alternatives with various design possibilities considering the
complex business environment to achieve the Exascale computing so that researchers and
manufacturers can have a better view of their development targets. In sum, the results
showed that the current development target of 2020 might entail technical risks
considering the rate of change toward the power efficiency observed in the past. It is
anticipated that either a Cray built hybrid system using Intel processors or an IBM built
Blue Gene architecture system using PowerPC processors will likely achieve the goal
between early 2021 and late 2022.
In addition, the results provided a systematic measure of technological change,
which can guide a decision on the new product target setting practice. Specifically, the
rate of change contains information not only about how much performance improvement
is expected to be competitive but also about how much technical capability should be
relinquished to achieve a specific level of technical capabilities in other attributes. One
can also utilize this information to anticipate the possible disruptions. As shown in the
HPC industry, the rate of change of the manycore system was found to be slightly faster
than that of the hybrid system. Although the arrival of the hybrid Exascale system is
forecasted earlier than a manycore system because of its current surpassing level of
118
performance, the fast rate of change of the manycore system implies that the performance
gap could be overcome, and Blue Gene architecture might accomplish the Exascale goal
earlier if the hybrid system development is unable to keep up with the expected progress.
119
V. CONCLUSION
5.1 CONTRIBUTIONS TO APPLICATION AREA

5.1.1 EXASCALE SUPERCOMPUTER DEVELOPMENT
Experts and recent literature commonly point out that technological progress of
supercomputers needs to be re-examined since power consumption and effective
parallelism are increasingly critical factors to build Exascale systems. Above all,
improvement of power efficiency by a factor of 23 is required to deploy and operate the
Exascale systems in practice. This unprecedented type of technology necessarily requires
choosing a high-level implementation strategy with respect to the tradeoffs between
system attributes. There are three leading architectureshybrid, multicore, and manycore
systemsto cope with related challenges, and manufacturers and researchers are
exploring their expected potentials to have an accurate view of the technological
advancement in high performance computing.
The newly developed extensions to TFDEA are well suited to this application to deal
with multiple tradeoffs between systems attributes. The results provide comparative
prospects of those three competing technology alternatives with various design
possibilities accounting for the business environment to achieve the Exascale systems.
Specifically, the results indicate that the first Exascale system is likely to be built by Cray
or IBM with their customized interconnects. Considering the recent business changes, it
is more likely that the first Exascale performance can be achieved by Intel processors and
Xeon Phi coprocessors as hybrid systems.
120
In addition, the identified rates of change can be used to give insights into the
estimation of the future performance levels for new product development target setting
purposes. Supercomputer manufacturers may have their own roadmaps based on past
performance improvement, which has been mostly driven by the computation speed.
However, as noted above, the transition toward Exascale demands considering both
power and core efficiency. This would necessarily require the established roadmap to be
modified. There are three alternatives, i.e., hybrid, multicore, and manycore systems,
each heading toward the same goal. Who then do manufacturers bet on to win the race?
In addition, how much performance improvement should be made by a certain point in
time to meet the planned delivery of the Exascale computer? The rate of change contains
information to better inform their decisions.
One can also utilize rates of change to anticipate the possible disruptions. For
example, the rate of change of manycore systems was found to be faster than that of
hybrid system. Although the arrival of a hybrid Exascale system is forecasted earlier
because of its current level of performance being superior to manycore systems, this
indicates that the performance gap could be overcome, and the Blue Gene architecture
might accomplish the Exascale goal earlier within the possible forecast deviation if Cray
and Intel are unable to keep up with performance advancement expected from the given
rate of change while they work out their new relationship.
121
5.2 CONTRIBUTIONS TO MANAGERIAL INSIGHT

5.2.1 RISK ANALYSIS
One of the motivations of this study stemmed from a practical question: When
might be the ideal timing to release the new product? To answer this question, one may
have to know not only what segments the current product concept is subject to but also
how competitive corresponding segments are. This subject can be translated into the
research topic of integrating product positioning with the assessment of performance
improvement over time, which has rarely been addressed in both new product
development and management science literature. The presented use of time series
benchmarking analysis makes it possible to estimate the auspicious time by which
proposed design concepts will be operational as competitive products by taking into
account the rate of performance improvement expected in a target segment.
The empirical illustration of commercial airplane development has shown that the
new method provides valuable information such as dominating designs, distinct segments,
and the potential rate of performance improvement, which can be utilized in the early
stage of new product development. In particular, six SOA airplanes characterizing
distinct segments were identified with corresponding local RoCs and, inter alia,
technological advancement toward long-range and wide-body airplanes represented very
competitive segments of the market with rapid changes. The resulting individualized
RoCs were able to estimate the arrivals of four different design concepts, which is
consistent with what has happened since 2007 in the commercial airplane industry.
Similarly, the case study of Exascale supercomputer development showed that the
current development target of 2020 might entail technical risks considering the rate of
122
change toward the power efficiency observed in the past. It is forecasted that either a
Cray built hybrid system using Intel processors or an IBM built Blue Gene architecture
system using PowerPC processors will likely achieve the goal between early 2021 and
late 2022. This indicates that the improvement of power efficiency by a factor of 23
would require the maximum delay of four years from the past performance curve.
5.2.2 NEW PRODUCT TARGET SETTING

Unlike market research methods or heuristic ideation techniques, this study employs a
product feature-based clustering approach. This engineering approach can be used to
identify homogeneous product groups by ordering products from broad to niche based on
the number of times each product appears in others reference sets. This information can
be very useful for decision makers to position their products by referring to product
designs in other clusters, which provides a direction to adjust the combination of its
product attributes so as to be assigned into a desired cluster.
The rate of technological advancement identified in each market segment can further
give insights into the target setting practices for a new product development planning.
That is, manufacturers may position their products within the current SOA frontier and
utilize the corresponding rate of change to see whether their design targets would be
located on the estimated future frontiers. Similarly, one could also utilize this information
for the pricing strategy if the price was included as an input variable.
123
5.3 CONTRIBUTIONS TO METHODOLOGICAL DEVELOPMENT

5.3.1 IDENTIFICATION OF LOCAL RATE OF CHANGE
As an answer to the first research question, a new model was developed to capture the
local RoCs from homogeneous technologies in the past. This enables each SOA
technology not only to characterize the corresponding segments but also to represent the
unique progress patterns. The resulting segmented rate of change extensions to TFDEA is
an objective data-driven process.
Table 25 Answer to research question 1
Research
Question
#1
How do we capture the local rate of change from past technologies?
Answer
to
Research
Question
#1
The efficiency changes of past products with respect to corresponding

segments are aggregated by taking the weighted average using the
reference information. Each local RoC therefore represents a growth
potential of adjacent frontier facets based on the technological
advancement observed from related past products.
5.3.2 IDENTIFICATION OF INDIVIDUALIZED RATE OF CHANGE

As an answer to the second research question, a new model was developed to identify
the appropriate progress potential, i.e., individualized RoC, for each forecasting target.
Similar to the process of calculating local RoCs from surpassed technologies in the past,
individualized RoCs can be obtained from SOA technologies that present similar
operating practices. Seven rolling-origin hold-out sample tests have shown that this
approach improves the overall forecasting results in comparison to the original TFDEA
model.
124
Research
Question
#2
How do we determine the individualized rate of change for future

technologies?
Answer
to
Research
Question
#2
Reference information of each forecasting target is used to combine

relevant local RoC(s) to identify an individualized RoC under which
each forecasting target is expected to arrive.
5.3.3 A FINITE FORECAST FOR AN INFEASIBLE TARGET

As an answer to the third research question, a new model was developed to make a
finite forecast for formerly infeasible targets. Specifically, this model identifies bidirectional distances, i.e., extremity and standard radial distance, for infeasible targets,
and therefore TFDEA computes both input-oriented and output-oriented RoCs and
applies them to estimate the arrival of target technologies. By virtue of the adopted
alternate super-efficiency measure, this approach yields the forecast results identical to
those from the original TFDEA model when the target is feasible. Consequently, a new
model can be fully integrated as an extension of the TFDEA model.
Research
Question
#3
How do we deal with infeasible targets?
Answer
to
Research
Question
#3
Bi-directional L1 distances are obtained and time spans for each

distance are estimated using corresponding RoCs, i.e., current
orientation RoC for standard radial distance and the opposite orientation
RoC for extremity (if any), from both orientations.
125
5.4 LIMITATIONS
This study adopts an engineering design perspective that a product is a complex
assembly of interacting components. Consequently, the term segment is being used to
indicate a set of engineering designs having a similar mix of product attributes. Note that
the academic community in marketing uses this term with a broader implication of shared
needs and value propositions determined by a meaningful number of customers. In this
study, the identified targets and competitors are derived purely based on technical
specifications. This attribute-based approach can be limited in its ability to represent the
overall appeal of products especially those for which other holistic product features that
are not reflected in technical specifications such as aesthetics are important.
In a similar vein, the DEA measurement is based on the relative performance of the
products, and therefore the state-of-the-art products may be the most advanced ones but
may not be the most successful ones in the market. This indicates that the resulting rate of
change is more likely to be reflected by the technological progress than market desires.
In addition, the estimation of release date from the proposed model doesnt take into
account externalities such as strategic postponement, financial conditions, market
acceptance levels, self-imposed delay due to the product portfolio management, etc.
Therefore, the resulting release date should be understood as a baseline for implementing
the tactical launch decision with respect to product attributes concerned rather than the
bottom line of decision. This also suggests that the estimated release date may have to be
further adjusted if the industrial market is less sensitive to the technological superiority
than to market strategies.
126
5.5 FUTURE WORK DIRECTIONS

This section suggests future work directions including methodological extensions of
proposed model as well as its applications to the managerial decision makings.
5.5.1 INNOVATIVENESS MEASURE

One might make use of a presented risk analysis approach to develop a measure of
innovativeness. When there is a need to quantify the innovativeness of products
independent from market factors, this method can suggest how much a certain product
has contributed to accelerating the rate of performance improvement or has moved up the
product release compared to the expected arrival. Alternate approaches could also
investigate the possible modifications in product designs to reduce the risk, thereby
meeting the release date determined.
5.5.2 ALTERNATE EFFICIENCY MEASURES

