Artificial Life Lecture 19

1
Artificial Life Lecture 19
Ragbag of suggestions received for mini-topics,
plus a personal prejudiced survey of hot topics.

• UK companies interested in EASy work
• How long should you wait for a GA to finish
  when to give up?
• Von Neumann and Alife
• Nano-technology and Alife, Real chemistry and
  Alife
• Current in-silico Alife theory
• Markov processes
• Modelling of environments.

2
Companies interested in EASy
BT a few years ago claimed they took 15 of all
computer science graduates in the UK surely no
longer true! BT RD labs at Martlesham (nr
Ipswich) used to have 4000 in Development and 600
in Research. Now slimmed down, sort of privatised
into cost centres and called BT Exact. EASy
contacts specifically with their Future
Technologies Group (formerly from 1996
Artificial Life Group) (Richard
Tateson) http//more.btexact.com/projects/ftg Evol
utionary computing, biological metaphors, fruit
flies and cell phones. Has funded DPhils
3
Companies interested in EASy

• Hewlett Packard - HP Labs Bristol
• BICAS Biologically-Inspired Complex Adaptive
  Systems http//www.hpl.hp.com/research/bicas/
• Dave Cliff (started the EASy MSc here)
• HP Labs Agent Research
• Complex Adaptive Systems Research Group
• Agent-based models in economics/market trading
• Immune system models for computer security
• hpDJ
• Take EASy students for Summer projects

4
Companies interested in EASy

• Many other big companies with research labs have
  an active interest in EASy issues eg British
  Aerospace, Logica, Nortel
• Smaller labs MASA, other businesses in Sussex
  Innovation Centre
• Natural Motion www.naturalmotion.com (CEO Torsten
  Reil is EASy graduate)
• Brighton media companies eg Runtime collective
  http//www.runtime-collective.com/
• Victoria Real http//www.victoriareal.com/
• Searchspace http//www.searchspace.com/index.shtml
  
• Etc etc

5
Recent robotics at Sussex

• Research mostly done in CCNR http//www.cogs.susse
  x.ac.uk/ccnr/
• Insect and robot navigation (eg Kyran Dale and
  Linc Smith)
• ER to investigate Learning (Tuci) Development
  (Wood) Collective Behaviour (Quinn) Multiple
  sensor modalities (Bird) GasNets (Tom Smith and
  Phil Husbands) Visually guided behaviour (Emmet
  Spier) Homeostasis (Di Paolo) Passive Dynamic
  Walking (Vaughan)
• Also eg Buehrmann Fernando Sojakka Macinnes
  Vickerstaff Bardeen
• Evolvable Hardware (Adrian Thompson, Garvie,
  Kenneally)

6
Alife simulation/modelling

• Real biological studies eg in population biology
  and ecology
• Traditional scepticism to modelling in biology
• Mathematical models Journal of Theoretical
  Biology
• Now all over the place !!
• Alife Agent-based model translates to/from
  IBM Individual Based Model
• Too many to list!

7
Miscellaneous

• Random comments

8
How long should you wait for a GA to finish?
and when should you give up? The bottleneck
is (a) the time it takes for a single evaluation
multiplied by (b) the minimum number of
evaluations likely to be needed.
Very rough rule of thumb for (b) (no guarantees,
derived from Speed Limit for Evolution)
Binary genotypes Number of bits (possibly make
allowance for junk) times 100 or 1000
(generations) times popn size. E.g. 100 bit
genotype, popn 30, 3 million evaluations.
9
how long ?
Real-valued genotypese.g. a Neural Net with 100
weights allow similar time as binary but with
say 8 bits per real weight, so 8 times 3 million
24 million evaluations.
Warning 1 This is a very speculative rule of
thumb, for a vague guideline only, a given real
problem may be 100 times easier or 1000 times
worse! Warning 2 Even if this speculation was as
true as could be for a particular problem, one GA
run could take 1/10 the estimate, another 10
times the estimate (very noisy).
10
Take-home lessons

• You need some luck!
• but also you need some common sense, dont
  expect unrealistic miracles
• Try and make your evaluations as fast as
  possible, within reason.
• Above all, start your experiments with minimal
  cut-down small versions and gain experience with
  these.

11
Von Neumann and Alife
Interesting observation that both major pioneers
of computing, Alan Turing and John von Neumann,
also had an active interest in what nowadays
would be called Alifey issues. Turing use of
evolution, morphogenetic issues, role of
embodiment. Von Neumann Cellular Automata,
self-reproduction and origin of life, role of
embodiment.
12
Von Neumann still relevant?
are von Neumanns contributions to Alife still
relevant, are there any remaining mysteries from
his work?
It is virtually always constructive and
instructive to go back and read early works by
such people, who come at the same questions with
often a different set of perspectives. Alife is
not an established science, it is an open field
of enquiry. There are solid pieces of work
(amongst the dodgy stuff), but there is no
established agreed foundation of Alife theory.
13
Nano-technology, real chemistry
Nano-technology is currently a promise with no
delivery date.

