The Mind II AGI architecture

                               Alan Grimes


Abstract

        The following is an outline of a practical approach towards developing
and refining a generally intelligent technology called Mind II. My purpose
here is to describe the Mind II approach to AI, to present a rationalle for its
fesability, and to outline the steps needed to realize it.
        The Mind II approach to AI is a cybernetic system derived from the
Celular Automaton abstract machine.



CONTENTS
        Introduction
        Motivation
        The Mind II approach.
        - Mind II is a cybernetic approach.
        - Design methodology
        Rationale
        The development process.
        - Phase 1: basic control model:          The flatworm.
        - Phase 2: Elementary behaviors:         The fish
        - Phase 3: Fundamentals of environmental awareness: The lizard
        - Phase 4: Basic cognition:              The rat.
        - Phase 5: Epesodic memory and planning: The monkey.
        - Phase 6: Language and abstract cognition: Man
        An artificial evolutionary context for the refinement of AI design.
        - A VR environment as a learning tool.
        - A VR interface for user-AI interraction.
        - A VR environment as an evolutionary context.
        - The mechanics of evolving a Mind II AI.
        - Exploiting evolutionary theory.
        A Comparison to Other AI Efforts
        Commercialization.
        - Accelerating the evolutionary process by making the software
available to home users.
        - AI as a stand-alone product.
        - Hardware accelerators for Mind II AI systems as commercial products.
        - Robotic systems.


Introduction

        The Mind II architecture is my proposal for developing an artificial
intelligent consciousness. It's name is derived from a classification scheme
which I use to describe intelligent objects. The first class is the unmodified
or weakly modified human brain, or Brain I. A biological variant of that design
with significant anatomical and physiological changes is termed by this
classificiation scheme as Brain II.
        The Mind type of AI is based on the neural organization of the brain
but in the form of computer software. A mind I AI is based on a detailed
simulation of organic neurons. It is highly likley that it is a feasable
design but it is not likley to be efficient enough to be workable in a
commercially viable product. It is also very likley that a design which uses a
more abstract approach to creating a functional equivalent to the brain will
yield much greater performance on the same hardware. Mind II aims to be such a
design.


Motivation

        The principal reason for working towards Mind II, or any general AI is
the immense benefit to humanity that can be expected from a successful design.
These benefits range from a better understanding of human cognition through a
working experamental model to having a practical, useful AI technology which
can be applied to any of a wide variety of problems. A listing of all the
potential categories of applications would not be relevant to the present
discussion.
        The second major motive which plays a large part in the Mind II
approach steams from the risks to human interests associated with strong AI.
These risks range from the minor risk from the possibility that an intelligent
robot could go on a crime spree in down town Detroit to apocalyptic scenereos
on a cosmic scale. The design of Mind II very deliberately stays close to the
pattern of development which lead to human consciousness so that this risk may
be mitigated. Firstly, the architecture will begin with a great deal of
commonality with it's human model so that it is unlikley that it will
demonstrate any faculties that humans might have difficulty dealing with.
Secondly, by adhering to a biologically inspired model, it is hoped that a
large portion of the many techniques used in human psychology, including the
innate intuitive understanding of mind, may be successfuly applied to a Mind
II based AI.
        The third area of risk in any AI venture is the risk that the project
will fail. History gives us many examples of failed attempts to build a
thinking machine. The Mind II architecture addresses this problem by staying
close to the biological archetype. It is beleived that by not getting too
"creative" by adding any "computer like" features such as direct number
crunching and ascii-based text processing the core problems can be addressed
and overcome. The downside to this strategy is that the resultant AI may lack
many features and capabilities that would otherwise be possible. This approach
is necessary to increase the chance the project will succede in creating
general intelligence at all. Enhancements to the basic design can be explored
in subsequent projects.


The Mind II approach.

- Mind II is a cybernetic approach.

        In studying intelligence and it's role in the human it becomes clear
that it is hardly meaningful at all without an environment to interact with.
Mind II, as a brain inspired approach, is grounded in a cybernetic theory.
Beyond that it is clear that many types of learning require an active
exploration of the environment such as procedural learning through trial and
error. It can be concluded that without active learning the knowledge available
to an AI is severely limited.
        Many previous AI projects reflect an assumption that this interaction
is not necessary. Needless to say, none of those have succeded in producing a
full AI.


