PerAc94 Conference Report, Hugo de Garis, ATR, Kyoto

[Hugo de Garis]

             PerAc94  Conference Report, Hugo de Garis, ATR, Kyoto

Tom Ray and I were invited by the Federal Poytechnic of Lausanne, Switzerland,
to give talks on our ATR work to an ALifey type tutorial, which lasted two
days, (Monday Sept 5 and 6). This was followed by a 3 day conference called 
"PerAc94", i.e. from Perception to Action, whose organiser asked me to write 
up and distribute a conference report over the relevant email networks. 
Following the PerAc conference was a 20 man brain storming session of invited 
experts on the future of the subject. I was originally invited to this, so I 
had expected to be able to report to you on the main issues, but when I asked 
the organizer (a rather prickly person) if it would be ok if I attended only 
half of it, (for reasons explained below) I got disinvited. A formal write up 
of this brain storming session will appear in the "Robotics and Autonomous 
Systems Journal", (Elsevier) in the next few months.

The following Monday was a seminar on the work of Professor Mange and his 
group on adaptible hardware using FPGAs (field programmable gate arrays) which
I believe to be of great importance. This report will focus on the second and 
fourth of these events.


PerAc94 Conference

This conference was a Swiss equivalent of the PPSN (Parallel Problem Solving
from Nature), SAB (Simulation of Adaptive Behavior), and ALife type of 
conferences, where the key words characterizing the conference can be taken 
from the session headings, namely, "collective intelligence, simple behavioral 
robots, genetic algorithms, active perception, building blocks and 
architectures for designing intelligent systems, complex architectures to 
control autonomous robots, cognition, collective intelligence, simple
behavioral robots, active perception".

When I accepted to speak at the tutorial it was for two reasons. One was to
talk with Professor Mange and his group, who are pioneering a new field that
I will talk about later in this report, and the other was to be in Switzerland,
for me the most beautiful country in the world. When I had to leave it was
with a real sigh. The PerAc conference itself I expected to be rather "small
beer", but in fact the world was there. Admittedly only about 100 people were
present the day I counted, but they were a specialised audience. If you are
working in the field of autonomous robots and their related problems of
control, and in particular the problems involved in converting perception
into action, then getting hold of the proceedings of this conference for you
is a must. Happily, that will not be difficult, because, thanks to 
characteristic Swiss efficiency, an IEEE (Computer Society Press) book was
ready and distributed to conference attendees. The book is entitled 
"Proceedings From Perception to Action Conference, Lausanne, Switzerland, 
September 7-9, 1994", IEEE Computer Society Press, 1994, ISBN 0-8186-6482-7,
edited by P. Gaussier and J-D. Nicoud.

The 3 day conference was small enough not to need split sessions, so some 30 
odd talks and a similar number of posters were available for all to hear and 
see. Unfortunately that was not true in my case. The great beauty of 
Switzerland attracted my wife to take a train down from Brussels (where she
was catching up on old friends), and she was on a tight schedule. Similarly,
a close Japanese friend of mine came up from Geneva on an equally tight 
schedule, and both of them wanted to yodel in the mountains with me. I ended 
up attending only 1 of the 3 conference days. Hence this report will be more 
limited and objective (based on the book) than subjective (based on 
impressions from someone who attended all the sessions). 


Conference Highlights 

The hero of the conference in my view was not a human being, but an (ice 
hockey) puck-shaped robot called "KHEPERA", developed by the Swiss, which I 
found myself in conversation referring to as the "puck robot". This flat, 
round, battery or mains driven, wheeled robot (containing infra red and other 
sensors), measures about 5 cms in diameter, about 2 cms high and is a
very versatile little tool to perform "evolutionary robotics" experiments on 
your desk. Many of the papers at the conference used this "puck robot" to 
obtain their results. These researchers were performing real time (i.e. 
mechanical) fitness measurements of the robot's motions controlled by the 
neural net systems they were evolving. It is far more practical to do such a 
thing on a puck robot than a Unimate (of car assembling size). If you are 
interested in buying one of these puck robots, then try emailing Mr. Gomi, the 
manufacturer and distributor of Brook's insect robots and others. He told me 
he has already sold some 40 of these little robots. Gomi is a Japanese 
Canadian, who speaks excellent English. His email address is 
71021.2755 at CompuServe.Com  These little puck robots, crammed with electronics 
and sensors, cannot be too expensive. 

At the conference itself there was a fair sprinkling of big names from around 
the world, e.g.

Pfeifer (Zurich Univ, Switzerland), the keynote speaker, made the claim that 
too much research effort is going into the unidirectional approach, i.e. from 
perception to action, whereas in the biological world, the reverse is also 
true, namely that what one perceives also depends on ones actions. He hoped 
future research in the field would be more bidirectional.

Deneubourg (Brussels Univ, Belgium) spoke about the self organization of 
transport systems in ants and robots.

