Report on the Second NASA/DoD Workshop
on
Evolvable Hardware
IEEE Transactions on
Evolutionary Computation
Vol. 5, No. 3, June 2001
Hugo de Garis
Head of the Brain Builder Group,
Starlab, Brussels, Belgium.
degaris@starlab.net
http://foobar.starlab.net/~degaris
The second annual NASA/DoD Workshop on Evolvable Hardware (EH-2000), took
place on
July 13-15, 2000, in Palo Alto, California, USA, sponsored by NASA and DARPA
(Defense Advanced Research Projects Agency), and co-hosted by NASA Ames
(Information Sciences and Technology Directorate), the JPL (Jet Propulsion
Lab)
Center for Integrated Space Microsystems (CISM), and the JPL Center for Space
Microelectronics Technology (CSMT), and with the cooperation of various other
divisions of NASA Ames.
The very fact that these evolvable hardware workshops are now an annual event
shows clearly that the field has been established solidly in the U.S., after
America's initial four-year delay behind Switzerland.
The Workshop
About 90 people from 11 different countries attended the workshop. (This was
down from the 130 who attended the prior workshop in 1999.) The three-day
workshop was held fully in plenary session. The invited talks were
spread over the three days, but most occured during the first morning. Once
the
invited talks concluded, most of which will be highlighted below, the
submitted
paper talks were divided into the following eight topics, which give an
indication
of how the EH field is evolving in human terms.
1. Algorithms
2. From Biology to Robotics
3. Evolvability
4. Evolution of Analog and Mixed Signal Circuits
5. Evolution of Digital Functions
6. Reconfiguration Architecture and Devices
7. Evolution of CA and Brain-Inspired Architecture
8. Real World Applications
Highlights of the Workshop
First, I will cover the invited speakers.
1. Steve Zornetzer (NASA) - "NASA's future will depend on EH"
What struck me initially about the 2nd NASA/DoD EH workshop was the very
impressive public relations job its organizers had done on their managements
in
the year since the 1st workshop. In particular, Adrian Stoica from JPL and
Jason
Lohn from NASA Ames (two of the driving forces of EH in the U.S.) and others,
had
persuaded their superiors of the importance of this new research field for
the
future of NASA and for the future of space travel in general. For example,
Dr.
Steve Zornetzer, Director of the NASA Ames Information Systems and Technology
Directorate, made a very strong statement in his kickoff speech, "Maybe
NASA's
future will depend on evolvable hardware." Zornetzer has the reputation of
supporting very innovative research, such as nanotechnology, which he did
several years before it became main stream. His brief introductory talk asked
the question, "Why is NASA interested in EH?" I found his answers very
persuasive and promising for the future development of EH in the U.S. in
general and for space in particular. He said that future planetary and deep
space exploration over the next 20-30 years will demand that the space
vehicles
have robust system architectures and be dynamically reconfigurable. Current
technologies cannot cope, he said, because they are just not robust enough.
Future space missions will send probes into the saline oceans under Europa's
ice
cap. How can such an unpredictable environment be explored? Obviously with
autonomous, real-time robots, which must be robust, dynamically
reconfigurable, and not brittle. It is impractical for NASA to send up triply
redundant systems to ensure reliability because the weight problem makes that
too expensive. Future architectures need to be adaptive to uncertain
environments. EH may have a major role to play in this regard. Zornetzer's
vision is to
develop three broad areas of competence, 1) space technology, 2) nanotech,
and
3) information technology. He has set up a newly funded, Intelligent Systems
(IS) program ($70M/yr), and a Space Biology program, which researches
biologically-based systems for space. He also set up a nanotech initiative.
IS
includes automated reasoning, human-centered computing, and revolutionary
computing (e.g. quantum computing, neural nets, evolvable hardware, etc.).
Space
biology includes understanding biological principles, the functioning of
nanotubes, EH, etc. The nanotech initiative is seen as a requirement for
future
NASA missions, to revolutionize electronics, for autonomous microrovers, mini
helicopters, etc. The IS initiative received 500-600 proposals. The space
biology initiative went out for January 2001, and the nano proposal will be
announced in 2001. There are other funding programs which are unsolicited.
