Last Modified: 3 January 2011; Last overhaul: 23 December
Research Related to Modular Artifical Intelligence, Action
Cognitive Systems, Cultural Evolution and Modelling Natural
This is work related to my own research.
Some of these projects are
discussed in my publications, while
others I've just stumbled across and haven't really had a chance to
look at. I try to keep this list fairly interesting and up-to-date.
me if you know of
anything else I should add on. Note this is very incomplete, but at least it's
a starting point.
Systems AI, `New' AI, Behavior Based (BB) AI and Cognitive
Systems More Generally
I do Systems AI research. That is, I am less interested in individual
algorithms e.g. for learning or planning and more interested in how to put
together a complete system. No one algorithm can overcome combinatorial
complexity. That's why brains don't have homogenous structures. In my 2001 PhD
dissertation I argued that New and BBAI
advanced the field because they emphasised maintainable design.
- Cognitive Systems research has the potential to pick up where New AI left
off and move AI forward, though sadly many CogSys researchers have phobias to thinking
about programming and systems engineering.
- Behaviour-based AI definitions:
- Other lists on new AI:
Arkin wrote the book about behavior-based AI, but his idea of
behaviors then were more like vector maps (reminding me more of Saphira.)
- I think behavior-based AI should be closely related to Object
Oriented Design in computer science, and to Modularity in psychology /
- Some "architectures" (also known as development methodologies)
Logic-Based Subsumption Architecture project at Stanford is working
on -- logic-based behavior-based AI.
- Mike Benjamin developed
Interval Programming for combining the desires of multiple
behaviors (as long as their plans / reward functions can be expressed
in the same representation.) PhD 2002
an excellent PhD in 1993 on comparing action selection mechanisms,
behavior-based ones. You can download his simulator from
here. The world-wide
project reimplemented this.
- The VUB AI Lab and The Sony Computer Science Laboratory,
Paris of Luc Steels is using their knowledge of
behavior-based AI. They are studying the evolution of language as well
as doing more work on robot software architectures.
Prescott came from this approach to solve some real neuroscience
- Mark Humphrys
likes to learn controllers for behavior based systems.
Animat Lab of Jean-Arcady Meyer emphasizes robots and simulations
for biological research.
- Owen Holland
always has wild cognitive robotics ideas.
- Very Related Research: A number of people have downloaded
or written about my architecture,
far as I know only
Torbjørn Semb Dahl has used a version of it in an ALife competition.
- I probably shouldn't say this on the web, but I think Multi-Agent
Systems (MAS) is largely a degenerate form of BBAI --- degenerated into
a focus on logic rather than applications or biology. But here
are a few links anyway:
Simulations, Games and Agent-Based
- about Agent Based Modelling (ABM):
Agents? On the Varied Motivations for Agent Computing in the
Social Sciences Robert Axtell,
CSED Working Paper No. 17,
- An extensive page on ABM by Craig Reynolds from 1999.
- A newer extensive page on
ABM by Jeff Schank (from 2011), emphasizing
interdiscipinliarity (as I type this it is so new it still
has bugs in it, but it looks like a fairly major effort that
will turn into a good resource).
- ABM simulators we (AmonI) use:
by Uri Wilensky (currently of Northwestern) seems to be my PhD and MSc
favorite, I guess because it's really easy to use. March 15 2004
the annoyingly named (esp. from a google perspective) but otherwise
impressive simulation environment from George Mason. We've used
little directly, but more in combination with our own POSH action
selection, see BOD/MASON.
Mason now also offers a PhD in the
Computational Sociaol Sciences.
- RePast does
multi-agent modelling like Swarm
but "move[s] beyond the representation of agents as discrete,
self-contained entities in favor of a view of social actors as
permeable, interleaved and mutually defining, with cascading and
recombinant motives." It's also apparently easier to learn and to
teach, according to Lars-Erik
Cederman. We've had a few students use it.
- breve is
to be good, but I don't know much about it. A couple students
have used it.
