Energy Policy 27 (1999) 247}280

Dynamics of energy technologies and global change

Arnulf GruK bler!, Nebojs\ a NakicH enovicH !, David G. Victor",*! Environmentally Compatible Energy Strategies Project, International Institute for Applied Systems Analysis, A-2361 Laxenburg, Austria

" Council on Foreign Relations, 58 E 68th Street, New York, NY 10021 USA

Received 13 November 1998

Abstract

Technological choices largely determine the long-term characteristics of industrial society, including impacts on the naturalenvironment. However, the treatment of technology in existing models that are used to project economic and environmental futuresremains highly stylized. Based on work over two decades at IIASA, we present a useful typology for technology analysis and discussmethods that can be used to analyze the impact of technological changes on the global environment, especially global warming. Ourfocus is energy technologies, the main source of many atmospheric environmental problems. We show that much improved treatmentof technology is possible with a combination of historical analysis and new modeling techniques. In the historical record, we identifycharacteristic &&learning rates'' that allow simple quanti"ed characterization of the improvement in cost and performance due tocumulative experience and investments. We also identify patterns, processes and timescales that typify the di!usion of newtechnologies in competitive markets. Technologies that are long-lived and are components of interlocking networks typically requirethe longest time to di!use and co-evolve with other technologies in the network; such network e!ects yield high barriers to entry evenfor superior competitors.

These simple observations allow three improvements to modeling of technological change and its consequences for globalenvironmental change. One is that the replacement of long-lived infrastructures over time has also replaced the fuels that power theeconomy to yield progressively more energy per unit of carbon pollution } from coal to oil to gas. Such replacement has&&decarbonized'' the global primary energy supply 0.3% per year. In contrast, most baseline projections for emissions of carbon, thechief cause of global warming, ignore this robust historical trend and show little or no decarbonization. A second improvement is thatby incorporating learning curves and uncertainty into micro scale models it is possible to endogenously generate patterns oftechnological choice that mirror the real world. Those include S-shaped di!usion patterns and timescales of technological dynamicsthat are consistent with historical experience; they also include endogenous generation of &&surprises'' such as the appearance of radically newtechnologies. Third, it is possible to include learning phenomena stylistically in macro-scale models; we show that doing so can yieldprojections with lessened environmental impacts without necessarily incurring negative e!ect on the economy. Arriving on that path by theyear 2100 depends on intervening actions, such as incentives to promote greater diversity in technology and lower barriers to entry for newinfrastructures that could accelerate historical trends of decarbonization. ( 1999 Elsevier Science Ltd. All rights reserved.

Keywords: Endogenous technological change; Modeling; Global warming

1. Introduction

Changes in products, devices, processes and practices} technology1 } largely determine the development andconsequences of industrial society. Technology hasallowed hunger to decline while the world populationmore than doubled since 1950 and cropland rose by only

*Corresponding author1A broad de"nition of &&technology'' is adopted here because it is

di$cult to separate the (economic and social) importance of physicalartifacts from the social and institutional processes that put thoseartifacts into practice. See, eg, Freeman (1982/1989).

one-third (Hayami and Ruttan, 1985). The appearance ofradically new technologies has eliminated some environ-mental problems while creating new ones } automobilesextinguished horses and manure stench from the roadbut now cause pervasive urban smog. Incremental ad-justments to existing technologies have also improvedenvironmental performance. Examples include theaddition of catalytic converters to automobiles, whichpartially cut smog and forced fuel suppliers to get thepoisonous lead out of gasoline. Even small technologicalchanges have had radical e!ects when compounded overmany years } thousands of hardware and managerialchanges allow today's airlines to deliver a seat-kilometer

0301-4215/99/$ - see front matter ( 1999 Elsevier Science Ltd. All rights reserved.PII: S 0 3 0 1 - 4 2 1 5 ( 9 8 ) 0 0 0 6 7 - 6

of service by burning less than half the fuel required threedecades ago.

Despite the centrality of technology, analytical tech-niques that are useful for analyzing technological changeand its impacts remain embryonic. We argue that it ispossible to do better by simultaneously applying a multi-tude of techniques developed and re"ned by colleaguesand ourselves over two decades at the InternationalInstitute for Applied Systems Analysis (IIASA). Our fo-cus is the impact of technology on &&global change'' } thelocal, regional and worldwide e!ects of industrial societyon the environment. We show that typical rates andpatterns of technological change can be identi"ed inhistory, and new analytical techniques and increasedcomputational power allow more appropriate (&&endo-genous'') modeling of technological change. Long-termtechnological forecasting still remains elusive; however,historical analysis and improved numerical modeling,together, can sharply increase the ability to anticipatetechnological changes and their environmental impacts.

Economic theory, historical evidence, and (neoclassi-cal) growth models con"rm that advancing technologicalknowledge is the most important single factor that con-tributes to long-term productivity and economicgrowth.2 Yet, in most models of long-term economicdevelopment and environmental change, technologycontinues to be treated as a quantity exogenous to theeconomy and society at large. Typically the only endo-genous mechanism of technological change in existingmodels is progressive resource depletion, such as runningout of oil, which results in increasing costs of resourceextraction that favor more expensive, but resource-frugaltechnologies. (Often these are modeled as &&backstop''technologies, a term coined by Nordhaus (1973).)However, technological change triggered by resourceconstraints is somewhat inconsistent with historicalexperience. (See Barnett and Morse (1967) and, for acurrent review of hydrocarbon resources, see Rogner(1997).)

Although analysts agree that technological change iscrucial, in practice most analysis proceeds as if most

2For reviews see Metcalfe (1987) and Freeman (1994). For empiricalevidence from economic history see Maddison (1991, 1995) and Mokyr(1990). For a review of the treatment of technological change as&&residual'' of long-run productivity growth see Griliches (1996). Thetwo classical papers sifting out the impact of technological change onproductivity growth based on neoclassical production function modelsare: Tinbergen (1942) and Solow (1957). For instance, in Solow's calcu-lations advances in knowledge (technology) account for 85% of eco-nomic growth per capita, and only 15% is accounted for by increases incapital. Alternative approaches, particularly those within &&new growththeory'' have extended the classical production function by adding anendogenous &&knowledge'' stock variable, measurement of which, how-ever, remains elusive (cf. Romer, 1990). As a rule, the factors enteringa production function (capital, labor, knowledge, technology) aretreated as independent of each other. This assumption is criticized byAbramovitz (1993).

technological change cannot be anticipated and modeled.Some studies largely ignore technological change; thus,by design, they typically yield Malthusian projections ofstarvation and ecological catastrophes as populationsand economies grow while not using "nite resourcesmore e$ciently (Meadows et al, 1972, 1992). Many stud-ies include only marginal and gradual technological cha-nges, often through an aggregate trend parameter } suchas the annual rate of e$ciency improvement } that is(exogenously) tuned according to historical experience.That technique is compatible with the highly aggregatedmacroeconomic modeling tools that are commonly em-ployed in global change studies. Such models do notrepresent the selection of particular technologies andthus are only able to include technological changes thatare marginal extensions of the present. Yet the historicalrecord is abundant with radical technological changes.Even models in the tradition of systems engineering} where detailed information on technological costs andperformance is used to calculate least cost technologicalsystems } have largely failed to address technologicalchange. Some ignore changes in costs and performance(and thus implicitly assume that technologies are static).Most impose those changes exogenously but have nomechanism } other than the intuition of the modeler } toensure that the assumptions imposed are plausible andinternally consistent. Very few systems models have in-cluded rudimentary endogenous mechanisms of techno-logical change, such as learning curves. (For exceptionssee Messner et al (1996), Fragniere and Haurie (1995) andMessner (1997); see also Nordhaus and van der Heyden(1993) and the discussion below.) Yet only rarely arethose models linked with macroeconomic tools to allowsystematic analysis of the technological aspects of globalchange problems. Here we present one of the "rst suchapplications of linked models.

Treatment of technological change is especially di$-cult yet crucial for analysis of the impacts of industrialsociety over long time periods } decades and centuries.On those time scales, which are characteristic of theenvironmental issues that constitute global change, evensmall technological changes can compound into radicallydi!erent technological systems and environmental ef-fects. Because the timescales are long and the existinganalytical techniques are imperfect, typically analystsbound a range of plausible futures and policy optionswith scenarios. Those scenarios are built with the samemodels that poorly represent technological change, andthus extant scenarios also typically ignore technologicalchange or mechanically extrapolate past trends into thefuture. (For a recent review see Alcamo et al (1994).)Some scenario-builders adopt radical visions for tech-nologies that will be invented and adopted (eg, Lazaruset al, 1993), which helps to de"ne boundaries for possiblefutures but o!ers little insight into the costs and prob-abilities of those utopic extremes. Indeed, extreme low

248 A. Gru( bler et al. / Energy Policy 27 (1999) 247}280

Table 1Stylized stages of technological development and typical characteristics.

Stage Mechanisms Cost CommercialMarket share

Learning Rate

Invention Seeking and stumbling upon newideas; breakthroughs; basic research

High, but di$cult toattribute to a particularidea or product

0% Unable to express inconventional learningcurve

Innovation Applied research, development anddemonstration (RD&D) projects

High, increasingly focusedon particular promisingideas and products

0% Unable to express inconventional learningcurve; high (perhaps'50%) in learningcurves modi"ed toinclude RD&D (see text)

Niche marketcommercialization

Identi"cation of special nicheapplications; investments in "eldprojects; &&learning by doing''; closerelationships between suppliersand users

High, but decliningwith standardization ofproduction

0}5% 20}40%

Pervasive di!usion Standardization and mass production;economies of scale; building ofnetwork e!ects.

Rapidly declining Rapidly rising(5}50%)

10}30%

Saturation Exhaustion of improvement potentialsand scale economies; arrival of moree$cient competitors into market;rede"nition of performancerequirements

Low, sometimes declining Maximum(up to 100%)

0% (sometimes positivedue to severecompetition)

Senescence Domination by superior competitors;inability to compete because ofexhausted improvement potentials

Low, sometimes declining Declining 0% (sometimespositive due to severecompetition)

Note: Also shown, in the right column, are three terms often used when classifying technologies that are marked by substantially di!erent relativeperformance at a given moment in time. Much of technological analysis for purposes of assessing environmental e!ects is aimed at examiningwhich new (radical and incremental) technologies will achieve what speed and level of penetration in commercial markets.

Q&&R

adical''

PQ&&In

cremental

''P

Q&&M

ature''

P

and high scenarios for emissions of carbon dioxide* which vary from about 2 GtC (gigatons, 1015 grams,of elemental carbon) to more than 30 GtC in the year2100 * di!er mainly due to their underlying assump-tions about technology.

We argue that technological futures are neitheropaque nor unbounded. Because global change problemsare numerous, we focus on the main source of manyenvironmental problems: the combustion of fossil fuels.Because technological changes can be most dramaticover the long term, we focus on greenhouse warming,which is principally caused by the accumulation of car-bon dioxide released during the combustion of fossil fuelsover many decades. With the use of comprehensive his-torical statistics compiled at IIASA, we argue that threerobust attributes of energy technologies and their green-house impacts are evident in the historical record. (1)Typical improvements in cost and performance of newtechnologies due to &&learning'' can be identi"ed. (2) Dy-namic competition between technologies to provide en-ergy services, such as mobility, yields predictable patternsfor the entry and exit of technologies in competitivemarkets. (3) Network e!ects and technological interde-pendence such as between petroleum re"neries, pipelines,

gas stations, and gasoline-powered automobiles result incharacteristic patterns of technological co-evolution. Weargue that these three attributes make possible the devel-opment of models with realistic and endogenous treat-ment of technological change, which we demonstraterigorously at the micro scale and stylistically for theworld's entire energy system. In sum, technological as-sessment is still imperfect; some aspects of technologicalfutures, especially related to the timing and character ofradical technological inventions, are still shrouded inmystery. But better analysis and modeling of the stages oftechnological change that follow invention is possible.Before we argue how, we present a consistent typologythat helps to classify the processes at work.

2. A typology for technology analysis

Technological change is a complex process. A simpletypology helps to identify the key mechanisms, conceptsand measures. We distinguish six stages (Table 1) in thelife-cycle of a technology. Following Shumpeter (1934)and Freeman (1982/1989) we distinguish between inven-tion, which is the creation of an idea, and innovation,

A. Gru( bler et al. / Energy Policy 27 (1999) 247}280 249

which is marked by the "rst practical application of aninvention. Many inventions are the product of funda-mental and applied research. Often innovation takes theform of development and demonstration projects in anindustrial laboratory. This de"nition of innovation isnarrow, although often the label is applied broadly to theentire process of commercialization that follows the "rstpractical application.

Following development and demonstration, the usefulservices of a new technology are typically "rst employedin niche markets, where a novel technology has substan-tial performance advantages over existing technologies.The "rst commercialization in niche markets allows sup-pliers and users to &&learn by doing'' and &&learn by using''which leads to further improvements in performance andcost. Use in a wider array of markets, or pervasive di+usion,follows. When those markets are exhausted saturation is theresult. Senescence follows when a better (new) competitortakes market share or rede"nes performance requirements.

Of all the aspects of technological development, theprocess of invention is least well understood, and evenless well predicted. The inventor's mind, which mustdeviate from predictable conventions to discover novelty,is intrinsically di$cult to model and anticipate. Evenrewards from invention may play only a small role,especially in the generation of radically new concepts anddevices. As the popular book on John Harrison's inven-tion of the accurate marine chronometer demonstrates,high hurdles face the person who de"es convention andthen seeks recognition and returns from radical invention(Sobel, 1995).

The supply of inventions is also di$cult to modelbecause it is loosely related to the background state ofknowledge } what some have called &&knowledge stock''(Romer, 1986, 1990). But quantitative measures of usefulknowledge, and their exact relationship to invention, aremurky. Similarly important but poorly understood is thecontribution of basic research.

For the other stages of technological change } frominnovation through senescence } it is increasingly pos-sible to make systematic observations and to model theprocesses at work, which are the tasks for the rest of thisessay. Investments needed to yield innovations and vi-able commercial products from an invention are princi-pally governed by market competition. Commercializ-ation of a technology often requires large organizationsand uniformity, which further increases the capacityof analysts to model relationships between inputs andoutputs.

Like bacteria vying for scarce food, the evolution oftechnology is a competitive process. In many "elds thecost of stumbling and searching for new ideas is relativelylow and thus the supply of inventions is abundant;thoughts and talk are cheap. But the Darwinian selectionmechanisms are stringent and few ideas ever take formoutside the laboratory in commercial markets.

Investments needed to sustain a technology in its early,pre-competitive stages are made because there is hope oflater returns. Once competitive, relative performance andcosts govern success, which can be measured by marketshare. Indeed, these three indicators } performance, costsand market share } are useful measures of a technology'sstage of development. Fig. 1 shows investment costdata for ten types of electricity generation technologies,drawn from IIASA's comprehensive energy technologydatabase (CO2DB) (Messner, 1996; SchaK fer et al, 1992;Messner and Strubegger, 1991). Technologies that arealready in pervasive di!usion and saturation (eg, conven-tional fossil fuel electricity generating plants) are, as ex-pected, less costly than those found only in niche markets(eg, photovoltaic cells).

The variance in investment cost for all technologies inFig. 1 is high because capital requirements and perfor-mance depend on many local factors. Installing the leastcostly coal-"red electricity plants at ideal sites withoutpollution control equipment costs about US$500 perkilowatt of generating capacity (kW(e)); typical costs forsuch plants are double that value, and modern plantswith sulfur- and nitrogen-removal are typically twiceagain as expensive (US$2000 /kW(e)). (For a more de-tailed assessment of power plant cost distributions seeStrubegger and Reitgruber (1995).) Niche markets areborn in the long tails and when performance require-ments change. Often new technologies provide a newservice or function that is not possible or cost-e!ectivewith the old technologies. Examples include the precisionof digital clocks compared with their analog counter-parts, the speed and range of jet airplanes over piston-powered models, and the a!ordability of solar photovol-taic power in remote mountain huts and road signswhere wire connections to the conventional electric gridwould be much costlier. Fear of global warming couldrede"ne the performance requirements for electricity gen-eration and thus create niche markets for low carbonelectricity production. Removing and disposing of CO

2from #ue gases would double or triple the cost of conven-tional fossil fuel-"red electricity plants, but solar, nuclear,and zero-carbon biomass could be largely una!ected andthus become more competitive.

Fig. 1 labels the ten technologies with useful common-sense terms that re#ect the di!erences in cost, marketshare, and stage of technological development. &&Mature''technologies have reached pervasive di!usion and havewell-known characteristics; often they can change or im-prove under competitive pressure, but in general perfor-mance and costs are stable. &&Incremental'' technologiesare found in niche markets } they are more costly buto!er some performance advantages and the potential forsigni"cant cost reduction with continued investment.The rates and direction of such performance and costimprovements in incremental technologies can beanticipated. Commercial enterprises can envision the

250 A. Gru( bler et al. / Energy Policy 27 (1999) 247}280

Fig. 1. Cost of selected electricity-generating technologies. Values (mean plus/minus one standard deviation) taken from IIASA's comprehensivetechnology database (CO2DB). Data are for approximately ceteris paribus conditions (eg, coal plants include de-SO

xand de-NO

xequipment). Mature

technologies in widespread use have lower costs with lower variance; the costs of radical new technologies are higher and more variable. Variability ofcosts is also an indicator of the uncertainty of technology costs. Radical technologies are little tried and their potentials for cost reductions areuncertain, and thus so are estimates of their costs.

Source: Strubegger and Reitgruber (1995).

possibilities and modalities for better performance ofa technology that is compatible with and incremental toexisting technological systems, especially as market con-ditions change. &&Radical'' technologies are more uncer-tain, both in their potentials for improvement and inwhether they will arrive on commercial markets at all.But the radical technology can also yield radical im-provements in performance and cost } often by a factorof 10 or more. By de"nition, the radical technology is notwidely employed, and thus only a small slice of its (poten-tial) di!usion history can be observed.

