(Image: Roya Hamburger)
ON A summer鈥檚 day in 1899, a bicycle mechanic in Dayton, Ohio, slid a new inner tube out of its box and handed it to a customer. The pair chatted and the mechanic toyed idly with the empty box, twisting it back and forth. As he did so, he noticed the way the top of the box distorted in a smooth, spiral curve. It was a trivial observation 鈥 but one that would change the world.
Advertisement
The shape of the box just happened to remind the mechanic of a pigeon鈥檚 wing in flight. Watching that box flex in his hands, Wilbur Wright saw how simply twisting the frame supporting a biplane鈥檚 wings would give him a way to control an aircraft in the air.
Serendipity and invention go hand in hand. The Wright brothers鈥 plane is just one of many examples. Take velcro: George de Mestral invented the material after he noticed the hook-covered seeds of the burdock plant sticking to his dog. And Harry Coover鈥檚 liquid plastic concoction failed miserably as a material for cockpit canopies, as it stuck to everything. But it had a better use: superglue.
It may be romantic, but it is an achingly slow way to advance technology. Relying on happenstance means inventions that could be made today might not appear for years. 鈥淭he way inventions are created is hugely archaic and inefficient,鈥 says Julian Nolan, CEO of Iprova, a company based in Lausanne, Switzerland, which specialises in generating inventions. Nothing has changed for hundreds of years, he says. 鈥淭hat鈥檚 totally out of sync with most other industries.鈥
But we are starting to make our own luck. Those eureka moments could soon be dialled up on demand as leaps of imagination are replaced by the steady steps of software. From algorithms that mimic nature鈥檚 way of producing the best designs to systems that look for gaps between existing patented technologies that new designs might fill, computer-assisted invention is here.
The impact could be huge. Some claim automated invention will speed up technological progress. It could also level the playing field, making inventors of us all. But what happens if the currency of ideas is devalued? To qualify for a patent, for example, an idea can鈥檛 be 鈥渙bvious鈥. How does that apply when ideas are found by brute force?
The first group to mimic evolution in patent design 鈥 pioneering the use of so-called genetic algorithms (see 鈥As nature intended鈥) 鈥 was led by John Koza at Stanford University in California in the 1990s. The team tested their algorithms by seeing if they could reinvent some of the staples of electronic design: the early filters, amplifiers and feedback control systems developed at Bell Labs in the 1920s and 1930s. They succeeded. 鈥淲e were able to reinvent all the classic Bell Labs circuits,鈥 says Koza. 鈥淗ad these techniques been around at the time, the circuits could have been created by genetic algorithms.鈥
In case that was a fluke, the team tried the same trick with six patented eyepiece lens arrangements used in various optical devices. The algorithm not only reproduced all the optical systems, but in some cases improved on the originals in ways that could be patented.
The versatility of this type of algorithm is clear from the showcase of evolved inventions at the annual Genetic and Evolutionary Computation Conference (GECCO). Innovations at this year鈥檚 event included efficient swimming gaits for a four-tentacled, octopus-like underwater drone 鈥 evolved by a team at the BioRobotics Institute in Pisa, Italy 鈥 and the most fuel-efficient route for a future space probe to clean up low-Earth orbits. Engineers at the European Space Agency鈥檚 advanced concepts lab in Noordwijk, the Netherlands, treated the task like a cosmic version of the famous travelling salesman problem 鈥 but instead of cities, their probe visits derelict satellites and dead rocket bodies to nudge them out of orbit.
However, the big prize at GECCO is the human competitiveness award, or 鈥淗umie鈥, for inventions deemed to compete with human ingenuity. The first Humie, in 2004, was awarded for an odd-shaped antenna, evolved for a NASA-funded project. It worked brilliantly even though it looked like a weedy sapling, with a handful of awkwardly angled branches, rather than a regular stick-like antenna. It certainly wasn鈥檛 something a human designer would produce.
That is often the point. 鈥淲hen computers are used to automate the process of inventing, they aren鈥檛 blinded by the preconceived notions of human inventors,鈥 says Robert Plotkin, a patent lawyer in Burlington, Massachusetts. 鈥淪o they can produce designs that a human would never dream of.鈥
This year鈥檚 Humie winner was a way to improve the accuracy of super-low-power computers. So-called approximate computers are built from simple logic circuitry that consumes very little power but can make a lot of mistakes. By evolving smart software routines for such computers, Zdenek Vasicek at Brno University of Technology in the Czech Republic was able to correct many of the errors introduced by the simple design. The result is a greener chip for use in applications where computational exactness doesn鈥檛 matter, like streaming music or video.
There鈥檚 just one problem with using genetic algorithms: you need to know in advance what you want to invent so that your algorithm can modify it in fruitful ways. 鈥淕enetic algorithms work well when you already know all the relevant features and can vary them until you get a solution that satisfies all your fitness constraints,鈥 says Tony McCaffrey, chief technology officer of Innovation Accelerator based in Natick, Massachusetts. Nolan agrees: 鈥淕enetic algorithms tend to be good at optimising pre-existing inventions but typically not ones of great commercial value.鈥 That鈥檚 because they don鈥檛 take big, inventive steps, he says, and so have less chance of making a commercially valuable hit.
Innovation Accelerator鈥檚 approach is to use software to help inventors notice easily missed features of a problem that, if addressed, could lead to a novel invention. 鈥淎n invention is something new that was not invented before because people overlooked at least one thing that the inventor noticed,鈥 says McCaffrey. 鈥淚f we can get people to notice more obscure features of a problem, we raise the chances that they will notice the key features needed to solve the problem.鈥
To do that, the firm has written software that lets you describe a problem in human language. It then 鈥渆xplodes鈥 the problem into a large number of related phrases and uses these to search the US Patent and Trademark Office database for inventions that solve similar problems. But similar is the operative word, says McCaffrey. The system is designed to look for analogues to the problem in other domains. In other words, the software does your lateral thinking for you.