- An even more difficult-to-google entry is the Multi-Agent
Suite (MASS). We haven't used it yet, but I've seen it
demoed and the guys who wrote it are very smart. They are taking on
RePast & NetLogo rather late in the game when a lot of people have
brand loyalty, and they are again introducing their own language, but
it does look like a very good one.
- ABM simulators we haven't used but have heard good things about:
Ties does massive social simulation with individual, social and
evolutionary learning --- which of course is a big mess, but they're
starting to get a handle on it all with modularity now. Action
selection is mostly through decision trees, parametisation mostly
through XML (eclipse plug in can help.)
(MASS) from AITIA / Gulyas Laszlo.
- Another Java ABS tool, used by real scientists: SeSAm
(Shell for Simulated Agent Systems)
was developed at Brookings but may be being moved to a
company called BiosGroup, at
least according to this bio of its creater, Miles T. Parker.
needs a new agent researcher, I'm liberal!
- Entorama - yet another
tool, though it is 3D. Looks like Microsoft-only, although on the
bright side it supports BOD-like behaviour modules.
- RVO Library:
for Real-Time Multi-Agent Simulation, another robotics-oriented
with navigation / road map built in.
- Games are some of the best simulations (since they aren't biased
for research and look great!)
- Biota - the digital biology
project. 13 Jan 2003
- Speaking of weird applications, here's a list of agent papers
(allegedly) relevant to terrorism.
(It looks like most of them are really about war, politics or
(ACE) at Iowa - includes a good review of current
- Framework for
Evaluation and Assesment of Regional Land Use Scenarios uses
agent-based social simulations for ecological policy.
- Jeffrey Ventrella has
software for evolution stuff. "A website of Artificial Life, Art, and
Georgia Tech Mobile Robot Laboratory home page has great robot
simulators on it under "software".
- Robocup is one of
the best things to happen to robotics or AI -- a great platform for
Works from Newtonium
is a pretty
good-looking piece of software. I haven't tried it yet, but it looks
interesting, and works with matlab and such.
- Online individual agents and platforms:
- Cybelle from
AgentLand (beware of pop-up ads!)
- Haptek's products make
3-D VR agents easy (but only for MS windows...).
- Elzware makes
online bots for other people, but that's one of their own. We've
used their technology in some of our talking avatars.
is a Multi-Agent System tool for simulations using BDI (belief,
desire and intention) agents. More about BDI is in the next section
Other AI Architectures
Note: `Cognitive archictures' are now under Modelling
Nature, not here.
- Ian Dickinson
of HP is working
on a cool new architecture which will hopefully be like PRS & JAM
next bullet) but actually work. Also, his model is that an agent
should be a library that you can link your behavior to, which I think
a very cool idea. But I can't find a link to his arch from his
page. Thu Jun 26 11:44:59 BST
- PRS is a well-know
from SRI. It has a GUI for creating
the action-selection scripts, which can also interface with or be
generated by a conventional planner. In theory, you should be able to
implement most of Edmund (my
architecture) in PRS, but there's no active support Edmund's behavior
approach to modularity and learning, and it's not trivial to
manipulate between-script priorities. I've had a bit more luck with
JAM from IRS, it
advantage of actively being used and supported. I'm not convinced any
of these BDI architectures are really designed for continuous,
ongoing behavior typical of embodied agents.
Laird's group (some of the big guys behind Soar) put together a Survey of
Cognitive and Agent Architectures in about 1992 (though it has
been updated in some places). This is an html document, and is useful
not only in describing the architectures (especially Soar, Brook's
Subsumption, Mitchell's Theo, Gat's ATLANTIS, and Hayes-Roth's
Adaptive Intelligent Systems) but in bringing out metrics of
- Erann Gat,
(before he changed his name) said of the AI for Deep Space 1's Remote Agent
It's one small step in the history of space flight.
But it may turn out to be one giant leap for computer-kind, with
artificial intelligence software being given primary command of a
spacecraft for the first time.
has implemented Norman and Shallice's cognitive architecture for
action sequencing, which Dave Glaspool
first told me was like mine. He's also written a great cognitive
modelling tool COGENT.