The stages of technological development are distin-guished by the mechanism at work, not time. For acompletely new technology it is possible to identify abeginning or particular stage change, such as the birth ofan idea or the "rst commercial sale. But mature, existingtechnologies can be the host for all stages of technolo-gical change simultaneously. Automobiles today, forexample, are the site for the incorporation of basic re-search on modeling of complex systems into the inven-tion of intelligent transportation systems that could, if

the technology is successful, optimize road tra$c #ows,akin to air tra$c control but without human intermedia-ries. Innovations in automobile technology includelong-distance electric-powered vehicles. On-board navi-gation systems } with local maps updated by globalpositioning system, and limited capacity to optimize anddirect trips } already exist in commercial niche markets,such as high-end rental cars designed for customers withpoor local knowledge or strong desire to demonstratetechnical prowess. The road data system (RDS), whichtransmits limited textual information and tra$c updates,is in pervasive di!usion in Western Europe } virtually allnew car radios are equipped, and all FM stations broad-cast the necessary signals. The car radio itself is anexample of saturation } practically 100% of new carshave one installed either by the factory or in the immedi-ate after market.

In short, even as a technological system enters into anddi!uses throughout a market, there is pervasive changewithin the system. The basic innovation that createsa new radical technology is followed by incremental

A. Gru( bler et al. / Energy Policy 27 (1999) 247}280 251

changes that accrete around the basic innovation, whichleads to improved performance and reduced costs. Likethe development of science itself, portrayed by Lakatos(1970), the core de"nes the basic functions of the techno-logy } it attracts compatible incremental changes anddeters radical incompatible change. New technologiesarise from within the economic system, but truly radicalchanges rarely come from the existing core. The processof innovation requires investments in innovations wellbefore a new technology becomes really competitive andpro"table. It also requires continued investments to fur-ther a technology's performance and to lower its costs tosustain successful di!usion. The process, if modeled in itsdetails, is enormously complicated. Thus the approach inthis paper is "rst to identify some basic patterns in thehistorical record that help to guide modeling and scen-ario building.

We use the term &&technology'' broadly. It denotes notonly physical devices but also production knowledge andprocesses, which typically combine physical equipment,ideas and even social institutions (eg, the Ford method ofproduction via standardization and assembly line). Theterm &&technology'' is also used to denote a system ofrelated hardware or production processes (eg, the &&auto-mobile''). Like a fractal, new layers of activity becomevisible as one looks closely at a technological system, butbasic patterns and processes remain similar. Even themundane windshield wiper, which is part of the automo-bile's technological system, is the technological amalgamof rubber blades, airfoils that hold the blades in place athigh airspeed, windshield-washing mechanisms, heatingand delay timers, and microprocessor controls. Eachcomponent has been the site of change, and thus thetechnological system as a whole has also changed.

In this essay we examine both components and techno-logical systems. When analyzing the environmentale!ects of technology it is usually crucial to examineparticular technologies or processes, which ultimatelydetermine environmental impacts. For example, the tech-nological components of today's automobile are unlikethose of the 1930s, and the environment is much betterfor the change. Emissions of carbon monoxide, nitrogenoxides and hydrocarbons totaled some 100 g per ve-hicle-km in the 1920s; today they have declined to only2 g (GruK bler, 1998). Part of the change re#ects the collec-tive e!ect of many technological changes to yield morecomplete combustion and higher e$ciency. Automobileengine e$ciencies in the 1920s were typically around10%, a value that has doubled today, thus halving energyuse per km driven. Lower pollution is also a consequenceof speci"c pollution control technologies, such as cata-lytic converters. Although environmental impacts aredetermined by particular technologies, we show thatthe selection of technologies is partially governed byattributes of the system, such as the availability of neces-sary infrastructures and other technological interdepen-

dencies. For example, hydrogen fuel cell vehicles couldallow practically zero emissions of air pollution, includ-ing carbon dioxide, but such technologies will not beviable without complementary hydrogen production anddistribution technologies and infrastructures.

3. Historical view of technological change

The systematic de"nition of stages and mechanisms oftechnological change helps to identify fundamental at-tributes of technological change in the historical recordand thus improve technological analysis and scenariobuilding. For two decades, scientists at IIASA have com-piled historical data on all major energy technologies forparticular markets, nations, regions and the globe(Marchetti and NakicH enovicH , 1979; NakicH enovicH , 1984;Marchetti, 1988; Ausubel et al, 1988; GruK bler, 1990;GruK bler and NakicH enovicH , 1991; NakicH enovicH , 1994;GruK bler, 1996). Those data help identify three robustattributes of technological change: (1) reductions in costand improvements in performance through learning; (2)regular patterns of dynamic competition between tech-nologies; and (3) the co-evolution of long-lived infrastruc-tures and technological clusters due to &&network e!ects''} the externalities and synergisms that make it costly forany single component to be incompatible with the whole.

3.1. Learning

The performance and productivity of individual tech-nologies and technological systems typically increases asorganizations and individuals gain experience with them.Long-studied in human psychology, technological learn-ing phenomena were "rst described for the aircraft indus-try by Wright (1936), who reported that unit labor costsin air-frame manufacturing declined signi"cantly withaccumulated &&experience'', measured as cumulative pro-duction (output). Technological learning has since beenanalyzed empirically for numerous manufacturing andservice activities (eg, ship building, petrochemicals, steamand gas turbines, farming of broiler chickens). Learningconcepts have also been applied in a wide range ofhuman activities, such as the success rates of new surgicalprocedures, productivity in kibbutz farming, and reliabil-ity of nuclear plant operation (Argote and Epple, 1990).In economics, &&learning by doing'' and &&learning by us-ing'' have been highlighted since the early 1960s (Arrow,1962; Rosenberg, 1982).

Learning phenomena are generally described in formof &&learning'' or &&experience'' curves, which typicallyshow the decline in unit costs of production as experienceis gained. Because learning depends on accumulation ofactual experience and not just on the passage of time,learning curves generally take the form of a power func-tion where unit costs decrease exponentially as a function

252 A. Gru( bler et al. / Energy Policy 27 (1999) 247}280

Fig. 2. Learning curve for gas turbines. Costs declined rapidly (&20% per doubling of installed capacity) when the technology was in the &&innovation''stage and learning was enhanced through applied research and development (R&D). Costs continued to decline, but at a slower rate (&10% perdoubling) when the technology was commercialized in niche markets, such as through demonstration projects. The incremental investment for thistechnology } from "rst application of the technology to electricity generation until the technology was fully competitive } was approximately US$5 billion. Data are only for one major "rm (GE).

Source: data adapted from MacGregor et al (1991).

of cumulative output. Other measures include cumulat-ive investments, installed hardware, or other proxies for&&experience''. The resulting curve is often plotted onlogarithmically scaled axes so that it becomes a straightline. The learning rate } the slope of the line } is thepercentage decline in costs per doubling of accumulatedexperience. Because each successive doubling requiresmore production volume, such straight-line plots shouldnot be misunderstood to imply that &&linear'' progress canbe maintained inde"nitely. The potential for cost reduc-tions becomes increasingly exhausted as the technologymatures.

The mechanisms for learning by doing are numerous.These include experience gained by individuals in per-forming routine tasks, improvement in the functioning oforganizations (eg, plant management, logistics, market-ing), and economies of scale (Cantley and Sahal, 1980).The causal mechanisms are well established, but learningis not the only means of reducing costs. Other factorsthat are external to learning by doing, such as improve-ments in upstream technologies, can also lower costs andare correlated with growing experience } thus the analystmust be careful when using learning curves that themechanisms of learning apply to the situation at hand. Atminimum, it appears that learning by doing requirescontinuous experience, not merely the accumulation ofoutput regardless of its time path. Unit costs of theLockheed L-1011 &&Tristar'' aircraft rose in the late 1970s

when production resumed after a drastic reduction thatincluded large-scale layo!s at production facilities. Ex-perience gained during the early 1970s was lost with thesta! turnover; as a result, the planes built in the early1980s were in real terms more expensive than those builtin the early 1970s (Argote and Epple, 1990).

Learning rates in manufacturing, including productionof energy-related technologies, mainly vary from 10 to30%. In some cases, typically at the early stages ofcommercialization of a technology, learning rates ap-proaching 50% have been observed (Argote and Epple,1990; Christiansson, 1995). A typical learning curve,shown in Fig. 2 for gas turbines used in electricity genera-tion, consists of two segments. The "rst segment corres-ponds with the innovation stage } from the invention(adapted from jet aircraft engines) in the 1950s to themiddle 1960s when the "rst gas turbine demonstrationprojects had been built and gas turbines entered nichemarkets. During this stage, cost reductions were rapid(some 20% per doubling of the small installed capacity ofdemonstration projects); gas turbines were a truly radicaltechnology } extremely expensive, but promising if sub-stantial investments were made. The second segment,from the middle 1960s until 1980, is marked by smallercost reductions for each doubling of experience, charac-teristic of expanding niche markets and early commer-cialization. Learning rates were approximately 10%.During that period the technology was costlier than

A. Gru( bler et al. / Energy Policy 27 (1999) 247}280 253

Fig. 3. Learning curves for several electricity generation technologies. Shown are gas turbines, which are an &&incremental'' technology on the cusp ofwidespread application (Fig. 2) and two higher cost technologies: photovoltaics and advanced windmills. These two more costly technologies are&&radical'' in that they are competitive only in special niche markets and thus market share is low, but both hold promise for lower costs with additionalinvestments. Learning rates for all three technologies in their pre-commercial stage are comparable (&20% per doubling of capacity).

Sources: see Fig. 2, and Christiansson (1995).

mature technology alternatives but became increasinglycompetitive through continued improvements that weresustained by substantial investments. Overall, the costper unit of capacity declined by a factor of 4 as cumulat-ive experience rose three orders of magnitude. Since 1980gas turbines have been an incremental technology } in-creasingly applied in commercial markets as one com-ponent of the electricity generation system. Today, gasturbines are the preferred technology for most electricitygeneration applications and are in the midst of pervasivedi!usion. Our estimate is that the total investments} R&D and commercial sales for niche market applica-tions } approached US$5 billion before the new techno-logy became economically competitive with alternativeelectricity generation technologies beyond special nichemarkets. However, the exact cost of applied R&D isdi$cult to estimate because statistics on such spendingby private "rms is typically not publicly available. More-over, this "gure does not include any of the originalpublic and private investments into aircraft jet enginesbefore their "rst derivatives were adapted for electricitygeneration.

Fig. 3 contrasts the learning curve of the gas turbineswith two new renewable electricity generation technolo-gies } wind and photovoltaics. Both technologies displayrapid learning that began at high cost; already both canbe competitive under special conditions. Wind power, forexample, is pro"table at sites with steady strong breezes,though often only with subsidies. Both are examples ofradical technologies that are competing within an exist-ing, mature technological system (i.e., electricity produc-tion and distribution). For producers of photovoltaics inboth Japan and the United States, costs fell by over 20%with each doubling of capacity } a learning rate similar tothe early history of the gas turbine.

A single learning curve, as in Figs. 2 and 3, helpsconvey the cost improvements that result from the com-plex processes of learning by doing in the commercialmarketplace. But such conventional learning curves areinadequate for modeling the relationship between invest-ments in a technology and consequences such as lowercosts and improved performance which a!ect the abilityof technologies to compete for market share. Such learn-ing curves include only investments that yield experience

254 A. Gru( bler et al. / Energy Policy 27 (1999) 247}280

in the commercial marketplace; they omit pre-commercial research and development (R&D) as well asnon-commercial demonstration projects that lead tocommercial demonstration projects in niche markets.Yet, together, research, development and demonstration(RD&D) are vital to the improvement of performanceand the lowering of costs in the early stages of technolo-gical development. For example, the cost of photovol-taics produced in Japan halved between 1973 and 1976,but none of this improvement is evident in Fig. 3 becauseit was prior to any installation of demonstration unitsand thus cumulative installed capacity was zero. SuchRD&D expenditures are a small factor in the cost im-provements of technologies that are already advanced tothe stage of commercial niche markets and are candi-dates for pervasive di!usion, such as gas turbines in 1980.But in the earlier stages, RD&D accounts for a largershare of performance improvements and cost reductions.Thus learning curves must be expanded from theirstandard formulation if they are to be of practicaluse in modeling technological change that results fromthe changing competitive position of technologies.Doing so requires accurate and comparable data onapplied research and development, which are usuallyscarce.

One of the few reliable sources of product-speci"cRD&D expenditures is Watanabe's analysis of theJapanese Ministry of International Trade and Industry(MITI) &&sunshine'' technology program to promotenew energy technologies such as solar photovoltaics(Watanabe, 1995, and personal communication). His ex-ceptionally comprehensive data include both public(MITI) and private RD&D expenditures. Watanabe esti-mated the parameters for a model that traced the directand indirect (feedback) relationships between MITI-in-itiated RD&D spending, private RD&D spending, pro-duction of photovoltaics, and changes in the unit cost ofphotovoltaics in Japan. In his model, public RD&Dspending and other incentives stimulated industryRD&D and increased the stock of knowledge speci"c tophotovoltaics, which led to performance and cost im-provements. Lower costs stimulated demand for theseimproved products, which increased the size of commer-cial niche markets and led to learning e!ects and furthercost reductions. Larger commercial markets yielded pro-duction increases and additional stimulus for industrialRD&D. In his terms, RD&D lubricated a &&spin cycle''.(He excluded inter-industry and cross-national R&Dspillover e!ects, as well as those from purchases of equip-ment.3 Such e!ects #ow into and out of "rms that areinvestigating particular photovoltaic technologies and

3On &&spillover'' e!ects see Mans"eld (1985); on the impact of equip-ment purchases see Organisation for Economic Cooperation and De-velopment (1996).

increase the stock of relevant knowledge, but they arenotoriously di$cult to quantify.4).

Parameters from Watanabe's plausible model can beused to plot a more comprehensive form of the &&learningcurve'', shown in Fig. 4. The independent variable iscumulative investments, which includes RD&D as wellas commercial consumer purchases. Over the period1973 to 1995 a total of 206 billion Yen (approximatelyUS$2.5 billion in 1995 prices and exchange rates) werespent on photovoltaics in Japan. 78% (162 billion Yen) ofthat amount were actual investments in commercialphotovoltaic capacity, and 22% (44 billion Yen) wasspent on RD&D proper. The "gure con"rms that oncea technology reaches the niche market stage that invest-ments in hardware (ie, installed capacity) dominate, butthat RD&D is signi"cant contributor to lower costs,especially in the early stages. Moreover, such investmentsand RD&D cannot be treated as separate, independentsources of technological improvement. Only when com-bined does a curve that is characteristic of a learningcurve materialize. Our understanding of the learningprocess is that it involves the interaction of both sup-pliers and users of technology.

This new &&learning curve'' has two important features.First, it is entirely economic and thus can be used inmodels that compute economic relationships betweenresource spending and changing prices of technologies.Such relationships are crucial for technological modelingbecause it is the anticipation of lower costs that leads toinvestments in immature technologies. In turn, thosecosts partially determine which technologies are selectedfor application, with environmental consequences. Thischain of relationships } from investments to learning tocost reductions to market application to environmentalconsequences } is not perfectly known ex ante. Thusmodels which use of learning as a driving force of techno-logical change must incorporate uncertainty, whichwe consider further below.5 Second, the curve isconceptually coherent because it includes both of themajor sources of technological improvement } RD&Dand commercial purchases. This approach thus also ad-dresses one of the critiques of conventional learningcurves } they do not include RD&D inputs. Similarcurves can be calculated for other technologies, indus-tries and countries, pending the availability of data.

4An analysis of inter-sectoral relationships of R&D expenditures forthe US and how these could be a!ected by climate policies (leading outto &&crowding out'' phenomena) is reported in Goulder and Schneider(1996) and Schneider and Goulder (1997).

5Postulating a relationship does not mean that the parameters ofthat relationship are known ex ante. Uncertainty therefore persists vis.the eventual outcome (level of improvement in cost and performance)and the level of investment (in R&D and niche markets) needed toachieve a particular outcome.

A. Gru( bler et al. / Energy Policy 27 (1999) 247}280 255

Fig. 4. Modi"ed learning curve for photovoltaic technologies. The curve shows decline in costs as a function of cumulative investments including notonly cumulative installations from demonstration projects and commercial niche markets but also R&D investments. Conventional learning curvesinclude only commercial installations and ignore other investments. At the early stages, when physical installations are few, R&D is relativelyimportant. The declining costs of photovoltaics correlate well with aggregate RD& D investments (R&D as well as pre-and commercial demonstrationprojects) and are comparable with a classic learning curve pattern with 54% reduction in costs for each doubling of cumulative investment. Thisformulation allows a single, simple learning curve to be used to model cost reductions from the innovation stage as well as later commercial stages.(Note that these curves are not directly comparable with Fig. 3 because the independent variables and the currency units di!er.) For more detail on themethodology used to produce this modi"ed learning curve see GruK bler and Gritsevskii (1998); data from Watanabe (1995, personal communication).

3.2. Technology dynamics: market competitionand diwusion

Learning curves help illuminate the dramatic reduc-tion in costs evident especially in the early stages ofa technology } innovation, niche market commercializ-ation, and the initial di!usion into widespread applica-tion. Learning curves help to identify technologies thatmight become competitive with adequate investment.But the tool is less powerful as the learning rate slows anda technology enters wider market application. In thosesituations, often many technologies compete and it isconceptually and empirically di$cult to develop com-plete and robust learning curve models that include alltechnologies and processes that determine the outcomeof dynamic competition between technologies in the mar-ketplace. In the ideal world the modeler would estimatethe cost and potential of all technologies and then deter-mine the outcome according to cost di!erences. Empiric-ally and conceptually that world does not exist (yet).Thus a multitude of techniques is needed.