2. Nikzad Toomarian (JPL) - "EH for 100+ year space probe survival"
The second institutional speaker was Toomarian of JPL, who made rather
similar
remarks to Zornetzer's, such as "EH is needed for deep space exploration in
extreme environments (-150 to 300 degrees)." He stressed that the various
planned
space missions, such as the Pluto express, launch date 2004, with a flight
time of 8-9 years, and later interstellar explorations, will need to
emphasise
long term survivability and evolvability. He wants to see furture space
hardware
systems based on nature's adaptibility that can re-evolve themselves in
seconds.
He wants EH for 100+ years survival of space systems with low power
and high intelligence.
3. Carver Mead - "I've been doing EH all my life"
Mead, who is in his mid 60s, is famous in electronics circles. He virtually
invented most of its
modern approaches. He's had several research careers in four fields,
1) device physics (smaller transistors, MOSFETs), 2) VLSI
design methodology (silicon boundaries), 3) biological electronics,
neuromorphic
scaling behavior (silicon cochlea and retina), and 4) collective
electrodynamics
(with a new book to come out on this). He said he had founded 23 startup
companies and that in a sense "I have been doing EH all my life." He designed
his first chip in 1971, a programmable logic array, PLA, while being involved
with Texas Instruments and Hewlett Packard at the same time. His 2nd
chip had an analog output. Then he was into CPLDs (complex programmable logic
devices), programmable state machines, data path elements, and the beginnings
of
RISC (reduced instruction set computer). He spoke of how the density of
wiring
can overwhelm chip design, and the need for good wire scaling. This was in
the
early 80s, when harbinging the FPGA (field programmable gate array). Also in
the
early 80s he got involved with Richard Feynman, the famous quantum
electrodynamics
theorist, and the physics of computation.
Mead got hooked on the idea of trying to get electronics to do what brains
do,
resulting in the development of a system that performs inner product learning
autonomously and in real time - hence his famous silicon retina and cochlea.
This led to his interest in reusable, configurable, and self-configurable HW
that can learn from the environment, performing analog processing on floating
silicon gates and do neuron-like things. He feels that Lamarckian processing
(i.e., genetic inheritance of socially acquired characteristics) is the way
to
go. We, that is humanity, cannot wait millions of years for slow Darwinian
evolution as a paradigm for building evolvable systems. He is very interested
in
brain functioning, asking how a human baby learns about the world. A baby's
nervous system proliferates early, and later the neurons/synapses die off.
Learning is the change in synaptic strengths. All these phenomena have analog
and digital properties. The mutation of genes is digital but the influence of
the environment on the genes is analog. There is lots of analog stuff in the
brain, he said. He spoke about the neural code, i.e., the relative time of
arrival of various neural signals. Neural signals are self clocking. There is
no
global clock. In dendritic trees, synapses form nonlinear analog interactions
of pulses. It's the coincidences that evolve. Certain combinations of pulses
at
the dendritic trees are reinforced, the rest are not significant. A fly's
brain
does more than the best computer. He felt that the evolution of the brain
follows a gradient into a higher-dimensional space. It uses a neural code and
is
robust. If 10% of the neurons die, that has no effect. The brain is not
brittle.
He spoke of the duplication of genes as a means of evolution. One gene is
modified by mutation while the other is intact. Sometimes the modified gene
is
used, so some experimentation is tried. Perhaps the brain areas were
duplicated
and experimented with. He felt that the so called "junk DNA" is not junk, but
rather a library of old ideas. He felt the discrete and the continuous analog
aspects of the brain play a role. In the discrete space of genes, one can
copy
and pass on the information and not lose information. In learning, one has
continuous gradients of computing structure. These two processes should be
integrated he felt. He even got into a bit of quantum theory. With electron
tunneling, he felt some systems could take advantage of wave functions
overlapping to create one large wave function. His views on the link between
QM
(quantum mechanics) and biology were not clear to me. He finished his talk by
saying that the electronics field could make some "awesome devices using bio
tricks."