A DEA measure is by definition an equiproportional ratio of how the DMU being
assessed can either reduce its inputs or augment its outputs to reach its virtual target. This
radial efficiency score may not account for all sources of inefficiency by having input
and/or output slacks that are not reflected in the collective proportion. Further, the
traditional DEA model is labeled as radial since it gives preemptive priority either to
conservation of the input or to expansion of the output, depending on model orientation.
This implies that the radial approach may not capture the technological advancement
within structural characteristics or functional improvements, while the technology
systems objective might often be the desire to change the mix of them. This suggests the
127
non-radial and/or non-oriented distance measures for estimating the frontier with
consideration of the furthest target, closest target, restricted targets, scale-efficient target,
or target located in a predefined direction that could set more realistic targets whereby
diverse patterns of technological advancement can be explored.
Table 28 Alternate non-radial efficiency measures

Oriented measure
Russell measure
Geometric distance function model (GDF)
Hyperbolic model
Non-oriented measure
Additive model
Range-adjusted model (RAM)
Slack-based model (SBM)
Proportional slack-based model (P-SBM)
Directional model
5.5.3 WEIGHT RESTRICTIONS

Another future research topic might consider exploring the product segmentation by
imposing weight restrictions on the model. The current model doesnt require an
axiomatic weighting scheme on any attributes being considered. This might cause an
overestimation of products having an extreme feature in certain attribute(s). To avoid this,
one can establish the boundary of weights attached to certain attributes whereby
significant value propositions of both extant and potential market segments can be
reflected into the analysis. This includes determining how much certain attributes should
be valued than others as well as how much maximum (or minimum) weight can be
assigned to certain attributes.
128
5.5.4 TIME VARYING RATE OF CHANGE

Another direction of future work could also investigate the varying RoCs over time.
The local RoCs can provide information about the number of distinct segments within
which differing rates of technological advancement have been captured. One can examine
this information in conjunction with maturity curves so that market dynamics among
identified segments can be studied.
5.5.5 DIRECTION OF TECHNOLOGICAL PROGRESS

Comparative results between the input-oriented model and the output-oriented model
could provide insights regarding whether the direction of innovation has been switched.
As previously discussed, an infeasible target in the input (output)-oriented model may
occur when the output (input) levels of the target technology are unprecedented in the
past. Hence, one might derive an indication of technological progress from an
increasing/decreasing number of infeasible targets over time. In a similar manner, the
magnitude of extremity may contain relevant information about the potential paths of
technological progress.
5.5.6 STOCHASTIC FRONTIER

Lastly, incorporating stochastic variation into the model must be a matter for
speculation. DEA is by definition a deterministic model which renders the presented
method confined to capturing the rate of performance improvement from the evolution of
the SOA frontier. This might provide an aggressive estimation unless the best performing
products on the market are sought. Hence, stochastic measurements would be able to
129
complement this aspect such that the rate of performance improvement can be obtained
from diverse levels of products, thereby yielding the risk distribution for each design
concept instead of a point estimation.
130
References
[1]
N. Taleb, The Black Swan: The Impact of the Highly Improbable Fragility. New
York: Random House LLC, 2010.
[2]
R. MacLean, The six mistakes executives make in risk management, Courtesy of

Richard MacLean & Associates, LLC, 2010.
[3]
E. Jantsch, Technological Forecasting in Perspective. Paris: OECD, 1967, p. 401.
[4]
P. Saffo, Six rules for effective forecasting, Harv. Bus. Rev., vol. 85, no. 7/8, pp.
122131, 2007.
[5]
R. Lenz, Technological Forecasting, 2nd ed. Virginia: Defense Documentation

Center for Scientific and Technical Information, 1962, p. 107.
[6]
A. L. Porter, S. W. Cunningham, J. Banks, A. T. Roper, and T. W. Mason,

Forecasting and Management of Technology, 2nd ed., vol. 18. John Wiley & Sons,
2011, p. 336.
[7]
D.-J. Lim, S. R. Jahromi, T. R. Anderson, and A. A. Tudorie, Comparing

technological advancement of hybrid electric vehicles (HEV) in different market
segments, Technol. Forecast. Soc. Change (In press), 2015.
[8]
N. D. Kondratiev, The major economic cycles, Vopr. Konjunktury, vol. 1, no. 1,

pp. 2879, 1925.
[9]
J. A. Schumpeter, Business cycles, migration and health., Soc. Sci. Med., vol. 64,
no. 7, pp. 14204, 1939.
[10] R. N. Foster, Innovation: The Attackers Advantage. Summit Books, 1986, p. 316.
[11] J. M. Utterback, Mastering the Dynamics of Innovation: How Companies can
Seize Opportunities in the Face of Technological Change, vol. 81, no. 1. Harvard
Business School Press, 1994, pp. 138140.
[12] G. Yokote and R. A. Utterback, Time lapses in information dissemination:
research laboratory to physicians office, Bull. Med. Libr. Assoc., vol. 62, no. 3,
pp. 251257, 1974.
[13] A. Sood and G. J. Tellis, Technological evolution and radical innovation, J.
Mark., vol. 69, no. 3, pp. 152168, 2005.
131
[14] D. Sahal, The farm tractor and the nature of technological innovation, Res.
Policy, vol. 10, no. 4, pp. 368402, 1981.
[15] C. Christensen, The Innovators Dilemma, vol. 1, no. 1. Harvard Business School
Press, 1997, p. 225.
[16] E. Danneels, Disruptive technology reconsidered: a critique and research agenda,
J. Prod. Innov. Manag., vol. 21, no. 4, pp. 246258, Jul. 2004.
[17] D. Yu and C. C. Hang, A reflective review of disruptive innovation theory, Int. J.
Manag. Rev., vol. 12, no. 4, pp. 435452, Dec. 2010.
[18] G. J. Tellis, Disruptive technology or visionary leadership?, J. Prod. Innov.
Manag., vol. 23, no. 1, pp. 3438, Jan. 2006.
[19] P. S. Cohan, The dilemma of the Innovators Dilemma: Clayton Christensens
management theories are suddenly all the rage, but are they ripe for disruption?,
Industry Standard, vol. 10, Industry Standard, 2010.
[20] C. M. Christensen and M. Overdorf, Meeting the challenge of disruptive change,
Harv. Bus. Rev., vol. 78, no. 2, pp. 6676, 2000.
[21] J. R. Doyle and R. H. Green, Comparing products using data envelopment
analysis, Omega, vol. 19, no. 6, pp. 6318, 1991.
[22] L. M. Seiford and J. Zhu, Context-dependent data envelopment analysis
Measuring attractiveness and progress, Omega, vol. 31, no. 5, pp. 397408, 2003.
[23] R.-W. Po, Y.-Y. Guh, and M.-S. Yang, A new clustering approach using data
envelopment analysis, Eur. J. Oper. Res., vol. 199, no. 1, pp. 276284, 2009.
[24] A. Amirteimoori and S. Kordrostami, An alternative clustering approach: a DEAbased procedure, Optimization. pp. 114, 2011.
[25] X. Dai and T. Kuosmanen, Best-practice benchmarking using clustering methods:
Application to energy regulation, Omega, vol. 42, no. 1, pp. 179188, 2014.
[26] P. Bogetoft and L. Otto, Benchmarking with DEA, SFA, and R, 1st ed. New York:
Springer, 2010, p. 367.
[27] L. M. Seiford and J. Zhu, Infeasibility of super-efficiency data envelopment
analysis models, INFOR, vol. 37, no. 2, pp. 174188, 1999.
132
[28] T. Modis, Strengths and weaknesses of S-curves, Technol. Forecast. Soc.

Change, vol. 74, no. 6, pp. 866872, Jul. 2007.
[29] T. S. Kuhn, The Structure of Scientific Revolutions, vol. 57, no. 1. Chicago, IL:
University of Chicago Press, 1996. 3rd ed., 1996, pp. 5975.
[30] H. I. Weisberg, Bias and Causation: Models and Judgment for Valid Comparisons,
1st ed. Needham: Wiley, 2010, pp. 1348.
[31] J. M. Utterback, Radical innovation and corporate regeneration, Res. Manag.,
vol. 37, no. 4, p. 10, 1994.
[32] T. W. Danner, A formulation of multidimensional growth models for the
assessment and forecast of technology attributes, Georgia Institute of Technology,
2006.
[33] C. Christensen, Exploring the limits of the technology S-curve. Part I: Component
technologies, Prod. Oper. Manag., vol. 1, no. 4, pp. 334357, 2009.
[34] M. A. Schilling and M. Esmundo, Technology s-curves in renewable energy
alternatives: analysis and implications for industry and government, Energy
Policy, vol. 37, no. 5, pp. 17671781, 2009.
[35] C. M. Christensen, A history of the rigid disk drive industry: commercial and
technological turbulence, Bus. Hist., vol. 67, no. 4, pp. 531588, 1993.
[36] A. Debecker and T. Modis, Determination of the uncertainties in S-curve logistic
fits, Technol. Forecast. Soc. Change, vol. 46, no. 2, pp. 153173, 1994.
[37] P. Young, Technological growth curves: a competition of forecasting models,
Technol. Forecast. Soc. Change, vol. 44, no. 4, pp. 375389, 1993.
[38] J. P. Martino, Technological Forecasting for Decision Making, 2nd ed. NorthHolland: McGraw-Hill, 1993.
[39] R. Fildes, An evaluation of bayesian forecasting, J. Forecast., vol. 2, no. 2, pp.
137150, 1983.
[40] S. Makridakis, A. Anderson, and R. Carbone, The Forecasting Accuracy of Major
Time Series Methods. Chichester (Wiley), 1984, p. 301.
[41] P. Young and J. Keith Ord, The use of discounted least squares in technological
forecasting, Technol. Forecast. Soc. Change, vol. 28, no. 3, pp. 263274, 1985.
133
[42] R. L. Chadda and S. S. Chitgopekar, A generalization of the logistic curves and

long-range forecasts (1966-1991) of residence telephones, Bell J. Econ. Manag.
Sci., vol. 2, pp. 542560, 1971.
[43] S. Makridakis, S. C. Wheelwright, and V. E. McGee, Forecasting: Methods and
Applications, Second. New York: John Wiley & Sons, 1983, p. 923.
[44] E. Mansfield, Technological change and the rate of imitation, Econometrica, vol.
29, no. 4, pp. 747766, 1961.
[45] M. N. Sharif and M. N. Islam, The Weibull distribution as a general model for
forecasting technological change, Technol. Forecast. Soc. Change, vol. 18, no. 3,
pp. 247256, 1980.
[46] F. M. Bass, A new product growth for model consumer durables, Manage. Sci.,
vol. 15, no. 5, pp. 215227, 1969.
[47] C. Easingwood, V. Mahajan, and E. Muller, A nonsymmetric responding logistic
model for forecasting technological substitution, Technol. Forecast. Soc. Change,
vol. 20, no. 3, pp. 199213, 1981.
[48] A. C. Harvey, Time series forecasting, J. Oper. Res. Soc., vol. 35, pp. 641646,
1984.
[49] H. Levenbach and B. E. Reuter, Forecasting trending time series with relative
growth rate models, Technometrics, vol. 18, no. 3, pp. 261272, 1976.
[50] J. P. Martino, A review of selected recent advances in technological forecasting,
[51]
R. U. Ayres, Toward a non-linear dynamics of technological progress, J. Econ.