In addition to learning curves, another approach is toidentify general patterns by which technologies di!usethrough competitive markets. Such patterns are oftenevident when plotting the fraction (F) of a useful productor service, such as electricity or mobility, supplied byeach major competing technology. The typical result is

an S-shaped curve, which is often termed a logistic substi-tution or di!usion curve. At the earliest stage of commer-cialization, growth in a technology's market share is slowas the technology is applied only in specialized nichemarkets and costs are high. Growth accelerates as earlycommercial investments lead to compounding cost re-ductions and standard-setting, which leads to imitationand adoption in a wider array of settings. As the potentialmarket is saturated and a product matures, growth inmarket share declines to zero. With the arrival of bettercompetitors, the market share of the senescent techno-logy declines. Such S-shaped curves are characteristic ofmany social and biological processes where the rate ofdi!usion or substitution depends on the probabilityof encounter between a supplier and a receptive host.Other studies of di!usion have applied these curves} pioneered by Lotka (1924) and Volterra (1927) } to thespread of infrastructures and technologies, as well as thedi!usion of epidemics, ideas and forms of social organiza-tion (HaK gerstrand, 1967; Fisher and Pry, 1971; Marchettiand NakicH enovicH , 1979; Marchetti, 1980; GruK bler, 1990;Astakhov et al, 1990; Buttner and GruK bler, 1995;NakicH enovicH , 1990; United States Department ofCommerce, 1975).

Perhaps the most famous case of technological sub-stitution is motor cars for horses. In this simple case,one technological artifact, the passenger car, replaced

256 A. Gru( bler et al. / Energy Policy 27 (1999) 247}280

Fig. 5. A simple case of technological substitution. With a time constant (*t) of 12 yr, motor cars replaced horse-drawn carriages for transportation inthe United States. The top panel shows this replacement in terms of number of units on the road. The data can be transformed into a format thathighlights technological di!usion and substitution processes. That is shown in the bottom left panel, where the fraction (F) of the total number of roadvehicles (horse carriages#automobiles) accounted for by each technology is shown. In this simple substitution process, two symmetrical S-shapedcurves are the result. The smooth lines are a "t from a logistic model (see GruK bler, 1990). The bottom right panel transforms the fraction data intoa format that converts S-shaped di!usion processes into straight lines by showing the logarithm of (F/(1!F)), ie, the ratio of the market share achievedby a technology over that remaining to be achieved.

Source: NakicH enovicH (1986).

another individual transport technology: the riding horseand the carriage. Looking at the absolute numbers ofdraft animals and cars in the USA (Fig. 5), we see that themillions of horses and mules used for transport practic-ally disappeared from the roads within fewer than threedecades. Interestingly, growth in transport services } ap-proximated by the growth in the sum of vehicles on theroad (horse carriages#cars) } rose smoothly and con-tinuously, largely una!ected by the "erce competition.The time constant (*t), which measures the time requiredfor a new technology to grow from 10 to 90% eventualmarket share, was only 12 yr, fast enough to traumatizethe displaced oat growers, coachmen, blacksmiths and(fatally) the horses. Similar time constants are observed

in other cases of single transportation technologies com-peting in a common infrastructure. For example, thedi!usion of modern low-emissions vehicles with catalyticconverters also occurred with the time constant of 12 yrin the USA (GruK bler, 1996; NakicH enovicH , 1986). Decade-long time constants also governed the replacement ofrailway rolling stock and substitution of steam by diesel-electric locomotives (GruK bler, 1990).

Typically many technologies compete, not just two.When competitors arrive in the market at di!erent mo-ments in time, the result is a sequence of S-shaped logisticcurves. For example, in steel manufacturing (Fig. 6) asmany as four technologies have competed simulta-neously, with varied time constants. The simple direct

A. Gru( bler et al. / Energy Policy 27 (1999) 247}280 257

Fig. 6. Competition among multiple technologies. Figure shows the share of steel production in the United States by "ve di!erent methods, 1850 tothe present, and curves (estimated from a logistic model) for the near future. In the 1940s and 1950s four di!erent technologies competedsimultaneously to supply the steel product.

Source: NakicH enovicH (1990).

replacement of crucible processes took less than twodecades. Its much superior replacement, the Bessemerprocess, enabled for the "rst time mass production ofhigh quality steel at low costs. Nearly seven decades wereneeded for the di!usion of electric arc steel in the USA,which was accompanied by other long-term changes inthe steel industry, such as increased use of (recycled)scrap in steel manufacture, the emergence of &&mini mills''and made-to-order small batch production. That longerreplacement process entailed far-reaching changes in in-dustrial structure, shifts to #atter economies of scale, andavailability of steel scrap as raw material instead of ironore.6

The S-shaped curves are useful for analyzing two sim-ilar but distinct processes } di!usion and substitution.The former refers to the spread of a technology intowholly new markets, often to provide services that pre-viously did not exist. Substitution is the replacement ofan existing technology with a new one, such as Bessemerfor crucible steel production or motorized cars for horse-drawn carriages. However, a "rm distinction between thetwo processes is often di$cult because both a!ect thesupply of and demand for a service. Often substitution ofan old technology by a new one yields lower prices andbetter performance; more e$cient supply of the serviceinduces greater demand. The expansion of automobile

6Analysis of economic determinants of di!erent di!usion rates foundthat ceteris paribus di!usion rates are higher (faster) for technologieswith higher relative economic advantage (performance, costs) and lowerinvestment requirements. See the classic studies of Mans"eld (1961,1968) and Mans"ed et al (1977).

travel in the United States, for example, is often analyzedas two overlapping processes. The "rst entailed the rapidsubstitution for horse carriages shown in Fig. 5, andsecond was the longer di!usion process during whichautomobiles co-evolved with the growing road networkand rede"ned American mobility. Similarly, the growthof railroads in the United States beginning in the 19thcentury is often cited as a case of pure di!usion } the newtechnology (railroad) and services it provided were novel.However, over a shorter period, the railroads also largelysubstituted a feeble competitor } stagecoaches } out ofthe business of inter-city transportation.

The competitive status of a technology is indicated notonly by changes in its market share but also by costs.Fig. 7 shows proxies for costs for the three main trans-port modes that have competed this century to supplyUS transportation services. As expected, costs declinedmost rapidly at the earliest stages of technological devel-opment when market di!usion was most rapid } forrailways in the 1870s (but data are scattered and timeseries are not available), roads for automobiles in the1900s and 1910s, and air travel in the 1930s and 1940s.(For more on transport infrastructures, see Fig. 10 anddiscussion below.) Fig. 8 illustrates the relationship inmore detail for one technology: the automobile and itssurfaced road infrastructure. The sales value of new auto-mobiles, which fell sharply as the automobile was com-mercialized, in tandem with the early growth of thenecessary surfaced road infrastructure. Ideally we wouldillustrate these principles with data on automobile oper-ating costs and levels of mobility, but neither was reliablymeasured before the 1950s.

258 A. Gru( bler et al. / Energy Policy 27 (1999) 247}280

Fig. 7. The cost of passenger transport in the United States has declined as the systems matured. As with individual technologies, the cost of transportsystems has declined most rapidly at the early stages. Nonetheless, continued cost reductions have been possible even in &&mature'' systems (eg,railroads) under severe competition from new, superior technologies (eg, air transport). (Operating costs, which are a better indicator of the economicattractiveness of the automobile, are available only since the 1950s; thus new car costs are used prior. Similarly, rail and air cost data would be betterindicators but are not available in comparable formats; thus revenue statistics are used as a good proxy.)

Source: GruK bler (1990).

Fig. 8. The co-evolution of automobile costs and the road network in the United States. Automobile and production methods improved, whichlowered the cost of new automobiles. At the same time the surfaced road network grew. These cross-enhancing developments occurred during theniche market commercialization and the early stages of pervasive di!usion (1905 to 1930). As automobiles became entrenched and served additionalfunctions (eg, icons of status); product life cycles shortened, and costs rose. Compared with niche market days, however, average sales costs have beenremarkably stable at approximately US$7000 to US$8000 (1990 dollars, adjusted with IMF GDP de#ators).

Source: GruK bler (1990).

Although a sharp drop in costs marks the onset ofpervasive di!usion, costs are not static once a technologyis in widespread use. For example, the length of therailway network reached 80% of its saturation by 1910;nonetheless, real costs per passenger-mile have since de-clined by half. Further improvements are still possible} for example, tra$c management systems are improvingrail speeds and load factors, thus sustaining competition

with roads and airways for passengers and freight. A dy-ing competitor does not exit without a "ght } a phenom-enon also called the &&sailing ship e!ect'' (Ward, 1967).When steam ships became competitive in the 1850s forocean-going freight and passengers, sail-powered boatsimproved signi"cantly. Indeed, the &&golden age'' ofClipper sailing ships emerged only once steam ships hadappeared as competitors; Montroll (1978) has estimated

A. Gru( bler et al. / Energy Policy 27 (1999) 247}280 259

from historical data that this e!ect may have delayed theextinction of sail ships by one decade. Although sailingships were slower and less competitive for passengertra$c they did preserve a niche in hauling lesstime-sensitive freight, including, ironically, coal forsteamers.

Despite variable time constants and substitution pro-cesses, the basic patterns of technological substitutionand di!usion are largely invariant across a large anddiverse set of historical examples. Theory suggests thatthe characteristic time constants of technology substitu-tion and di!usion vary depending on characteristics in-trinsic to new technologies (such as their costs, perfor-mance, social acceptability) as well as a function of moregeneral &&systemic'' characteristics of technology. For in-stance, ceteris paribus, di!usion can proceed faster whenreplacing an artifact that previously provided a similarfunction; technologies that provide an entirely new ser-vice di!use more slowly. The replacement of black-and-white TV sets by color ones, for example, proceeded in allmarkets much faster than the initial di!usion of black-and-white TVs.7

The di!usion model can be extended from competitionin a single market to separate or only loosely linkedmarkets. That phenomenon is common in technologicalanalysis because technologies are developed and tested insmall, core markets from which they di!use globally. Animmature new technology is more costly and requiresgreater investments to be competitive in its "rst market;the di!usion time is thus longer when compared with theperiphery. However, the longer period of establishmentallows the technology to have a larger e!ect on settingstandards that bar competitors and thus the penetrationlevel is higher in core markets. Because much technolo-gical knowledge is a free good peripheral regions can&&catch up'' rapidly to the core, as the technologies havebeen tried out already in core markets, and initial highcosts have already declined through R&D e!orts andlearning e!ects.

Historical analysis con"rms that the periphery gener-ally reaches lower adoption levels when compared withcore markets (E Schmidt, cited in Marchetti, 1983;GruK bler, 1991). For example, the automobile di!usedinto widespread passenger service "rst in the UnitedStates with a time constant (*t) of 60 yr. At its 1960s peakin this core market automobiles supplied over 90% of allmotorized passenger kilometers. In contrast, the di!u-sion rate was more rapid in Europe (*t about 30 yr), buttoday European automobile travel is saturating at about70% of total passenger kilometers. In countriesstarting motorization even later, the di!usion rate for

7Characteristic di!usion rates of black-and-white TV sets are 30 yrin OECD countries, whereas the replacement of black-and-white bycolor TV sets took typically only about half that time. See Steward(1982).

automobiles has been even more rapid } about twodecades in Japan and in Latin America, for example } butthe saturation level will probably also be lower. In theperipheral regions, automobile shares are lower in partbecause the next generation of transport technologies} high speed trains and aircraft } are already di!usingand substituting rapidly. In core regions, entrenchmentor &&lock in'' of dominant technologies can also limitcompetition and the development of new alternatives.8Hence, market shares are higher and saturation takeslonger.

Viewed together, early and late markets exhibita strong focusing of di!usion } as the technology reachessaturation the fraction (F) of saturation level narrowsbetween early and late markets. This phenomenon isillustrated in Fig. 9 for the major transport infrastruc-tures in industrialized countries during recent centuries.Ultimate saturation occurs in all markets over a relative-ly short common time period. Such synchronicity may beevidence and cause of world economic cycles (GruK bler,1991).

3.3. Infrastructures and technological clusters

Learning curves help modelers estimate performanceand cost improvements in the early stages of a techno-logy, and the di!usion model helps model the dynamicsthat result from selection of technologies in competitivemarkets. Individual and "rm-level decisions that result inthe selection of technologies are often so numerous andcomplicated that they can not be modeled individually.Learning curves and substitution/di!usion models allowsimpli"ed, but powerful and conceptually coherent,means of modeling technological changes and dynamics.But by themselves neither technique is complete. Neitherfully explains which technologies attract investments andsurvive to the market; nor does either tool internallydetermine the rates of change, such as *t, and extent ofdi!usion that prevails in the marketplace.

The fuller picture is partially completed by analyzingclusters of related technologies, not only the individualcomponents. Individual technologies cause externalities;as technologies grow synergistically, the network e!ectsincreasingly bind the fates of all technologies in thecluster. At the center of these clusters are &&infrastruc-tures'' } technologies that serve mainly the purpose ofsupplying externalities. The most obvious examples areenergy, transport, and communication infrastructures,without which numerous end-use technologies could notfunction. Since the cluster partially determines the selec-tion of its technological components, it also stronglyin#uences environmental impacts.

8For a more detailed discussion and model of technological &&lock-in'' see Arthur (1983, 1989) and David (1985).

260 A. Gru( bler et al. / Energy Policy 27 (1999) 247}280

Fig. 9. Focusing of di!usion between leading and laggard markets, industrialized countries. Technologies and infrastructures have co-evolved acrosscountries in industrial society. Three eras are evident, each de"ned by its principal transport infrastructure } canals, railways, and roads (cars). Thoseinfrastructures also correspond with three di!erent principal energy sources } wood, coal and oil, respectively. The initial stages of di!usion of eachtransport infrastructure are marked by high diversity in starting dates and di!usion speeds, which focus as the technology cluster matures. Thecountries indicated are those industrial nations for which di!usion was earliest (and slowest) and latest (and fastest). The >-axis shows the respectivesize of transport infrastructures as a fraction of their historical saturation levels (forecasted in the case of cars). The next technology cluster might bede"ned by aircraft and trains that operate at similar speeds, electricity and natural gas.

Source: GruK bler and NakicH enovicH (1991).

Fig. 10. The evolution of infrastructures in the United States. The "gure shows the length of the three major transportation infrastructures (canals,railways and hard surfaced roads) as a fraction of their total saturation levels. The total length grew approximately ten-fold for each new infrastructure.At their peak in 1851 canals extended over 6500 km. Railroads peaked at 490,000 km in 1929. Growth in the road network has slowed considerably,and with a logistic model we estimate that it will soon saturate at approximately 6 million kilometers. Analysis in terms of total saturation is the sameas application of the fraction concept (F ) introduced in Fig. 5. Also evident in the "gure is the co-evolution of compatible infrastructures: railways andtelegraphs; roads and oil pipelines.

Source: GruK bler and NakicH enovicH (1991).

The importance of infrastructures and of networks forlong-term technological evolution and economic growthis a well covered "eld (Schumpeter, 1934; Isard, 1942;Hughes, 1983; Anderson et al, 1984; Teubal et al, 1996).

Infrastructures, like other technologies, can be analyzedwith the di!usion model. Fig. 10 shows the case of trans-portation infrastructures. Canals, rails and roads haveeach evolved along a similar S-shaped dynamic pattern.

A. Gru( bler et al. / Energy Policy 27 (1999) 247}280 261

In the "gure the size of individual networks has beennormalized by the saturation level for each infrastruc-ture; the values shown are the fraction (F) of that level. Inabsolute values, surfaced roads were an order of magni-tude longer than the rail network, which was one orderof magnitude larger than the maximum length of thecanal system. In time, the three infrastructures arespaced rhythmically apart by a half-century or so. Theduration of growth (*t) has risen slightly with each new,longer infrastructure. Today, surface roads are saturat-ing; aircraft infrastructures are growing rapidly, thoughcomparable measures are not available: modern air-craft navigation, tra$c control and &&airways'' are adi!erent kind of infrastructure compared with railwaysand roads.

Infrastructures yield externalities } &&network e!ects''} that allow compatible technologies and infrastructuresto prosper while raising high barriers for products andprocesses that are incompatible. Components of thesenetworks thus co-evolve. Such externalities are also oftenfound between infrastructures. Fig. 10, for example, sug-gests that the railway and the telegraph evolved together} both used the same strips of land (rights of way).Synchronization is also evident in the evolution of theroad network and the oil pipelines that deliver fuel topower cars and trucks. Such technological interdepen-dence } or technology &&clusters'' } are responsible for thepervasive e!ects of technological change throughout theeconomy and society. For example, the emergence ofpipelines and petrol retailing has reinforced the domi-nance of petroleum-fueled automobiles and made it in-creasingly di$cult for incompatible electric vehicles tosupply automobility at the turn of the century. Only withthe rede"nition of automobile performance by requiringlower speci"c vehicle emissions has electricity re-emergedas a possible automobile propulsion. Technologicalinterdependence is also a source of considerable inertiain technological systems. In that respect, numeroustechnological systems are &&locked in'' to particularcon"gurations which are inherently di$cult todislodge within a short period of time. The constrainedevolution of technology clusters over time is also referredto in the literature as &&path dependence''. Models ofinduced innovation often describe technological changeas a cumulative process and thus can explain the phe-nomenon of path dependency (Binswanger, 1978; Ruttan,1996).

From a macro perspective, the network e!ects of in-frastructures create technological clusters. Infrastruc-tures lie at the center of technological clusters that de"nemajor eras in economic development } for example, theage of &&railways'' or the &&automobile era''. The di!usionof technologies shown earlier in Fig. 9 (for the industrial-ized countries) and in Fig. 10 (for the USA) are theinfrastructures that de"ne those eras } they are bothindicators and causes of the network e!ects that create

the cluster.9 Analysis at the level of individual countriesand markets is important because it is crucial to examinethe di!usion of technologies within the boundaries oftheir network e!ects, which typically coincide with theboundaries of coherent social system.