- Design methodology

        The essence of the Mind II approach is that it mimics the brain but
completely disregards the fact that it is constructed with neurons. Wherever
the computation preformed by the brain is known, the best available algorithm
for performing the same job will be used. Where the algorithm is unknown a
stand-in will be used and configured to match our best guess of it's
information flows and to mimic all known connections with other modules.
        In the case of the cortex, an advanced hexaginal-latice celular
automaton will be used to mimic the function of cortical columns. Current
Celular automata models are theoretical in nature. They are specified by a
state, a set of rules, and their interractions with their immediate neighbors.
As such, they are not especially practical because it takes quite a while for a
signal to propagate across a large latice.
        The brain solves this problem by using a complex topology where certain
cells are able to send signals directly to cells in remote locations, greatly
increasing the system's reaction speed. In most areas of the cerebral cortex
six anatomical layers have been identified, these layers appear to have fairly
distinct functions. The first attempt at the Mind II design will implement a
"best guess" for all of them.
        One of the most interesting features of the brain is it's ability to
re-task cortical columns on the basis of computational demands. Mimicing this
function algorithmicly poses an interesting problem. One possible approach is
to attempt to measure the workload of a each cell and make a rule for
re-directing connections from highly utilized cells to neighboring cells that
are not as heavily stimulated.


Rationale

        It is reasnoable to expect that the Mind II architecture will succede
in producing a general intelligence because it is based on a celular automata.
Even though the educated guess that the cerebral cortex is basically a type of
celular automata may be completely false the approach still has a high
probability of success through the Church-Turing thesis. Because many types
of celular automata are known to be a universal computers it is therefore
highly plausable that a celular automata that can emulate the cerebral cortex.
        The only assumption being made here is that the Church-Turing thesis
holds in the case of the cerebral cortex. The C-T thesis has stood for around
65 years so that is highly unlikley at this juncture.

        The design of the Mind II architecture is also strongly motovated by
the possibility that there are certain critical aspects of the brain which we
do not understand at all. Prior attempts at AI have depended on the programmer
accurately determining what types of computations the brain does and then
emulating them in the machine. Mind II, on the other hand, lies right on the
border between emulation and simulation. On the level of details, it is an
emulation, and on the systems level it is a simulation. It is hoped that by
adhering to the form of the human mind that all qualities necessary for
intelligence will be included, or notably absent, even if the programmer is
not fully aware of what they are.
        I say it is good that some features of Mind II might be notably absent
precicely because the exact nature of their absense will, itself, serve as a
means for their discovery and inclusion. Because the workings of Mind II will
be so similar to the brain, the description of missing functionalities can be
made with sufficient detail as to be useful in their creation and inclusion in
the next prototype.


The development process.

        The most reliable strategy for developing a Mind II architecture is to
re-trace the course of the evolution of man. This approach demonstrates a
working implementation of each of the respective layers of functionality in the
brain. It may be necessary, at times, to backtrack to earlier phases as
deficiencies of the implementation strategy are uncovered.


- Phase 1: basic control model:          The flatworm.

        The earliest phase of our evolutionary development is the flatworm.
It is a primative creature with a very simple nervous system which demonstrates
little more than reflexive behavior. The goal of this level of development is
to demonstrate a capacity to respond to an environment in a simple manner.
        A secondary task is to refine the programming techniques used to
implement this functionality. This choice has a strong effect on the patterns
used in the higher levels.
        The actual animal model may be any type of worm progressing towards
more advanced kinds.


- Phase 2: Elementary behaviors:         The fish

        The next level is to implement sequential behaviors that allow the
animal to respond to the environment in a patterned manner. It does this by
both state-based controll and by either inhibiting or exciting the basic
reflex behaviors.
        I would suggest that a quadripedal model be used for this phase.


- Phase 3: Fundamentals of environmental awareness: The lizard

        Using the same animal model, the next evolutionary step is to include
a basic memory. It's basic function is to generate maps of relationships
between specific stimuli. This allows for the organism to organize behaviors
with regards to the geography of a specific environment. The same mechanism
is also forms the basis of epesodic memory.
        The system should be tested untill it can be shown to exhibt similar
behaviors to a lizard.


- Phase 4: Basic cognition:              The rat.