• Technological advances at constructing tinier and
  tinier mechanisms is proceeding inexorably
• But the promised clever design seems mostly
  fantasy, assuming the AI guys deliver on
  unrealistic promises

This seems a ripe area for Alife people and
evolutionary approaches to design of lifelike
systems
14
real chemistry
Artificial chemistry approaches (such as Pietros
Guest lecture) are formally very interesting, but
are disembodied. There are a few people looking
at the practical questions of creating real
artificial (I.e. synthetic) living cells Pier
Luigi Luisi synthesising an autopoetic
cell Rebek et. Al., Scripps Institute, emergent
autocatalysis. Gerald Joyce, Scripps Institute,
artificially evolving RNA. Rich Lenski.
15
Current in-silico Alife theory
Is there any current in-silico Alife theory
Alife is not an established science, it is an
open field of enquiry. There are solid pieces of
work (amongst the dodgy stuff), but there is no
established agreed foundation of Alife theory.
16
Markov processes
what are they ? Roughly speaking, a formal way
of modelling sequences of events, where each
event is one of a number of discrete
possibilities. e.g. text is a stream of letters
from an alphabet. You could model the production
of text as ltspacegt has a 20 chance, ltegt has a
13 chance, lttgt has a 7 chance etc of being
produced at the next step, and a program set up
to deliver this would produce vaguely text-like
material
17
Markov
Such a program would, at each time-step, generate
a new symbol independently of what happened
before. OR you could start taking account of the
immediately preceding letter- e.g. ltqgt is
followed by ltugt 99 of the time, and by ltspacegt
1 of the time (Iraq ) OR you could take
account of 2, 3 or more preceding letters, and
the generated text would look more and ore
plausible as you do so.
18
Markov
So as with so many things, Markov processes are a
formal well-defined model, that behaves
(stochastically) according to precisely defined
rules and then there is the very different
question of to what extent is this formal model
a good model of some natural phenomenon?
19
Modelling of environments
Horses for courses. For e.g. minimal cognition
experiments (e.g. Beer) then one deliberately
uses some abstract minimal environment the
Frictionless Plane/Brain For agents moving
around, big step between grid-world and
real-valued-world. Physics engines ODE Role of
noise Jakobis minimal simulations.
20
Hot research topics
Personal prejudiced list Could supervise summer
projects in these areas
21
Passive Dynamic Walking
This is a very embodied, dynamical systems
approach. So far PDWs have just gone down gentle
slopes under gravity. There is scope for adding
small amounts of power input for more general
walking. Matt Williamsons work, associated with
Brooks COG project, using coupled oscillators,
seems a very promising lead that could be applied
here. Eric Vaughan now doing this.
22
Neutral Networks
Wide open for research. Barnetts work shows that
in a formally defined class of (binary) fitness
landscapes full of NNs in a particular fashion,
best strategy is a population of size 11, with a
fixed number of mutations based on getting
expected proportion that are neutral as close as
possible to 1/e 37. Adrian Thompsons hardware
evolution supports this Extensions noisy fitness
evaluations ? Real valued genotypes?
23
Gaia/Maximum entropy?
Relationship between Daisyworld models of
homeostasis, and Thermodynamic ideas that
- Systems try to organise themselves to
produce entropy as fast as possible. Cf. Kay and
Schneider, 4th Law of Thermodynamics. Organised
systems can dissipate junk entropy faster, and
( speculation) Life does it better than
anything else and hence should naturally occur (
given the right circumstances)
24
Homeostasis
Ezequiels work on looking at issues of
homeostasis in minimally cognitive models seems
very important and interesting. As also does Jim
Stones work on ( roughly speaking) what
perceptual systems (eg ANNs) have to do in order
to sift out higher-level invariants from all the
noise. Roughly- (1) output of a system should
not fluctuate wildy (just echoing the noise is
pointless) But (2) Should not stay still it
should fluctuate over the long term if it is to
reflect real things happening in the world.
25
Economy
So, Jim Stone points out that outputs of of
perceptual systems should minimise (Short Term
Variance) divided by (Long Term
Variance) Minimise STV/LTV where short/long refer
to appropriate time scales. This is one way of
ensuring that an output neuron is earning its
keep. Related to Ezequiels interpretation of
homeostasis keepng neurons activations in
the middle of their sigmoids
26
A Speculation
If you evolve CTRNNs for an agent to do a task
(eg phototaxis) with an extra fitness factor
related to making sure each neuron earns its
keep e.g. Stones criteria, minimising
STV/LTV then you can hope to see homeostais
??? (variant on Ezequiels upside-down glasses
experiments)
27
The End
end of ragbag of speculations time for more
discussion ??