Three rather clear clusters can be distinguished, al-though the individual technologies entered the market atdi!erent dates and di!used with di!erent time constants.The "rst cluster saturated around 1865, the secondaround 1930, and the third is saturating now. Due to theoverlap of technologies between clusters, perhapsa fourth cluster of technologies has been already laun-ched, but is not yet clearly perceivable due to the insigni-"cant degree of di!usion achieved to date by its keyconstituent technologies. Looking back from the future,the new emerging cluster might be one day characterizedby dematerialization, the dominance of natural gas andelectricity infrastructures, and worldwide communica-tions that are too cheap to meter.

Technology clusters partially determine the time con-stants (*t) that govern the di!usion processes. Fig. 11shows the cumulative frequency distribution of *t for 265di!usion processes in the USA, based on studies per-formed at IIASA (117 cases) and 148 other well-documented examples from the literature (GruK bler,1991). The case studies include di!usion of energy, trans-port, manufacturing, agriculture, consumer durables,communication, and military technologies, as well aseconomic and social processes, such as literacy, reductionof infant mortality, and changes in job classes. The timeconstants of change range from very short-term processesof only a few years to processes that extend over two tothree centuries. The mean value of the time constants is41 yr, with a standard deviation of about equal size. Fewdi!usion processes extend more than one century andthus span more than one cluster. Half of the di!usionprocesses have *t of less than 30 yr; about three quartershave *t of less than 50 yr; and 93% of the sample exhibits*t of less than 100 yr. Analysis that distinguishes thetypes of technologies in the data sample suggests that theappropriate time constants for di!usion of infrastruc-tures and of entire technology clusters span up to a cen-tury (Fig. 10), with the main period of growth extendingbetween 5 to 6 decades. Most technologies di!usewithin the time-period of a single cluster. Only a few

9Technology clustering and network e!ects are also referred to astechnological interdependence in the literature. Technological interde-pendence can delay di!usion of new technologies that are either incom-patible with existing network e!ects, or need considerable time toestablish their own. The classic paper (and model) on this topic remainsFrankel (1955). Recently, similar concepts have been included in form ofa &&technological inertia'' variable in traditional macro-economic modelformulations with exogenous technology to analyze its impact onclimate policies and their timing (Ha-Duong et al, 1997), althoughapplication of the model is di$cult since the phenomenon of &&inertia''has not been crisply de"ned nor is it currently measured.

262 A. Gru( bler et al. / Energy Policy 27 (1999) 247}280

Fig. 11. The cumulative distribution of di!usion rates (*t) for 265 di!usion processes. The sample is the superset of di!usion processes documented inthe literature and those analyzed at IIASA. The mean rate is 41 yr and the standard deviation is 42 yr.

Source: GruK bler (1990).

technologies can tunnel from one to another cluster suchas some infrastructures.

In sum, technical change remains an extremely com-plicated process. The di!usion/substitution model canhelp simplify the modeling of that process and thus canaid in the projection of market shares and global changee!ects of speci"c technologies. It remains impossible todetermine } from theory or statistical analysis of histori-cal data } exactly what rates (*t) will prevail in particularcircumstances. But both theory and history can helpmodelers by bounding a range of plausible rates, whichTable 2 summarizes and illustrates. Those rates can beused for future projections; statistical methods and pack-ages exist for estimating di!usion parameters for a var-iety of models, from simple one-on-one competition tocompetition between multiple technologies (GruK bleret al, 1988; Marchetti and NakicH enovicH , 1979; Meyer,1994).

In general, revolutionary changes score high on each ofthe factors summarized in Table 2 and thus are markedby long di!usion constants. Such changes can be robustlyidenti"ed only with hindsight of several decades, whenthe cluster, including its infrastructures, has substantiallydi!used into use and the interdependencies are evident inpractice. A future hydrogen-powered economy will re-quire not only new end use technologies (eg, fuel cells) butalso infrastructures (eg, hydrogen pipelines) and interde-pendent technologies (eg, upstream hydrogen productionmethods). Individually some of those elements may beable to di!use rapidly }for example, today there is sub-stantial production of hydrogen for use in re"neries } butthe system as a whole can emerge only at the pace of thenecessary infrastructure, ie over a time span of several

decades. Indeed, the historical record shows that thelargest technological systems di!use with long time con-stants (6 to 8 decades).

4. Implications for global change

The typology and basic observations on patterns oftechnological change evident in the historical recordallow at least three improvements to the way thatanalysts model technological changes and their environ-mental impacts. One improvement is based on a long-term pattern that is directly evident in the historicalrecord } the existence and pace of decarbonization. Theother two concern modeling. At the micro level it ispossible now to endogenize rigorously the process oftechnological change due to (uncertain) learning. At themacro level it is possible to employ modeling techniquesthat stylistically endogenize technological change, butonly if conventional macroeconomic models are linkedwith models that allow resolution in the choice of par-ticular technologies. None of these three improvements isfully re#ected in the mainstream models and scenariosthat are being used to analyze global change.

4.1. The historical record: decarbonization

The existence of technological clusters partially deter-mines the selection of energy sources and thus environ-mental e!ects. When fuel wood and animal feed were theprime sources of energy, people traveled mainly by foot(walking and horses) and boat (canals). During the age ofcoal, railways provided most mobility; today, in the age

A. Gru( bler et al. / Energy Policy 27 (1999) 247}280 263

Tab

le2

Exa

mpl

esof

Di!

usion

pro

cess

es

Dura

tion

del

ta¹

Exa

mpl

eD

i!usion

(D)

Ref

eren

ceR

elat

ive

Sca

leIn

fras

truct

ure

Tec

hni

cal

(yr)

Subst

itution

(S)

adva

nta

genee

dsin

terd

epen

den

ce

Long

110

Coa

lvs

.tr

aditio

nal

ener

gy,W

orld

SN

akicH

enovi

cH(1

984)

##

##

##

##

##

#

'40

yr80

Coal

vs.tr

aditio

nal

ener

gy,U

SA

SN

akicH

enovi

cH(1

984)

##

##

##

##

##

#

60G

row

thofra

ilw

ays,

World

DG

ruKble

r(1

990)

##

##

##

##

##

##

47G

row

thofra

ilw

ays,

Fra

nce

DG

ruKble

r(1

990)

##

##

##

##

##

#

55St

eam

vs.s

ailship

s,W

orld

SG

ruKble

r(1

991)

##

##

##

##

#

70O

pen

-hea

rth

stee

lmak

ing,

Wor

ldS

Nak

icHen

ovi

cH(1

990)

##

##

##

#

55O

pen

-hea

rth

stee

lmak

ing,

USA

SN

akicH

enovi

cH(1

990)

##

##

##

44Ele

ctri"ca

tion

ofho

mes

,U

SA

DU

SD

OC

(197

5)#

##

##

##

##

#

Med

ium

25%

house

hold

sw

ith

radi

o,U

SAD

US

DO

C(1

975)

##

##

##

##

20}40

yr26

Air

vs.r

ailin

inte

rcity

trav

el,U

SAS

Fisher

and

Pry

(197

1)#

##

##

##

#

28M

echa

niza

tion

inco

alm

inin

g,U

SSR

SA

stak

hov

etal

(199

0)#

##

##

#

25Bas

icox

ygen

stee

lfu

rnac

e,W

orld

SN

akicH

enovi

cH(1

990)

##

##

##

20Bas

icox

ygen

stee

lfu

rnac

e,U

K,U

SAS

Nak

icHen

ovi

cH(1

990)

##

##

#

27R

ailw

aytr

ack

elec

tri"

cation

,U

SSR

DG

ruKble

r(1

991)

##

##

##

#

26C

hem

ical

pres

erva

tion

ofra

ilway

ties

USA

SG

ruKble

r(1

991)

##

##

##

Fas

t18

Car

air

condi

tion

ers,

USA

DN

akicH

enovi

cH(1

986)

##

##

#

(20

yr19

Air

condi

tion

ers

inhom

es,Ja

pan

DButtner

ran

dG

ruble

r(1

995)

##

##

#

16A

utom

obi

les

vsca

riag

es,U

KS

GruKble

r(1

990)

##

##

##

##

15C

ars

vs.hor

ses,

Fra

nce

SG

ruKble

r(1

990)

##

##

##

##

12C

ars

vshor

ses,

USA

SN

akicH

enovi

cH(1

986)

##

##

##

##

16C

ars

vshor

ses,

UK

&F

rance

SG

ruKble

r(1

990)

##

##

##

#

Die

sel/el

ectr

icvs

.st

eam

12Loco

motive

s,U

S,U

SSR

,U

KS

GruKble

r(1

990)

##

##

15Tra

nsisto

rsvs

.va

cuum

tube

sin

radi

os,

USA

SSt

ewar

d(1

982)

##

##

#

15Bla

ckan

dw

hite

vs.co

lor

TV

,U

SA

SSte

war

d(1

982)

##

##

#

9W

ashin

gde

terg

ent

vs.s

oap,

USA

SFisher

and

Pry

(197

1)#

##

#

Not

e:St

yliz

edsu

mm

ary

ofd

i!usion

rate

san

dth

eirde

term

inan

ts.T

he

rate

ofd

i!us

ion

(*t)

ofte

chnol

ogi

esis

det

erm

ined

by

man

yfa

ctors

.Fourm

ajorones

aresu

mm

ariz

edin

the

tabl

eth

at,o

ther

thin

gsbei

ng

equal

,hav

eth

efo

llow

ing

e!ec

ts.(

a)R

elat

ive

adva

ntag

e,co

mpr

isin

ga

variet

yof

dim

ensionsin

clud

ing

engi

neer

ing

(eg,

perfor

man

ce,p

rodu

ctiv

ity)

,eco

nom

ic(e

g,re

lative

cost

s,pro"ta

bili

ty),

and

soci

al(e

g,ea

seofa

doption

and

use

).The

high

erth

ere

lative

adva

ntag

eof

ane

wte

chno

logy

com

pare

dto

existing

ones

,the

high

er(fa

ster

)will

cete

rispar

ibusbe

itsdi!

usion

rate

.(b)&&S

ize,''

com

prisin

ga

variet

yofd

imen

sionslike

geogr

aphic

alsp

read

(eg

loca

lvs.

global

),m

arket

size

and

adopt

ion

insp

ecia

lized

appl

icat

ions(e

g,pr

oce

sste

chno

logi

esin

spec

ializ

edin

dust

ries

)vs.

perv

asiv

ead

option

(eg,

inev

ery

hous

ehold

).C

eter

ispa

ribus,

larg

ersize

impl

ieslo

nge

r(slo

wer

)di!

usio

nra

tes.

(c)I

nfra

stru

cture

nee

ds(n

etw

ork

exte

rnal

itie

s),w

hic

him

ply

cete

rispar

ibuslo

nge

r(slo

wer

)di!

usion

rate

s.A

lthou

ghfreq

uently

anim

port

antove

rlap

exists

betw

een

infras

truc

ture

san

dla

rge

tech

nic

alsy

stem

s,not

allla

rge-

scal

ete

chnol

ogi

esre

quire

new

infras

truct

ures

.Forin

stan

ce,n

ucl

earpow

erw

asab

leto

di!use

rela

tive

lyra

pid

lyby

pro"ting

from

anel

ectr

icity

tran

sport

and

distr

ibut

ion

infras

truct

ure

larg

ely

alre

ady

inpla

ce.(

d)I

nter

dep

enden

cew

ith

other

tech

nolo

gies

.Cet

eris

paribu

s,th

ehig

herte

chno

logi

cal

inte

rdep

enden

ce,th

eslow

erin

div

idual

tech

nolo

gies

di!

use

.The

latt

ertw

ofa

ctors

that

are

often

inte

rrel

ated

are

freq

uen

tly

term

ed&&ne

twork

e!ec

ts.''

For

each

illus

trat

ive

exam

ple

ofte

chno

logi

cal

di!

usion

orsu

bstitu

tion,

the

auth

orsha

vequ

alitat

ivel

ysc

ore

dea

chofa

bove

in#ue

ncin

gfa

ctors

rang

ing

from

low

(#),

med

ium

(##

),to

high

(##

#).

Cet

eris

paribus,

mor

e&&#''

sym

bols

lead

tolo

nge

r(slo

wer

)di!

usion

rate

,but

also

indic

ate

more

per

vasive

impac

tsofte

chnol

ogi

calch

ange

.Our

appro

ach

and

clas

si"ca

tion

ishi

ghly

styl

ized

,as

few

ofth

ese

conc

epts

have

bee

nre

pre

sent

edby

com

para

ble

indic

ators

forw

hic

hdat

aar

eav

aila

ble

.Forea

chill

ustr

atio

nth

eta

ble

also

indi

cate

sw

heth

erth

epr

oce

ssis

pred

om

inan

tly

di!

usion

(D)o

fthe

tech

nol

ogy

into

anew

serv

ice,

orsu

bstitu

tion

(S)fo

ran

existing

tech

nolo

gy.H

ow

ever

,in

mos

tca

ses"rm

clas

si"ca

tion

isdi$

cult

(see

text

for

expla

nat

ion).

Sub

stitut

ion

pro

cess

esar

eoften

more

rapid

than

di!

usio

nbe

cause

they

typic

ally

occ

urw

ithin

ate

chno

logi

calsy

stem

that

isco

mpat

ible

with

bot

hth

eold

and

new

(subst

ituting

)te

chnolo

gy.

264 A. Gru( bler et al. / Energy Policy 27 (1999) 247}280

Fig. 12. Selection of fuels progressively lighter in carbon. Figure shows the primary energy supply in the United States in familiar units (million tons ofoil equivalent, Mtoe). For each major fuel its fraction (F) of total energy supply was calculated, and then transformed as in Fig. 5 so that logisticsubstitution processes appear as straight lines. The result, shown in the bottom "gure, makes it easier for the human eye to identify those processes atwork. US data are used because they are the most complete.

Sources: NakicH enovicH (1984) and GruK bler and NakicH enovicH (1991).

of oil, automobiles, buses and aircraft provide unprece-dented levels of individual mobility. When people andmaterials traveled by foot, most energy was consumed atthe point of gathering; in contrast, today, most energy isitself transported. Increasingly, energy is converted intoa readily usable form (eg electricity) prior to transporta-tion. The share of "nal energy supplied by electricity isgrowing rapidly in most countries and worldwide, withno sign of saturation (Ausubel and Marchetti, 1996). Theenvironmental problems of at-source energy gatheringwere localized, such as deforestation adjacent to roads

and settlements. Today the reach is global, and part ofenvironmental concern is focused on the transport offuels themselves, such as oil spills and electromagnetic"elds.

Fig. 12 shows the shares of total primary energy in theUnited States supplied by di!erent fuels (top panel) andestimates for the consumption of major primary fuelsworldwide (bottom panel). We use US energy data oftenin this survey and review paper because the statisticallydocumented record is longest and thus technologicaldynamics most evident. As with the evolution of

A. Gru( bler et al. / Energy Policy 27 (1999) 247}280 265

Fig. 13. Decarbonization of the energy supply. The ratio of carbon per unit of energy has steadily declined in the United States and the World. Datafor this "gure and Fig. 12 include estimates for fuel wood and feed (biomass) that were major energy sources until the middle/late 19th century. We aremindful that their carbon intensity depends on the mode of production; for the purpose of calculating decarbonization we make the reasonableassumption that production of these fuels (mainly fuel wood) was not through sustainable harvesting and thus net released carbon. Even if these fuelsare omitted and only fossil fuels are included in the analysis, the strong trend of decarbonization since the rise of coal in the early 19th century is stillevident. Source: GruK bler and NakicH enovicH (1996).

particular technologies and infrastructures, the di!usionof energy sources follows a similar pattern. The marketshare of a fuel expands initially slowly as gradually ex-panding niche markets are "lled; a more lengthy processof pervasive di!usion follows, eventually saturating anddeclining as a superior competitor enters and di!uses.Despite the simultaneous competition and interactionamong many di!erent sources of energy, each historicalperiod is characterized by a clear dominance of a singleenergy source, which corresponds with the main techno-logical clusters: fuelwood (and feed), followed by coal,and later by oil and natural gas. The rates for di!usionalso correspond } all of the major fuels (wood, coal, oil,gas) required a century to achieve their ultimate satura-tion. The time constant for nuclear is not yet evidentbecause it has low market share. Planned nuclear instal-lations have fallen radically, which suggests that theenergy source may be saturating. A new pulse for di!u-sion of nuclear power may require a new concept forreactors and their attendant cost and waste disposalproblems. The existing nuclear power capacity di!usedinto service relatively rapidly because it could employ anexisting large-scale infrastructure (ie, the electrical grid).

The consequence of shifting fuels is one of the moststriking and consistent trends in the energy system: de-carbonization. Each of the main fossil fuels has beenprogressively lighter in carbon per unit of energy releasedwhen burned. Coal, the dominant fuel from 1880 to 1950,contains approximately one hydrogen per carbon atom.Today's dominant fuel, oil, contains two hydrogen atomsfor every carbon. The hydrogen to carbon ratio in gas,the primary fuel whose share is rising most rapidly today,

is four to one. More recently energy sources that have nodirect carbon emissions such as hydropower and nuclear"ssion lighten our carbon diet. Hydrogen-rich fuels re-lease more energy for every carbon atom that is oxidizedto CO

2during combustion.

As shown in Fig. 13, the carbon to primary energyratio for the USA has declined about 0.25% per yearsince 1800. Worldwide, decarbonization of energy sour-ces has occurred at a similar pace (0.3% per year since1850). In some countries decarbonization has been rela-tively rapid } in France, for example, since 1950 nuclearpower has replaced fossil fuels as the supplier of nearly allelectricity. Decarbonization of the economy has beeneven more dramatic because the decarbonization of en-ergy sources (grams carbon per unit energy) has beenmultiplied by the decline in energy intensity of the econ-omy (unit energy per dollar of economic output). Nearlyevery country's economy has shifted from energy-inten-sive industries to services; in addition, e$ciency withwhich useful energy is generated from primary fuel sour-ces has also risen in most countries over most periods oftime. In the United States, the speci"c carbon dioxideemissions per unit of economic output has declined byalmost an order of magnitude during the last two centu-ries, from about 2.5 kg of elemental carbon (C) per $1GDP (at 1990 prices) in 1800 to about 0.3 kg C per $1GDP in 1990. In other terms, the economic &&productiv-ity'' of carbon increased by more than 1% per yearduring the last two centuries.