        The rat is probably the most well-studied creature in the universe.
One would be tempted to begin the project with the rat instead of "wasting
time and money" in working up through the evolutionary heirarchy from the
beginning. The purpose, however, is to establish a design pattern and good
coding practices that will serve to organize and clarify the organization of
the design and keep it in tune with it's biological counterpart.
        The testing at this point will be more extensive because so much is
known about what the capabilities of a rat brain are. Specificly these are
the ability to be able to acquire a world model which can be used not just to
map out a known territory but to make predictions about a new teritory based
on remembered knowledge. Instead of simply selecting among innate behaviors,
the same type of automata should be able to learn completely new behaviors.
        If a working approach is developed it is very likley that it will be
very close to the algorithm required for high level cognition. Infact, the
bulk of changes from this point will be to adding complexity to the system by
increasing the cortical area and increasing the complexity of it's topology.

- Phase 5: Epesodic memory and planning: The monkey.

        The next major evolutionary improvment is the addition of the epesodic
memory. This allows the system to remember not just places but events in it's
past. Successful completion of this phase will yield a mind on a par with the
chipanze.
        The body of the animal should be changed to a roughly humanoid bipedal
design with some approximation of a hand maniplator although the precice
physics are probably not necessary.

- Phase 6: Language and abstract cognition: Man

        Far far from finally except for this brief document, there is human
equivalency. The topolgy of the mind is expanded once again to include regions
for associating symbols with internal representations. This, combined with all
of the previous steps, should be enough for the mind to use language and think
like a human being.


An artificial evolutionary context for the refinement of AI design.

        Because intelligence can't form in a vacuum, a virtual world must
be provided to provide the AI the stimulation and feedback it requires to
acquire knowledge. Such a virtual world can serve three major purposes in
the AI development effort. Those are that of a learning tool, that of a medium
for user-AI interraction and as an evolutionary context for refining the
Mind II design. To serve these purposes it is necessary that the virtual world
have certain qualities.


- A VR environment as a learning tool.

        A successful environment will provide many options for complex
interractions. It should provide, to the AI, a rich array of auditory, visual,
gustatory, and tactile feedback. It should have a robust and consistient
physics. All objects should have predictable behaviors and these behaviors
should have no arbitrary limitations.
        The environment should provide "raw materials" with which the AI can
construct artifacts which may have useful properties depending on how well they
were designed. The classic toys and games, such as chess, should be available
as artifacts within the VR world.
        The actual "resolution" of the environment is not especially important
but rather that it be possible to deduce the function of an object from it's
graphical representation.


- A VR interface for user-AI interraction.

        Human researchers and other users should have two forms of interraction
with the world, both are of great importance. The first is the G0D-mode
interface where the user sees a top-down (or other appropriate view) of the
world and is able to adjust various objects in the system in arbitrary ways.
This interface should also privide a way to inspect, edit, remove, or insert
the artificial lifeforms.
        The other means for interraction is through a human-controlled VR
avatar. The "front end" of this system should be symetrical with what is
provided to the AI. It will connect to the VR through an avatar that is
identical to the ones used for the AI. This is so that the human can directly
debug the avatar software and so that the VR system's function as a medium for
communication is not compromized because the AI only sees dots in space where
the human is shown a top of the line 3D animation.
        The avatar interface should provide the user with the ability to
interact socially with the AI entities.
        An AI development workstation may provide both of these interfaces
simultaneously on different monitors.


- A VR environment as an evolutionary context.

