中国科技网4月15日报道（张微 编译）在线学习产业发展的有多快?风力涡轮机是一个有投资潜力的行业吗?价格低廉的磁悬滑板什么时候能上市?

要回答这些问题就要知道相关技术发展和改良的速度。现在麻省理工学院的工程师们设计出了一个公式，通过从相关专利收集到的信息，来估算技术推进的速度。

研究人员确定了28种不同技术的改进率，包括太阳能光伏发电、3D打印、燃料电池技术和全基因组测序。他们通过美国专利数据库，搜索了每个领域的专利，共有超过50万项之多，并通过一种创新方法快速准确地选择出每个技术领域最有代表性的专利。

一旦这些专利被确定，研究人员就开始分析各领域专利的某些指标，并发现那些最能预测技术进步率的指标。尤其是，专利的前向引证——一项专利被后续专利引用的次数，这是一个很好的预测指标，还有专利的公布日期：最新的技术专利比旧的专利创新速度快。

研究团队设计了一个方程式，包括专利集的平均前向引证数量和专利平均公布日期，并计算每个技术领域的进步速度。他们的计算结果与一项耗费人力的方法——找到的每个技术众多的历史绩效数据相匹配。

在这28项技术领域的分析中，研究人员发现发展速度最快的技术包括光学，无线电通讯，3D打印，以及MRT技术，而电池、风力涡轮机和内燃机发展速度较慢。

麻省理工学院机械工程系的研究生克里斯 本森说，风险资本家、初创企业、政府以及开发新技术的实验室会对新的预测工具感兴趣。

“我们的方法有独特之处，我没有把它看作是向公众发布的娱乐工具，”参与开发预测工具的本森说。“我更多地把它当作是，我们与别人合作，帮助他们了解他们感兴趣的技术的未来发展趋势。我们可能更像是房地产经纪人，而不是Zilow(注：zillow是一家提供免费房地产估价服务的网站，在美国一上线就造成大轰动。Zillow创建于2006年，主要向网民提供各类房地产信息查询服务。)

2003年，麻吉开始研究各种技术的进步率。当时，他很好奇摩尔定律相关技术是如何发展的，摩尔定律说的是计算机发展趋势，计算机芯片上的晶体管数量每两年增加一倍。

“很多行业没有摩尔定律那样快的发展速度，我开始想办法测度它们的速度，”麻吉说。

麻吉最初通过个案分析接触这个问题，确定一个特定的领域里哪些指标最能代表生产率。然后，他为每个指标收集数据，例如价格和产品生产速度，并用数据来计算总的改进率。在2010年，他才意识到美国专利数据是技术层面最全面的资源之一。

“我们认为，也许有足够的信息让我们了解技术变革的动力，”麻吉说。

几年来，他和他的团队花费的大量时间，通过阅读成千上万专利文献，发现科技领域最相关的专利数据。这个方法不是最可靠的，因为两个人可以选择完全不同的专利集来代表相同的技术。

2012年，麻吉和本森通过调查美国和国际专利分类系统的重叠部分，提出了一个更加有效而且可重复的方法用于识别相关专利集。

对于美国专利局接受的每一项专利，专利审查员都会将专利文件归档到专利分类系统的不同类别下。例如，太阳能光伏专利可以进入到美国专利的“电池”以及“固态元件”类别下，在国际分类系统中则作为“半导体器件”类别。

研究团队发现，通过寻找两个分类系统的专利重叠部分，，他们就能够在几个小时内而不是几个月，重复性地识别出最能代表一个技术领域的相同专利集。

一旦他们发现了相关专利集，研究人员就能够找到他们用于计算技术改进率的专利指标。他们发现，一个专利集发布前三年的平均前向引证数量，以及平均发布期是最好的技术改进率预测指标。本森说，他们还能够清除作用不大的专利信息。

“如果一项技术拥有多项专利，那么它应该发展的更快，但事实并非如此，”本森说。“3D打印技术只有300到500项专利，但是它的发展速度与半导体产业相同，而半导体产业拥有15万项专利，因此相关率是零。”

研究团队设计了一个引入了前向引证和专利发布日期的简单方程式，并用该种方法预测了28项技术的进步率。研究人员将计算结果与他们之前花费大量时间，用历史数据方法计算的结果进行了对比，发现两种方法的计算结果吻合。

然后，他们利用开发出的有效方法预测了11种新兴技术在未来10年的技术改进率。其中，增长最快的领域是在线学习和数字表达，而发展较慢的是食品工程和核聚变。

#p#分页标题#e#

麻吉希望这个方法可以当作一个评级系统，类似于标准普尔和其它股票市场指数。这个评级系统为投资者寻找下一个市场突破，实验室确定新的研究方向大有益处。麻吉说了解各项技术在未来十年如何发展，能够让创新者知晓为之奋斗的技术什么时候能够成熟，并让我们有更多异想天开的想法，如大批量生产磁悬滑板和飞行汽车。

New method uses patent data to estimate a technology's future rate of improvement

How fast is online learning evolving? Are wind turbines a promising investment? And how long before a cheap hoverboard makes it to market?