Decarbonization has not autonomously eliminatedcarbon from the economy, but it has steadily softened theeconomy's impact on the atmosphere. For example, if the

266 A. Gru( bler et al. / Energy Policy 27 (1999) 247}280

fuel mix of the 1920s powered the American economytoday then annual carbon emissions would be more than300 million metric tons (24%) higher. For comparison,those averted emissions total more than today's annualindustrial carbon emissions from Argentina, Brazil,Canada and Mexico combined.10

Although decarbonization is strongly evident in thehistorical record, most baseline scenarios project the op-posite } stagnation or even re-carbonization with risingmarket shares for coal, which is the most carbon-inten-sive primary fuel. We return to those scenarios, especiallythe widely used IS92a baseline scenario of the Intergover-nmental Panel on Climate Change (IPCC), after discuss-ing the types of models that are often the foundation ofscenario analysis. Technological change is poorly ad-dressed in those models, and thus it is no surprise thatmodel estimates of future decarbonization rates } whichis one indicator of technological change } are not consis-tent with the historical record.

4.2. Modeling the selection of technologies andtheir environmental impacts

Historical data indicate possible futures } such asdecarbonization } but projecting the future by merelyextending past historical trends is rarely adequate, espe-cially when the underlying causal process are complexand liable to change. Thus causal models are often usedto project scenarios for the future and to analyze policychoices. However, such models depend on good know-ledge of the underlying processes. Because most energyanalysts assume that technological change is hopelesslycomplex and shrouded in uncertainty, few rigorouscausal models have been applied. Here we argue that theprocess of technological change is su$ciently well under-stood to make possible plausible models that endogenizetechnological change. Mathematical and computationalbarriers remain, but we illustrate the argument with onepromising micro-scale approach that rigorously en-dogenizes the dynamic evolution of three electricity gen-eration technologies. In the next section we address thearguments on a macro scale } in modeling and buildingscenarios for the world energy system } and show thateven there better, though much more stylized, treatmentof technological change is possible.

The controversy that surrounds modeling of technolo-gical change is endemic not only to energy and environ-mental assessments but also in economic analysis moregenerally. Treatment of technological change lies at thecenter of the debate between advocates of &&old'' and

10Data from the Carbon Dioxide Information Analysis Center(CDIAC), Oak Ridge National Laboratory, Oak Ridge, USA. Valuesare expressed in tons of carbon; because those emissions are in the formof carbon dioxide, some studies report data in tons of carbon dioxide.To convert, multiply by 44/12 (3.67).

&&new'' growth theory.11 Both sides agree that technologyis central to long-term growth, but the mechanisms, andhow to include them in formal models continue to bedisputed. As Joseph A. Schumpeter observed long ago (in1934), technological change arises from within the eco-nomic system and is central to its growth. In contrast,most models consider technological change as exogenous} in an expression coined by Bill Nordhaus, these modelsassume that new technologies fall into service like &&leavesfrom autumn trees''. In the "eld of global change, resolv-ing these controversies and including endogenous tech-nological change in models is especially important sincethe objective of this research } projecting plausible levelsof future emissions, and assessing policies for controllingthem } all depend principally on technology.

While the processes that lead to inventions remainlargely opaque, it is now possible to model the factorsthat a!ect which innovations are selected for investmentsthat lead to niche market applications, commercializ-ation, and widespread adoption of technologies. Theapproach rests on two fundamental reasons for whyprivate "rms and the public (through government) investin the pursuit of new technologies. (1) Investments intoRD&D and practical experience in commercial nichemarkets leads to learning, which improves costs andperformance and thus makes technologies more competi-tive and o!ers pro"t opportunities for economic agents.(2) Uncertainty about the future } such as level of de-mand for technology services (eg, electricity generation),costs, and learning rates } lead "rms and societies tohedge risks by investing in portfolios of new technologieswith potentially useful attributes (Mans"eld et al., 1972,1977).

Despite overwhelming empirical evidence and solidtheoretical underpinnings, learning phenomena } as dis-cussed above } have been explicitly introduced into onlya few models of intertemporal choice. A "rst detailedmodel formulation was suggested by Nordhaus and vander Heyden (1983) to assess the potential bene"ts ofenhanced R and D e!orts in new energy technology (thefast breeder reactor). Computational limitations at thattime precluded the application of the model to a widerportfolio of technologies in the model formulation. A "rstfull-scale operational optimization model incorporatingsystematic technological learning was "rst developed atIIASA (Messner, 1997; NakicH enovicH , 1996). Conventionallearning curves for a number of advanced electricitygenerating technologies were introduced into a linearprogramming model of the global energy system. Be-cause learning rates were assumed ex ante, future techno-logy costs in the model depended solely on the amount ofintervening investments in installed capacity. Given spe-ci"c (known) learning rates, the model determinedthe optimal investment pro"le that would yield future

11See references in note. 2.

A. Gru( bler et al. / Energy Policy 27 (1999) 247}280 267

technology improvements via learning by doing. (RD&Dthat did not directly result in an increase in installedcapacity was excluded as a source of technologicalchange in the model.) That model con"rmed the expecta-tion that it can be economically optimal to invest intocostly, immature technologies in anticipation of improve-ments through technological learning. Absent such priorinvestment, new technologies did not become competi-tive in the future } rather, the technological system re-mained &&locked-in'' to existing technologies, productivitydid not increase, and the portfolio of available cost-e!ec-tive technologies remained limited and insu$cient toaddress possible long-run resource or emissions con-straints. In such cases, the only options for addressingthose futures in such a technology optimization model(and in real society) entailed resorting to currently knownexpensive technological alternatives (&&backstops''), lead-ing to economic losses, such as lower consumption andwelfare.

Today, few of the main models used to assess the leveland costs of controlling future carbon emissions rigor-ously employs a learning curve formulation. Thus few ofthe models internally generate the investments needed toyield new commercial technologies. Rather, future costsare assumed as exogenous model inputs and are indepen-dent from any intervening e!orts and costs.12 We suggestthat if technological change is to be adequately includedthen models must explicitly link investment decisions tochanges in technology characteristics. The learning curveconcept is one simple device for doing this. But the simpleapplication of learning curves confronts two shortcom-ings. One is that only traditional learning curves havebeen employed in these models and thus cost improve-ments due to RD&D have been ignored. As shownearlier, especially at the stages of innovation and nichemarket commercialization, RD&D is a signi"cant con-tributor to improved cost and performance. The other isthat the empirical literature suggests a range of learningrates that might be expected, but exact rate for particulartechnologies and processes remains of course uncertain,and thus the learning rate } ie, returns from investments} must be modeled as uncertain.

Uncertainty is pervasive when "rms and societiesmake technological choices and thus must be re#ectedthroughout models used in technological forecasting.The importance of technological uncertainty has beenrecognized and explored ever since the earliest days ofglobal environmental modeling (eg, Nordhaus, 1973;Starr and Rudman, 1973). Di!erent approaches havebeen followed for analyzing the impacts of technological

12Some models deploy additional exogenous constraints to &&force''models into earlier investments into expensive technologies beforecheaper ones become available. See, for example, Manne and Richels(1997).

uncertainty including the formulation of alternative scen-arios (eg, NakicH enovicH et al, 1995), analysis of modelsensitivity to input parameters (eg, Nordhaus, 1973; Nor-dhaus, 1979), and sensitivity analysis based on expertpolls or Delphi-type methods (eg, Manne and Richels,1994). Yet in practice these studies were principally basedon adaptations of the deterministic optimizationmodels.13

That modeling approach has allowed assessment ofthe sensitivity of models outcomes to variations in uncer-tain model input parameters or some exploration ofoptimal management strategies, such as strategies thatemphasize adjustment of policy decisions over time asuncertainties are resolved (Hammit et al, 1992; Lempertet al, 1996) (more generally see Holling, 1978, and Wal-ters, 1986). But the deterministic framework is unable toidentify investment strategies that are robust or &&opti-mal'' when decision-makers simultaneously face manyuncertain choices because uncertainty (stochasticity) isnot built into the decision framework. In the model ofendogenous technological change explored here, uncer-tainty translates into both economic risks and oppor-tunities, and both are directly endogenized into themodel's decision rules and the resulting technology strat-egies.

True endogenization of uncertainty and learning phe-nomena has required the application of models that aremathematically cumbersome involving simultaneouslystochasticity and recursive formulations. Moreover, tech-nological learning is a classical example of increasingreturns, ie, the more learning takes place, the bettera technology's performance. Thus the mathematical solu-tions are non-convex } the more investment the lower thecosts } which is especially di$cult to handle in tradi-tional optimization models and algorithms. A generalmethodology for handling uncertainty in optimizationproblems through a stochastic sampling technique wasdescribed in Ermoliev and Wets (1988) and an improvedalgorithm was developed by Ermoliev (personal com-munication). That stochastic optimization frameworkhas been applied in the MESSAGE technological optim-ization model for several technologies by the IIASAenergy research group.14 Similar mathematical and

13For example, Manne and Richels (1994). Examples of modelsincorporating stochasticity include Kolstad (1993) and Fragniere, andHaurie (1995).

14See Golodnikov et al (1995), Messner et al (1996) and Messner(1997). Fragniere and Haurie (1995) have applied a similar frameworkusing the MARKAL optimization energy model. The two stochasticmodeling approaches are however di!erent. The stochastic MARKALmodel takes the classical stochastic approach towards decisions underuncertainty: assuming that at some future date, uncertainty (on techno-logy costs, environmental limits, etc.) becomes resolved. The stochasticMESSAGE model does not rely on the assumption that at a given dateuncertainty disappears altogether. Rather the model calculates an opti-mal diversi"cation strategy in face of persistent uncertainty by integrat-ing stochastically drawn samples into an overall objective function.

268 A. Gru( bler et al. / Energy Policy 27 (1999) 247}280

conceptual problems also have been partially addressedin work on &&path-dependence,'' in which initial decisionslead to multiplying network e!ects and increasing re-turns from the adoption of a particular technologicalsolution (Cowan, 1991).

4.2.1. A model of technological change through uncertainlearning

The model with endogenous technological change de-veloped at IIASA overcomes the main failings of existingapproaches } the inadequate representation of learningas a consequence of both RD&D investments and cumu-lative commercial experience (learning by doing), as wellas the failure to incorporate uncertainty in models oftechnological learning (GruK bler and Gritsevskii, 1997).The intention is to extend the traditional approach tomodeling energy technologies in which technical changeis induced only by relative resource and factor endow-ments and price changes.15 In contrast, the IIASAapproach yields an optimization model that can makeforward-looking investments that are needed to yield(uncertain) learning and cost reductions for particulartechnologies. Thus new technology may be renderedcompetitive and is selected for widespread commercialapplication. Traditional optimization models make nosuch anticipatory investments.16

The conceptually simple model represents a demandfor one homogenous good, electricity. Total demandrises over time in proportion to total demand in thewidely used IPCC and IIASA/World Energy Council(IIASA/WEC) &&high demand'' scenarios.17 There existsone primary resource (coal), whose extraction costs in-crease over time as a function of resource depletion, whilebeing su$ciently large that absolute resource scarcitiesare not encountered over the entire simulation horizon(200 yr).

Three technologies } &&Mature'', &&Incremental'', and&&Radical'' } are available that can supply the demand forelectricity. The stages of technological development forthose correspond with the same-named stages in Table 1.

15 eg, Hayami and Ruttan (1985); for overviews see Binswanger(1978), Jorgenson and Fraumeni (1981) and Ruttan (1996).

16Linear programming models typically include constraints on therate at which a new technology can enter the marketplace; without suchconstraints the model would instantly switch from one technology toanother as soon as the latter is less costly (for additional discussion, seeFragniere and Haurie (1995)). Thus the optimal solution to thesemodels often entails earlier investments into new technologies so thatrate constraints are met. Such investments are the consequence of themodel formulation, not because the model anticipates (as does a pro"t-seeking "rm) that an investment will make a technology more competi-tive in the future.

17Demand is scaled to economic growth in the Series A scenarios ofthe World Energy Council (WEC) and the IS92e and f scenarios of theIntergovernmental Panel on Climate Change (IPCC). For IPCC seePepper et al (1992); for IIIASA-WEC see NakicH enovicH et al (1995).

The current costs of each technology are assumed to beknown perfectly, but the potentials for future learningvary markedly as a function of the uncertainty in learningrates.18 The Mature technology is assumed to be inwidespread di!usion; its characteristics (costs and re-source conversion e$ciency) do not change over time.The Incremental technology has a slight e$ciency ad-vantage over the mature technology, is initially morecostly (by a factor of 2), and has a potential mean learn-ing rate assumed at 10% for each doubling of cumulativeproduction capacity. The Radical technology hardly re-quires any resource inputs and thus o!ers a substantiale$ciency premium. Because it is a radical innovation it isalso a factor of 40 more costly than the Mature techno-logy. However, as common with radical new technolo-gies, it has high potential for technological learning } weassume a mean learning rate of 30% (per doubling ofcapacity), which is consistent with historical examples ofradical technological change reviewed earlier. For boththe Incremental and Radical technologies the rates oflearning are treated as uncertain, represented by a log-normal distribution function around the mean value. Forthe Radical technology the variance of the distributionfunction is three times that of the Incremental techno-logy, re#ecting the much higher spread of uncertaintyassociated with radical technologies.

These three technologies correspond with three prom-inent electricity technologies available today. Currentcoal-"red power stations, with thermal e$ciency of 30%,and costs of US$1000 /kW(e), are a typical mature tech-nology. An incremental improvement is represented byan advanced coal-"red power station (eg, with #uidizedbed boilers and all basic environmental equipment suchas precipitators and SO

xand NO

xremoval units, but not

CO2

removal), with thermal e$ciency of 40% and initialcosts of US$2000/kW(e). The coal resource in our modelis consistent with known world reserves and resources ofcoal, which exceed 200 years and exhibit rising extractioncosts as the most accessible and highest quality reservesare depleted. The Radical technology corresponds withphotovoltaic (PV) cells, which use a basically free re-source (sunlight) but are initially very costly (we usea starting value of US$40,000 USD/kW(e) characteristicof PV costs in the early 1970s).

For each run in a large sample N the stochastic modelsamples a value for each of the uncertain parameters. The

18Current prices of new technologies (which are costly but declinewith learning) are assumed to be known perfectly (the market providesfor the mechanism to reduce any uncertainty on current costs: actualinvestment). The model treats future learning rates as uncertain. Hencefuture technology costs are a function of (uncertain) learning rates andthe resulting intertemporal optimum investment pro"le the modeldetermines for any given (stochastic draw) of the learning rate. Nofurther exogenous cost inputs enter the model. The model also does notrequire exogenous speci"cation of a future &&#oor'' price, ie a lowerbound to which costs decline as a result of technological learning.

A. Gru( bler et al. / Energy Policy 27 (1999) 247}280 269

Fig. 14. A simple optimization model with endogenous technological change. The "gure shows the fraction of total installed generating capacitysupplied by alternative new electricity generating technologies: Incremental (dashed lines) and Radical (solid lines). These di!er in their current costs aswell as possibilities for future cost reductions (learning). Four simulation runs are shown: (1) static technology costs and performance for alltechnologies, which yields technological change only in response to resource depletion; (2) exogenously given cost improvements for Incrementaltechnologies; (3) certain (deterministic) learning rates for Incremental and Radical technologies; and (4) uncertain learning rates for Incremental andRadical technologies. Without endogenous learning, the Radical technology remains at zero market share. When the rate of learning is certain theoptimal solution is to invest heavily and early into new technologies because the resulting cost declines render the technology quickly competitive.When learning rates are uncertain, the optimal solution is more cautious. Market shares for the existing, Mature technology are not shown on this"gure. Source: GruK bler and Gritsevskii (1997).

model then calculates the minimum cost to build theinstalled capacity for each of the three technologiesneeded to satisfy the total demand for electricity. (Forsimpli"cation, we assume 100% capacity utilization forall three technologies.) That calculation includes the opti-mal investment pro"le that yields future cost reductionsin the Incremental and Radical technologies, which isnecessary to make those new technologies competitivewith the existing Mature technology. Because the learn-ing rates are stochastic, the future RD&D and invest-ment costs needed to make the incremental or radicaltechnology competitive in the future is also a stochasticfunction of the intervening cumulative investments. Thusactual costs in any run of sample N may deviate from theexpected costs, which arise by applying the mean learn-ing rate.

The optimal solution for the entire sample is calculatedby "nding the minimum of the overall objective function,which integrates all the realized outcomes from each ofthe stochastic draws in sample N. Because each indi-vidual draw deviates from the expected model outcomewe need a weighting procedure for integration of all thesedi!erent realized individual model solutions into theoverall objective function. The overall objective functionincorporates the objective function of the mean (ex-pected) value expanded to include two additional terms.(1) If the realized costs in a draw are less than expected,the resulting cost di!erence (savings) is added to theobjective function (i.e. subtracted). (2) If costs exceedexpectations, the resulting cost di!erence (cost overrun)

is also added to the objective function, but quadratically.This re#ects our assumption that underestimating futurecosts is penalized more heavily in competitive marketsthan cost overestimation. Cost underestimation (leadingto higher costs than expected and higher than one'scompetitors) risks the very survival on the market,whereas cost overestimation, which leads to lower futuretechnology costs than expected yields merely higherpro"ts than expected. In other words, our model is con-servative } it assumes that commercial decision-makersgive paramount value to survival. The model performsa simultaneous sampling of parameter values. All realiz-ed outcomes (plus and minus the two additional termsexplained above) of the individual draws of sample N areintegrated into the overall objective function. The result-ing overall minimum solution represents the optimaltechnological diversi"cation strategy vis aH vis the (uncer-tain) economic returns on learning that result from R&Dand investment into technology demonstration and nichemarkets.19

Fig. 14 shows results } the share of electricity suppliedby each technology } for four of the simulations. Iftechnology is treated as static (no learning) then generallyneither the Incremental nor the Radical technology isselected, and 100% of the market is supplied by the

19For each draw, the model solves simultaneously the non-convexand non-smooth optimization problem by applying a combination ofa simple global procedure, a modi"ed Nelder-Mead algorithm, anda BFGS quasi-Newton minimization (see GruK bler & Gritsevskii, 1997).