Behav. Organ., vol. 24, no. 1, pp. 3569, 1994.
[52] C. M. Christensen and J. L. Bower, Customer power, strategic investment, and

the failure of leading firms, Strateg. Manag. J., vol. 17, no. 3, pp. 197218, 1996.
[53] C. M. Christensen and M. E. Raynor, The Innovators Solution, vol. 25, no. 11.
Harvard Business School Press, 2003, p. 304.
[54] H. Chesbrough, Assembling the elephant: A review of empirical studies on the
impact of technical change upon incumbent firms, Comp. Stud. Technol. Evol.,
vol. 7, pp. 136, 2001.
134
[55] A. C. Cooper and D. Schendel, Strategic responses to technological threats, Bus.

Horiz., vol. 19, no. 1, pp. 6169, 1976.
[56] M. L. Tushman and P. Anderson, Technological discontinuities and
organizational environments, Adm. Sci. Q., vol. 31, no. 3, pp. 439465, 1986.
[57] R. M. Henderson and K. B. Clark, Architectural innovation: the reconfiguration
of existing product technologies and the failure of established firms, Adm. Sci. Q.,
vol. 35, no. 1, pp. 930, 1990.
[58] R. Henderson, Underinvestment and Incompetence as Responses to Radical
Innovation: Evidence from the Photolithographic Alignment Equipment Industry,
RAND J. Econ., vol. 24, no. 2, p. 248, 1993.
[59] H. Chesbrough, Arrested development: the experience of European hard disk
drive firms in comparison with US and Japanese firms, J. Evol. Econ., vol. 9, no.
3, pp. 287329, 1999.
[60] L. H. Lynn, From silicon valley to singapore: location and competitive advantage
in the hard disk drive industry, J. Eng. Technol. Manag., vol. 19, no. 2, pp. 217
219, 2002.
[61] M. Tripsas and G. Gavetti, Capabilities, cognition, and inertia: evidence from
digital imaging, Strateg. Manag. J., vol. 21, no. 1011, pp. 11471161, 2000.
[62] C. E. Helfat and M. B. Lieberman, The birth of capabilities: market entry and the
importance of pre-history, Ind. Corp. Chang., vol. 11, no. 4, pp. 725760, 2002.
[63] A. A. King and C. L. Tucci, Incumbent entry into new market niches: the role of
experience and managerial choice in the creation of dynamic capabilities,
Manage. Sci., vol. 48, no. 2, pp. 171186, 2002.
[64] S. Klepper and K. L. Simons, Dominance by birthright: entry of prior radio
producers and competitive ramifications in the U.S. television receiver industry,
Strateg. Manag. J., vol. 21, no. 1011, pp. 9971016, 2000.
[65] R. K. Chandy and G. J. Tellis, The incumbents curse? incumbency, size, and
radical product innovation, J. Mark., vol. 64, no. 3, pp. 117, 2000.
[66] M. Tripsas, Unraveling the process of creative destruction: complementary assets
and incumbent survival in the typesetter industry, Strateg. Manag. J., vol. 18, no.
S1, pp. 119142, 1997.
135
[67] F. T. Rothaermel, Incumbents advantage through exploiting complementary

assets via interfirm cooperation, Strateg. Manag. J., vol. 22, no. 67, pp. 687699,
2001.
[68] M. R. Darby, L. G. Zucker, and A. B. A. H. C. Robert, Change or die: the
adoption of biotechnology in the Japanese and U.S. pharmaceutical industries,
Comp. Stud. Technol. Evol., vol. 7, pp. 85125, 2001.
[69] G. M. Schmidt, Low-end and high-end encroachment strategies for new products,
Int. J. Innov. Manag., vol. 8, no. 2, pp. 167191, Nov. 2011.
[70] A. Martinelli, An emerging paradigm or just another trajectory? Understanding
the nature of technological changes using engineering heuristics in the
telecommunications switching industry, Res. Policy, vol. 41, no. 2, pp. 414429,
2012.
[71] S. Kassicieh and N. Rahal, A model for disruptive technology forecasting in
strategic regional economic development, Technol. Forecast. Soc. Change, vol.
74, no. 9, pp. 17181732, 2007.
[72] R. Phaal, E. OSullivan, M. Routley, S. Ford, and D. Probert, A framework for
mapping industrial emergence, Technol. Forecast. Soc. Change, vol. 78, no. 2, pp.
217230, 2011.
[73] A. Keller and S. Hsig, Ex ante identification of disruptive innovations in the
software industry applied to web applications: The case of Microsofts vs.
Google's office applications, Technol. Forecast. Soc. Change, vol. 76, no. 8, pp.
10441054, 2009.
[74] D. Barber-Toms and D. Consoli, Whatever works: uncertainty and
technological hybrids in medical innovation, Technol. Forecast. Soc. Change, vol.
79, no. 5, pp. 932948, 2012.
[75] S. Husig, C. Hipp, and M. Dowling, Analysing disruptive potential: the case of
wireless local area network and mobile communications network companies,
R&D Manag., vol. 35, no. 1, pp. 1735, 2005.
[76] K. Adamson, Calculating the price trajectory of adoption of fuel cell vehicles,
Int. J. Hydrogen Energy, vol. 30, no. 4, pp. 341350, 2005.
[77] R. Letchumanan and F. Kodama, Reconciling the conflict between the `pollutionhaven hypothesis and an emerging trajectory of international technology transfer,
Res. Policy, vol. 29, no. 1, pp. 5979, 2000.
136
[78] H. Koh and C. L. Magee, A functional approach for studying technological

progress: Application to information technology, Technol. Forecast. Soc. Change,
vol. 73, no. 9, pp. 10611083, 2006.
[79] J. Paap and R. Katz, Anticipating disruptive innovation, IEEE Eng. Manag. Rev.,
vol. 32, no. 4, pp. 1322, 2004.
[80] V. Govindarajan and P. K. Kopalle, The usefulness of measuring disruptiveness
of innovations ex post in making ex ante predictions, J. Prod. Innov. Manag., vol.
23, no. 1, pp. 1218, 2006.
[81]
D. S. Doering and P. Roch, Identification and assessment of emerging

technologies, in Wharton on Managing Emerging Technologies, New York:
Wiley, 2000, p. 480.
[82] S. T. Walsh, Roadmapping a disruptive technology: A case study, Technol.

Forecast. Soc. Change, vol. 71, no. 12, pp. 161185, 2004.
[83] D. Padgett and M. S. Mulvey, Differentiation Via Technology: Strategic
Positioning of Services Following the Introduction of Disruptive Technology, J.
Retail., vol. 83, no. 4, pp. 375391, 2007.
[84] X. Huang and G. Soi, Analysis of industry equilibria in models with sustaining
and disruptive technology, Eur. J. Oper. Res., vol. 207, no. 1, pp. 238248, 2010.
[85] D. C. Kaslow, A potential disruptive technology in vaccine development: genebased vaccines and their application to infectious diseases., Trans. R. Soc. Trop.
Med. Hyg., vol. 98, no. 10, pp. 593601, 2004.
[86] B. Rao, B. Angelov, and O. Nov, Fusion of disruptive technologies:: lessons from
the Skype case, Eur. Manag. J., vol. 24, no. 23, pp. 174188, 2006.
[87] W. G. Bradley, MR-guided focused ultrasound: a potentially disruptive
technology., J. Am. Coll. Radiol., vol. 6, no. 7, pp. 5103, 2009.
[88] H. C. Lucas and J. M. Goh, Disruptive technology: How Kodak missed the digital
photography revolution, J. Strateg. Inf. Syst., vol. 18, no. 1, pp. 4655, 2009.
[89] F. Madjdi and S. Hsig, The response strategies of incumbent mobile network
operators on the disruptive potential of public W-LAN in Germany, Telecomm.
Policy, vol. 35, no. 6, pp. 555567, 2011.
[90] S. T. Walsh, R. L. Boylan, C. McDermott, and A. Paulson, The semiconductor
silicon industry roadmap: Epochs driven by the dynamics between disruptive
137
technologies and core competencies, Technol. Forecast. Soc. Change, vol. 72, no.
2, pp. 213236, 2005.
[91] P. N. Figueiredo, Discontinuous innovation capability accumulation in latecomer
natural resource-processing firms, Technol. Forecast. Soc. Change, vol. 77, no. 7,
pp. 10901108, 2010.
[92] J. P. Caulkins, G. Feichtinger, D. Grass, R. F. Hartl, and P. M. Kort, Two state
capital accumulation with heterogenous products: Disruptive vs. non-disruptive
goods, J. Econ. Dyn. Control, vol. 35, no. 4, pp. 462478, 2011.
[93] M.-C. Belis-Bergouignan, V. Oltra, and M. Saint Jean, Trajectories towards clean
technology: example of volatile organic compound emission reductions, Ecol.
Econ., vol. 48, no. 2, pp. 201220, 2004.
[94] J. C. Ho, Technology evaluation in Taiwans technology industries: Strategies,
trajectories, and innovations, Technol. Forecast. Soc. Change, vol. 78, no. 8, pp.
13791388, 2011.
[95] S. H. Werfel and A. B. Jaffe, Induced innovation and technology trajectory:
Evidence from smoking cessation products, Res. Policy, vol. 42, no. 1, pp. 1522,
2012.
[96] H. J. No and Y. Park, Trajectory patterns of technology fusion: Trend analysis
and taxonomical grouping in nanobiotechnology, Technol. Forecast. Soc. Change,
vol. 77, no. 1, pp. 6375, 2010.
[97] Y. E. Spanos and I. Voudouris, Antecedents and trajectories of AMT adoption:
The case of Greek manufacturing SMEs, Res. Policy, vol. 38, no. 1, pp. 144155,
2009.
[98] K. Frenken and L. Leydesdorff, Scaling trajectories in civil aircraft (19131997),
Res. Policy, vol. 29, no. 3, pp. 331348, 2000.
[99] C. Watanabe, S. Lei, and N. Ouchi, Fusing indigenous technology development
and market learning for greater functionality developmentAn empirical analysis
of the growth trajectory of Canon printers, Technovation, vol. 29, no. 4, pp. 265
283, 2009.
[100] M. Hobo, C. Watanabe, and C. Chen, Double spiral trajectory between retail,
manufacturing and customers leads a way to service oriented manufacturing,
Technovation, vol. 26, no. 7, pp. 873890, 2006.
138
[101] C. Watanabe, J. Y. Hur, and K. Matsumoto, Technological diversification and