It was wonderful to hear this brilliant famous man, with a solid electronics
background, talking with such passion and knowledge about the brain and
evolution. He had obviously spent a lot of his time in recent years studying
these fields, providing further proof of how interdisciplinary the most
creative minds are becoming lately.
4. Michael Conrad (Wayne State) - "The father of molecular computing"
Conrad, who passed away in December 2000 due to an illness with cancer,
was the pioneer of molecular computing and was famous for that.
He was head of the Biocomputing Group at Wayne State University, in Detroit.
One
of his major goals was to get molecules to compute.
Conrad gave a talk on his latest work with his grad students on
getting molecules to compute. It was not an easy talk to follow for someone
unfamiliar with the field. It was factually dense, and the listener sensed
that
the man was trying to summarise a life's work in half an hour. He mentioned
that
molecules perform according to a "shape-based pattern recognition" paradigm.
Shape-based systems are now buildable he said. Molecules need to be big
enough
for lock and key pattern recognition. They explore each other thanks to
Brownian
motion. Molecules are not sloppy, they are discrete, with lots of nonlinear
dynamics. They are built in a precise manner. The way they fold depends on
their
milieu. Molecules allow for conformational signal processing and
conformational
signal fusion, which can lead to macroscopic action, e.g., a neuron firing.
One
can send in chemical signals to molecules which cause conformational change,
which results in an output. These ideas can be implemented in an enzymatic
pattern processing flow system, using the enzymes as the pattern recognizers.
He
used such a system (and I had trouble with the details, not being a molecular
biologist nor a biochemist) to perform Boolean operations such as OR, XOR,
NAND,
etc. The processing is context sensitive. He admitted that programmability is
a
conceptual bottle neck at the moment in molecular- or bio-computing. He also
said that a lot of evolutionary work is needed to make molecular computing
useful, and that it will take years of researchers time to develop it.
5. Tetsuya Higuchi (ETL) - "E-Hard finds its killer-ap"
Higuchi is one of the pillars of the EH community, having been one of
the field's founding fathers. He
began his EH work by evolving GAL chips, but extrinsically, not
intrinsically,
i.e., he simulated the evolution in software. He did not evolve directly in
the chip
itself. Higuchi's approach to EH has been very applied for the past few
years.
He is getting MITI (the Japanese Ministry of International Trade and
Industry)
money to pursue his research within his "Evolvable Hardware Systems" group at
ETL. (http://www.etl.go.jp/~ehw/) His talk was entitled "Evolvable Hardware
for
Industrial Applications" which reflects his primary interests. For several
years
Higuchi has been saying that "for the EH field to survive it will need to
find
its "killer ap" (big selling application)."
I think it's fair to make several
remarks here. The first is that the field is doing just fine, judging by the
number of researchers and groups listed in Adrian Thompson's EH list
(http://www.cogs.susx.ac.uk/users/adrianth/EHW_groups.html) I counted 48
entries
(22 Jan 2001) (UK 16, USA 11, Germany 5, Italy 3, Canada/Japan/Norway 2,
Denmark/Czech/Holland/Switzerland/Romania/Australia/Belgium 1). The second is
that Higuchi himself has recently found a "killer ap" in the form of evolving
analog intermediate frequency (IF) filters for cellular phones that NEC will
manufacture at a rate of half a million per month! Higuchi spoke twice at the
workshop, firstly as an invited speaker, and secondly as an regular
speaker. In his invited speech, he gave examples of EH being applied by his
group at ETL to industrial tasks, including work on a prosthetic hand, an
autonomous robot, analog chips in cell phones, and femptosecond laser system
alignment.
Session Speaker Highlights
There were some 30 session speaker slots. Due to space limitations,
I review only three here.