Fukuda (Nagoya Univ, Japan) showed how his cellular robots could connect 
together to show group behavior.

McFarland (Oxford Univ, England) used his expertise in the ethological field 
to advise roboticists what qualities autonomous robots need, particularly in 
regard to motivation and cognition.

Cruse (Bielefeld Univ, Germany) presented a new neural net controller for a 6
legged walking system which reacts adaptively to the environment.

Ferrell (MIT, USA) is Brooks' chief grad student who is helping him coordinate
research on MIT's "COG" (torso) robot. She introduced her "Hannibal" hexapod 
robot (similar to Genghis) and put it thru its paces.

Steels (Brussels Univ, Belgium) spoke on a mathemetical analysis of behavior
systems where the main idea is that a behaving system is in fact a dynamical
system whose state reaches equilibrium once the behavior it controls is 
attained.

3 Musketeers (i.e. Husbands, Harvey, Cliff, Sussex Univ, England) presented
their latest results in automated mechanical fitness measurement in neural net
based evolutionary robotics. Evolving a single neural net module takes them
about a day or so. Interestingly they had an M.Sc. student who simulated the 
puck robot and evolved the simulation at much higher speeds. The resulting
simulated elite chromosome was then down loaded to the KEPHERA which then
performed in the real world as predicted by the simulation. Interesting.
This "fast simulation vs. slow reality" issue I will pursue later.

Nolfi et al (NRC, ROME, Italy) spoke on plasticity (i.e. learning) in
phenotypic neural nets. The new idea is that the mapping from GA chromosome
to neural net does not occur instantaneously, but takes place over the
lifetime of the individual and is sensitive to the environment.

Taylor (Kings College, England) (ab?)used his brow beating style and actor's 
resonating voice to explain to his audience his theories of the relational
mind, i.e. "consciousness arises due to the active comparison of ongoing brain 
activity stemming from external inputs in the various modalities, with 
somewhat similar past activity stored in semantic and episodic memory". Taylor,
an ex theoretical physicist, with already an encyclopaedic knowledge of his 
new field, was probably the smartest man at the conference. Unfortunately,
he succombs too easily to the temptation to let everyone know it. The 
considerable respect for his abilities that everyone has would only be
enhanced if he he were more low key. His constant, rather condescending 
interjections got on ones nerves after a while.


Comments

The conference felt like a mini SAB conference, especially judging by the 
overlap of the conference committees of SAB94 and this one. As mentioned
above, the KEPHERA puck robot featured strongly, but to my mind, I found
myself becoming irritated (beyond the usual 7 hour jet lag) as speaker after
speaker presented his (nearly always his) results using the puck. My gut 
feeling is that this is not the way to go. My dream is to build artificial
brains. Dave Cliff wants to build artificial brains. Lots of people want to
build artificial brains, but to do so will probably require the evolution of
probably millions of neural modules. Simple math says that with a one day
evolution per neural net module (e.g. using the Sussex "gantry robot"), one 
will die before building even the brain's retina. Somehow, the evolutionary 
process so vital to this field, needs to be speeded up significantly. I 
believe the way to go is to evolve neural circuits in hardware at electronic 
speeds. I hope this was one of the issues discussed at the post-conference 
brain storming session by the 20 odd invited experts. If not, I would be
surprised. Its seems to me to be a critical issue. Keep an eye out for the
Robotics and Autonomous Systems Journal for a writeup of the conclusions coming
out of this brain storming session.


				     ***


Professor Mange's "Embryological Electronics"

The first I heard about Professor Mange's work (pronounced as in the french 
word "mange(r)" = to eat) was when my colleage at ATR, Tom Ray, passed me a 
copy of a paper he had to review for ECAL93 (Euro Conf on ALife). Tom was 
really impressed. So was I. Mange's dream is to achieve von Neumann's universal
calculator, universal constructor, using his specially designed FPGAs (field 
programmable gate arrays). These FPGA "cells" are relatively simple and can
be made in large numbers on wafer scale silicon slices. They have the ability
to reproduce the circuit of any computable function, and of self repair. In its
present design, when a cell becomes dysfunctional, its row and column in the 
grid of cells are switched off. I asked Professor Mange if it would be possible
to switch off only the faulty cell. Yes, he said. This blew my mind, because
the consequences of this alone are profound, let alone those following from the
achievement of Mange's dream. One of the really big problems in VLSI is "yield"
i.e. the percentage of chips that are fault free. This factor limits the size
of chips made from wafers. If a faulty piece of a circuit can be switched off
and its function somehow "moved" elsewhere on the wafer using reproduction,
then circuits of wafer size become possible. The yield becomes irrelevant. Wow!
Mange's et al's papers are now starting to appear regularly in the 
GA/NN/ALife/... conference circuit. In the few pages of these papers, it is 
difficult for a non electronics digital hardware specialist to follow his 
work easily. I made this point to him, saying that John Holland had only 
himself to blame that his invention of genetic algorithms took 20 years to 
become hot, because his book was so unreadable. GAs really only became 
popular after Goldberg wrote a text that everyone could understand. Mange got 
the point and will either write a solid (100+ page) technical report spelling
out all the details so people can understand and copy, or he will write a book
based on a course he will be teaching at the EPFL.