firms techno-economic structure: An assessment of Canon's sustainable growth
trajectory, Technol. Forecast. Soc. Change, vol. 72, no. 1, pp. 1127, 2005.
[102] S.-H. Chen, M.-H. Huang, D.-Z. Chen, and S.-Z. Lin, Detecting the temporal
gaps of technology fronts: A case study of smart grid field, Technol. Forecast.
Soc. Change, vol. 79, no. 9, pp. 17051719, 2012.
[103] M. Epicoco, Knowledge patterns and sources of leadership: Mapping the
semiconductor miniaturization trajectory, Res. Policy, vol. 42, no. 1, pp. 180195,
2012.
[104] J. L. Funk, The future of the mobile phone Internet: an analysis of technological
trajectories and lead users in the Japanese market, Technol. Soc., vol. 27, no. 1, pp.
6983, 2005.
[105] R. P. Raven, Towards alternative trajectories? Reconfigurations in the Dutch
electricity regime, Res. Policy, vol. 35, no. 4, pp. 581595, 2006.
[106] F. Castellacci, Technological paradigms, regimes and trajectories: Manufacturing
and service industries in a new taxonomy of sectoral patterns of innovation, Res.
Policy, vol. 37, no. 67, pp. 978994, 2008.
[107] D. E. Kash and R. W. Rycoft, Patterns of innovating complex technologies: a
framework for adaptive network strategies, Res. Policy, vol. 29, no. 78, pp. 819
831, 2000.
[108] P. Arqu-Castells, How venture capitalists spur invention in Spain: Evidence
from patent trajectories, Res. Policy, vol. 41, no. 5, pp. 897912, 2012.
[109] W.-J. Kim, J.-D. Lee, and T.-Y. Kim, Demand forecasting for multigenerational
products combining discrete choice and dynamics of diffusion under technological
trajectories, Technol. Forecast. Soc. Change, vol. 72, no. 7, pp. 825849, 2005.
[110] C.-Y. Lee, J.-D. Lee, and Y. Kim, Demand forecasting for new technology with a
short history in a competitive environment: the case of the home networking
market in South Korea, Technol. Forecast. Soc. Change, vol. 75, no. 1, pp. 91
106, 2008.
[111] V. Krishnan and K. T. Ulrich, Product development decisions: a review of the
literature, Management Science, vol. 47, no. 1. pp. 121, 2001.
139
[112] M. V. Mansfield and Wagner. K., Organizational and strategic factors associated
with probabilities of success in industrial R&D, J. Bus., vol. 48, pp. 179198,
1975.
[113] G. Bacon, S. Beckman, D. C. Mowery, and E. Wilson, Managing product
definition in high-technology industries: a pilot study, Calif. Manage. Rev., vol.
36, no. 2, pp. 3256, 1994.
[114] V. Srinivasan, W. S. Lovejoy, and D. Beach, Integrated product design for
marketability and manufacturing, J. Mark. Res., vol. 34, no. 1, pp. 154163, 1997.
[115] E. Dahan and V. Srinivasan, Predictive power of Internet-based product concept
testing using visual depiction and animation, J. Prod. Innov. Manag., vol. 17, no.
2, pp. 99109, 2000.
[116] A. Schulze and M. Hoegl, Organizational knowledge creation and the generation
of new product ideas: A behavioral approach, Res. Policy, vol. 37, no. 10, pp.
17421750, 2008.
[117] J. Kim and D. Wilemon, Focusing the fuzzy front-end in new product
development, R&D Manag., vol. 32, no. 4, pp. 269279, 2002.
[118] M. Engwall and A. Jerbrant, The resource allocation syndrome: the prime
challenge of multi-project management?, Int. J. Proj. Manag., vol. 21, no. 6, pp.
403409, Aug. 2003.
[119] G. R. Schirr, Flawed tools: the efficacy of group research methods to generate
customer ideas, in Journal of Product Innovation Management, 2012, vol. 29, no.
3, pp. 473488.
[120] N. Dalkey and O. Helmer, An Experimental Application of the DELPHI Method
to the Use of Experts, Manage. Sci., vol. 9, no. 3, pp. 458467, 1963.
[121] F. Zwicky, Morphology of propulsive power, Society for Morphological
Research, California, pp. 1382, 1962.
[122] E. von Hippel, Lead users: a source of novel product concepts, Management
Science, vol. 32, no. 7. pp. 791805, 1986.
[123] A. Griffin and J. Hauser, The voice of the customer, Mark. Sci., vol. 12, no. 1,
pp. 127, 1993.
140
[124] G. S. Lynn, J. G. Morone, and A. S. Paulson, Marketing and discontinuous

innovation: the probe and learn process, Calif. Manage. Rev., vol. 38, no. 3, pp.
837, 1996.
[125] D. Leonard and J. F. Rayport, Spark innovation through empathic design., Harv.
Bus. Rev., vol. 75, no. 6, pp. 102113, 1997.
[126] A. J. M. Jetter, Educating the guess: strategies, concepts and tools for the fuzzy
front end of product development, in PICMET-Technology Management for
Reshaping the World, 2003, pp. 261273.
[127] J. Surowiecki, The Wisdom of Crowds. New York: Anchor, 2005, p. 336.
[128] R. Sethi, S. Pant, and A. Sethi, Web-based product development systems
integration and new product outcomes: A conceptual framework, J. Prod. Innov.
Manag., vol. 20, no. 1, pp. 3756, 2003.
[129] M. Fisher, K. Ramdas, and K. Ulrich, Component sharing in the management of
product variety: a study of automotive braking systems, Manage. Sci., vol. 45, no.
3, pp. 297315, 1999.
[130] S. L. Brown and K. M. Eisenhardt, Product development: past research, present
findings, and future directions, Acad. Manag. Rev., vol. 20, no. 2, pp. 343378,
1995.
[131] S. Ahmad, D. N. Mallick, and R. G. Schroeder, New product development:
Impact of project characteristics and development practices on performance, J.
Prod. Innov. Manag., vol. 30, no. 2, pp. 331348, 2013.
[132] C. Moorman and A. S. Miner, The impact of organizational memory on new
product performance and creativity, J. Mark. Res., vol. 34, no. 1, pp. 91106,
1997.
[133] L. Cassia, A. De Massis, and E. Pizzurno, Strategic innovation and new product
development in family firms: An empirically grounded theoretical framework, Int.
J. Entrep. Behav. Res., vol. 18, no. 2, pp. 198232, 2012.
[134] E. Hultink, Industrial new product launch strategies and product development
performance, J. Prod. Innov. Manag., vol. 14, no. 4, pp. 243257, 1997.
[135] Y. Yin, S. Qin, and R. Holland, Development of a design performance
measurement matrix for improving collaborative design during a design process,
Int. J. Product. Perform. Manag., vol. 60, no. 2, pp. 152184, 2011.
141
[136] M. Perkmann, A. Neely, and K. Walsh, How should firms evaluate success in
university-industry alliances? A performance measurement system, R&D Manag.,
vol. 41, no. 2, pp. 202216, 2011.
[137] M. D. Morelli, S. D. Eppinger, and R. K. Gulati, Predicting technical
communication in product development organizations, IEEE Trans. Eng. Manag.,
vol. 42, no. 3, pp. 215222, 1995.
[138] S. B. Modi and V. A. Mabert, Supplier development: Improving supplier
performance through knowledge transfer, J. Oper. Manag., vol. 25, no. 1, pp. 42
64, 2007.
[139] R. J. Slotegraaf and K. Atuahene-Gima, Product development team stability and
new product advantage: the role of decision-making processes, J. Mark., vol. 75,
no. 1, pp. 96108, 2011.
[140] D. Robertson and T. J. Allen, CAD system use and engineering performance,
IEEE Trans. Eng. Manag., vol. 40, no. 3, pp. 274282, 1993.
[141] M. Muethel, F. Siebdrat, and M. Hoegl, When do we really need interpersonal
trust in globally dispersed new product development teams?, R&D Manag., vol.
42, no. 1, pp. 3146, 2012.
[142] J. E. Burroughs, D. W. Dahl, C. P. Moreau, A. Chattopadhyay, and G. J. Gorn,
Facilitating and rewarding creativity during new product development, J. Mark.,
vol. 75, no. 4, pp. 5367, 2011.
[143] N. Ale Ebrahim, S. Ahmed, and Z. Taha, Modified stage-gate: a conceptual
model of virtual product development process, African J. Mark. Manag. Vol 1 No
9 pp 211219 December 2009, vol. 1, no. 9, pp. 211219, 2009.
[144] K. B. Kahn, G. Barczak, J. Nicholas, A. Ledwith, and H. Perks, An examination
of new product development best practice, J. Prod. Innov. Manag., vol. 29, no. 2,
pp. 180192, 2012.
[145] R. G. Cooper and E. J. Kleinschmidt, Winning Businesses in Product
Devhopmbiit: the Critical Success Factors, Res. Technol. Manag., vol. 50, no. 3,
pp. 5266, 2007.
[146] A. C. Edmondson and I. M. Nembhard, Product development and learning in
project teams: The challenges are the benefits, J. Prod. Innov. Manag., vol. 26, no.
2, pp. 123138, 2009.
142
[147] R. T. Keller, Transformational leadership, initiating structure, and substitutes for