In one example, McCaffrey asked the system to come up with a way to reduce concussion among American football players. The software exploded the description of the problem and searched for ways to reduce energy, absorb energy, exchange forces, lessen momentum, oppose force, alter direction and repel energy. Results for how to repel energy led the firm to invent a helmet that contained strong magnets to repel other players鈥 helmets, lessening the impact of head clashes. Unfortunately, someone else beat them to the patent office by a few weeks. But it proved the principle, says McCaffrey.
One algorithm found a way to make helmets safer using magnets (Image: Brian Kersey/Press Association)
In another case, the software duplicated a ski-maker鈥檚 recent innovation. The problem was to find a way to stop skis vibrating so skiers could go faster and turn more safely. The manufacturer eventually stumbled upon an answer, but Innovation Accelerator鈥檚 software was able to find it quickly. 鈥淎 violin builder had a method to produce purer music by reducing vibrations in the instrument,鈥 says McCaffrey. 鈥淭he method was applied to the skis and made them vibrate less.鈥
鈥淣inety per cent of problems have already been solved in some other field,鈥 says McCaffrey. 鈥淵ou just have to find them.鈥 He now plans to use IBM鈥檚 supercomputer Watson, which draws inferences from millions of documents, to help his system understand patents and technical papers far more deeply.
The technology at Nolan鈥檚 firm, Iprova, also helps inventors to think laterally 鈥 but with ideas derived from sources far beyond patent documents. The company is unwilling to reveal exactly how its Computer Accelerated Invention technique works, but in a 2013 patent, Iprova says it provides clients with 鈥渟uggested innovation opportunities鈥 by interrogating not only patent databases and technical journals, but also blogs, online news sites and social networks.
Of particular interest is the fact that it alters its suggestions as tech trends on the internet change. The result seems to be extremely productive. 鈥淲e use our technology to create hundreds of high-quality inventions per month, which we then communicate to our customers,鈥 says Nolan. 鈥淭hey can then choose to patent them.鈥 If their wide range of customers in the healthcare, automotive and telecommunications industries is anything to go by, Iprova appears to have hit paydirt. One of its clients is Philips, a major technology multinational. Such firms don鈥檛 add outside expertise to their R&D teams lightly.
All this means that algorithm-led discovery is likely to be the most productive inventing process of the future. 鈥淗uman inventors who learn to leverage computer-automated innovation will leapfrog peers who continue to invent the old-fashioned way,鈥 says Plotkin.
鈥淗uman inventors who use automated innovation will leapfrog those who don鈥檛鈥
But where do we draw the line between the two? 鈥淚 don鈥檛 think there is a clear separation between human and algorithm,鈥 says , founder of Icosystem, a company based in Cambridge, Massachusetts. 鈥淭he key is to find the right division of labour.鈥 Icosystem uses genetic algorithms to optimise everything from inventions to business processes 鈥 an approach Bonabeau calls 鈥渆nhanced serendipity鈥.
However, if the division of labour is too much on the computer鈥檚 side, it could undermine the patent system itself. Currently a 鈥減erson having ordinary skill in the art鈥 must believe that an invention isn鈥檛 obvious if it is to be granted a patent. But if inventors are only tending a computer, the inventions that arise could be deemed an obvious output of that computer, like hot water from a kettle.
These concerns have already been raised with drug discovery, says Gregory Aharonian, a consultant based in San Francisco, who specialises in patents. 鈥淚f drug discovery tools become so powerful that a researcher is just overseeing the tools鈥 activity, does that make the whole process obvious and so not patentable? Industry could be shooting itself in the foot by developing such technology.鈥
Another concern is that broad access to smart invention tools could speed up human technological development. Making the resulting gadgets may consume Earth鈥檚 resources all the quicker. McCaffrey is more optimistic. 鈥淚 am really impressed with engineers who are creating ways to improve housing, food storage, crop growth, water purification and transportation in the developing world,鈥 he says. 鈥淚 sincerely hope we use this emerging invention assistance technology to address the really important problems faced by humanity.鈥
Chance favours the prepared mind. If Wilbur Wright hadn鈥檛 been thinking about his problem, he may never have had his eureka moment. 鈥淎utomating that amounts to making accidental encounters orders of magnitude more efficient,鈥 says Bonabeau. 鈥淚n other words, outsource serendipity to the algorithm.鈥
Read more: 鈥Eureka by machine: How computers will be the mother of invention鈥
As nature intended
Genetic algorithms tackle the problem of design by mimicking natural selection. Desired characteristics are described as if they were a genome, where genes represent parameters such as voltages, focal lengths, or material densities, say.
The process starts with a more or less random sample of such genomes, each a possible, albeit suboptimal, design. By combining parent genomes from this initial gene pool 鈥 and introducing 鈥渕utations鈥 鈥 offspring are created with features of each parent plus potentially beneficial new traits. The fitness of the offspring for a given task is tested in a simulation. The best are selected and become the gene pool for the next round of breeding. This process is repeated again and again until, as with natural selection, the fittest design survives (see diagram below).
As well as evolving new designs, evolutionary algorithms can be used to evolve 鈥減arasites鈥 that inflict maximal damage to test safety or security features. 鈥淣ature has been very good and very creative at finding loopholes in every possible complex system,鈥 says Eric Bonabeau of Icosystem of Cambridge, Massachusetts, who has used this technique to improve the design of ships for the US navy.
This article appeared in print under the headline 鈥淓ureka machines鈥