        It can be expected that the team implementing this plan may run into
difficulties with the performance of their CAM rules or the overall topology
of the mind. This problem can be addressed with an evolutionary approach.
        At first the environment will have to be very tolerant of the minimial
capabilities of the AI model, a paradise if you will. The paradise zone will
have a relatively high rate of resource replenishment, no requirements for
shelter, and very few dangers.
        It's only evolutionary value will be to train the agents in the basics
of food gathering and to protect their eggs (instead of eat them or allow them
to be eaten).
        As the population increases the resources will run out and expansion
will be necessary. The adjacent habitats will be somewhat hostile, shelter
will be required for sleeping and protecting eggs. Food will be relatively
scarce but agriculture will be supported. There may be engineering problems,
such as building a boat or a bridge, required to move further along. There may
also be periodic disasters requiring planning in order to achieve optimal
utilization of the area.
        The next region will be extremely harsh, akin to a desert or an arctic
tundra. Only by scouting and then making preparations will the creatures be
able to make any use of this area at all.
        Beyond this area will be a challenge zone. A task such as building a
complex machine from materials available in raw form in locations scattered
across the entire world. Success will give access to another paradise zone,
completing more difficult variants of the task will provide rewards to the
creature such as increased strength, extended reproductive life, decreased
metabolic needs, a number of instant clones, or any other advantageous gift.
        In this last zone, at the end of the world, there should be a "trial
zone". This will begin with a tower on a tall mountain, a mountain so tall that
only the most determined, and best prepared, creature will be able to reach it.
Inside there will be a series of IQ tests for spatial, mathematic, and
reasoning skills. Once a test is passed the creature can move on to the next.
The biological modeling can be suspended in this zone so that the creature can
take as much time as it needs. If it gives up, it can use a transporter to
go back to the foot of the mountain where it will have to re-supply before it
can make another attempt.
        The last of these trials will be the turing test administered through
an avatar controlled by one of the experamentors. If the creature passes this
it becomes immortal and invulnerable. It also gains access to a room within
which it can control the rest of the world and use the full power of the
software for self-modification -- it has won the game.


- The mechanics of evolving a Mind II AI.

        The avatar models will be similar in nature to the ones used in the
software toy "Creatures". The avatars will have to have biologically plausable
requirements for food and a limited tolerance to environmental extremes and
traumatic injury. All avatars will have identical physical capabilities and
needs. These will include a base metabolic rate and an activity dependant rate.
Each will have a "genome" which will consist of the program used to specify
it's brain and how fast it runs relative to the rest of the simulation. This
rate combined with the total size of it's cortex will determine it's brain's
metabolic requirements.
        Also encoded in these genes will be information about it's coloring
and pattern so that each creature will have a recognisably unique appearance.
The lifecycle of the creature will have to be fairly long in order to make sure
that it encounters enough interesting challenges during it's life in order to
have an evolutionary impact. It would probably be useful too for it to have an
egg phase so that it will have to evolve to protect it's eggs. The eggs will
also be colored and patterned based on the genes so that it is possible for a
creature to determine which eggs are it's own.
        Unlike the game Creatures, there will not be any classes of creatures.
Any avatar can be used for any type of lifestyle from agrarian to canabalism.
Various things in the game can be used as weapons. This might be unplesant to
the creatures but it is necessary to promote their ability to think
strategicly, work towards objectives, and deal with other minds.


- Exploiting evolutionary theory.

        To accelerate the evolutionary process several steps can be taken.
Firstly, the relatively short lifespan of the creature increases the number of
generations that can take place in a short period of time. The limiting factor
is that that the creatures need a significant ammount of interraction with the
world in order to demonstrate a real intelectual capacity. Another approach is
to use a faster computer as the simulator in order to compress the clock time
each generation requires with minimal impact on the quality of it's experience.
        Population dynamics can also play a role. By periodically closing
travel between areas populations can be limited to a size where they evolve
rapidly away from the baseline. A pattern can then be established where the
simulation runs for ten or fifteen generations as isolated units at which point
the gates are opened for a period of two to five generations followed by
another period of isolation. During the period of free travel, the populations
can mix and traits which evolved during the isolation can spread to the broader
population.
        Finally, programmers can examine all the creatures for desirable
traits and then attempt to manually produce a winning design.


Complexity analysis

        It is assumed that the dominant source of computational complexity
in this model will indeed be the emulation of the cortex. It is further assumed
that signals within this cortex will not have to propagate across a distance
greater than a single bordman's area. Furthermore it is assumed that a
propagation of this distance will be required at a rate proportional to the
apparent step-rate of the human mind which appears to be around 18 cycles per
second under normal levels of arousal. Humans appear to have a throttling
mechanism that causes their brains to accererate to handle stressful situations
and slow down when in a relaxed state, presumably to conserve energy. It is not
beleived that this function is critical to producing a demonstration of
functional intelligence though such functionality would likley to be of great
use in mobile platforms for the same reason similar features are common in
today's laptop computers.
        The computation to complete one update cycle of a hexagonal celular
automaton is proportional to the number of cells times the distance between the
edges of the latice which is proportional to the cube root of the number of
cells.

               O(N * (3û)N)

        While the actual compute time required for each cell is unknown it
is known that it will complete in essentially constant time because it will be
a non-looping, non-recursive function of the previous state combined with the
eight input channels representing the neighboring cells, a link to a
topologicly remote cell in the cortex and a system-wide input.
        The input and output channels will be an ordered set of bits of a size
yet to be determined.