leadership: a longitudinal study of research and development project team
performance., J. Appl. Psychol., vol. 91, no. 1, pp. 202210, 2006.
[148] R. Pike and S. Ho, Risk analysis in capital budgeting: barriers and benefits,
Omega, vol. 19, no. 4, pp. 235245, 1991.
[149] N. . Archer and F. Ghasemzadeh, An integrated framework for project portfolio
selection, Int. J. Proj. Manag., vol. 17, no. 4, pp. 207216, 1999.
[150] N. R. Joglekar and D. N. Ford, Product development resource allocation with
foresight, Eur. J. Oper. Res., vol. 160, no. 1, pp. 7287, 2005.
[151] K. J. Petersen, R. B. Handfield, and G. L. Ragatz, Supplier integration into new
product development: Coordinating product, process and supply chain design, J.
Oper. Manag., vol. 23, no. 34, pp. 371388, 2005.
[152] J. Jiao, T. W. Simpson, Z. Siddique, and J. Jiao, Product family design and
platform-based product development: a state-of-the-art review, J. Intell. Manuf.,
vol. 18, no. 1, pp. 529, 2007.
[153] X. A. Koufteros, T. C. Edwin Cheng, and K. H. Lai, Black-box and gray-box
supplier integration in product development: Antecedents, consequences and the
moderating role of firm size, J. Oper. Manag., vol. 25, no. 4, pp. 847870, 2007.
[154] D. B. Hertz and H. Thomas, Risk analysis and its applications. New York: Wiley,
1983, p. 340.
[155] D. H. Stamatis, Failure mode and effect analysis: FMEA from theory to execution,
vol. 38. 2003, p. 80.
[156] J. A. Keizer, J. P. Vos, and J. I. M. Halman, Risks in new product development:
Devising a reference tool, R&D Manag., vol. 35, no. 3, pp. 297309, 2005.
[157] T. Markeset and U. Kumar, Integration of RAMS information in design
processes-a case study, in Annual Reliability and Maintainability Symposium,
2003, pp. 220225.
[158] P. E. Green and V. Srinivasan, Conjoint analysis in consumer research: issues and
outlook, J. Consum. Res., vol. 5, no. 2, p. 103, 1978.
[159] G. R. Amin, A. Emrouznejad, and S. Rezaei, Some clarifications on the DEA
clustering approach, Eur. J. Oper. Res., vol. 215, no. 2, pp. 498501, 2011.
143
[160] G. Rowe and G. Wright, The Delphi technique as a forecasting tool: issues and
analysis, Int. J. Forecast., vol. 15, no. 4, pp. 353375, 1999.
[161] M. T. A. Wentholt, S. Cardoen, H. Imberechts, X. Van Huffel, B. W. Ooms, and L.
J. Frewer, Defining European preparedness and research needs regarding
emerging infectious animal diseases: Results from a Delphi expert consultation,
Prev. Vet. Med., vol. 103, no. 23, pp. 8192, 2012.
[162] J. Pearce, C. Jones, S. Morrison, M. Olff, S. van Buschbach, A. B. Witteveen, R.
Williams, F. Orengo-Garca, D. Ajdukovic, A. T. Aker, D. Nordanger, B. LuegerSchuster, and J. I. Bisson, Using a delphi process to develop an effective trainthe-trainers program to train health and social care professionals throughout
Europe, J. Trauma. Stress, vol. 25, no. 3, pp. 337343, 2012.
[163] J. Landeta and J. Barrutia, People consultation to construct the future: A Delphi
application, Int. J. Forecast., vol. 27, no. 1, pp. 134151, 2011.
[164] O. Lindqvist, G. Lundquist, A. Dickman, J. Bkki, U. Lunder, C. L. Hagelin, B. H.
Rasmussen, S. Sauter, C. Tishelman, and C. J. Frst, Four essential drugs needed
for quality care of the dying: a Delphi-study based international expert consensus
opinion., J. Palliat. Med., vol. 16, no. 1, pp. 3843, 2013.
[165] H. Kahn, On alternative world futures: issues and themes, New York, 1965.
[166] M. Nowack, J. Endrikat, and E. Guenther, Review of Delphi-based scenario
studies: Quality and design considerations, Technol. Forecast. Soc. Change, vol.
78, no. 9, pp. 16031615, 2011.
[167] J. J. Winebrake and B. P. Creswick, The future of hydrogen fueling systems for
transportation: An application of perspective-based scenario analysis using the
analytic hierarchy process, Technol. Forecast. Soc. Change, vol. 70, no. 4, pp.
359384, 2003.
[168] K. Kok, M. van Vliet Mathijs, I. Brlund Ilona, A. Dubel, and J. Sendzimir,
Combining participative backcasting and exploratory scenario development:
Experiences from the SCENES project, Technol. Forecast. Soc. Change, vol. 78,
no. 5, pp. 835851, 2011.
[169] A. Jetter and W. Schweinfort, Building scenarios with Fuzzy Cognitive Maps: An
exploratory study of solar energy, Futures, vol. 43, no. 1, pp. 5266, 2011.
[170] T. A. Tran and T. Daim, A taxonomic review of methods and tools applied in
technology assessment, Technol. Forecast. Soc. Change, vol. 75, no. 9, pp. 1396
1405, 2008.
144
[171] I. Linkov, F. K. Satterstrom, G. Kiker, C. Batchelor, T. Bridges, and E. Ferguson,

From comparative risk assessment to multi-criteria decision analysis and adaptive
management: recent developments and applications., Environ. Int., vol. 32, no. 8,
pp. 107293, 2006.
[172] C. S. Kolar and D. M. Lodge, Ecological predictions and risk assessment for alien
fishes in North America, Science, vol. 298, no. 5596, pp. 12331236, 2002.
[173] P. Morris and R. Therivel, Methods of Environmental Impact Assessment, vol. 2,
no. 3. Spon Press, 2001, p. 378.
[174] R. Ramanathan, A note on the use of the analytic hierarchy process for
environmental impact assessment., J. Environ. Manage., vol. 63, no. 1, pp. 2735,
2001.
[175] H. M. G. Van Der Werf and J. Petit, Evaluation of the environmental impact of
agriculture at the farm level: A comparison and analysis of 12 indicator-based
methods, Agric. Ecosyst. Environ., vol. 93, no. 13, pp. 131145, 2002.
[176] C. W. Churchman, R. L. Ackoff, and E. L. Arnoff, Introduction to Operations
Research. New York: John Wiley & Sons, 1957.
[177] D. Ghiculescu, N. Marinescu, D. Ghiculescu, and C. Pirnu, Customer matrix
with different time horizons applied to electrodischarge machines, Appl. Mech.
Mater., vol. 371, pp. 220224, 2013.
[178] G. Manuel and J. Pretorius, The significance of relevance trees in the
identification of artificial neural networks input vectors, J. Energy South. Africa,
vol. 24, no. 1, pp. 2734, 2013.
[179] X. Feng and L. Fuhai, Patent text mining and informetric-based patent technology
morphological analysis: an empirical study, Technol. Anal. Strateg. Manag., vol.
24, no. 5, pp. 467479, 2012.
[180] S. Jun, S. S. Park, and D. S. Jang, Technology forecasting using matrix map and
patent clustering, Ind. Manag. Data Syst., vol. 112, no. 5, pp. 786807, 2012.
[181] B. Yoon, I. Park, and B. youl Coh, Exploring technological opportunities by
linking technology and products: Application of morphology analysis and text
mining, Technol. Forecast. Soc. Change, vol. 86, pp. 287303, 2014.
[182] B. Mazlish, The railroad and the space program - An exploration in historical
analogy. Cambridge: The MIT press, 1965, pp. 1223.
145
[183] P. Goodwin, K. Dyussekeneva, and S. Meeran, The use of analogies in

