123 Compute.Net

Dreaming SiteMap


Evolving a Conscious Machine

by Gary Taubes

Discover Magazine, June 1998, pp.73-79

THE PLACE: The Swan, a pub in Falmer, outside Brighton, England, an hour from London, few hundred meters from University of Sussex.

University of Sussex houses COGS, the School of Cognitive & Computing Sciences.

Also supports a Center for Computational Neuroscience and Robotics, which is COGS + Biology.

Someone there is also studying the mechanisms of ant navigation.

THE PEOPLE: Inman Harvey, Adrian Thompson, and the ghost of Hugo de Garis.

Hugo de Garis created programmable circuits in 1992. He's now a visiting scientist at Advanced Telecommunications Research Institute in Kyoto. Also affiliated with George Masson University in Virginia. Dabbled in a programming technology called genetic algorithm.

de Garis: "I am a builder of brains".

Inman Harvey: computer scientist become "evolutionary roboticist". Spent 20 years in Afghanistan in the import/export business.

Adrian Thompson married evolution to FPGA chips independently of De Garis, also around 1992.


According to Harvey: We know it when we see it...

What do you trust more- a veering car or a loose boulder?

If we don't understand our own consciousness, so how can we exclude its existence elsewhere?

Once we CAN'T understand how computers compute, we'll attribute Consciousness to them!

According to Thompson: "I don't think the work I'm doing says anything about it."

It makes little sense to ask whether something IS conscious. What matters is... does it ACT conscious?


The Hardware is FPGA cells.

The Software is evolution.

The Premise: "Species evolve because the offspring best suited to thrive in their environment are those most likely to breed successfully and pass on their genes to the next generation."

Faster race horses

Bats that can "see" in the dark

"After several thousands or millions of years, the result will be creatures uniquely adapted for living in particular environments."

Start with a bit string of yeses and nos that are potential "solutions" to problems; these are the "chromosomes" of the solution to be evolved.

Test offspring against a fitness scale- more like animal husbandry than evolution because the scientist knows exactly what's wanted.

Parts of each HiScore bitstring are mixed, mutations added...

Thompson's Project: Evolve hardware by having it solve problems.

Using programmable chips that can reconfigure on the fly.

FPGA's: Field Programmable Gate Array chips. Thompson uses the Xilinx SC216.

Config bits can be changed: OR gates to AND or NOT gates, for example. INPUT to OUTPUT

Silicon processor changes its wiring in billionth of a second

Processor graded on how well it can distinquish between "Yes' and "No"

High scorers are mated...and so it goes until

After 1000's of generations, a circuit becomes "flabbergastingly efficient".

Flexible but not fast: Manufactures can ship FPGA's first, reconfigure, then ship final. Used now for products such as Routers.

Once the FPGA is debugged, make a special purpose chip to do same job.

Extrinsic Evolution: FPGA cells, circuits, breeding, testing can be simulated on computer.

Intrinsic Evolution uses chips directly, bypassing simulation. Thompson uses this method in order to take advantage of "peripheral phenomema".

...in order to investigate what he doesn't understand....

"...he decided he didn't want to constrain himself by making assumptions about how evolution should work on a computer based on how it works in nature."

Didn't tell his genetic algorithm it was dealing with a digital devise whose circuitry could only have specific states of on/off...

Let evolution work on the circuitry as though it were an analog device, in which signals that pass down a wire can take on any value between 0 and 1.

There could be varying degrees of MAYBE, if that's the way evolution wanted it.

Thompson's Task: Evolve circuits that distinquish between two tones.

Use 100 of the chip's 4096 cells (logic elements)

No clock, no timing components.

Logic elements work quickly, tones' frequency a million times slower

"..a task something like trying to evolve a human who could tell the difference between a year and a decade while using only the second hand of his watch and without counting to himself as he did it."

THE ALGORITHM: individuals, bursts, parenting, elitism

The number of config bits needed to describe fully the wiring of those 100 cells is: 50 random bit strings, each 1800 bits long.

These are the 50 individuals to be run through the evolutionary process.

Pump in 5 bursts of 1 kilohertz, and 5 of10 kilohertz, in random order. Look at the output for each individual. Find maximum difference.

Choose parents for next generation at random, but with bias toward best scorers

Copy single best individual over unchanged to next generation. (elitism)

After 5000 generations and two weeks of computer time, the computer was distinquishing between the two tones.


Thompson doesn't know exactly how the evolved circuits are doing it.

Only 32 of the 100 cells are in use. Holding voltage constant on the other 68 doesn't affect performance.

Human design would have need 10 to 100 times as many logic elements, or at least access to a clock, to perform same task.

Performance: Current feeds back and forth through the gates, "swirling around", says Thompson

Some of the 32 chips seem out of the loop altogether, yet affecting the performance of the chip.

OUT of LOOP circuits: How are they affecting performance?

Electromagnetic coupling: Broadcasts signals without sending current down the wiring.-Put this to work!

Communicating through power supply wiring? (Each cell was hooked up independently to same supply .)

and/or through silicon substate in which they lie?

Is the process even taking advantage of the computer that's hooked up to the system to run the algorithm?

Evolved circuitry only survives in the environment in which it is born.

Change temperature and the circuits won't work

Change algorithm program on the PC- and circuits won't work.

Ending up with machines as different from each other as two humans, and whose thinking processes are equally inscrutable.

"What should be done next?"

Coordinate all the computer's thought processes to come up with an answer.

In doing so, it would ..have to differentiate between all the myraid FPGA's and circuits that constituted its self and the external world.

Evolve a higher level of autostimulation, using the language in which it communicated with its programmers to communicate with itself.

Thompson now working on ROBUSTNESS: Evolve circuits that can work on more than an single FPGA

Evolve circuits on 5 FPGA's at once: all Xilink, but made at different plants. Some are even seconds.

Change the temperature during the evolutionary process, to develop tolerance.

"The result might be not just a sense of self, but an inner voice to go with it."

Once again-It makes little sense to ask where something IS conscious. What matters is... does it ACT conscious?

Is human consciousness inevitable, or some perverse accident of nature (as this author claims most students of evolution believe)?

Evolving a conscious computer would take a very long time... Then, says Thompson, there's "the problem of us dying during the experiment".

"I'm really exploring what evolution can do that human designers cannot."


Email comments or questions to webmaster@123compute.net

Dreaming SiteMap

This page was last built on Mon, Oct 26, 1998 at 5:30:12 PM with Frontier on a Macintosh.