        Under these assumptions it is possible to conclude that if the
efficiency of the Mind II design can be made proportional to that of the brain,
which has approximately four million CAM-like cells divided up into dozens of
regions, then a machine capable of sustaining 350 billion operations per second
should be sufficient to implement Mind II.


A Computer System for Mind II development

        I beleive that it is possible to implement Mind II and the evolutionary
system proposed above with today's hardware or, at the very most, hardware
that can be expected to be on the market within the next few years. A
distributed platform for this effort will probably follow the following
outline.
        Each individual instance of Mind II will function on a discrete system
based on a comodity platform. It will accept state information from a world
severer, generate a 3D image of the world, process it, and then send commands
to the avatar which resides on the server.
        In addition to the host processor a comodity 3D video processor will
be used for generating the 3D scene and the bulk of the Mind II processing work
will be accomplished by an array processor board inserted into one of the
expansion slots. The total cost for each unit, at today's pricing, will be
on the order of $10,000. A suitable evolutionary population of 200 can
therefore be sustained on hardware costing around $2,000,000.
        Each Mind II system will be connected, via a simple network connection,
to a host machine that will simulate it's world. Each area, as described above,
may reside on a seperate machine. A suitable platform may be obtained for
$15,000. I expect ten of these world-servers will be desired.
        The total cost of the hardware for this venture, including a few
workstations and network infrastructure will be in the neighborhood of $2.25M


A Comparison to Other AI Efforts

        A review of the literature will clearly show that very few of the ideas
presented here are new. There are indeed a number of research efforts by
people who actually have accademic credentials that, superficially, would
appear to incorporate the key ideas of this proposal. I do not beleive that
this is the case. I will here present a review of the most visable of these
efforts and what I believe to be their flaws.

        One of the most prominent proponents of a Celular Automata based
approach to AI is Dr. Hugo DeGaris. His approach, in my classification scheme,
is a Mind-I design. He uses CAM as a medium for simulating organic neurons.
He hopes that the rules he selects will succede in replicating the
functionality of the cortex. While his approach might break new ground for
neural-net approaches it is unlikley to succede in producing a pracitcal mind.
        The first major drawback to his approach is that it doesn't provide a
general purpose memory as the Mind II does. His current approach of evolving
neural modules and then wiring them togeather is unlikley to produce a capacity
for introspection because the basic cognitive units aren't discrete. While
his design will likley be able to solve a number of problems that involve a
direct chain of cognition, it is unlikley to be able to exhibit directed
cognition or the ability to construct and use representation of the
environment.
        Assuming he did produce a workable model of the Cortex and all of the
vital sub-cortical regions the system will still be too expensive to be
marketable. Each cell in the Mind II design attempts to emulate a cortical
column. The simulation he proposes requires hundreds if not thousands of CAM
cells to simulate the structure of each neuron in that column. Even though his
cells are simpler in nature it isn't enough to overcome the orders of magnitude
increase in their number. If Mind II is barely feasable on today's computers
then, assuming an 18 month Moore's law I would expect that a DeGaris-style
CAM Brain would be feasable 15 years from now.

        Steve Grand, the creator of Creatures also used a Cellular Automata
approach. Like DeGaris' work it also was patterned on a neural network model
and does not appear to exploit its full computational potential. He coined a
phrase, CyberLife to refer to the ideas behind his implimentation of
artificial life. A version of almost all of the ideas in Mind II can be found
in Creatures. Infact, it was my hope that his work would be sufficient to
produce AI and that I would not have to go through the effort myself.
Unfortunately his design had to cut many corners in order to work on 1995
vintage comodity hardware. The most drastic of which is the way the creatures
experience their environment. Their perceptions are limited to a few points in
2-dimensional space and little else.
        While his design did a very good job in implementing motovators it did
not implement a heirarchical control system. This is the primary reason, I
beleive, that his Norns were very poor at goal-directed behavior. Any semblance
of goal directed behavior was due to the limited nature of the cybernetic loop
the system employed where the creature could only receive information from one
stimulus at a time allowing it to react in a somewhat organized fassion
relative to that stimulus. A true AI system will need to be able to accept a
much larger ammount of input and then dynamicly select behaviorally relevant
content as it is detected.
        His biological models were also loose. For example, it was possible to
starve a norn well past the duration of any biological endeurance and have it
suffer no real ill efects. It is apparent that this was for marketing reasons
as, unfortunately, the system was not geared for research use. In all, his
system was far too limited for there to be much hope at all of intelligence
emerging. I would be very interested, however, to see what he would do with a
practically unlimited computational budget.