forecasting the annual sales of new electronics products, IMA J. Manag. Math.,
vol. 24, no. 4, pp. 407422, 2012.
[184] S. C. Bhargava, A generalized form of the fisher-pry model of technological
substitution, Technol. Forecast. Soc. Change, vol. 49, no. 1, pp. 2733, 1995.
[185] K. Christodoulou, K. Jensen, and K. Vlahos, Using object-oriented simulation to
explore substitution between technologies: An application to the UK mobile
telecommunications industry, Technol. Forecast. Soc. Change, vol. 62, no. 3, pp.
203217, 1999.
[186] M. Dekimpe, P. Parker, and M. Sarvary, Globalization: modeling technology
adoption timing across countries, Technol. Forecast. Soc. Change, vol. 63, no. 1,
pp. 2542, 2000.
[187] J. Goldenberg and S. Efroni, Using cellular automata modeling of the emergence
of innovations, Technol. Forecast. Soc. Change, vol. 68, no. 3, pp. 293308, 2001.
[188] T. Islam and N. Meade, The diffusion of successive generations of a technology:
A more general model, Technol. Forecast. Soc. Change, vol. 56, no. 1, pp. 4960,
1997.
[189] D. M. Kiefer, Winds of change in industrial chemical research, Chem. Eng.
News, vol. 42, no. 23, pp. 88109, 1964.
[190] W. H. Pound, Research project selection: Testing a model in the field, IEEE
Trans. Eng. Manag., vol. EM-11, no. 1, pp. 1622, 1964.
[191] R. E. Beckwith, A stochastic regression model for proposal success evaluation,
IEEE Trans. Eng. Manag., vol. EM-12, no. 2, pp. 5962, 1965.
[192] J. Ondrus and Y. Pigneur, An assessment of NFC for future mobile payment
systems, in International Conference on the Management of Mobile Business
ICMB 2007, 2007, pp. 4343.
[193] T. U. Daim, E. Bayraktaroglu, J. Estep, D. J. Lim, J. Upadhyay, and J. Yang,
Optimizing the NW off-shore wind turbine design, Math. Comput. Model., vol.
55, no. 3, pp. 396404, 2011.
[194] M. T. Tang and G.-H. Tzeng, A hierarchy fuzzy MCDM method for studying
electronic marketing strategies in the information service industry, in
International Academy for Information Management Annual Conference, 1998, no.
1, pp. 116.
146
[195] B. V. Dean and S. S. Sengupta, Research budgeting and project selection, IRE
Trans. Eng. Manag., vol. 4, pp. 158169, 1962.
[196] K. Moslehi and R. Kumar, A reliability perspective of the smart grid, IEEE
Trans. Smart Grid, vol. 1, no. 1, pp. 5764, 2010.
[197] M. Friedewald and O. Raabe, Ubiquitous computing: An overview of technology
impacts, Telemat. Informatics, vol. 28, no. 2, pp. 5565, 2011.
[198] B. J. W. Forrester, Industrial dynamics. A major breakthrough for decision
makers, Harv. Bus. Rev., vol. 36, no. 4, pp. 3766, 1958.
[199] T. U. Daim, G. Rueda, H. Martin, and P. Gerdsri, Forecasting emerging
technologies: Use of bibliometrics and patent analysis, Technol. Forecast. Soc.
Change, vol. 73, no. 8, pp. 9811012, 2006.
[200] F. H. Maier, New product diffusion models in innovation managementa system
dynamics perspective, Syst. Dyn. Rev., vol. 14, pp. 285308, 1998.
[201] E. Suryani, S. Y. Chou, and C. H. Chen, Air passenger demand forecasting and
passenger terminal capacity expansion: A system dynamics framework, Expert
Syst. Appl., vol. 37, no. 3, pp. 23242339, 2010.
[202] J. Martino, A comparison of two composite measures of technology, Technol.
[203] N. Gerdsri, An Analytical Approach on Building a Technology Development
Envelope (TDE) for Roadmapping of Emerging Technologies, Portland State
University, 2005.
[204] A. J. Alexander and J. R. Nelson, Measuring technological change: Aircraft
turbine engines, Technol. Forecast. Soc. Change, vol. 5, no. 2, pp. 189203, 1973.
[205] D.-J. Lim, T. R. Anderson, and J. Kim, Forecast of wireless communication
technology: A comparative study of regression and TFDEA Model, in PICMETTechnology Management for Emerging Technologies, 2012, pp. 12471253.
[206] G. C. Taylor, Methodologies for characterizing technologies, Electric Power
Research Institute, Palo Alto, p. 120, 1978.
[207] R. E. Klitgaard, Measuring technological change: Comments on a proposed
methodology, Technol. Forecast. Soc. Change, vol. 6, pp. 437440, 1974.
147
[208] D. J. Aigner and S. F. Chu, On estimating the industry production function, Am.
Econ. Rev., vol. 58, no. 4, pp. 826839, 1968.
[209] H. O. Fried, C. A. K. Lovell, and S. S. Schmidt, The Measurement of Productive
Efficiency and Productivity Growth, 1st ed. New York: Oxford University Press,
2008, p. 638.
[210] D. Aigner, C. A. K. Lovell, and P. Schmidt, Formulation and estimation of
stochastic frontier production function models, J. Econom., vol. 6, no. 1, pp. 21
37, 1977.
[211] E. N. Dodson, Measurement of state of the art and technological advance,
[212] E. N. Dodson, A general approach to measurement of the state of the art and
technological advance, Technol. Forecast. Soc. Change, vol. 1, no. 4, pp. 391
408, 1970.
[213] J. P. Martino, Measurement of technology using tradeoff surfaces, Technol.
[214] B. F. Cole, An Evolutionary Method for Synthesizing Technological Planning
and Architectural Advance, Georgia Institute of Technology, 2009.
[215] M. J. Farrell, The measurement of productive efficiency, J. R. Stat. Soc. Ser. A,
vol. 120, no. 3, pp. 253290, 1957.
[216] A. Charnes, W. W. Cooper, and E. Rhodes, Measuring the efficiency of decision
making units, Eur. J. Oper. Res., vol. 2, no. 6, pp. 429444, 1978.
[217] R. Fare, S. Grosskopf, and C. A. K. Lovell, Production Frontiers. Cambridge
University Press, 1994, p. 316.
[218] A. Charnes, C. T. Clark, W. W. Cooper, and B. Golany, A developmental study
of data envelopment analysis in measuring the efficiency of maintenance units in
the U.S. air forces, Annals of Operations Research, vol. 2, no. 1. pp. 95112,
1984.
[219] T. Sueyoshi, Comparisons and analyses of managerial efficiency and returns to
scale of telecommunication enterprises by using DEA/WINDOW, Commun. Oper.
Res. Soc. Japan, vol. 37, pp. 210219, 1992.
148
[220] W. W. Cooper, L. M. Seiford, and K. Tone, Introduction to Data Envelopment

Analysis and Its Uses: with DEA-solver Software and References. Springer New
York, 2006, p. 354.
[221] A. Emrouznejad and E. Thanassoulis, Measurement of productivity index with
dynamic DEA, Int. J. Oper. Res., vol. 8, no. 2, pp. 247260, 2010.
[222] S. Malmquist, Index numbers and indifference surfaces, Trab. Estad., vol. 4, no.
2, pp. 209242, 1953.
[223] O. L. Inman, Technology Forecasting using Data Envelopment Analysis,
Portland State University, 2004.
[224] J. O. Kim, J. H. Kim, and S. K. Kim, A Hybrid Technological Forecasting Model
by Identifying the Efficient DMUs: An Application to the Main Battle Tank,
Asian J. Technol. Innov., vol. 15, no. 2, pp. 83102, 2007.
[225] A.-M. Lamb, T. R. Anderson, and T. Daim, Research and development targetsetting difficulties addressed through the emergent method: technology forecasting
using data envelopment analysis, R&D Manag., vol. 42, no. 4, pp. 327341, 2012.
[226] K. C. Land, C. A. K. Lovell, and S. Thore, Chance-constrained data envelopment
analysis, Manag. Decis. Econ., vol. 14, no. 6, pp. 541554, 1993.
[227] P. L. Brockett, A. Charnes, W. W. Cooper, Z. M. Huang, and D. B. Sun, Data
transformations in DEA cone ratio envelopment approaches for monitoring bank
performances, Eur. J. Oper. Res., vol. 98, no. 2, pp. 250268, 1997.
[228] A. Charnes and W. W. Cooper, Chance-constrained programming, Manage. Sci.,
vol. 6, no. 1, pp. 7379, 1959.
[229] J. K. Sengupta, Transformations in stochastic DEA models, J. Econom., vol. 46,
no. 1, pp. 109123, 1990.
[230] O. Inman, T. R. Anderson, and R. Harmon, Predicting U.S. jet fighter aircraft
introductions from 1944 to 1982: A dogfight between regression and TFDEA,
[231] T. R. Anderson, R. Fare, S. Grosskopf, L. Inman, and X. Song, Further
examination of Moores law with data envelopment analysis, Technol. Forecast.
Soc. Change, vol. 69, no. 5, pp. 465477, 2002.
149
[232] D.-J. Lim, N. Runde, and T. R. Anderson, Applying technology forecasting to

new product development target setting of LCD panels, in Advances in Business
and Management Forecasting, 9th ed., Emerald, 2013, p. 340.
[233] W. W. Cooper, L. M. Seiford, and K. Tone, Data Envelopment Analysis: A
Comprehensive Text with Models, Applications, References and DEA-solver
Software. New York: Springer, 2007, p. 512.
[234] R. D. Banker, A. Charnes, and W. W. Cooper, Some models for estimating
technical and scale inefficiencies in data envelopment analysis, Management, vol.
30, no. 9, pp. 10781092, 1984.
[235] C. A. K. Lovell and A. P. B. Rouse, Equivalent standard DEA models to provide
super-efficiency scores, J. Oper. Res. Soc., vol. 54, no. 1, pp. 101108, 2003.
[236] W. D. Cook, L. Liang, Y. Zha, and J. Zhu, A modified super-efficiency DEA
model for infeasibility, J. Oper. Res. Soc., vol. 60, no. 2, pp. 276281, 2009.
[237] H.-S. Lee, C.-W. Chu, and J. Zhu, Super-efficiency DEA in the presence of
infeasibility, Eur. J. Oper. Res., vol. 212, no. 1, pp. 141147, 2011.
[238] H.-S. Lee and J. Zhu, Super-efficiency infeasibility and zero data in DEA, Eur. J.
Oper. Res., vol. 216, no. 2, pp. 429433, 2012.
[239] P. Andersen and N. C. Petersen, A procedure for ranking efficient units in data
envelopment analysis, Manage. Sci., vol. 39, no. 10, pp. 12611264, 1993.
[240] D.-J. Lim, T. R. Anderson, and O. L. Inman, Choosing effective dates from
multiple optima in Technology Forecasting using Data Envelopment Analysis
(TFDEA), Technol. Forecast. Soc. Change, vol. 88, pp. 9197, 2014.
[241] A. A. Ali and L. M. Seiford, Computational accuracy and infinitesimals in data
envelopment analysis, INFOR, vol. 31, no. 4, pp. 290297, 1993.
[242] A.-M. Lamb, T. U. Daim, and T. R. Anderson, Forecasting airplane technologies,
Foresight, vol. 12, no. 6. pp. 3854, 2010.
[243] D. A. Irwin and N. Pavcnik, Airbus versus Boeing revisited: international
competition in the aircraft market, J. Int. Econ., vol. 64, no. 2, pp. 223245, 2004.
[244] Anonymous, Technical charateristics - Boeing 747-400, Boeing, 2013.
[245] D. L. Greene, Energy-efficiency improvement potential of commercial aircraft,
Annu. Rev. Energy Environ., vol. 17, no. 1, pp. 537573, 1992.
150
[246] D. Briere and P. Traverse, AIRBUS A320/A330/A340 electrical flight controls - A

