ICBMTO : N48 10'07'' E011 33'53'' http://www.lrz.de/~ui22204
57F9CFD3: ED90 0433 EB74 E4A9 537F CFF5 86E7 629B 57F9 CFD3
---------- Forwarded message ----------
Date: Tue, 10 Apr 2001 21:26:24 -0700
From: DS2000 <firstname.lastname@example.org>
To: isml <email@example.com>
Subject: [isml] Computers that improve themselves
>From The News Observer,
Published: Monday, April 9, 2001 12:13 a.m. EDT
Computers that improve themselves
At first glance, Darwin's ideas on evolution don't seem to have much to do with computers. But if a line of computer code doesn't remind you at least vaguely of a chromosome -- both are essentially stored information -- you might want to look into the new field of evolvable hardware, where chips redesign themselves for optimum efficiency. This is evolution with a silicon flair.
Hot ideas come and go, but I know of no technology more likely to reshape our relationship with computers than this one.
A computer that evolves may redesign itself in such a way that even its inventors don't know how it's functioning. They just know that it works better than ever before, and future generations may work even better.
Something like this has already happened in the laboratory of Adrian Thompson at the University of Sussex in England. There, at the Center for Computational Neuroscience and Robotics, Thompson has spent the past four years working with computer chips that mutate. Chips can manipulate their own logic gates within nanoseconds, try the new design, and choose the configurations that work the best.
All of this takes place not in software but hardware. The chips are called Field Programmable Gate Arrays. The ones Thompson uses come from San Jose chip-maker Xilinx. The transistors of the chip appear as an array of "logic cells," which can be changed in value and connected to any other cell on the fly. By reprogramming a chip's memory, its logic cells can be tuned for any task at hand.
The work draws on the insights of Hugo de Garis, a computer scientist now working in Brussels, Belgium, who spent several years building neural modules -- software units that could be assembled to create artificial nervous systems.
About that project, de Garis, sounding almost like a biologist, said: "I was very conscious of the idea of using bit strings as codable mutatable instructions ('chromosomes') in evolutionary algorithms."
Let's untangle this. An algorithm is a way of getting something done through computer code, something our PCs do every time we run a program. But an evolutionary algorithm (also called a "genetic" algorithm) is different. It generates slight variations to its own code and then puts these changes through a series of mutations to see what works best. Couple evolutionary algorithms with an FPGA and amazing things happen.
You can run through thousands of generations quickly with this technology, saving code that works well, rejecting ideas that don't contribute and breeding in mutations to keep the mix dynamic. At Sussex, Adrian Thompson evolved a circuit that could distinguish between two different audio tones. It took more than 4,000 generations of algorithm evolution and roughly two weeks of computer time and produced results that were, well, strange.
Thompson's chip was doing its work preternaturally well. But how? Out of 100 logic cells he had assigned to the task, only a third seemed to be critical to the circuit's work. In other words, the circuit was more efficient by a huge order of magnitude than a similar circuit designed by humans using known principles.
And get this: Evolution had left five logic cells unconnected to the rest of the circuit, in a position where they should not have been able to influence its workings. Yet if Thompson disconnected them, the circuit failed. Evidently the chip had evolved a way to use the electromagnetic properties of a signal in a nearby cell. But the fact is that Thompson doesn't know how it works.
And that's the weird promise of using computers that evolve. These algorithms take us into an era where accepted design rules break down, where components get smaller and the properties of materials are only sketchily understood. At this level, pushing into the realm of nanotechnology, it may take evolutionary algorithms to work out their own best practice because we don't know how to proceed ourselves.
Imagine the philosophical problem this creates. What if you build a critical system for, say, a nuclear power plant. It works and works well, but you don't know how to explain it. Can you implement it? Can you rely on it?
If this sounds theoretical, consider that NASA's Langley Research Center has just announced that it is buying a HAL hypercomputer from Star Bridge Systems of Midvale, Utah. This computer is no larger than a regular desktop machine, yet it's roughly 1,000 times faster than traditional commercial systems because it uses Field Programmable Gate Arrays like those Thomson used in his work. Surely the name HAL of 2001 fame is no coincidence.
HAL, after all, was the machine that could think almost as well as a person, certainly well enough to threaten the entire crew he was in charge of. And though a Star Bridge hypercomputer might not be conscious in any sense we would recognize, it's able to use an operating system called Viva to continually reconfigure itself, adapting specifically to deal with computing situations it's handed.
We're just exploring the possibilities of evolutionary algorithms, but already applications are apparent in areas such as image recognition, in which a PC might continually refine its methods of identifying what it sees, leading to machines that can recognize a human face. And evolvable hardware means future computers might be able to upgrade their core circuitry simply by downloading new code.
In Japan, Tetsuya Higuchi and his fellow computer scientists at the Electro-Technical Laboratory are using genetic algorithms to build analog circuit components that will go into new cellular telephones. Adaptive hardware is also being studied at the Jet Propulsion Laboratory in Pasadena, Calif., to create adaptive sensors for spacecraft. Evolvable computers aren't yet front-page stuff, but I think they will take us in directions too potent to ignore.
Paul A. Gilster can be reached at firstname.lastname@example.org
-- Dan S
[ISML] Insane Science Mailing List
- To subscribe: http://www.onelist.com/subscribe.cgi/isml
Your use of Yahoo! Groups is subject to http://docs.yahoo.com/info/terms/
This archive was generated by hypermail 2b30 : Mon May 28 2001 - 09:59:45 MDT