For new readers and those who request to be “好友 good friends” ,  please read my 公告栏 first.

I thank Professor Guanrong Chen of the City University of HongKong for bringing  the following article to my attention.

NATURE PHYSICS | VOL 12 | MARCH2016 | www.nature.com/naturephysics

Generalizing Moore’s Law  by MARK BUCHANAN

IT’S POSSIBLE TO MAKE PREDICTIONS OF TECHNOLOGICAL PROGRESS  USING HISTORICAL DATA

Over the past few years, several independent teams of researchers have  noticed something surprising in historical data on a broad set of technologies.  Everyone, of course, knows about Moore’s Law — for decades, the density  of transistors on integrated circuits has doubled every two years,  with  computational speed advancing even faster. This spectacular record of  improvement shows up in just about any metric. Much less known, however,  is that this pattern of exponential advanceisn’t actually limited to electronics; it  applies just as well to technologies ranging from cars or batteries to  beer or nuclear power.

One study, for example, looked at data for these and other technologies, in 28  domains in all, stretching backover 50 years, and considered the rate of improvement for measures such as performance and energy efficiency.  The data for all these technologies follows a similar exponential improvement,  although the time scale for doubling can bevery different, with  annual improvement rates ranging from 3 to 65 percent. LED performance, for instance, has been improving by a factor of about ten each decade, whereas batteries  have improved more slowly — ten-fold improvement requiring about 40 years. Even so, it’s the same pattern: (seeC. L. Magee et al., available at Moore’sLaw is a law of technology in general http://go.nature.com/sxJbnv; 2014).

Technology is close to discovery and invention, and we’re used to thinking  of these as largely unpredictable processes. They’re chancy, success being  furthered by effort and investment, of course, but also requiring some luck.  So the regularity of this pattern acrossthe whole domain of technology is surprising. It’s not yet clear what it means. But it may make it possible to make actual  predictions of technological rogress using historical data alone,  rather than relying on experts.

That idea comes from the work of physicist Doyne Farmer and economist Francois Lafond, who have built on these empirical studies. Their idea was to model the data on technology improvement as a stochastic process, using a  geometric random walk, which naturally exhibits an exponential, long-term growth (or drift) decorated with erratic noise. Fitting such a process to  the data on a broad class of technologies, they then tried to stand at one  moment in time and to use past data to predict future trends. They found good  success in making suchout-of-sample forecasts. They don’t claim these  are the best forecasts possible; only that this simple method works quite well.

Anyone can make predictions, of course. And historically, many predictions of  technological futures have been spectacularly wrong. The really important  thing is having some knowledge of the likely accuracy of a prediction. Indeed,  as Farmer and Lafond point out,predictions of low accuracy can even be dangerous if they’re trusted. Because he mathematical properties of the geometric random  walk are well studied, they were able to go further in deriving a closed form expression for each particular technology and to get an estimate of the expected  distribution of forecast errors for projections over any interval of time.  For the specific case of solar photovoltaic modules, for example, the method suggests that theprice of such modules will most likely continue to drop at about 10% annually,but that there’s still a 5% chance that prices in 2030 will be higherthan today.

This new capability should be valuable to anyone charged with making decisions  on which technologies to nvest in, especially policy makers aiming to invest public funds wisely in response to problems such as climate change.  In pursuing alternative energy sources, for example, we should expect a lot more from investments inphotovoltaics than from alternatives such as biofuels or wind. It’s possible tomake such claims not because we know anything about why technologies work thisway, only because the data implies that they do. Yet this data may also offerhints regarding the mechanisms that make one technology grow faster than another.

Many people have likened technological advance to an evolutionary process,  advancing as older techniques, components or ideas get combined in new ways.  Biologists know that some organisms evolve and adapt more rapidly than others  due to features that make it relatively easy to alter some elements —  cell surface receptors in bacteria, for example — without undermining other  underlying functions.Such independent flexibility enables fast, profitable experimentation, and creates the capacity for rapid evolution. In a recent study,  Subarna Basnet and Chris Magee at MIT find evidence that something very  similar seems to betrue with technologies (preprint at http://arxiv.org/abs/1601.02677; 2016). Thefaster evolving ones seem to have fewer interactions or complex interdependences between their elementary components.

In the case of technology,interactions among components or properties happen in many ways. A spring mechanism might improve a device’s performance  if it were made stiffer. Making it bigger might do that, yet would also increase  the device weight, which might be bad. Steps to speed up transistors may be  good for computing rates, yet also create more heating and the need for cooling  systems. Engineers encounter these interactions all the time. In a clever way,  Basnet and Magee found a way to geta rough ranking of technologies  by the number of such interactions theyinvolve. They looked at the patent literature and measured the frequency of sixspecific keywords — prevent, undesirable, requirement, fail, disadvantageand overcome — they suspected might reflect important interactions among omponents or processes. Across the board,  in technologies ranging from milling machines to superconductors to 3D  printers, they found that the rate at which a technology has improved is  significantly  correlated, inversely, with the frequency of these words in relevant  patents. The more interactions described in the reports, the slower the technology  advances, apparently because finding beneficial changes is just harder to do.  Complexity slows discovery and advance.

This doesn’t explain, of course,why some technologies are more complex than others. Naively, one might think that older technologies, after long development,  would grow more complex, but that’s not consistent with Moore’s Law  holding for individual technologies overmany decades. For now, we just don’t know.

Note added by this blogger: It is important not to read too much  into the result of this article and to distinguish between

1.   predicting technological advance once“something” has been invented.

2.   predicting the occurrence of an invention.

Moore’s law accurately predicted the advance of integrated  circuit technology, but it did not predict the invention of the integrated  circuit itself. Similarly, very few of the great discoveries of science  and technology were actually predicted. “Invention” as opposed  to “technological progress”  still seems to be a random event.

Notes added 3/8/2016: It also worth pointing out that we are not necessarily good at recognizing the economic impact of an invention leading to major investment mistakes http://blog.sciencenet.cn/blog-1565-530569.html .