family of fault-tolerant systems. IEEE Comput. Soc. Press, 1993, pp. 616623.
[247] Anonymous, 777-200LR/-300ER/-freighter airplane characteristics for airport
planning, Boeing, Boeing, pp. 1166, 2009.
[248] J. Michaels, The Airbus A380: The giant on the runway, The Economist, vol.
385, no. 8550, pp. 8183, 2007.
[249] M. Berrittella, L. La Franca, V. Mandina, and P. Zito, Modelling strategic
alliances in the wide-body long-range aircraft market, J. Air Transp. Manag., vol.
13, no. 3, pp. 139148, 2007.
[250] J. C. Williams and E. A. Starke, Progress in structural materials for aerospace
systems, Acta Mater., vol. 51, no. 19, pp. 57755799, 2003.
[251] D. A. Senzig, G. G. Fleming, and R. J. Iovinelli, Modeling of terminal-area
airplane fuel consumption, J. Aircr., vol. 46, no. 4, pp. 10891093, 2009.
[252] Anonymous, Technical characteristics - Boeing 747-8, Boeing, 2013. [Online].
Available: http://www.boeing.com/boeing/commercial/747family/7478_fact_sheet.page.
[253] W. Steve, Reports hint at what may be behind Boeing 747-8 delay, Puget Sound
Business Journal, 2011. [Online]. Available:
http://www.bizjournals.com/seattle/news/2011/09/16/boeing-monday-747-8freighter-delivery.html?page=2.
[254] S. Mary and R. Susanna, Boeing affirms delivery target for first 787-9
Dreamliner, Bloomberg, 2012. [Online]. Available:
http://www.bloomberg.com/news/articles/2012-10-17/american-says-first-boeing787-9-dreamliner-delayed-beyond-2014.
[255] M. Harriet, Boeing delays decision on 787-10 production site, Reuters, 2014.
[Online]. Available: http://www.reuters.com/article/2014/03/10/us-boeingsouthcarolina-idUSBREA291LA20140310.
[256] G. Dominic, Simmering Boeing strike scorching both sides, The Seattle Times,
2008. [Online]. Available: http://www.seattletimes.com/business/boeingaerospace/simmering-boeing-strike-scorching-both-sides.
[257] R. Andrea and B. Thomas, Airbus blinking with A350 helps Boeing plot 777
successor, Bloomberg, 2012. [Online]. Available:
151
http://www.bloomberg.com/news/articles/2012-05-29/airbus-blinking-first-witha350-helps-boeing-plot-777-successor.
[258] H. Scott, Smaller seats, fee rises and new planes? 2014: The year ahead in air
travel, CNN, 2014. [Online]. Available:
http://www.cnn.com/2014/01/01/travel/2014-aviation-year/.
[259] K. Max, PICTURE: 10-abreast A350 XWB would offer unprecedented operating
cost advantage, Flight, 2008. [Online]. Available:
http://www.flightglobal.com/news/articles/picture-10-abreast-a350-xwb-39wouldoffer-unprecedented-operating-cost-223853/
[260] Anonymous, A350XWB family A350-1000, Airbus, 2014. [Online]. Available:
http://www.airbus.com/aircraftfamilies/passengeraircraft/a350xwbfamily/a3501000/.
[261] L. J. Tashman, Out-of-sample tests of forecasting accuracy: an analysis and
review, Int. J. Forecast., vol. 16, no. 4, pp. 437450, 2000.
[262] S. Makridakis, S. C. Wheelwright, and R. J. Hyndman, Forecasting: Methods and
Applications, 8th ed. New York: John Wiley & Sons, 2008, p. 656.
[263] J. S. Armstrong, Principles of Forecasting: A Handbook for Researchers and
Practitioners, 2001st ed. New York: Springer, 2001, p. 849.
[264] R. J. Hyndman and A. B. Koehler, Another look at measures of forecast accuracy,
Int. J. Forecast., vol. 22, no. 4, pp. 679688, 2006.
[265] S. Makridakis and R. Winkler, Sampling distributions of post-sample forecasting
errors, Appl. Stat., vol. 38, no. 2, pp. 331342, 1989.
[266] N. R. Swanson and H. White, Forecasting economic time series using flexible
versus fixed specification and linear versus nonlinear econometric models, Int. J.
Forecast., vol. 13, no. 4, pp. 439461, 1997.
[267] J. S. Armstrong and F. Collopy, Error measures for generalizing about forecasting
methods: Empirical comparisons, Int. J. Forecast., vol. 8, no. 1, pp. 6980, 1992.
[268] S. Jahromi, T. Anderson, and A. Tudori, Forecasting hybrid electric vehicles
using TFDEA, in PICMET-Technology Management in the IT-Driven Services,
2013, pp. 20982107.
[269] B. S. Yoon, A. Charoensupyan, N. Hu, R. Koosawangsri, M. Abdulai, and X.
Wang, Digital single-lens reflex camera technology forecasting with technology
152
forecasting using data envelopment analysis, in PICMET-Technology

Management in the IT-Driven Services, 2013.
[270] D. J. Lim and D. F. Kocaoglu, China - can it move from imitation to innovation?,
in PICMET-Technology Management in the Energy Smart World, pp. 113.
[271] D. G. Feitelson, The supercomputer industry in light of the Top500 data,
Comput. Sci. Eng., vol. 7, no. 1, pp. 4247, 2005.
[272] J. Dongarra, Visit to the national university for defense technology changsha,
china, Oak Ridge National Laboratory, pp. 118, 2013.
[273] National Research Council, Getting Up to Speed: The Future of Supercomputing,
1st ed. Washington, D.C: National Academies Press, 2005, p. 308.
[274] J. Dongarra, H. Meuer, and E. Strohmaier, TOP500 supercomputer sites,
Supercomputer, vol. 13, pp. 89111, 1997.
[275] M. Peckham, RIKEN plans Exascale supercomputer 30 times faster than
todays fastest in six years, Time, New York, 2013.
[276] S. Ashby, P. Beckman, J. Chen, P. Colella, B. Collins, D. Crawford, J. Dongarra,
D. Kothe, R. Lusk, P. Messina, T. Mezzacappa, P. Moin, M. Norman, R. Rosner,
V. Sarkar, A. Siegel, F. Streitz, A. White, and M. Wright, The opportunities and
challenges of Exascale computing, U.S Department of Energy (Office of Science),
pp. 171, 2010.
[277] H. Simon, T. Zacharia, and R. Stevens, Modeling and simulation at the Exascale
for energy and the environment, Dep. Energy Tech. Rep., 2007.
[278] S. Kamil, J. Shalf, and E. Strohmaier, Power efficiency in high performance
computing, in 2008 IEEE International Symposium on Parallel and Distributed
Processing, 2008, pp. 18.
[279] C.-H. Hsu, J. A. Kuehn, and S. W. Poole, Towards efficient supercomputing: A
quest for the right metric, in The third joint WOSP/SIPEW international
conference on Performance Engineering, 2012, p. 157.
[280] F. Robert, Who will step up to exascale?, Science, pp. 264266, 2013.
[281] T. Geller, Supercomputings exaflop target, Commun. ACM, vol. 54, no. 8, p. 16,
2011.
153
[282] H. Simon, Why we need exascale and why we wont get there by 2020, in
Optical Interconnects Conference, 2013.
[283] J. D. Owens, M. Houston, D. Luebke, S. Green, J. E. Stone, and J. C. Phillips,
GPU computing, Proc. IEEE, vol. 96, no. 5, pp. 879899, May 2008.
[284] Sheng Li, J.-H. Ahn, R. D. Strong, J. B. Brockman, D. M. Tullsen, and N. P.
Jouppi, McPAT: An integrated power, area, and timing modeling framework for
multicore and manycore architectures, in MICRO-42. 42nd Annual IEEE/ACM
International Symposium on. IEEE, 2009, pp. 469480.
[285] J. P. Panziera, Five challenges for future supercomputers, Supercomputers, pp.
3840, 2012.
[286] R. Vuduc and K. Czechowski, What GPU computing means for high-end
systems, IEEE Micro, pp. 7478, 2011.
[287] N. Rajovic, P. Carpenter, I. Gelado, N. Puzovic, and A. Ramirez, Are mobile
processors ready for HPC?, in Supercomputing Conference, 2013.
[288] J. J. Dongarra, P. Luszczek, and A. Petitet, The LINPACK benchmark: past,
present and future, Concurr. Comput. Pract. Exp., vol. 15, no. 9, pp. 803820,
2003.
[289] P. Wood, Cray Inc. replacing IBM to build UI supercomputer, The News Gazette,
Illinois, p. 1, 2011.
[290] L. Lawrence, Cray jumped from AMD for the flexibility and performance of Intel
chips, The Inquirer, 2012.
[291] C. Murariu, Intel signs deal with Cray, marks AMDs departure from Crays
supercomputers, Softpedia, 2012.
[292] F. Wang, C.-Q. Yang, Y.-F. Du, J. Chen, H.-Z. Yi, and W.-X. Xu, Optimizing
Linpack benchmark on GPU-accelerated Petascale supercomputer, J. Comput. Sci.
Technol., vol. 26, no. 5, pp. 854865, Sep. 2011.
[293] A. Cole, Intel and InfiniBand: pure HPC play, or is there a fabric in the works?,
IT Business Edge, 2012.
[294] Anonymous, InfiniBand strengthens leadership as the high-speed interconnect of
choice, Mellanox Technologies, pp. 131, 2012.
154
[295] Z. Whittaker, Intel buys QLogics InfiniBand assets for $125 million, ZD Net,
2012. [Online]. Available: http://www.zdnet.com/article/intel-buys-qlogicsinfiniband-assets-for-125-million/.
[296] IntelPR, Intel acquires industry-leading, high-performance computing
interconnect technology and expertise, Intel Newsroom, 2012. [Online]. Available:
http://newsroom.intel.com/community/intel_newsroom/blog/2012/04/24/intelacquires-industry-leading-high-performance-computing-interconnect-technologyand-expertise.
[297] T. Morgan, Intel came a-knockin for Cray super interconnects, The Register,
2012. [Online]. Available:
http://www.theregister.co.uk/2012/04/25/intel_cray_interconnect_followup/.
[298] R. Haring, M. Ohmacht, T. Fox, M. Gschwind, P. Boyle, N. Chist, C. Kim, D.
Satterfield, K. Sugavanam, P. Coteus, P. Heidelberger, M. Blumrich, R.
Wisniewski, A. Gara, and G. Chiu, The IBM Blue Gene/Q compute chip, IEEE
Micro, vol. 32, no. 2, pp. 4860, 2011.
[299] C. Lazou, Should I buy GPGPUs or Blue Gene?, HPCwire, 2010.
[300] M. Feldman, Blue Genes and GPU clusters top the latest Green500, HPCwire,
2011.
[301] H. Kimihiko, Exascale supercomputer project, RIKEN Advanced Institute for
Computational Science, Kobe, 2013.
[302] L. Tung, After X86 servers, is IBM gearing up to sell its chips business too?,
ZDNet, 2014. [Online]. Available: http://www.zdnet.com/article/after-x86-serversis-ibm-gearing-up-to-sell-its-chips-business-too/.
[303] Davis Nick, Cray CS300 cluster supercomputers to feature new Intel(R) Xeon
Phi(TM) coprocessors x100 family, Cray Media, 2013. [Online]. Available:
http://investors.cray.com/phoenix.zhtml?c=98390&p=irolnewsArticle&ID=1830424.
[304] Davis Nick, Cray to acquire appro international, Cray Media, 2012. [Online].
Available: http://investors.cray.com/phoenix.zhtml?c=98390&p=irolnewsArticle&ID=1756786.
[305] P. Thibodeau, Coming by 2023, an exascale supercomputer in the U.S.,
Computerworld, 2014. [Online]. Available:
http://www.computerworld.com/article/2849250/coming-by-2023-an-exascalesupercomputer-in-the-us.html.
155
[306] D. Takahashi, IBM and Nvidia win $425M to build two monstrous
supercomputers for the Department of Energy, VentureBeat, 2014.
[307] Anonymous, Department of energy awards $425 million for next generation
supercomputing technologies, Department of Energy, 2014.
[308] Y. Kawaguchi, Japans policy toward exascale computing, Ministry of
Education, Culture, Sports, Science, and Technology (Japan), 2014.
[309] R. V. Aroca and L. M. G. Gonalves, Towards green data centers: A comparison
of x86 and ARM architectures power efficiency, J. Parallel Distrib. Comput., vol.
72, no. 12, pp. 17701780, 2012.
[310] A. Maruccia, US intelligence signs up IBM to build an eco-friendly
supercomputer, The Inquirer, 2014.
[311] S. Anthony, US intel agency is developing a superconducting exascale computer
and cryogenic memory, Extreme Tech, 2014.
[312] D.-J. Lim and T. R. Anderson, Technology trajectory mapping using data
envelopment analysis: the ex-ante use of disruptive innovation theory on flat panel
technologies, R&D Manag.(In press), 2015.
[313] D.-J. Lim, Technology forecasting using DEA in the presence of infeasibility,
Int. Trans. Oper. Res.(In press), 2015.
[314] A. Charnes, W. W. Cooper, and R. M. Thrall, A structure for classifying and
characterizing efficiency and inefficiency in Data Envelopment Analysis, J.
Product. Anal., vol. 2, no. 3, pp. 197237, 1991.
[315] T. R. Anderson and L. Inman, Resolving the issue of multiple optima in
Technology Forecasting using Data Envelopment Analysis, in PICMETTechnology Management in the Energy Smart World, 2011, pp. 15.
[316] S. Durairajan, M. I. Prado, N. Rahimi, and S. R. Jahromi, Forecasting
microprocessor technology in the multicore era using TFDEA, in PICMETTechnology Management in the IT-Driven Services, 2013, pp. 21082115.
[317] D.-J. Lim and T. R. Anderson, Improving forecast accuracy by a segmented rate
of change in technology forecasting using data envelopment analysis (TFDEA),
in PICMET-Infrastructure and Service Integration, 2014, pp. 29032907.
156
[318] D.-J. Lim, T. R. Anderson, and T. Shott, Technological forecasting of