270 A. Gru( bler et al. / Energy Policy 27 (1999) 247}280

Fig. 15. Endogenous technological change with carbon constraints and other uncertainties. Figure shows new capacity additions for three simulationruns: (1) Base Case with cost improvements due to uncertain rates of learning (same as run 4 in Fig. 14); (2) addition of an uncertain carbon constraint(see text); and (3) addition of the possibility that demand for electricity will be substantially higher or lower than expected in the Base Case simulation.

Source: GruK bler and Gritsevskii (1997).

Mature technology. The "rst simulation (labeled &&statictechnology'') portrays technological change as in modelsthat employ the concept of a &&backstop'' technology. TheIncremental technology di!uses eventually not by im-proved performance but as a result of resource depletionthat raises energy prices, making the Incremental techno-logy commercially viable due to exogenous develop-ments. A second run, which we term &&exogenous learn-ing'', assumes that the cost of the Incremental technologyfalls at an exogenously determined rate. The result isa pattern that is again typical of optimization modelsthat employ exogenous technological change: at somepoint in the future a new technology massively enters thepicture due to an exogenously assumed cost reduction.Large-scale technology optimization models, which arewidely used to assess the costs of abating various envir-onmental problems, display similar &&#ip-#op'' behavior.Published runs typically do not illustrate such behavioronly because additional constraints or restrictions on therate and pattern of technological di!usion, tuned accord-ing to the modelers sense of the historical record, areapplied to make the outputs appear more realistic. Likesausage, the "nal product is evidently wholesome but themethod of producing tasty results is best left shrouded inmystery.

A third run in Fig. 14 shows the e!ect of addinglearning with zero uncertainty }what we term &&determin-istic (certain) learning''. A fourth run illustrates &&un-certain learning'' } it fully incorporates uncertainty(stochasticity) and learning, and thus is the full model ofendogenous technological change. Note that the addition

of learning in both of these runs leads the Radical techno-logy to enter the market. In the deterministic case with nouncertainty a new technology enters the market earlierand di!uses faster. When learning is uncertain, di!usionis more gradual and market entry is later. Only the lattercase exhibits the S-shaped di!usion patterns evident inthe historical record. Both learning and uncertainty areessential to building a model that endogenously producesresults that re#ect how technologies enter and di!useinto markets in the real world.

Fig. 15 shows the results when an uncertain possibilityof a constraint on emissions } represented by a carbontax } is added to the model. For illustration, we assumethat there is a one in three chance that some tax would beimplemented sometime in the future; if the tax is imple-mented, we assume that the probability of introductionby 2050 is 50%, rising to 99% by the year 2100. The levelof the tax is also uncertain } we draw a Weibull distribu-tion around the mean expected value of the tax (US$50per ton carbon), with a 99% probability that it would notexceed US$125 per ton carbon. As expected, the existenceof an uncertain environmental constraint alters the pat-terns of technological change substantially. Investmentsin new technologies are shifted earlier in time to preparefor the possibility of facing a costly future with environ-mental constraints. For comparison, Fig. 15 also showsthe results if future demand is treated as uncertain } high-er or lower } than in the base case runs in Fig. 14. Thesethree types of uncertainty shown in Figs. 14 and 15} learning rates, emissions constraints, and demand } arethe most important unknowns for the energy sector. In

A. Gru( bler et al. / Energy Policy 27 (1999) 247}280 271

all three, earlier and longer investments in new technolo-gies is the result when the need for technologies with newattributes looms large.

The goal with this exercise is not a full-blown realisticmodel of technological change for all aspects of theenergy system. Rather, at this initial stage, the purposehas been to develop a simple model that demonstratesthe feasibility of a mathematical formulation. Even withthis simple model, three important points emerge. First,radical technological change does not occur in the modelif there is no endogenous mechanism for improving fu-ture performance and costs (unless one ventures intoimposing daring exogenous technological discontinuitiesinto the model). It is hardly surprising that if technolo-gical change is modeled as a certain, gradual and auton-omous process } as it is in most energy technologymodels and nearly all macroeconomic models } then thecomputed technological dynamics evolve slowly, if at all,and the future is assumed to look much like the present.(Nonetheless, not all parameters are endogenous to themodel; the sensitivity of the results to the exogenousparameters is examined elsewhere.20)

Second, only by including uncertainty and learningdoes the model display S-shaped di!usion curves that aresimilar to those that have been observed empirically.Uncertain learning leads to risk minimization and hedg-ing strategies*ie to initial investments that make newtechnologies more competitive and allows them to enterniche markets. Modeling uncertainty in other para-meters, such as the possibility of higher energy demand,

20A model that incorporates increasing returns is obviously highlysensitive to the parameter values adopted in the simulations. Sensitivityanalyses showed that model outcomes depended most heavily on theassumed learning rates (treated endogenously in the model), followedby uncertainties in energy demand, and the possible (uncertain) exist-ence of environmental constraints. Conversely, varying initial costsmattered only for some combinations. For instance, varying the initialcosts from 10,000 to 30,000 for a mean learning rate of 30% delays theeconomic break-even point of the Revolutionary technology by morethan "ve decades Initial costs were not treated as uncertain, } as inreality uncertainty can be reduced immediately by placing of an order.The group also investigated sensitivity to changes in the discount rate(results in "gure are based on a 5% discount rate). As expected, higherdiscount rates result in postponed technological investment, experi-mentation and learning.

Maintaining the same variance in the uncertainty of the learning ratebut changing the distribution (eg, from the symmetric lognormal toasymmetric distributions with long tails like Weibull or Gamma distri-butions) were also found to in#uence the model outcomes. Even withthe same variance, the model is sensitive to the existence of evenextremely low-probability outlier values. For example, including thevery small possibility of an extreme high learning rate, the model movesin direction of earlier and higher investments in new technology. Typi-cally the penetration curve for &&uncertain learning'' (see Fig. 15) isshifted left by one to two decades. Even a slight chance of a big payo!results in accelerated investment. (The converse is not fully symmetricalbecause the objective function gives special aversion to under-estima-ting actual costs.)

leads to similar up-front investments and S-shapeddi!usion patterns. When full uncertainty is included (e.g.,the runs in Fig. 15) the results can be interpreted as anoptimal technology diversi"cation strategy in the face ofpervasive uncertainty.

Third, the model autonomously produces &&surprising''results. Even without carbon or signi"cant resource con-straints, plausible assumptions in the model lead to thewidespread use of a zero carbon (photovoltaic) energysource. These surprises are even more abundant whenexamining individual runs rather than the optimal solu-tion for the whole sample runs N. As the treatment oftechnology in models improves, the results become lessdeterministic. As in the real world, the future can easilyhold dramatic changes in technological regimes thatcould eliminate or create whole classes of environmentalproblems. While the potential for surprise has long beenrecognized, until now plausible models could not simul-taneously generate realistic technology dynamics andautonomous surprises. The approach thus addresses animportant critique of existing models and other analyti-cal methods } the tendency to ignore factors that causediscontinuities and surprises (Brooks, 1986).

4.3. Models and scenarios for the world energy system

In previous sections we have shown that it is possibleto simulate how changing costs a!ect technological choi-ces at the micro level } i.e., to model the selection amonga handful of competing technologies in an isolated mar-ket. But at larger scales } the energy system of a wholeeconomy, region or the globe } models that fully andrealistically endogenize technological change are still im-possible. Relationships between the wide array of factorsthat determine technological choices are poorly under-stood, and the need to model decision making by thou-sands of agents and for hundreds of possible technologycombinations exceeds available modeling and computa-tional capacity.

Because uncertainties abound at the macro scale,almost every model-based analysis employs scenarios.Scenarios bound possible futures, and they make it pos-sible to focus attention on analytically tractable issues,such as how the adjustment of one or more &&policy''variables causes changes in emissions from a baselinescenario. Because scenarios are the backbone of macroanalysis, here we focus on ways to incorporate technolo-gical dynamics into scenarios by incorporating technolo-gical change into the macro modeling tools that are usedto create scenarios. We argue that the existing model-ing tools fail to re#ect well-established patterns oftechnological change, and thus the scenarios thatthey generate are often inconsistent with plausible tech-nological futures. The resulting policy analysis is mis-leading.

272 A. Gru( bler et al. / Energy Policy 27 (1999) 247}280

Most scenario analysis begins with a &&business asusual'' baseline scenario } the result of a model run thatincorporates no particular drastic changes in technolo-gies or policies. The IS92a scenario of the Intergover-nmental Panel on Climate Change (IPCC), the mostwidely used such baseline scenario, illustrates the genericapproach (see Table 3). Worldwide economic growth ishigh (2.2% per year on average from 1990 to 2100);developing countries, especially in Asia, grow morerapidly than the industrialized countries. Radical tech-nological change does not occur. Rather, the future isassumed to be merely the result of compounded marginalchanges to today's energy system. For example, just ascoal was the principal fossil energy source some 50 yrago, in 2100 the world of IS92a is also principally (47%)powered by coal. The ratio of sulfur emissions to carbonis also high, and thus the environmental e!ects of sulfur,such as acid rain, would be severe in the world of IS92a.Energy consumption per unit of economic output de-clines over 70% because the economy as a whole isprojected to become more e$cient and structural changehas favored less energy-intensive services over industrialmanufacturing, but the energy delivered is just as car-bon-intensive as today. Decarbonization, though strong-ly evident in the historical record, is practically nonexist-ent: in 1990 every 106 J (MJ) of primary energy released17.4 g of carbon, which decreases slightly to 14.1 g ofcarbon per MJ of primary energy by 2050. Thereafter,decarbonization stagnates and reaches a value of 13.6 gC/MJ by 2100). This static view of technological changeis hardly surprising since the models used to create theIS92a scenario do not capture the dynamics of technolo-gical change.

To illustrate the consequences of improving the repres-entation of technological change, working with collabor-ators at IIASA we have linked a conventional macroeco-nomic model with a model of regional and global energysystems. The macroeconomic model } an 11 region ver-sion of the widely used MACRO model (Manne andRichels, 1992) } computes the size of the economy, invest-ment #ows, and demand for electric and non-electricenergy. Such a macroeconomic model is broadly similarto other macro models used for scenario analysis } itsstrength is that it treats the economy of coherent regionsof the world in an integrated fashion and estimates de-mand for energy. Its weakness, also shared by the modelstypically used in macroeconomic scenario analysis, isthat it has little resolution of technological choices. Yetthe choice of technologies determines emissions and en-vironmental impacts. The IIASA team has addressedthat weakness by linking MACRO with MESSAGE IV,the latest version of a widely used linear programmingmodel that optimizes the technology system needed tosatisfy a given demand for electric and non-electric en-ergy (Messner and Strubegger, 1995). MESSAGE IVselects from a large portfolio of current and future tech-

nologies, whose cost and system characteristics are de-rived from the IIASA CO2DB technology inventorydatabase that comprises some 1600 technologies. MESS-AGE IV computes the minimum (discounted 5% for allruns presented here) total systems costs (including invest-ment and operating costs) needed to satisfy the projecteddemand from the MACRO model. The approach par-tially solves the classic &&top-down'' and &&bottom-up''dichotomy of energy models. It allows estimation of theeconomic losses that result, e.g., when carbon constraintsare applied to the economy (a strength of top-down,macroeconomic models) as well as the minimum costsuite of technologies needed to meet a given constraint (astrength of bottom-up technological system models). Formore on the methodology of linking the models see Wene(1995). Because it employs a model of whole energysystems from technologies for production of primaryenergy through "nal energy consumption by end-usetechnologies } the approach accounts for the synergiesand infrastructures that de"ne technological clusters andgiven choices of individual technologies.

The details of the models, resource constraints, popu-lation estimates and other factors are described elsewhere(NakicH enovicH et al., 1998). Here we use the linked modelto illustrate a simple point: a plausible change in therepresentation of technology in such models can lead toradical changes in the computed environmental e!ects.That change is impossible to implement in conventionalmacroeconomic models that ignore the competition be-tween technologies to supply energy.

First we generated a baseline scenario by mimickingthe approach in existing macroeconomic models and&&business as usual'' scenarios. We assume that tradeliberalization continues and thus the world economygrows rapidly } in 2100 the economy is nearly twentytimes larger than in 1990. As in other high demandbaseline scenarios, we allow gradual and modest im-provements in the performance of end-use technologies.The economy as a whole continues the structural changefrom manufacturing to services as income rises. Thus, asin other models, the energy intensity of production de-clines: in 1990 13.1 MJ of "nal energy were needed toproduce one dollar of economic output; by 2100 onlyone-"fth the level of energy (2.5 MJ) would be needed inour scenario. Global energy needs thus increase less thana factor of 4 while the global economy grows nearly20-fold. Consistent with this trend towards higher e$-ciency, the shift to higher quality "nal energy carriersalso continues } the share of "nal energy supplied byelectricity rises from 13% in 1990 to 35% by 2100.Overall, the e$ciency of converting primary energy toGDP improves 1.3% per year, comparable with the his-torically observed rate and also with the 1% per year rateof &&autonomous energy e$ciency improvement (AEEI)''that is used in many macroeconomic models (Manne andRichels, 1992). As in other baseline scenarios such as

A. Gru( bler et al. / Energy Policy 27 (1999) 247}280 273

IS92a, the scenario does not assume any radical techno-logical changes in energy supply technologies and acces-sibility of fossil energy resources. Hence, as in IS92a, thefuture world is mainly powered by coal, as conventionaloil and gas become depleted and new, alternative energytechnologies are not su$ciently developed and remaintoo expensive to compete with coal. By 2020 coal hassurpassed oil as the single largest source of primaryenergy. In 2100 nearly half (42%) of primary energyis from coal, and coal is the dominant fuel for elec-tricity generation. The environmental consequences areprofound: emissions of carbon dioxide and sulfur are4 times 1990 levels. In Asia, sulfur emissions, whichalready are causing health and environmental problems,rise "ve-fold.

Next the IIASA team made one revision: the additionof technological learning. The results show more rapidtechnological dynamics and di!erent technological choi-ces than in the baseline scenario. As in the earlier modelthat endogenized technological learning for a singleagent selecting between three technologies, we did this bydividing all technologies considered in the MESSAGEmodel into three categories }mature (existing), incremen-tal and radical. The investment cost for existing technolo-gies does not vary; the cost of investing in incrementaltechnologies declines 15% for every doubling of installedcapacity; and the cost of radical technologies declines30% for every doubling. For example, in this &&dynamictechnology (DT)'' scenario advanced coal power plants(an &&incremental technology'') decline in cost fromUS$1650 per kW(e) of installed capacity to US$1200/kW(e) while total installed capacity more than doubles.Solar photovoltaics (a &&radical'' technology) decline incost from US$5000/kW(e) in 1990 to approximatelyUS$500/kW(e) as installed photovoltaic capacity risesfrom practically zero today to some 100 GW by the endof the 21st century.

(At present it is only possible to endogenize learningcurves in a one-region model that includes a few tech-nologies (Messner, 1997). Adding learning curves toa model of 11 regions with hundreds of technologiesremains practically impossible. Thus for the full modeldiscussed here the IIASA team added illustrative &&learn-ing'' through an iterative process that was consistent withthe principle of learning but did not confront the math-ematical and computational barriers to a hard-wired fulllearning model. We "rst used a simple spreadsheet modelto calculate the investment pro"les and decline in costsfor each of the incremental and radical technologiesconsistent with the assumed learning curve parameters.With these cost data the MESSAGE model was runagain and the resulting investment pro"les of the modelwere compared with those calculated by the learningcurve model. For all of the incremental and radical tech-nologies, the actual investment costs made by MESS-AGE were within 3}5% of that needed to sustain the

appropriate learning rate (15% for incremental and 30%for radical technologies21).

The result of these simple adjustments to the repres-entation of technology has little e!ect on the macroeco-nomic outputs of the linked model. Because energy costs(per unit of economic output) are lower and investment ishigher, the world economy is slightly larger than in thebaseline scenario, which already envisioned even morerobust economic growth than in the bullish IS92a scen-ario. But unlike our baseline and IS92a baseline scenario,the wealthy future of the DT scenario is also relativelyclean. Despite declining costs for advanced coal powerplants, MESSAGE shifts to even less costly advanced gasplants (which are already abundant in today's energymarkets) and then to solar and other zero carbon fuels.Until 2060 the share of gas as a transition fuel for electricgeneration rises. Coal and biomass, which supply a grow-ing share of transport fuel (methanol) and electricity inthe baseline scenario, give way to solar-based energysources, including hydrogen, and new nuclear powertechnologies such as inherently safe, modular, hightemperature reactor designs. By 2100 decentralized(ie, non-grid) energy generation technologies supplytwo-thirds of all electricity demand. This category con-sists of on-site solar photovoltaics (30%), hydrogen co-generation (25%), and hydrogen fuel cell vehicles thatgenerate household electricity when parked (45%). Hy-drogen itself can be generated in many ways } the moste$cient emit little carbon and sulfur (eg, steam reformingof natural gas) or no carbon at all (eg, when based onsolar or nuclear primary energy).

