Editing IPCC:AR6/WGIII/Chapter-16 (section)

==== 16.2.2.1 Learning by Doing and Research and Development ====

<div id="h3-5-siblings" class="h3-siblings"></div>

Learning by doing and R&amp;D efforts are two factors commonly used by the literature to explain past and projected future speed of technological change ( [[#Klaassen--2005|Klaassen et al. 2005]] ; [[#Mayer--2012|Mayer et al. 2012]] ; [[#Bettencourt--2013|Bettencourt et al. 2013]] ). Learning by doing is the interaction of workers with new machines or processes that allows more efficient use ( [[#Arrow--1962b|Arrow 1962b]] ). R&amp;D effort is dedicated to looking for new solutions (e.g., blueprints) that could increase the efficiency of existing production methods or result in entirely new methods, products or services ( [[#16.2.1.1|Section 16.2.1.1]] ).

Learning by doing and R&amp;D are interdependent. [[#Young--1993|Young (1993)]] postulates that learning by doing cannot continue forever without R&amp;D because it is bounded by an upper physical productivity limit of an existing technology. R&amp;D can shift this limit because it allows for replacing the existing technology with a new one. On the other hand, incentives to invest in R&amp;D depend on the future cost of manufacturing, which in turn depends on the scale of learning by doing. The empirical evidence for virtuous circle between cost reduction, market growth and R&amp;D were found in the case of the photovoltaic (PV) market ( [[#Watanabe--2000|Watanabe et al. 2000]] ) (Box 16.4), but could also lead to path dependency and lock-in ( [[#Erickson--2015|Erickson et al. 2015]] ). Sections 16.4.4 and 13.7.3.1 discuss how simultaneous use of technology push and pull policies could amplify the effects of research and learning.

The benefits of R&amp;D and learning by doing are larger at the economy level than at the firm level ( [[#Arrow--1962b|Arrow 1962b]] ; [[#Romer--1990|Romer 1990]] ;). As a result, when left to its own, the market tends to generate less investment than socially optimal. For instance, if the cost of a technology is too high before a large amount of learning by doing has occurred, there is a risk that it will not be adopted by the market, even if it is economically advantageous for the society. Indeed, initially new technologies are often expensive and cannot compete with the incumbent technologies ( [[#Cowan--1990|Cowan 1990]] ). Large numbers of adopters could lower this cost via learning by doing to a level sufficient to beat the incumbent technology ( [[#Gruebler--2012|Gruebler et al. 2012]] ). However, firms could hesitate to be the first adopter and bear the high cost ( [[#Isoard--2001|Isoard and Soria 2001]] ). If this disadvantage overwhelms the advantages of being a first mover [[#footnote-001|1]] and if adopters are not able to coordinate, it will lead to situation of a lock-in ( [[#Gruebler--2012|Gruebler et al. 2012]] ).

The failure of markets to deliver the size of R&amp;D investment and learning by doing that would be socially optimal is one of the justifications of government intervention. Policies to address these market failures can be categorised as technology-push and demand-pull policies. The role of these policies is explained in Table 16.3.

'''Table 16.3 | Categories of policies and interventions accelerating technological changes, the factors promoting them and slowing them down, illustrated''' '''with examples.'''

{| class="wikitable"
|-
!
! What it refers to

! What promotes technological change

! What slows down technological change

! Examples

|-
| Technology push

| Support the creation of new knowledge to make it easier to invest in innovation

| Research and development (R&amp;D), funding and performance of early demonstrations ( [[#Brown--2009|Brown and Hendry 2009]] ; [[#Hellsmark--2016|Hellsmark et al. 2016]] )

| Inadequate supply of trained scientists and engineers ( [[#Popp--2012|Popp and Newell 2012]] ); gap with demand pull ( [[#Grübler--1999b|Grübler et al. 1999b]] )

| Japan’s Project Sunshine, the US Project Independence in the 1970s. Breakthrough Energy Coalition and Mission Innovation, respectively private- and public-sector international collaborations to respectively focus energy innovation and double energy R&amp;D, both initiated concurrently with the Paris Agreement in 2015 ( [[#Sanchez--2017|Sanchez and Sivaram 2017]] )

|-
| Demand pull

| Instruments creating market opportunities

| Enlarging potential markets, increasing adoption of new fuels and mitigation technology

Digital innovations

Social innovation and awareness

| Willingness of consumers to accept new technology

Policy and political volatility can deter investment

| Subsidies for wind power California, the German feed-in tariff for photovoltaic, quotas for electric vehicles in China (F. [[#Wang--2017|]] [[#Wang--2017|Wang et al. 2017]] ) and Norway ( [[#Pereirinha--2018|Pereirinha et al. 2018]] )

Biofuels (Brazil)

Social innovation with wind energy (Denmark, Germany)

|}

[[#16.4|Section 16.4]] discusses individual policy instruments in greater detail.

