A digital simulation of natural selection, taking place in scores of internet-linked personal computers, is being used to evolve superior electronic circuits.

The calculations used to improve circuit design would normally be performed on a single powerful computer or a large cluster of machines. But Miguel Garvie, a research student at the University of Sussex in the UK, has developed software that lets ordinary computer users contribute their spare processing power to create a virtual evolutionary environment for the project.

Such "distributed computing" is already providing cheap but substantial computer power to the search for alien messages in radio signals from space and to the quest for the largest prime numbers.

In the five days since the project was launched, Garvie says he has evolved circuits that outperform commercial designs on standard tests by 100 per cent but are only 50 per cent larger.

"It's gone as far as conventional circuits and beyond," he told New Scientist. "The plan is to go with bigger and bigger circuits, which is why I started the distributed project."

Random mutations

It is crucial for certain types of circuits to be able to raise an alarm if there has been an error in the fabrication process. This removes the need for costly and time-consuming testing and inspection.

Mimicking evolution via natural selection to come up with efficient solutions to specific problems is already an established, if relatively experimental, method of designing electronic circuits.

Evolving a new circuit design begins with a population of simple circuits with slight and random differences in their design. A hardware simulator tests each circuit design to identify the ones that come closest to the producing the desired output.

The best designs are then combined - a simulation of sexual reproduction - to produce an offspring population with further mutations, and the selection process begins over again.

Evolutionary 'islands'

Garvie's software uses internet-connected computers to perform these simulations. Each machine evolves its own population of circuit designs and periodically contacts a central server to upload a couple of its best designs.

These are then distributed at random to other machines on the network where they are added to the local population. Dividing the evolutionary process up into different "islands" in this way guarantees greater "genetic diversity" and better overall results, Garvie says.

Peter Young, a biologist at the University of York who also works on evolutionary computing, says this principle has proven itself in biology. "The idea of distributing it to a lot of computers makes a lot of sense," he told New Scientist.

The best solutions to complex problems often result from the occasional combination of individuals that have evolved in many diverse populations, he says. Single populations can become stuck in an evolutionary niche that is too highly specialised.

Even though the project has already produced some promising results, Garvie admits that it is not yet ready for commercial use. The circuit designs generated do not have a standard die size, for example. "It's still a bit green," he says. But he believes it should be relatively simple to refine the process.

Will Knight
NewScientist.com news service