Mange's colleague, Sanchez, wants to use these FPGA chips to perform evolution
in hardware, perhaps by using Koza's Genetic Programming approach, and 
binary decision trees. Since any Boolean circuit can be expressed in terms of
binary decision diagrams and hence binary trees, which in turn can be
implemented on Mange's FPGAs, then since binary trees can be evolved with
Koza's GP, it follows that any Boolean circuit should be evolvable on Mange's 
FPGAs. If so, then this will be the first example (as far as I know) of what I
call "intrinsic evolvable hardware". Maybe I should explain.


Evolvable Hardware (EHW)

Since 1992, I have been pushing the idea of evolvable hardware. Evolution and
(what I call) "evolutionary engineering" (or applied evolution) is one of the
major research themes in computer science in the 90s, but it is virtually all
software based. My dream is to see a machine, (i.e. hardware) which evolves.
The basic idea is to conceive the software bit string that is used to
configure PLDs (programmable logic devices) as a genetic algorithm chromosome,
and thus evolve the configuration (architecture) of the circuit at electronic
speeds. Unfortunately, PLDs are not designed with evolution in mind. If they 
are not RAM based, they are not "infinitely" rewritable, so they need to be
RAM based. Personally I use cellular automata machine hardware to evolve
CA based neural nets, but this is cheating in a way, because strictly speaking,
the hardware of the CA machine remains fixed in its architecture. So far, 
noone has evolved circuits directly in hardware. If I am wrong here, please 
email me. I would love to be shown up on this point. There are however, several
groups of people around the world who claim to be doing evolvable hardware.
I will list them and their email addresses below. These groups are doing what
I call "extrinsic evolvable hardware", i.e. they take a software simulated
description of a hardware circuit, evolve it in software, take the elite
chromosome (circuit description) and down load it into the rewritable hardware.
That is, the hardware gets written to just once. The evolution occurs outside
(extrinsic to) the circuit, by using the software simulated description. 
Intrinsic evolvable hardware would rewrite (reconfigure) a hardware circuit
for each chromosome for each generation of the genetic algorithm. The evolution
occurs inside (intrinsic to) the circuit. The Mange team in Switzerland hopes
to do this, and if so, they will open up a new era in electronics and 
evolutionary engineering. Circuits will be grown/evolved rather than be
designed. Hence they can become more complex and hopefully more performant.

The following groups of people around the world that I know of are doing
(extrinsic) evolvable hardware. 

Mange and Sanchez (EPFL, Lausanne, Switzerland) - described above. 
mange at di.epfl.ch     sanchez at di.epfl.ch

Hemmi (ATR, Kyoto, Japan) - Hemmi is a colleague in the same group as Tom Ray
and myself. He uses an HDL (hardware description language) in the form of 
trees, to which he applies Koza's GP approach. He has evolved digital counter 
circuits and the like.
hemmi at hip.atr.co.jp

Higuchi (ETL, Tsukuba, Japan) - Higuchi is an ex colleague of mine from my
postdoc days at ETL (Electro Technical Lab). He used a simulated GAL PLD chip
description which he evolved in software to perform various digital funtions.
Lately he has been working on a hardware design to perform genetic algorithms
- a GA machine.
higuchi at etl.go.jp

Cliff (Sussex Univ, England) - Dave Cliff has a new grad student who is a
VLSI designer who wants to evolve hardware. I met Dave at this conference and
asked him whether his student would be doing intrinsic or extrinsic EHW. Dave
said that the student is trying to make a deal with a Silicon Valley company
in California. If the deal comes thru, then he wants to do intrinsic EHW,
otherwise he will do extrinsic EHW like everybody else.
davec at cogs.susx.ac.uk


If there are other people doing EHW, please let me know. I believe the era of
EHW and "Darwin Machines" is upon us, and should be vigorously supported. 
We will never have truly performant artificial nervous systems and artificial 
brains until we can overcome the "slowness of evolution" problem. 
Evolution at electronic speeds i.e. EHW, is the key breakthrough here.


Cheers,
          Hugo de Garis.


Dr. Hugo de Garis,
Brain Builder Group, Evolutionary Systems Department.
ATR Human Information Processing Research Laboratories,
2-2 Hikaridai, Seika-cho, Soraku-gun,
Kansai Science City, Kyoto-fu, 619-02, Japan.
tel. (+ 81) (0)7749 5 1079, fax. (+ 81) (0)7749 5 1008,
email.  degaris at hip.atr.co.jp