Attempting to answer such questions requires knowing something about the rate at which a technology is improving. Now engineers at MIT have devised a formula for estimating how fast a technology is advancing, based on information gleaned from relevant patents.

The researchers determined the improvement rates of 28 different technologies, including solar photovoltaics, 3-D printing, fuel-cell technology, and genome sequencing. They searched through the U.S. Patent Office database for patents associated with each domain—more than 500,000 total—by developing a novel method to quickly and accurately select the patents that best represent each technology.

Once these were identified, the researchers analyzed certain metrics across patents in each domain, and found that some were more likely to predict a technology's improvement rate than others. In particular, forward citations—the number of times a patent is cited by subsequent patents—is a good predictor, as is the date of a patent's publication: Technologies with more recent patents are likely innovating at a faster rate than those with older patents.

The team devised an equation incorporating a patent set's average forward citation and average publication date, and calculated the rate of improvement for each technology domain. Their results matched closely with the rates determined through the more labor-intensive approach of finding numerous historical performance data points for each technology.

Among the 28 domains analyzed, the researchers found the fastest-developing technologies include optical and wireless communications, 3-D printing, and MRI technology, while domains such as batteries, wind turbines, and combustion engines appear to be improving at slower rates.

Chris Benson, a former graduate student in MIT's Department of Mechanical Engineering, says the new prediction tool may be of interest to venture capitalists, startups, and government and industry labs looking to explore new technology.

"There's a lot of nuance to our method, and I don't see it as something to hand out to the masses to play with," says Benson, who helped developed the prediction tool. "I see it more as something where we work with somebody to help them understand what the future technological capabilities that they're interested in are. We're probably more like a real estate agent, and less like Zillow."

Technological dynamism

In 2003, Magee began determining the improvement rates of various technologies. At the time, he was curious how technologies were developing relative to Moore's Law—an observation pertaining originally to computers, in which transistors on a computer chip double every two years.

"There were a lot of things that weren't going as fast as Moore's Law, and I started trying to get measures of them," Magee recalls.

Magee initially approached the problem on a case-by-case basis, determining which metrics best represent productivity for a given domain. He then compiled data for each metric, such as the price and speed of manufacturing a product, and used the data to calculate the overall rate of improvement. In 2010, he realized that one of the most comprehensive resources on technology lay in the U.S. patent record.

"We thought, 'Maybe there's enough information there that we can do something about linking it to the dynamism of technical change,'" Magee says.

For several years, he and his group identified the most relevant patents in a technological domain, by literally reading through thousands of patents—an incredibly time-intensive process. The approach was not very reliable, as two people may choose entirely different sets of patents to represent the same technology.

A "Standard & Poor's" for technology

In 2012, Magee and Benson came up with a more efficient, repeatable method for identifying relevant patent sets, by looking at the overlap between the U.S. and international patent-classification systems.

For each patent accepted by the U.S. Patent Office, a patent reviewer will file the patent under several classes within both classification systems. For instance, a solar photovoltaics patent may be entered under the U.S. classes "batteries" and "active solid-state devices" and within the international system as "semiconductor devices."

#p#分页标题#e#

The team found that by looking for patent overlap between both classification systems, they could repeatedly identify the same set of patents that best represent a technology, within a matter of hours, rather than months.

Once they identified a relevant set of patents, the researchers looked for metrics within patents that they could use to calculate a technology's rate of improvement. They found that a patent set's average forward citations within the first three years after publication, and the average date of publication, were the best predictors of technological improvement. Benson says they were also able to weed out less-helpful patent information.

"If a technology has more patents in general, it should be moving faster, but that turns out not to be the case," Benson says. "3-D printing only has 300 to 500 patents, and that's improving at the same rate as semiconductors, which have about 150,000 patents. So there's almost zero correlation."

The team devised a simple equation incorporating forward citation and publication date, and used the method to predict improvement rates for 28 technologies. The researchers then compared the rates with those they previously obtained using their more time-intensive, historical data-based approach, and found the results from both methods matched closely.

They then used their more efficient approach to predict the improvement rates of 11 emerging technologies in the next 10 years. Among these, the fastest-growing domains appear to be online learning and digital representation, while slower technologies include food engineering and nuclear fusion.

Magee hopes the method may be used much like a rating system, similar to Standard & Poor's and other stock-market indices. Such ratings could be useful for investors looking for the next big breakthrough, as well as scientific labs that are contemplating new research directions. Magee says knowing how various technologies may improve in the next decade could give innovators an idea of when "feeder technologies" may mature, and enable more pie-in-the-sky ideas, like mass-produced hoverboards and flying cars.