Kortenkamp seems to have everything going for him, including the
hybrid/layered robot architecture 3T and mobile robots that see
- 3T used to include James Firby's Reactive Action Packages, which
influenced my work, but James is long gone from academia.
- Similarly, Marcel
Schopper's Universal Plans (UP) turned out not to
be. I'm not surprised you can't generate AI without a lot of
normally much of it by design.
- Henry Hexmoor's
a 3 layer architecture,
his thesis talks about learning between layers/representations, in
learning between skilled and declarative forms of knowledge.
- Ian Horswill
architecture these days, and anyway just does great stuff.
- INRIA project BIP
a humanoid biped project, I think they are trying to solve AI. Their ORCCAD
for designing the behavior is interesting.
Humans must all have some kind of architecture (at least the
- For the Linux hacker, the AI-Alife-HOWTO.
- For any programmer, there's the Portland
Pattern Repository which seems stunningly useful. Optimize Later!
- Building probabalistic models from natural subjects may be a key
of creating AI. Here's a weird short list of some people doing
- University of Hertfordshire's Adaptive
Systems Research Group, with Kerstin Dautenhahn
Cañamero, among others.
Project of ERATO, JST ended in 2001 (though their
collaborators from USC, The Computational
Control Laboratory are still going.) It was great stuff, that
totally scooped Cog: their robot has legs, and learned
folk-dancing by imitation.Cambridge's Behaviour and Evolution
Natural Intelligence: Cultural
Evolution & Memetics
I'm interested in where intelligent behaviour comes from.
How much work is done by the individual, how much by genetic evolution,
and how much by cultural evolution. This area is exploding right
now and so its impossible to keep these all up-to-date, but here's a
few links anyway...
- Memetics is the study of culturally transmitted knowledge.
As with genetics, it makes the so-far unsustained assumption that there
is an identifiable unit of transmission, the meme. I'll start
worrying about what a meme might be after we figure out what genes
really are. For now, think of it as aether -- a supposed medium
- Cultural evolution is a topic of study regardless of memetics.
- If social learning and memetics are important in non-human
primates, then they must have some culture. Here are the primate
culture websites I know about.
- Other cool primate work (e.g. primate social stuff)
Flack works evolution and primate social structure -- have a look
at her publications.
Silk is deeply cool too even though she doesn't have a culture page.
- Imitation could explain a lot of culturally-transmitted
learning. Roboticists noticed this a fair few years ago.
- Andy Whiten
used to worry about defining imitation, but now works on demonstrating
cultural transmission in non-human animals.
Page. This is about how learning affects
evolution. I'm also interested in the interaction of different learning
rates and types within a single agent, e.g.
Beecher's work with sparrows shows stable (rather than
evolutionary) cultural (song) change.
- Elhanan Borenstein
researches how learning, evolution & memetics (cultural evolution)
the way genes are expressed can be determined by social outcomes.
think this is an incredible new contribution to understanding action
in nature. Here's some labs working on this:
- The Robinson Lab
at UIUC says "from society to genes" but surely the other direction is
interesting! Anyway, they are working with honey bees.
- I first heard about this in a talk by Hans Hoffman (here's the Hoffman Lab) who is
working with cichlids (a
surprisingly smart sort of fish.)
- Biology groups working on the evolution of behaviour:
- Human evolution / biological anthropology is leaking into biology
& primatology these days
- Worrying about things like the biological evolution of cultural
evolution, phenotypic plasticity and the Baldwin Effect is apparently
now called EvoDevo.
Other Natural Intelligence
I'm also interested in other more conventional topics of natural
- A web page on transitive
inference / performance and the hippocampus. This is
something I model, because a) its an interesting example of complex
task learning and b) the hippocampus is thought to be involved in
episodic memory, learning to learn, and memory consolidation.
- Here's a bunch of people doing transitive inference in animals
(there are many, many more...):
- Some nice lists...
- Here are some sites of communities:
- Pages related to some new AI interests:
- I'm doing some work on the evolution of emotions and how they
interact with action selection and social behaviour. Here's a few
- The Emotion
at Cambridge Brain Sciences with Phil Barnard have an elaborate theory
one of the researchers at Bath, David Duke, is involved with.