The impact of the di!erent technologies in the DTscenario on environmental quality is dramatic. Sulfuremissions decline six-fold from 1990 levels by 2100, andworld carbon emissions rise only slightly (20%). By 2100,only 4 g C are released per MJ of primary energy supply,which is an average rate of decarbonization of 1.3%per year (four times the historical value). With thewidely used MAGICC climate model22 we have

21 It should be noted that because MESSAGE is a linear program-ming model that it includes typical constraints on the rate of marketpenetration, without which the model would rapidly switch from onetechnology to another (sometimes overnight) as soon as the new tech-nology is cost-competitive. Because MESSAGE has perfect foresight,these constraints lead to pre-competitive investments in new techno-logy so that future market penetration constraints can be met. Thus thetime path of actual technology investments, which is broadly consistentwith that observed in the real world, is more an artifact of the assump-tions that underpin the linear programming method. Consequently wehave used the simple learning model to verify not only that the annualinvestment costs are consistent but, more importantly, that the totaland multi-decade costs are consistent with the learning rates that weemploy. These penetration constraints are not needed if a learningcurve is used directly in the optimization model (see Messner, 1997) andin the micro model presented in this paper, but for the larger scalemodel such modeling tricks remain essential.

22The carbon model is from Wigley (1991, 1993); the climate model isfrom Wigley and Raper (1987, 1992).

274 A. Gru( bler et al. / Energy Policy 27 (1999) 247}280

Table 3Four global scenarios and the importance of technological change. Conventional baseline (&&no policy'') scenarios re#ect only gradual changes or statictechnology } here, we show the widely cited IPCC IS92a and WEC/IIASA Series A2 (high coal) baseline scenarios. In these scenarios the future energysystem appears similar to the present and is thus heavily based on coal. Linking a macroeconomic model with a model of technological choice (eg,MESSAGE) can generate radically di!erent results. We show a baseline scenario from the MACRO-MESSAGE model that is designed to mimic otherbaseline scenarios. Then we show a &&DT'' scenario in which learning curves have been stylistically added with the result that technology is moredynamic and emissions are lower, even in the absence of carbon constraints.

Scenario attributes 1990 levels(observed)

Conventional baseline scenarios MACRO-MESSAGE scenarios(this paper)

IPCC IS92a WEC/IIASA Series A2 Baseline Dynamic technology(DT)

World income (1012 1990 USD,market exchange rates)

20.9

2050 (% growth)! 92 (2.6%) 100 (2.7%) 117 (2.9%) 118 (2.9%)2100 (% growth)! 240 (2.3%) 310 (2.5%) 384 (2.7%) 393 (2.7%)

Final energy intensity(106 j per 1990 USD)" 13.1

2050 6.8 7.2 5.8 5.72100 3.6 3.5 2.5 2.8

Primary energy (1018 J) 3812050 930 1040 1100 9902100 1500 1900 1700 1800

Primary energy, Share of coal andzero carbon# 24%/22%

2050 37%/21% 37%/27% 31%/32% 4%/50%2100 47%/30% 38%/51% 42%/46% 0.4%/77%

Carbon emissions (1015 g C)$ 6.22050 13 15.1 16 9.52100 20 22 25 8.5

Sulfur eissions (1012 g S)$ 602050 101 85 170 172100 123 160 240 11

Sulfur/carbon ratio (g S per kg C) 102050 10 5.7 11 1.82100 6.2 7.4 9.6 1.3

Radiative forcing (W/m~2)% 2.6}1.3"1.3&

2050 5.7!2"3.7 6!1.5"4.5 6.5!2.4"4.1 5.6!0.9"4.72100 8.2!2"6.2 9!2.3"6.7 10.3!2.9"7.4 6.9!0.8"6.1

!Average global annual growth rate since 1990." Intensity of "nal energy is a measure of the economy's e$ciency; it re#ects direct improvements in the e$ciency with which energy is used to makea given economic output (product or service) as well as structural change (eg, shifts from energy-intensive manufacturing to service-orientedeconomy). IPCC values are for &&secondary energy,'' which is comparable in de"nition to &&"nal energy'' in the WEC and IIASA studies (ie, heatcontent of energy after conversion and distribution).

#These two parameters are indicators of change in technology and the fuel mix. For IPCC, &&coal'' primary energy is not reported; the values here arefor &&solid'' energy. &&zero carbon'' is assumed to be nuclear and renewable energy; in none of the scenarios is carbon &&scrubbed'' from #ue gases.

$Carbon and sulfur from industrial processes only, mainly the burning of fossil fuels and biomass for energy and as feedstocks. IPCC reports slightlyhigher (70]1012g S) anthropogenic emissions of sulfur.

% Increase in radiative forcing since preindustrial era (&1765). Cells show net increase in radiative forcing from carbon and other greenhouse gases(including indirect e!ects such as stratospheric water vapor and ozone) minus the negative forcing from sulfur (direct and indirect e!ects). Valuescomputed with MAGICC model and added to 1990 forcing levels to yield total increase in forcing.

&Values for 1990 used in calculations here; however, note that sulfur forcing is uncertain. IPCC (1995, Fig. 6.19) reports less intense negative sulfurforcing (!0.8 W/m~2).

estimated the future concentration of carbon dioxide andgreenhouse forcing. In the baseline scenario carbon con-centrations rise to 800 parts per million by volume (ppmv)from 355 ppmv in 1990. Carbon concentrations also rise inthe DT scenario, but more slowly}by 2100 the concentra-tion is only 556 ppmv, or approximately equal to the550 ppmv target that many ambitious policy makers haveargued should be the long-term goal of climate policy.

Interestingly, greenhouse forcing in the DT scenario re-mains high } only 20% less intense than in the baselinescenario. High sulfur in the baseline scenario masks nearlyhalf the greenhouse forcing from carbon dioxide. The ratioof sulfur to carbon drops to one-tenth that of the baselinescenario as coal is nearly extinguished from the economy.

Table 3 summarizes the salient characteristics ofthe IS92a scenario and our baseline and DT scenarios.

A. Gru( bler et al. / Energy Policy 27 (1999) 247}280 275

Obviously the results of our model, including the DTscenario, depend on the assumptions that drive the mod-els. For example, the decentralized energy future in theDT scenario, in contrast with the centralized grid-basedelectricity system of today, is not the only possibility.Here we intend only to illustrate a simple but crucialpoint: plausible changes in the treatment of technologycan yield models and scenarios with dramatically di!er-ent environmental projections. Whereas other macromodels have induced technological change through re-source constraints, our approach is more consistentwith the emerging view that fossil resources areabundant. Technological changes, in our view, emergefrom new demands for energy services, which createsincentives for innovation and application in nichemarkets, which lead to rapid learning and decliningcosts; in turn, less costly new technologies di!use intowidespread use.

In our DT scenario, as in history, radically newtechnologies emerge and dominate on the timescale ofa century, often with dramatic changes in environmentalworries. In our view the future environmental problemsmay be less with carbon, sulfur and other externalities offossil fuels and more with the problems (if any) of hydro-gen and electricity energy carriers.

In sum, the use of a technology systems model can helpto ensure that the computed future primary and end useenergy technologies are internally consistent. A hydrogeneconomy, for example, must include a compatible hydro-gen technology cluster } generation and transport tech-nologies as well as end-use fuel cells and other consumersof hydrogen fuel. Although full endogenous macro scalemodeling of technological change is still elusive, our viewis that the most promising direction towards that goal isto build models that resolve the building blocks of indi-vidual technologies and their interdependencies. Alterna-tive approaches } such as AEEI and other aggregatetrend parameters applied to models that have been tunedto present conditions } do not resolve individual techno-logical choices. At best, they are adequate only over shorttime horizons during which technological change con-sists predominantly of incremental adjustments to tech-nologies that already exist in the market place (3}4decades for energy technology systems). Over the longerperiods that are characteristic of global change, thatapproach is inadequate unless (by chance) the aggregatequalities of the new system are like the old. We haveshown that some aggregate trends have extended overmany successive energy infrastructures. Gradual decar-bonization, for example, is evident over two centuriesduring which more than three energy infrastructureshave been in place. Yet even in with that aggregateindicator, existing modeling approaches yield inconsist-ent results over long time periods because they fail toresolve individual technologies and technological clus-ters. The IIASA approach shows a way forward, but it

does not solve all extant problems. Notably, radical in-ventions } the building blocks of future radically di!erenttechnological clusters } remain di$cult to project.

5. Conclusions

Across a wide range of intellectual and political viewsone view is shared: in the long run, technology is thefactor that most governs growth of the economy, the costand availability of services such as mobility and food,and the impacts on the environment. Some herald tech-nological innovation as the savior against ecological de-struction while others view it as the extension of whatcaused ecological concerns. Examining the role and sour-ces of technological change, and analyzing potential pol-icies, requires coherent concepts, historical data sets, andmodels that have not been much applied in the debateover the environmental e!ects of technology and techno-logical change.

Since technological change is crucial, analytical tech-niques must be improved. The analysis here suggests atleast six "ndings.

First, while progress on technological analysis hasbeen enormous, much remains to be done. The crucialprocess of invention, especially the invention of radicaltechnologies, remains largely opaque to systematic anal-ysis. Perhaps an even greater inadequacy of existingtechniques, which might be more tractable to solve, is toidentify how inventions are selected for commercial in-vestments. The most critical shifts in energy technology,such as the rise of wood- and coal-powered railroads orthe emergence of the automobile, can be identi"ed his-torically, but robust identi"cation of future technologicalregimes is poor. Most key technologies in the next era areprobably found in the marketplace today, but at lowlevels of penetration; identifying them, and the clusterthey comprise, remains beyond existing techniques. Inthe absence of techniques for perfect forecasting, creativescenario writing and continued debate on alternativefutures must continue as the principle means to analyzealternative technological futures.

Second, several patterns in the changing properties oftechnologies can be identi"ed in the historical record,and these can greatly aid in the modeling of technologicalchange. Learning curves, and characteristic learningrates, can be identi"ed for many technologies. Suchcurves characterize the expected decline in cost due toinvestments. We suggest that such curves should bemodi"ed to include not only commercial investments} &&learning by doing'', which begins in niche markets} but also RD&D that is crucial during the earlier in-novation stage. The result is a convenient formulationthat allows improved endogenous modeling of technolo-gical change. We demonstrate both micro and macromodel applications.

276 A. Gru( bler et al. / Energy Policy 27 (1999) 247}280

Third, also evident in the historical record are predict-able patterns and rates of technological di!usion, whichcan help modelers estimate the rate and character ofchange once a technology has entered the marketplace.In general, di!usion and substitution of compatible tech-nologies within an existing technological system or in-frastructure occurs more rapidly (time constants of oneto two decades) than di!usion of infrastructures and thetechnologies that are clustered together with numerousother technologies and infrastructures (5 to 10 decades).These time constants can be applied in simple but power-ful logistic models of substitution and di!usion.

Fourth, clusters of compatible technologies are evidentin the historical record. Such clusters help determinerates of di!usion and which technologies survive frominvention to the commercial marketplace. The co-evolu-tion of technologies in clusters becomes more intense asthe cluster ages and synergies of compatible technologiesdeepen. Such co-evolution is evident not only in particu-lar technologies but also the fuels that supply primaryenergy. The dominant fossil fuels have become progress-ively lighter in carbon and thus decarbonization is evi-dent as a continuous, long-term historical phenomenon.In the United States, as for the world as a whole, the fuelmix is decarbonizing at approximately 0.3% per year.

Fifth, it is feasible to solve the non-convex, stochasticproblems that result when modeling the (uncertain) in-creasing returns resulting from learning processes. In ourview the causal mechanisms for technological change areinvestments (RD&D and learning by working with infanttechnologies in commercial niche markets) that are madefor two reasons } expectation that new technologies willbe more competitive, and to hedge against many uncer-tain risks, including the risk that investments into newtechnologies will not yield expected improvements incost and performance. A simple model that includesthose drivers } expected cost reductions and uncertainty} yields S-shaped patterns of technological dynamics thatare consistent with the historical record. With plausibleassumptions, the model endogenously generates radicaldepartures from existing technological practices } a zerocarbon future without policy constraints on carbon, andeven more rapid decarbonization when modest and un-certain carbon constraints are included. Neither outcomeis the result of resource constraints, which is the mainmechanism for radical change in conventional modelsthat do not endogenize the process of technologicalchange. We do not contend that the carbon problem willthus autonomously solve itself. But the model outputsunderscore that new technologies can penetrate the mar-ket even if they are initially a factor of 40 (or more) moreexpensive than the existing dominant technology. Themodel suggests that a strategy of investing in radicaltechnologies is socially rational and optimal when risksmust be diversi"ed, especially when there is even a smallprobability of a signi"cant emissions constraint in the

future. If the model included the varied other bene"ts ofdiversi"cation and the costs of environmental problemssuch as urban smog which are caused by the same pro-cesses that cause carbon emissions, the estimated bene"tsof diversi"cation should grow further

Sixth, we have also demonstrated that it is possible toinclude a stylized representation of technological changein macro scale models of the world's energy system.Doing so requires coupling traditional macroeconomicmodels with models that allow more detailed representa-tion of technological choice } failure to do so forces themodeler to include only gradual and aggregate patternsof technological change (or implausible parameters forsudden, discontinuous radical technological change).Thus conventional models, which are widely used fordeveloping the scenarios that are employed in policyanalysis of global change, typically project that the futurewill be much like the present. Coal retains a high (evenincreasing) share of primary energy, and carbon emis-sions are thus also high. In contrast, a coupled model} which we illustrate here } can plausibly yield technolo-gical dynamics and lower carbon emissions. Our ap-proach still does not fully endogenize technologicalchange at a macro scale, but conceptually that is nowpossible. Computational and mathematical bottlenecksshould soon be solvable.

These six "ndings strongly suggest that existing base-line scenarios, which are widely used as the starting pointfor policy analysis, are incompatible with historical ex-perience, notably decarbonization. (Those scenarios donot have su$cient resolution to determine whether theyare incompatible with learning, S-shaped technology dy-namics, and co-evolution of technological clusters, all ofwhich are also evident in the historical record.) For futurescenarios, the burden is on modelers to justify such sharpdeviations from historical experience.

Finally, we o!er a few speculations on the policy im-plications of this work. Our intention has been to showthat substantially improved treatment of technologicalchange is possible and mandatory for improved policyanalysis. Although we have not analyzed speci"c policies,our analyses underscore several facts that are relevant forcurrent policy debates. In particular, normal technolo-gical dynamics are yielding decarbonization, but at a ratethat is slower than the rise in demand for carbon-basedfossil fuels. Thus, overall, carbon emissions are rising, butless rapidly than the growth of primary energy consump-tion. Studies of future energy technologies should takemore seriously the need to analyze the technologies thatwill be in place the future, not to assume that the futurewill look much like the present. Those future technolo-gies surely include variants of the present vintage, butthey are also likely to include many others, including lowcarbon technologies that are already entering the marketthrough normal technological dynamics that are occur-ring even without signi"cant constraints on carbon. Thus

A. Gru( bler et al. / Energy Policy 27 (1999) 247}280 277

the policy task may be less to promote zero carbontechnologies from the laboratory bench to the marketand more to explore ways to ensure that network e!ectsenhance rather than bar those low-carbon technologiesthat on their own will become innovations and commer-cialized in niche markets.

The results from the micro-level and global optimiza-tion models suggest that when confronted with uncer-tainties } such as whether stringent action to slow globalwarming or other environmental externalities will beneeded } that it is socially rational to diversify technolo-gies. But in market societies, decisions are principallymade by market agents rather than social optimizers;those agents respond to individual rather than socialincentives. That fact has long justi"ed social (e.g., govern-ment) policy to promote research. Insofar as societiestoday increasingly consider the likely the need for long-term constraints on carbon they should focus attentionon incentives to diversify the portfolio of especially rad-ical technologies that will be required if it proves neces-sary to cut carbon emissions sharply in the future. As wehave shown, there is some chance that, by surprise, thosetechnologies will emerge autonomously. It is plausible toproject a century-long transition to a practically zerocarbon emissions that appears to eliminate (today's) fearsabout energy-related externalities without any cost. Butit remains di$cult to assess the probability of such posit-ive surprises, or of their negative counterparts. It is clearthat increasing the probability that markets will selecttechnologies with particular characteristics requires notonly investment into invention but also more costly in-vestments that carry inventions through developmentand early demonstration in commercial niche markets.Because the bene"ts of those investments accrue publiclyand globally, technology policy for global change isa global public goods problem. To date, nations haveused devices such as treaties to coordinate policies tolimit emissions that cause environmental externalities,which has indirectly coordinated technology policy. Buta frontal approach to global change problems may re-quire more direct attention to technology policy.

Acknowledgements

This work is based on the authors' long associationwith the International Institute for Applied SystemsAnalysis (IIASA) in Laxenburg, Austria. We thank ourcolleagues Yuri Ermoliev, Andrei Gritsevskii, and SabineMessner for contributing to work on which this paper isbased. We hope that they will absolve us from anymisrepresentations in methods and ideas developedjointly. We also bene"ted from discussions at theIIASA seminar on induce technological change heldJune, 1997.

References

Abramovitz, M., 1993. The search for the sources of growth: areasof ignorance, old and new. Journal of Economic History 52,217}243.

Alcamo, J., Bouwman, A., Edmonds, J., Grubler, A., Morita, T.,Suganddhy, A., 1994. An evaluation of the IPCC IS92 emissionscenarios. Climate Change 1994, Reports of Working Group Iand III of the IPCC. Cambridge University Press, Cambridge,pp. 251}305.

Anderson, A.E, Isard, W., Puu, T., 1984. (Eds.), Regional and IndustrialDevelopment: Theories, Models and Empirical Evidence. Elsevier,Amsterdam.

Argote, L., Epple, D., 1990. Learning curves in manufacturing. Science247, 920}924.

Arrow, K., 1962. The economic implications of learning by doing.Review of Economic Studies 29, 920}924.