supercomputer development: The march to Exascale computing, Omega, vol. 51,
pp. 128135, 2015.
[319] L. Kilian and M. P. Taylor, Why is it so difficult to beat the random walk forecast
of exchange rates?, J. Int. Econ., vol. 60, no. 1, pp. 85107, 2003.
157
Appendix. Model Building Guidelines
Step 1: Selection of the input(s) and the output(s)

The classification of product features into suitable input(s) and output(s) is difficult,
but it is of vital significance. Theoretically, the inputs should capture all resources that
significantly affect the outputs. The outputs should reflect all relevant outcomes on which
one wishes to assess the technologies. A common pitfall of treating undesirable outputs
as inputs should be avoided, and in such a case it is advised to construct the model with
weak disposability. Moreover, any environmental factors that directly affect the
transformation of inputs into outputs should also be reflected in the model either by
including nondiscretionary factors or by normalizing the inputs and/or the outputs.
Step 2: Selection of the orientation

The orientation selection is purely based upon which direction one wishes to
measure the technological progress. For example, the input-oriented model captures the
technological progress in a way that it demonstrates how many input reductions have
been made while attaining the same levels of outputs, whereas the output-oriented model
captures the technological progress in a way that it demonstrates how many output
augmentations have been achieved while maintaining the same levels of inputs. Note that
non-oriented models could also be implemented by specifying the desired direction of the
measurement. Thus, the analyst needs to articulate the purpose of the analysis, whether
input reduction, output expansion or both.
158
Step 3: Selection of the returns to scale (RTS)

The returns to scale assumptions determine the shape of the frontier by constructing
the linear segments based on observed technologies. The simplest form is the constant
returns to scale (CRS) assuming the ratio of output(s) to input(s) is not affected by scale
size. Therefore, the CRS would be appropriate should the virtual technologies scaled up
(or down) from the observed technologies be feasible. The increasing RTS (IRS or NDRS)
assumes that it is possible to introduce virtual technologies scaled up but not down from
the observed technologies. In contrast, decreasing RTS (DRS or NIRS) introduces virtual
technologies scaled down from the observed technologies only. In a situation where
virtual technologies generated as a mix of (or scaled from) observed technologies are less
convincing as benchmarks than actually observed technologies, free disposal hull (FDH)
would be suitable to construct the nonconvex set.
Step 4: Selection of the frontier year (T)

The frontier year determines the time in which the rate of change is obtained. This
may be the most recent time in the dataset when the predictive analysis can benefit from
updating the rate of change by up-to-date data or a certain point in time such as a
forecasting origin where the time series hold-out sampling begins. In the latter case, the
rule of thumb often recommends picking the forecasting origin in such a way that the
training period can account for at least two-thirds of the dataset.
159
Step 5: Selection of the frontier type and second goal

The current model has two options for the frontier type: static and dynamic. The
static frontier fixes the time of the current frontier as a predefined frontier year. That is,
the model treats all the technologies on the current frontier as being concurrent in the
frontier year. In contrast, the dynamic frontier allows having varying points of time on
the current frontier whereby each forecasting target would have a different starting point
of forecast. Therefore, the practical decision factor may be influenced by whether the past
technologies still located on the current frontier should be regarded as current
technologies (in terms of frontier year). This choice necessarily affects the calculation of
the rate of change as well as the arrival of the forecasting target since it determines
elapsed time of technological progress.
When the dynamic frontier is selected, a secondary objective function should also be
specified to choose the individual target year (i.e., effective date) from multiple optima.
In particular, the maximization option will choose the benchmark technology presenting
in the farthest time horizon by identifying the maximum sum of effective dates, whereas
the minimization option will choose the benchmark technology presenting in the closest
time horizon by identifying the minimum sum of effective dates. As a result, the
maximization (minimization) option tends to compute the rate of change more
conservatively (aggressively) but makes more aggressive (conservative) forecasts than
the minimization (maximization) option. Note that depending on the application area,
slack maximization may be preferred to prevent weakly efficient (non-pareto)
technologies from setting the effective date.
160

Technological Forecasting Based On Segmented Rate of Change

Uploaded by

Copyright:

Available Formats

Technological Forecasting Based On Segmented Rate of Change

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Technological Forecasting Based On Segmented Rate of Change

Uploaded by

Copyright:

Available Formats

Portland State University

Dissertations and Theses

Technological Forecasting Based on Segmented

Let us know how access to this document benefits you.

Technological Forecasting Based on Segmented Rate of Change

A dissertation submitted in partial fulfillment of the

Portland State University

2015 Dong-Joon Lim

In addition, the case study of the Exascale supercomputer development is presented to

II. Background ...............................................................................................................12

III. Model Development ..................................................................................................58

IV. Validating through Case Studies.............................................................................80

5.5.3 Weight Restrictions ..................................................................................128

Table 1. Research gap 1 .......................................................................................................6

Figure 1. Complexity of technology assessment .................................................................4

Analytical Hierarchy Process

Advanced Micro Devices

Cryogenic Computer Complexity

Capital Asset Pricing Model

Corrected Ordinary Least Square

Central Processing Unit

Comparative Risk Assessment

Constant Returns to Scale

Data Envelopment Analysis

Decision Making Unit

Department of Energy (U.S.)

Decreasing Returns to Scale

Digital Video Disk

Environmental Impact Analysis

Entry Into Service

Fuzzy Cognitive Map

FLoating-point Operations Per Second

Failure Mode and Effects Analysis

Geometric Distance Function Model

Graphics Processing Unit

Hierarchical Decision Model

High Performance Computing

Intelligence Advanced Research Projects Activity

Increasing Returns to Scale

Life Cycle Cost

Liquid Crystal Displays

Multi-Criteria Decision Analysis(Method)

Multi-Dimensional Growth Model

Malmquist Productivity Index

Non-Decreasing Returns to Scale

Nippon Electric Company

Near Field Communication

Non-Increasing Returns to Scale

National Nuclear Security Administration (U.S.)

National Science Foundation

National University of Defense Technology

Original Equipment Manufacturer

Organic Light Emitting Diodes

Passenger Fuel Efficiency

Production Possibility Set

Proportional Slack-Based Model

RIKEN Advanced Institute for Computational Science

Revealed Comparative Advantage

Maximal LINPACK performance achieved in Gigaflops