Jordan Pollack -- "Throwaway Robots"
Pollack of Brandeis University, USA, presented his "Golem Project"
(Genetically Organized Lifelike Electro Mechanics) which co-evolves
simultaneously both robot bodies and their neural net controllers, or as he
describes it, the "fully automated design of throwaway robots". He used an
evolutionary algorithm to co-evolve 1) a neural net that was applied to
motors to
drive ball-and-socket jointed bars, and 2) the configuration of the bars to
form
a body, or robot. These bars were evolved to form a 3D "graph" of nodes (the
joints) and lines (the bars). The linear motors made the lengths of the bars
change, hence generating motion of the whole body. The fitness definition was
the distance covered by the robot body in a given time. The elite individual
after an EA run could then be built physically using a "3D fax" or "3D
printing"
technique called "rapid prototyping" in the commercial sector. This involves
a
machine that uses a temperature-controlled head to extrude thermoplastic
material layer by layer to build arbitrary 3D shapes. This device would build
the robot body with its bars and joints. The process was semi-automatic. The
only task requiring human intervention was the insertion of the motors. What
was
new in Pollack's work relative to Karl Sims work, who similarly co-evolved
bodies (interconnected blocks) and their neural controllers, was the
semi-automated physical construction of the robots based on the evolved
design.
A colorful multimedia website presenting this work can be found at
http://www.demo.cs.brandeis.edu/golem. Pollack's work was published
in Nature [1].
Adrian Thompson -- "Evolutionary Nano-Electronics"
Thompson of Sussex University, U.K., is well known amongst evolvable
hardware (E-Hard) researchers for his work on intrinsic EH. Thompson is
now helping to pioneer the application of
evolvable hardware techniques to the area of nanometer scale electronics or
"nanoelectronics." This development is inevitable as Moore's law forces the
scale of electronics to become ever smaller. Electronic processing
speeds increase as the distance between components decreases, given the fixed
and finite speed of light. Unfortunately, he was unable to attend the
workshop,
so he sent a video presentation
instead. One of Thompson's main messages in his work over the last few years,
is
that EH can exploit the physics of the devices one evolves, for example, in
his
case, exploiting the inherent analog behavior of digital devices to perform a
given task. In his "talk" (video) he tried to show how this "exploiting the
physics" idea could be applied to the evolution of behavior in "single
electronics" systems, i.e. those with very few electrons, e.g., quantum dots,
SETs (single electron transistors), etc. The case study he chose was to
evolve a
"single electron NOR gate" in which he felt that "the particular thermal
energies of the electrons are exploited." Hence hopefully, it may be possible
to
do in nanoelectronics what he did in microelectronics (FPGAs), i.e., to
"explore
beyond the scope of conventional design". Thompson had access to a software
simulator of a single electron NOR gate circuit so he used this simulator as
the
basis of his investigation. The "gate" Thompson worked with consisted of an
array of interconnected capacitors and tunneling "junctions." A junction
consists of a thin layer of insulating material sandwiched between two slices
of
conducting material. Single electrons can "quantum tunnel" across the
insulator
junction.
A genetic algorithm was used to evolve arrays that consisted of a graph of
nodes, arranged in a 7 by 4 matrix, whose links could contain a junction, a
capacitor, a wire, or nothing. Links were only local, i.e., between
neighboring
nodes only. Numerical values for the capacitance of the capacitors and the
tunnel resistance of the junctions were also evolved. The top row of nodes
were
supplied with a given voltage, and the bottom row of nodes kept at 0 volts.
No
components were allowed between the nodes in the top or bottom rows. The two
input nodes and the output node were fixed. The fitness was defined to be
proportional to how well a target output curve matched the actual. The target
curve was a simple NOR function of two straightforward semi-square wave
inputs.
The evolution first took place in the simulation at 0 degrees Kelvin, and
evolved well, but the fitness crashed if the temperature was increased by a
mere
30 mK. Thompson then decided on an "incremental" or "stepwise" evolutionary
approach to overcome the temperature sensitivity. Once the elite member of
the
population, of 30, reached a given fitness level, the temperature of
simulation
was raised by 10mK, and the resulting population at the lower temperature was
used as the starting population for the higher temperature run. This
procedure
was followed until 340 mK. What he was then very surprised to discover was
that
not only would the fitness drop off as the temperature of simulation went
above
340 mK, as would be expected, but that it also dropped off as the temperature
decreased. Thompson interpreted this odd behavior to mean that the "circuit
exploits or relies upon the particular thermal energies of the electrons at
around 340 mK." This phenomenon may be a further example of the type of
exploitation of the (nanoelectronic) physics that Thompson is so fascinated
by.