Arthur, W.B., 1983. On competing technologies and historical smallevents: the dynamics of choice under increasing returns. WorkingPaper 83}90, International Institute for Applied Systems Analysis,Laxenburg, Austria.

Arthur, W.B., 1989. Competing technologies, increasing returns, andlock-in by historical events. The Economic Journal 99, 116}131.

Astakhov, A., GruK bler, A., Mookhin, A., 1990. Technology di!usion inthe coal-mining industry of the USSR: an interim assessment. Tech-nological Forecasting and Social Change 50, 113}134.

Ausubel, J.H., GruK ebler, A., NakicH enovicH , N., 1988. Carbon dioxideemissions in a methane economy. Climatic Change 12, 245}263.

Ausubel, J.H., Marchetti, C., 1996. Elektron: electrical systems in retro-spect and prospect. Daedalus 125, 139}169.

Barnett, H.J., Morse, C., 1967. Scarcity and Growth: The Economics ofNatural Resource Availability. Johns Hopkins Press, Baltimore.

Binswanger, H.P., 1978. Induced technical change: evolution ofthought. In: Binswanger, H.P., Ruttan, V.W. (Eds.), InducedInnovation: Technology, Institutions and Development. JohnsHopkins University Press, Baltimore, pp. 13}43.

Brooks, H., 1986. The typology of surprises in technology, institutions,and development. In: Clark, W.C., Munn, R.E. (Eds.), SustainableDevelopment of the Biosphere. Cambridge University Press, Cam-bridge, pp. 325}348.

Buttner, T., GruK bler, A., 1995. The birth of a &Green' generation?generational dynamics of resource consumption patterns. Techno-logical Forecasting and Social Change 50, 113}134.

Cantley, M.F., Sahal, D., 1980. Who learns what? a conceptual descrip-tion of capability and learning in technological systems. ResearchReport 80}42. International Institute for Applied Systems Analysis,Laxenburg, Austria.

Christiansson, L., 1995. Di!usion and learning curves of renewableenergy technologies. Working Paper 95}126. International Institutefor Applied Systems Analysis, Laxenburg, Austria.

Cowan, R., 1991. Tortoises and hares: choice among technologies ofunknown merit. The Economic Journal 101, 801}814.

David, P., 1985. CLIO and the economics of QWERTY. AmericanEconomic Review 75, 332}337.

Ermoliev, Y., Wets, R. (Eds.), 1988. Numerical Techniques for Stochas-tic Optimization. Computational Mathematics. Springer, Berlin.

Fisher, J.C., Pry, R.H., 1971. A simple substitution model of techno-logical change. Technological Forecasting and Social Change 3,75}88.

Fragniere, E., Haurie, A., 1995. A stochastic programming model forenergy/environment choices under uncertainty. International Jour-nal of Environment and Pollution, 6, 587}603.

Frankel, M., 1955. Obsolescence and technological change in a matur-ing economy. American Economic Review 45, 296}319.

Freeman, C., 1982/1989. The Economics of Industrial Innovation, 2nded. MIT Press, Cambridge.

278 A. Gru( bler et al. / Energy Policy 27 (1999) 247}280

Freeman, C., 1994. The economics of technical change. CambridgeJournal of Economics 6, 587}603.

Goldnikov, A., Gritsevskii, A., Messner, S., 1995. A stochastic version ofthe dynamic linear programming model MESSAGE III. WorkingPaper 95} 94. International Institute for Applied Systems Analysis,Laxenburg, Austria.

Goulder, L., Schneider, S., 1996. Induced technical change, crowdingout, and the attractiveness of CO2 emissions abatement. StanfordUniversity, Stanford, Mimeo.

Griliches, Z., 1996. The discovery of the residual: a historical note.Journal of Economic Literature 34, 1324}1330.

GruK bler, A., 1990. The Rise and Fall of Infrastructures. Physica Verlag,Heidelberg.

GruK bler, A., 1991. Di!usion: long-term patterns and discontinuities.Technological Forecasting and Social Change 39, 159}180.

GruK bler, A., 1996. Time for a change: on the patterns of di!usion ofinnovation. DcH dalus 125, 19}42

GruK bler, A., 1998. Technology and Global Change. Cambridge Univer-sity Press, Cambridge.

GruK bler, A., Gritsevskii, A., 1997. A model of endogenous technologicalchange through uncertain returns on learning (R&D and Invest-ments). Economic Journal (submitted).

GruK bler, A., NakicH enovicH , N., 1991. Long waves, technology di!usionand substitution. Review 14, 313}342.

GruK bler, A., NakicH enovicH , N., 1996. Decarbonizing the global energysystem. Technological Forecasting and Social Change 53, 97}110.

GruK bler, A, NakicH enovicH , N., Posch, M., 1988. Methods for estimatingS-shaped growth functions: algorithms and computer programs.International Institute for Applied Systems Analysis, Laxenburg,Austria, Mimeo.

Ha-Duong, M., Grubb, M.J., Hourcade, J.-C., 1997. In#uence of socio-economic inertia and uncertainty on optimal CO2-emission abate-ment. Nature 390, 270}273.

HaK gerstrand, T., 1967. Innovation Di!usion as a Spatial Process. Uni-versity of Chicago Press, Chicago.

Hammitt, J.K., Lempert, R.J., Schlesinger, M.E., 1992. a sequential-decision strategy for abating climate change. Nature 357, 315}318.

Hayami, Y., Ruttan, V., 1985. Agricultural Development: An Interna-tional Perspective, 2nd ed., Johns Hopkins Press, Baltimore.

Holling, C. S., 1978. Adaptive Environmental Assessment and Manage-ment. Wiley, Chichester.

Hughes, T.P., 1983. Networks of Power: Electri"cation in WesternSociety 1880-1930. John Hopkins Press, Baltimore.

Isard, W., 1942. A neglected cycle: the transport building cycle. Reviewof Economics and Statistics 24, 149}158.

Jorgenson, D.W., Fraumeni, B,M., 1981. Relative prices and technicalchange. In: Bendt, E., Fields, B. (Eds.), Modeling and MeasuringNatural Resource Substitution. MIT Press, Cambridge.

Kolstad, C.D., 1993. Looking vs. leaping: the timing of CO2

control inthe face of uncertainty and learning. In: Kaya, Y., NakicH enovicH , N.,Nordhaus, W.D., Toth, F. L. (Eds.), Costs, Impacts, and Bene"ts ofCO2 Mitigation. International Institute for Applied Systems Analy-sis (CP-93-2), Laxenburg, Austria, pp. 63}82.

Lakatos, I., 1970. Falsi"cation and the methodology of scienti"c re-search programmes. In: Lakatos, I., Musgrave, A. (Eds.), Criticismand the Growth of Knowledge. Cambridge University Press,Cambridge.

Lazarus, M. et al., 1993. Towards a Fossil Free Energy Future: TheNext Energy Transition. Stockholm Environmental Institute,Boston.

Lempert, R.J., Schlesinger, M.E., Bankes, S.C., 1996. When we don'tknow the costs or the bene"ts: adaptive strategies for abatingclimate change. Climatic Change 33, 235}274.

Lotka, A., 1924. Elements of Physical Biology, reprinted., 1956).Elements of Mathematical Biology. Dover Publishers, New York.

MacGregor, P.R., Maslak, C.E., Stoll, H.G., 1991. The market outlookfor integrated gasi"cation combined cycle technology. Proceedings

of International Exhibition and Conference for the Power Genera-tion Industries } Power-Gen 7}8, 1298}1325.

Maddison, A., 1991. Dynamic Forces in Capitalist Development:A Long-run Comparative View. Oxford University Press, Oxford.

Maddison, A., 1995. Monitoring the World Economy 1820}1992,OECD Development Centre Studies, OECD, Paris.

Manne, A.S., Richels, R.G., 1992. Buying Greenhouse Insurance: TheEconomic Costs of Carbon Dioxide Emission Limits. MIT Press,Cambridge.

Manne, A., Richels, R., 1994. The cost of stabilizing global CO2

emis-sions: a probabilistic analysis based on expert judgement. TheEnergy Journal, 15, 31}56.

Manne, A., Richels, R., 1997. Modeling endogenous technology di!u-sion in MERGE 3. Paper presented at the Conference on InducedTechnical Change and the Environment. International Institutefor Applied Systems Analysis, Laxenburg, Austria, 26}27 June1997.

Mans"eld, E., 1961. Technical change and the rate of imitation. Econo-metrica 29, 741}766.

Mans"eld, E., 1968. The Economics of Technological Change. W.W.Norton, New York.

Mans"eld, E., 1985. How rapidly does new industrial technology leakout? Journal of Industrial Economics 34, 217}223.

Mans"eld, E., Rapoport, J., Romeo, A., Villani, E., Wagner, S., Husic,F., 1977. The Production and Application of New Industrial Tech-nology. W.W. Norton, New York.

Mans"eld, E., Rapoport, J., Schnee, J., Wagner, S., Hamburger, M.,1972. Research and Development in the Modern Corporation.Norton, New York.

Marchetti, C., 1980. Society as a learning system } discovery, inventionand innovation cycles revisited. Technological Forecasting and So-cial Change 18, 267}282.

Marchetti, C., 1983. The automobile in a systems context. Technolo-gical Forecasting and Social Change 23, 3}23.

Marchetti, C., 1988. The future of natural gas: a darwinian analysis. In:Lee, T.H., Linden, H.R., Dreyfus, D.A., Vasko, T. (Eds.), The Meth-ane Age. Kluwer, Dordrecht.

Marchetti, C., NakicH enovicH , N., 1979. The dynamics of energy systemsand the logistic substitution model. Research Report 79}13, Inter-national Institute for Applied Systems Analysis, Laxenburg,Austria.

Meadows, D.H., Meadows, D.L., Randers, J., Behrens, W.W., 1972. TheLimits to Growth. A Report for the Club of Rome's Project on thePredictament of Mankind. Signet, New York.

Meadows, D.H., Meadows, D.L., Randers, J., 1992. Beyond the Limits:Global Collapse or a Sustainable Future? Earthscan, London.

Messner, S., 1996. The CO2DB inventory and its application in a com-parative assessment of electricity end use options. Proceedings ofa Symposium on Electricity, Health and the Environment: Com-parative Assessment in Support of Decision Making Vienna,Austria, International Atomic Energy Agency (IAEA-SM-338/57),16}19 October 1995.

Messner, S., 1997. Endogenized technological learning in an energysystems model. Journal of Evolutionary Economics 7, 291}313.

Messner, S., Strubegger, M., 1991. User's Guide to CO2DB: The IIASACO

2technology data bank, version 1.0. Working Paper 91} 31a.

International Institute for Applied Systems Analysis, Laxenburg,Austria.

Messner, S., Strubegger, M., 1995. User's Guide for MESSAGE III.Working Paper 95}69. International Institute for Applied SystemsAnalysis, Laxenburg, Austria.

Messner, S., Golodnikov, A., Gritsevskii, A., 1996. A stochastic versionof the dynamic linear programming model MESSAGE III. Energy21, 775}784.

Metcalfe, S., 1987. Technical change. In: Eatwell, J., Milgate, M.,Newman, P. (Eds.), The New Palgrave, A Dictionary of Economics,vol. 4. Macmillan, London, pp. 617}620.

A. Gru( bler et al. / Energy Policy 27 (1999) 247}280 279

Meyer, P., 1994. Bi-logistic growth. Technological Forecasting andSocial Change 47, 89}102.

Mokyr, J., 1990. The Lever of Riches: Technological Creativity andEconomic Progress. Oxford University Press, Oxford.

Montroll, E.W., 1978. Social dynamics and the quantifying of socialforces. Proceedings National Academy of Sciences 75, 4633}4637.

NakicH enovicH , N., 1984. Growth to Limits: Long Waves and the Dynam-ics of Technology. Dissertation, Vienna, Austria, Sozial- undWirtschaftswissenschaftliche FakultaK t, UniversitaK t Wien.

NakicH enovicH , N., 1986. The automotive road to technological change:di!usion of the automobile as a process of technological substitu-tion. Technological Forecasting and Social Change 29, 309}340

NakicH enovicH , N., 1990. Dynamics of change and long waves. In: Vasko,T., Ayres, R., Fontvieille, L. (Eds.), Life Cycles and Long Waves.Springer, Berlin, Lecture Notes in Economics and MathematicalSystems, pp. 147}192.

NakicH enovicH , N., 1994. Energy gases: the methane age and beyond. TheFuture of Energy Gases. Washington, DC, US Geological SurveyProfessional Paper 1570.

NakicH enovicH , N., 1996. Technological change and learning. In:NakicH enovicH , N., Nordhaus, W.D., Richels, R., Toth, F.L. (Eds.),Climate Change: Integrating Science, Economics and Policy. Inter-national Institute for Applied Systems Analysis (CP-96-1), Laxen-burg, Austria.

NakicH enovicH , N., Amann, M., Fischer, G. principal authors, 1998. Glo-bal energy supply and demand and their environmental e!ects,Tokyo, Final Report to the Central Research Institute of the Elec-tric Power Industry, 28 February 1998.

NakicH enovicH , N., GruK bler, A., Je!erson, M., McDonald, A., Messner, S.,Rogner, H.-H., Schrattenholzer, L., 1995. Global Energy Perspect-ives to 2050 and Beyond, Laxenburg, Austria, International Insti-tute for Applied Systems Analysis, and London, World EnergyCouncil.

Nordhaus, W.D., 1973. The allocation of energy resources. BrookingsPapers on Economic Activity 3, 529}576.

Nordhaus, W.D., 1979. The E$cient Use of Energy Resources. YaleUniversity Press, New Haven, (Cowles Foundation Monograph 26).

Nordhaus, W.D., van der Heyden, L., 1983. Induced technical change:a programming approach. In: Schurr, S.H., Sonenblum, S., Wood,D.O. (Eds.), Energy, Productivity, and Economic Growth. Oelges-chlager, Gunn and Hain, Cambridge, MA, pp. 379}404.

Organisation for Economic Cooperation and Development, 1996.Technology and Industrial Performance: Technology Di!usion,Productivity, Employment and Skills, International Competitive-ness. OECD, Paris.

Pepper, W., Leggett, J., Swart, R., Wasson, J., Edmonds, J., Mintzer, I.,1992. Emissions Scenarios for the IPCC: an update. prepared for theIntergovernmental Panel on Climate Change, Working Group I,May, 1992.

Rogner, H.-H., 1997. An assessment of world hydrocarbon resources.Annual Review of Energy and the Environment 22, 217}262.

Romer, P.M., 1986. Increasing returns and long-run growth. Journal ofPolitical Economy 94, 1002}1137.

Romer, P.M., 1990. Endogenous technical change. Journal of PoliticalEconomy 98, 570}601.

Rosenberg, N., 1982. Inside the Black Box: Technology and Economics.Cambridge University Press, Cambridge.

Ruttan, V.W., 1996. Induced innovation and path dependence: a reas-sessment with respect to agricultural development and the environ-ment. Technological Forecasting and Social Change 53, 41}59.

SchaK fer, A., Schrattenholzer, L., Messner, S., 1992. Inventory of green-house-gas mitigation measures: examples from IIASA technologydata bank. Working Paper 92}85 Laxenburg, Austria, InternationalInstitute for Applied Systems Analysis.

Schneider, S.H., Goulder, L.H., 1997. Achieving low-cost emissionstargets. Nature 239, 13}14.

Schumpeter, J.A., 1934, transl. of 1912 ed. The Theory of EconomicDevelopment. Harvard University Press, Cambridge.

Sobel, D., 1995. Longitude: The True Story of a Lone Genius WhoSolved the Greatest Scienti"c Problem of His Time. Walker, NewYork.

Solow, R.M., 1957. Technical change and the aggregate productionfunction. Review of Economics and Statistics 39, 312}320.

Starr, C., Rudman, R., 1973. Parameters of technological growth.Science 182, 358}364.

Steward, H.B., 1982. Technology Innovation and Business Growth.Nutveco, San Diego.

Strubegger, M., Reitgruber, I., 1995. Statistical analysis of investmentcosts for power generation technologies. Working Paper 95}109Laxenburg, Austria, International Institute for Applied SystemsAnalysis.

United States Department of Commerce, 1975. Historical Statistics ofthe United States: Colonial Times to 1970, Washington, US DOC,Bureau of the Census (2 volumes).

Teubal, M., Foray, D., Justman, M., Zuscovitch, E. (Eds.), 1996. Tech-nological Infrastructure Policy. Kluwer, Dordrecht.

Tinbergen, J., 1942. Zur Theorie der langfristigen Wirtschaftsentwick-lung. Weltwirtschaftliches Archiv 1, 511}549.

Volterra, V., 1927. Una teoria matematica sulla lotta per l'estistenza.Scientia 41, 850}102.

Walters, C.J., 1986. Adaptive Management of Renewable Resources.Macmillan, New York.

Ward, W.H., 1967. The sailing ship e!ect. Buttelin Institute of Physicsand Physics Society 18, 169.

Watanabe, C., 1995. Identi"cation of the role of renewable energy.Renewable Energy 6, 237}274.

Wene, C.-O., 1995. Energy-economy analysis: linking the macroeco-nomic and systems engineering approachs. Energy}The Interna-tional Journal 21, 809}824.

Wigley, T.M.L., 1991. A simple inverse carbon cycle model. GlobalBiogeochemical Cycles 5, 373}382.

Wigley, T.M.L., 1993. Balancing the carbon budget: implications forprojections of future carbon dioxide concentration. Tellus 45B,409}425.

Wigley, T.M.L., Raper, S.C.B., 1987. Thermal expansion of sea waterassociated with global warming. Nature 330, 127}131.

Wigley, T.M.L., Raper, S.C.B., 1992. Implications for climate and sealevel of revised IPCC emissions scenarios. Nature 357, 293}300.

Wright, T.P., 1936. Factors a!ecting the costs of airplanes. Journal ofthe Aeronautical Sciences 3, 122}128.

280 A. Gru( bler et al. / Energy Policy 27 (1999) 247}280