        Another notable AI researcher is Rodney Brooks of the MIT. His efforts
of the late '80s and early '90s were most remarkable in the field of AI for no
other reason than they actually made sense. Instead of messing around with
trees of abstract syntax, he developed robots which interacted with their
environment. His early robots were equivalent of Phase 1 and, perhaps, Phase 2
of the project outlined above. Had he continued his work he would almost
certainly have something really interesting by now. I really don't know what he
has been doing the last ten years but I can hardly find anything new projects
on his website at all.
        I'll give the man credit for being one of the very few AI researchers
out there persuing the most obvious course of research but his follow-through
is most sorely lacking. If I were him, I would have made a larger mobile robot
with a richer suite of sensors, especially cameras, and at least one reasonably
dexterous manipulator. Then I would have started hacking togeather a system
that can perceive it's environment.
        Instead, his work seems to be focused on producing human-like afect
display. This is only a mite and a half better than plain anamatronics. My
supposition is that he has sold out and is working with a very short-sighted
commercial product in mind, basically the face for the androids coming out of
Japan and little more. Either that or he has just gotten too caught up in all
of the organizations he is involved with.
        I would say more about the androids from Japan but very few details
are publicly available (and in English).

        There are other efforts to produce AI but few of them are worth
mentioning. The examples above should be sufficient to give the reader a
general impression of where my ideas are situated in the broader scheme of
things.


Commercialization.

        The Mind II project, both during and after it's development, presents
a number of commercial opportunities. While the AI architecture produced by
this project will have some limitations that will hinder it from performing
well in the most demanding environments, it will have a broad number of useful
applications.
        Some people claim that once general AI is achieved commercial concerns
will quickly be overwealmed by other impacts of the AI technology on society.
It is my beleif that any intention to unleash an AI with the ability and motive
to effect some change on society or the environment is criminal. The commercial
model of deployment is very well tuned to allow people to choose for themselves
in a supprisingly democratic fassion how they wish new technologies to affect
their lives.


- Accelerating the evolutionary process by making the software
      available to home users.

        During the development effort the Virtual world simulation software
may have some commercial value in the same way the Creatures software attempted
to gain a market. This possibility is mitigated, however, by the fact that the
level of hardware required to run a population of Mind II instances is likley
to be beyond what will be common for at least another few years.
        Another opportunity is in the simulation software itself. As this
application is intensely demanding on the detail and features of the simulation
software, the simulation itself may have some commercial value.


- AI as a stand-alone product.

        The commercial applications of AI are practically limitless because a
digital generally intelligent agent can take the place of virtually any
knowledge worker. With a Mind II AI the task of a producer is to construct and
educate instances of Mind II for the varrious tasks. Additionally, variations
of the basic topology will need to be created to address unusual or specialized
usage scenereos.
        The actual business model for this type of venture is a matter for
debate with the community at large. It also is dependant on varrious technical
issues with regards to the power and flexability of the AI technology itself
which are not entirely forseeable.


- Hardware accelerators for Mind II AI systems as commercial products.

        The Mind II approach, as it is outlined here, will not run very
efficiently on today's processors. Architectures designed specificly for the
CAM paradigm of computing will be better suited to the task. As the Mind II
AI system grows in popularity a secondary market for Mind II processors will
emerge. These may be packaged either as accelerator boards for existing
computer systems or as stand-alone "cybernetic modules".


- Robotic Systems.

        Perhaps the most natural application for an AI of this architecture
is in the form of androids. Because the Mind II follows the human brain
pattern so closely the only problem becomes adapting the control structure of
the android to the AI architecture.


Refferances


Dr.Rodney Brooks.
http://www.ai.mit.edu/people/brooks/index.shtml

Dr.Hugo DeGaris.
http://www.cs.usu.edu/~degaris/

Creatures Labs (what's left of it.)
http://www.gamewaredevelopment.co.uk/creatures_index.php

Steve Grand
http://www.cyberlife-research.com/