He is continuting his work on this phenomenon. Nano electronics people ought
to
study this work closely, as it opens up new avenues. In my view, this paper
was
the best at the workshop, and potentially the most significant for the
future,
which explains why I have devoted so much space to it.
I should note that Thompson is not the first to work on the theme of
evolutionary nanoelectronics. Adrian Stoica et al. published a paper entitled
"Evolutionary Design of Electronic Devices and Circuits" in 1999 [2].
Quoting from the abstract of that paper "This paper addresses the use of
evolutionary algorithms in the
design of electronic devices and circuits. In particular, the paper
introduces
the idea of evolutionary design of nanodevices, and illustrates it with the
design of a resonant tunneling diode... ." Stoica and his colleagues at JPL
also
published a short paper called "Genetically Engineered Nanoelectronics" at
the
1st NASA/DoD EH workshop in 1999, on page 247 of the proceedings.
Gary Fehr -- "Brain Building Machine, CBM"
Gary Fehr of Genobyte, Inc. of Boulder, Colorado, USA, develops software for
the CAM-Brain Machine (CBM), The CBM is a piece of specialised evolvable
hardware
that evolves a 3D cellular automata-based neural network circuit module in a
few
seconds. The elite modules are downloaded into a gigabyte of RAM, one at a
time,
for 64000 modules, which are then interconnected according to the
architectural
decisions of human "BAs" (brain architects) to build artificial brains. Fehr
showed off the capacities of the CBM. For example, it uses Xilinx's XC6264
FPGA
chips (which have now been taken off the market because they were not popular
with electronic engineers, and did not sell well) to update the billion 3D CA
cells in the CBM at a rate of 130 billion a second. The CBM can execute a
complete genetic algorithm run with a population of 100 modules over 100
generations in a second or so. There are maximum 1152 neurons per module, so
with the 64000 modules that the CBM can handle, thats implies an artificial
brain with the order of 100 million artificial neurons. Such a brain could be
updated fast enough by the CBM to enable real-time control of robots. The CBM
is
the most powerful and one of the most significant examples of evolvable
hardware
on the planet. (See the Guinness Book of World Records, 2001, p126). Its
design
and construction was contracted to Michael Korkin, the president of
Genobyte, by the de Garis' previous research lab in Kyoto, Japan (ATR) to
implement de Garis' ideas on building artificial brains. There are four CBMs
in
the world: one in Japan at ATR, two in Belgium (one at Starlab, the author's
present lab), and one at Genobyte in the U.S. Starlab obtained a $1M
grant from the Brussels government in 2000 to build an artificial brain using
its CBM.
Panel Discussion -- "Impact of Industry on E-Hard"
An industrial panel consisting of people from Virtual Computer Company (VCC),
HP
Labs, Triscend Inc, Xilinx, etc. discussed how industry could impact the EH
field at four different levels: 1) user, 2) silicon level, 3) tools for EH,
and 4)
systems. The official title of the discussion was "The Future of
Reconfigurable
Computing Technologies: Industry Viewpoint." The HP Lab's representative said
that current design methodologies don't scale well. For example, with HP
Lab's
microprocessor 200X, the gates are effectively free, the wires are expensive,
design of symmetry is easy, and the design of irregularity is hard. He spoke
about HP's Teramac project which consisted of a million gates operating
at a megahertz with faulty components. One of the aims of the project was to
create a functional system from imperfect components. The team found 220,000
defects, but they got their machine working nevertheless, a sort of genetic
algorithm supercomputer.
The question was asked, "What can the (US) government do for EH?" The general
answer, which I heartily agree with (see below in my general remarks), was
the
need for the government to fund the development of a large FPGA with definite
dynamic reconfigurability characteristics for general use by the whole of the
EH
community. Such a chip could replace the Xilinx 6200 family of FPGAs which
have
traditionally been the intrinsic (in chip) EH chips of preference (see
below).