The size of the learning-by-doing effect is quantified in literature using learning rates, that is estimates of negative correlation between costs and size of deployment of technologies. The results from this literature include estimates for energy technologies ( [[#McDonald--2001|McDonald and Schrattenholzer 2001]] ), electricity generation technologies ( [[#Rubin--2015|Rubin et al. 2015]] ; [[#Samadi--2018|Samadi 2018]] ), for storage (Schmidt 2017), for end-of-pipe control ( [[#Kang--2020|Kang et al. 2020]] ) and for energy demand and energy supply technologies ( [[#Weiss--2010|Weiss et al. 2010]] ). Meta-analyses find that learning rates vary across technologies, within technologies, and over time ( [[#Nemet--2009a|Nemet 2009a]] ; [[#Rubin--2015|Rubin et al. 2015]] ; [[#Wei--2017|Wei et al. 2017]] ). Moreover, different components of one technology have different learning rates ( [[#Elshurafa--2018|Elshurafa et al. 2018]] ). Central tendencies are around 20% cost reduction for each doubling of deployment ( [[#McDonald--2001|McDonald and Schrattenholzer 2001]] ).

Studies of correlation between cumulative deployment of technologies and costs are not sufficiently precise to disentangle the causal effect of increase in deployment from the causal effects of R&amp;D and other factors ( [[#Nemet--2006|Nemet 2006]] ). Numerous subsequent studies attempted to, among others issues, separate the effect of learning by doing and R&amp;D ( [[#Klaassen--2005|Klaassen et al. 2005]] ; [[#Mayer--2012|Mayer et al. 2012]] ; [[#Bettencourt--2013|Bettencourt et al. 2013]] ), economies of scale ( [[#Arce--2014|Arce 2014]] ), and knowledge spillovers ( [[#Nemet--2012|Nemet 2012]] ). Once those other factors are accounted for, some empirical studies find that the role of learning by doing in driving down the costs becomes minor ( [[#Nemet--2006|Nemet 2006]] ; [[#Kavlak--2018|Kavlak et al. 2018]] ). In addition, the relation could reflect reverse causality: increase in deployment could be an effect (and not a cause) of a drop in price ( [[#Nordhaus--2014|Nordhaus 2014]] ; [[#Witajewski-Baltvilks--2015|Witajewski-Baltvilks et al. 2015]] ). Nevertheless, in some applications, learning curves can be a useful proxy and heuristic ( [[#Nagy--2013|Nagy et al. 2013]] ).

The negative relation between costs and experience is a reason to invest in a narrow set of technologies; the uncertainty regarding the parameters of this relation is the reason to invest in wider ranges of technologies ( [[#Fleming--2001|Fleming and Sorenson 2001]] ; [[#Way--2019|Way et al. 2019]] ). Concentrating investment in narrow sets of technologies (specialisation) enables fast accumulation of experience for these technologies and large cost reductions. However, when the potency of technology is uncertain, one does not know which technology is truly optimal in the long run. The narrower the set, the higher the risk that the optimal technology will not be supported, and hence will not benefit from learning by doing. Widening the set of supported technologies would reduce this risk ( [[#Way--2019|Way et al. 2019]] ). Uncertainty is present because noise in historical data hides the true value of learning rates, and due to unanticipated future shocks to technology costs ( [[#Lafond--2018|Lafond et al. 2018]] ). Ignoring uncertainty in integrated assessment models implies that these model results are biased towards supporting a narrow set of technologies, neglecting the benefits of decreasing risk through diversification ( [[#Sawulski--2020|Sawulski and Witajewski-Baltvilks 2020]] ).

<div id="16.2.2.2" class="h3-container"></div>

<span id="knowledge-spillovers"></span>