Subscribe to ideas and commentary from Z + Partners!
Recommend a site or give us feedback at blog@zpluspartners.com.
Browse past postings in the Z+Blog archive
Get the RSS Feed for Z+Blog
Essential Voices:
|
Welcome to the Z+Blog!
Pointers and commentary on emerging futures issues collected by Z + Partners
Moore's Law, Darwin's Method
Our ever-increasing dependence on the microchip is enabled, in no small part, by a near-universally held idea about computers: no matter how complex and powerful microprocessors might become, in principle they are always rational, predictable, and 'debug-able' by the engineers who design them. As opposed to the messiness, ambiguity and complexity of the natural world, computers are structured, transparent and orderly - features that also make them trustworthy. We admit them into our lives because we believe (or assume) that we know what they’re going to do and how to fix them if they don’t.
This tidy state of affairs is so obvious that it usually goes unremarked upon. Yet there is evidence that these cherished values in computing might soon cease to be. That’s because the chip designs of tomorrow will be influenced as much by principles of Darwinian evolution as by traditional forms of engineering - and the results might not be understandable, even in principle, to the engineers that fashion them.
Recently at University of Sussex, Adrian Thompson and his colleagues engaged in a series of experiments to create a relatively simple circuit that would distinguish between two tones - the sort of function that’s routinely used in devices like cell phones. Rather than design the circuit by traditional means, Thompson simulated biological evolution by testing, mating and mutating more than 4000 generations of circuit designs. A set of instructions called a ‘fitness function’ killed off inefficient designs and mated efficient ones, mimicking the effects of natural selection.
The resulting, 'evolved' circuit design was not only more efficient than any designed by human beings, but it was so unorthodox that it took months for Thompson and his colleagues to determine how it worked. The evolutionary approach exploited what was there in ingenious and intricate ways, including properties of the chip its inventor couldn't even measure at the time of the experiment. Ultimately, to figure out what was going on, he had to use the kinds of techniques that biologists use to understand the nervous systems of simple animals.
New rounds of experiments have shown that these kinds of evolutionary approaches may benefit even smaller "logical" structures within a chip, such as ‘single electron’ logic gates. These rudimentary experiments, using a mere whiff of evolutionary ideas and processes, have generated startlingly innovative and counterintuitive designs that upend decades of conventional wisdom. How much more powerful - and bizarre - are the designs that will be created by more robust evolutionary approaches? No one really knows, but we’ve certainly only scratched the surface.
On a parallel path, there has been a great deal of work creating ‘reconfigurable’ hardware – chips that can be reprogrammed on the fly. These new kinds of chips are the perfect physical ecosystems for evolving hardware designs. Already, companies like Chameleon Systems are selling such chips. The joining of evolutionary design algorithms to such reconfigurable hardware is the first step toward the larger goal of truly "evolvable hardware" - computer chips that can adapt, reproduce and heal themselves instead of being programmed in the traditional sense.
All of this raises a lot of interesting questions: how do computer engineers, and the rest of us, operate in a world where some computer systems are fundamentally incomprehensible? After all, a chip that evolves its own instruction set doesn’t also grow a manual to explain its behavior. How do we 'fix' such a system, if we can’t tell in principle how it operates? In what contexts would the performance gains of such systems outweigh their inherent risks? How does the idea of engineering change when ‘selection’, natural or otherwise, is doing some or all of the work?
The first answer is, ironically, that things get simpler. Evolutionary approaches don’t require traditional design and engineering activities like modeling, abstraction, or analysis. These components of the human engineer’s process are thrown out in favor of the most basic measure of all – physical performance. This simplification allows for evolutionary processes to work in truly novel ways, utilizing aspects of a given circuit such as shape, material, and placement relative to other circuits - factors that might never inform the traditional design process. Anything that provides an edge, however miniscule, is fair game: evolution can be a more ruthless discoverer of efficiencies than any team of Intel or Microsoft engineers.
As engineers act less on planning the specifics of the design, and more as artificial selectors of the end product, engineering begins to look less and less like architecture, where one designs from the foundation up, and more and more like gardening, where one directs forces only partially under one’s direct control. In the evolutionary approach, engineers “prune” design iterations as they evolve and grow under the force of random mutations, directing the path of evolution in the same way one might clip a Bonsai tree to achieve a desired shape, or breed a new form of tulip.
This kind of breeding is already working its way into the practice of chip design. Today, companies like Analog Design Automation have developed software that enables an engineer to generate a large number of potential chip designs based on up to ninety different variables, and then select the one that best fits the desired goals/tradeoffs. Such tools will almost certainly develop into full-blown 'evolutionary' design testbeds in the coming years.
Engineers will also need entirely new kinds of training that are less mechanistic and more organic, and a new mindset that accommodates evolution’s strange and 'irrational' forms. And they’re likely to get them: Nature’s engineering methods so clearly trump our own that mimicking them will prove irresistible, whether we understand the results or not.