The Triscend Inc.'s representative spoke mostly about "configurable systems
on a
chip" ("CSoC"s) mainly in a communications application context - for example
his
company's "E5" devices. (See http://www.triscend.com.) He felt that such
chips
accelerate the time to market, will have about 10 million gates, and create a
market of about $50B in 10 years. He mentioned a list of companies doing the
same kind of thing (Triscend, Atmel, Quick Logic, Chameleon, Altera, Lucent,
LSI
Logic, etc). His company's E5 chip is an eight-bit microcontroller, a
byte-wide SRAM
chip, with a processor and DMA controller on a single die. It can access any
flip flop, and is fast. He made a remark that I really picked up on, namely -
"The Triscend E5 has features useful for reconfiguration". E-Harders who are
curious might like to check out their web site (and the others) to see if
such
"CSoC"s might be suitable for EH.
The Xilinx representatives (Delon Levi and Bill Carter) said that EH tools
need
to be created based on basic principles, for example, 1) people evolve first
the
functional FPGA nodes and then evolve the routing between them, in the
context
of CAs (cellular automata), NNs (neural networks), DSP (digital signal
processors) filters, etc. 2) use main stream products so that the know-how of
the experts can be used, 3) try to solve problems that engineers cant solve
or
that can be done more cheaply than paying an engineer, e.g. voice/video
recognition etc. As for application areas, Levi suggested, pattern
recognition,
video, evolving the feature detectors, evolving DSP filters, cleaning up
signals
in noisy channels etc. Carter said that EH was not a big deal at Xilinx, that
it
was confined largely to academics. He felt that it was NASA that was driving
the
need for reconfigurability, for its long duration space voyages.
The VCC representative, John Schewel, with a colorful personality, asked
"What I
want is an artificial brain, but I get Kephera robots." "Is a digital soul
possible?" He wants to incorporate the dynamics into computing architectures
rather than simulating the dynamics. He wants to build a machine with
billions
of gates of reconfigurable logic that can be used in 2001.
Demonstration -- "CAM-Brain Machine CBM"
The conference featured one major demonstration, namely the hardware
unveiling
of the "CAM-Brain Machine (CBM)" mentioned above in Fehr's talk. Korkin and
Fehr
transported their own CBM to the workshop and showed it off to a crowd of
onlookers in the lobby. Close to the machine was a large poster which read -
"Genobyte, CAM-Brain Machine (CBM), A large scale evolvable hardware research
platform; Cellular Automata Update Rate: 131 billion cells/second; Number of
Supported Cellular Automata Cells: 893 million; Number of Supported Neurons
(max., per neural module): 1152; Number of Supported Neural Modules: 64,640;
Number of Supported Neurons (max. per brain): 74.5 million; Neural Module
Chromosome Length: 91,000 bits; Number of FPGAs: 72 (Xilinx XC6264BG560);
Number
of FPGA Reconfigurable Function Units: 1,179,648; Phenotype/Genotype Memory:
1.18Gbytes; Power Consumption: 1.5KWatt (5V, 300A); Computational Power:
10,000
Pentium III 500 MHz PCs." Hopefully in the summer of 2001, de Garis and
Genobyte will be able to show off what they have been able to build with the
CBMs, namely an artificial brain architecture, or at least to show
substantial
progress along that route.
General Remarks
The first thing that surprised me about the workshop was that the definition
of the word
"hardware" had changed. As a father of the field in a sense, having had the
original idea of considering the configuring bitstring of FPGAs as a genetic
algorithm chromosome, I had the unconscious assumption that the term
"evolvable
hardware" meant "evolvable electronic hardware," but Pollack's paper, which I
thought was one of the highlights of the workshop, showed that the workshop's
organizers had generalized the definition to mean "evolvable matter." This
seems
to me to be a good thing, and opens up new avenues for research. For example,
Thompson's evolvable nanoelectronics work makes the distinction between
electronics and matter fuzzy, because at a small enough scale, e.g.,
nanometers,
everything, whether electronics or matter, is made up of molecules.