- Do Fish
Feel Pain? good question, great answer by James D. Rose of Wyoming.
Knowns and Unknowns from 1995, a report in response to The Bell
Unfortunately, many of these only point to papers about the models, not
the models themselves. I feel strongly that programs are like
data, and should be archived with their associated publications and
freely available to anyone who asks. The AmonI software page
has links to our code, and my
publications page makes links to it too.
models of disease spreading & conservation.
Very impressive, used by the British Government.
Hemelrijk models of primate social behaviour.
hopefully one day put his brilliant political science research models
Goodwin uses particle individual based models and ABM to help
manage aquatic species for the US government.
- Michael North's work on modelling power use / distribution is
supposed to one day make ABM the next big thing (like OR), but you have
to hunt hard to find it on the web.
Based Land Use Simulations 19 November 2006.
Diermeier does a lot of excellent modelling, including ABM &
including recently cultural evolution. Sat Oct 29 18:27:18 BST 2005
- The Edinburgh Language
Evolution and Computation Research Unit used to have their (and
other peoples!) code on the web, but apparently took it down. FOR
Fugo of the Mathematical
Biology group of Nara
is modelling a number of things, including the co-evolution of bird
of host & parasite species (e.g. the Cuckoo). Most of his
are mathematical `deterministic' but he's recently started doing
models - I'm eagerly awaiting a paper comparing these models he's
- CMU's Center for
Computational Analysis of Social and
Organizational Systems (CASOS) have a list of tools,
tools & data.
- The Lotto Lab, among
other things, does A Life models of the evolution of vision.
- Cool VR model of Gorilla social life (for education) Gorillas
the bits (not my joke!) from Georgia Tech.
Modelling' architectures (that is, tools / architectures
psychologists use to model human task performance).
- The aforementioned COGENT.
- The aforementioned Soar. From the
same university, EPIC
(there's also an EPIC-Soar...)
- ACT-R (not mentioned
yet, but it's starting to feature in some of our research.)
- Pat Langley has a newish semi-reactive architecture that shows
obvious Stanford & CMU influence, Icarus.
It has short-term & long-term memory (which looks rather like
code :-). He hopes to take on ACT-R with it.
Consciousness, a web page run (as far as I know) by Owen Holland as
part of his half a million pound grant to achieve that.
- Dialogue Systems for
Embodied Agents, from Edinburgh / Ewan Klein. December 08 2003
- Nature Neuroscience special issue on Computational
- Understanding the Evolution of Language would help us understand
thought - see the UIUC
Language Evolution and Computation Resources page (thanks to Les Gasser.)
- A field
guide to the philosophy of mind from Nani and Marraffa.
Computing is another approach to distributed intelligence
(there's (was?) a great
links page there too.)
- Similarly, the Collective
Intelligence group at NASA, under David Wolpert does a lot of hard
maths and maybe something more.
- (Originally from Sante Fe) The
of Artificial Life seems to want to be Slashdot. They seem to have flushed
their agent web page, which is fine because...
the Cortex by Edelman and
Code by Calvin both talk about the "evolution" of competing
thoughts -- something that seems a bit too slow, but still
Based Reasoning might be something like behavior-based AI (honest!)
- Vision might
have something to do with cognition, too.
- The Principia
Cybernetica Project (PCP), is "an international organization
[whose] aim is the computer-supported collaborative development of an
evolutionary-systemic philosophy." They study epistemology,
metaphysics, ethics, concepts and principles, memetics, and the history
and future of evolution. They have interesting web pages on, among
- Laboratory for Autnomous
Robotics and Artificial Life asked to be present
on my links page.
- CMU's Skinnerbots
project is intended to combine robots and animal learning. (The
last entry is from 1997, but their picture is funny.)
- The (former) Wellcome
of Cognitive Neurology now says it does Neuroimaging instead.
- MIT Brain and Cognitive
- Intelligent Systems from
- The Assocition for the
Scientific Study of Consciousness
page author: Joanna