My surprise at the redefinition of the term "evolvable hardware" was however
somewhat unjustified, because in fact Stoica had alread said in the preface
of
the proceedings of the 1st NASA/DoD EH workshop in 1999 that "While this
(electronics) is still the main vein of research, the semantic envelope has
expanded to include various forms of hardware, from sensors to antennas to
complete evolvable space systems... ."
Another remark which I think needs to be made was my impression that the
workshop was "reinventing the wheel" to some extent. Some of the papers were
definite "evolutionary computation" (EC) conference-type papers that would
have been
serious candidates for publication in a "pure EC" conference. Since these EH
workshops use evolutionary computational techniques as the main tool to
evolve
its electronics (or hardware "matter") there is a case to be made in allowing
EC
type papers into an EH workshop, provided that they are not repititions of
work
that has already been done by the EC research community. There is already a
large literature of EC work that has accumulated over the decades. My feeling
was listening to some of the EC-type papers that they were just reinventions
of
earlier EC work that these EH researchers could have consulted in the EC
journals and proceedings.
Some of the papers, especially towards the end of the workshop, were of
questionable relevance. For example, the very last paper was on the use of a
genetic
algorithm
to minimize the drag of fluid flow around a cylinder by placing actuators on
the
cylinder's surface. If such a paper is considered evolvable hardware, then so
would many papers in the applications sections of EC conferences, thus
diluting
the meaning of the term "evolvable hardware" to almost nothing.
At a broader level, I was concerned that few remarks were made at the
workshop
at the lack of a replacement for the Xilinx 6400 FPGA family which has been
the
intrinsic EH workhorse for the field for several years. Xilinx's emphasis on
its
current "Virtex" FPGA family creates difficulties for the EH field. The most
obvious one is that the detailed architecture of the Virtex is a Xilinx
company
secret, unlike the public disclosure of the XC6400 family architecture. EH
researchers cannot write software for the evolution of the Virtex because
they
dont know the detailed circuitry. At the moment the only people who can do
that
are the EH researchers working inside Xilinx, lead by Delon Levi, who have
in-house secret knowledge. This monopolistic situation is not healthy for the
EH
field. EH researchers can obtain the "JBits" software from Levi
(Delon.Levi@xilinx.com) which acts as an interface between EH researchers who
desire to evolve a Virtex chip and the secret Virtex internal architecture.
What
is needed is that the EH community develops its own evolvable chips that the
field can use, and make them public, so that researchers all over the planet
can
contribute to their development, rather analogous to Linux, the "freeware"
operating system. The problem however is cost. Xilinx spent millions of
dollars
developing their Virtex.
Fortunately, there are moves afoot to overcome this problem. Stoica and his
colleagues at JPL have designed three generations of evolvable chips already.
The 1st was fabricated in Oct 1998, the 2nd in Oct 2000, and the 3rd in March
2001. This work has been funded by DARPA and NASA, and the designs are
public.
The Stoica team does not support these chips in the sense of baby-sitting
users,
but it has given access to the chips via the internet to other organisations,
e.g. NASA Ames and Natural Selection Inc. The intention of this
government funding at JPL is that the chip designs will migrate towards
industry. If industry takes up these designs, they are welcome.
Didier Keymeulen, a colleague of Adrian Stoica at JPL, feels that their
analog
reconfigurable chip is an alternative to the Xilinx XC6200 family of FPGAs.
He
says that one can "design your own chip on a laptop" due to the use of
increasingly affordable design tools.
Acknowledgments
I would like to thank Adrian Stoica and Didier Keymeulen, both at JPL,
for their suggested improvements and corrections of misconceptions in an
earlier
draft of this report.
References
1. H. Lipson and J. B. Pollack, "Automatic Design and Manufacture of Robotic
Lifeforms", Nature 406, pp. 974-978, 2000.
2. A. Stoica, G. Klimeck, C. Salazar-Lazaro, D. Keymeulen, A. Thakoor,
"Evolutionary Design of Electronic Devices and Circuits," Proceedings of the
1999 Congress on Evolutionary Computation, IEEE Press, Vol. 2, pp 1271-1278,
1999.