Content uploaded by Edward Rietman
Author content
All content in this area was uploaded by Edward Rietman on Sep 17, 2012
Content may be subject to copyright.
Evolution in Materio: Exploiting the Physics of Materials for Non-Classical
Computation
Simon Harding and Julian F. Miller
Department of Electronics
University of York
York, UK
slh,jfm@evolutioninmaterio.com
Edward A. Rietman
Physical Sciences Inc
20 New England Business Center
Andover, MA 01810, USA
rietman@psicorp.com
Abstract
In this position paperwe report on our work on program-
ming materials for non conventional computing, using an
evolutionary algorithm as the programming technique. The
aim is to use the complexity of the physical world to allow
sophisticated computation, and in particular as a platform
for non-Von Neumann computation. We have demonstrated
this technique using liquid crystal for signal processing and
robot control, however we believe that there are many mate-
rials that could be programmed in a similar way. It is hoped
that such a methodology will provide a general technique
for extracting useful computation from matter, possibly at a
molecular level.
1 Introduction
There are many physical processes that can be described
as a computation. For example, crystal growth from nucle-
ation, corrosion-dendrites on an electrochemical electrode,
a drop of ink dispersing in a glass of water, are all physi-
cal/chemical processes of increasing complexity that can be
thought of as a computation. Further, since biological sys-
tems are part of the physical universe then the development
of an organism from a fertilized egg is also a computational
process. The common element in each of these processes is
the fact that the computation is taking place only between
nearest neighbours. There is no global clock, or central pro-
cessor to distribute tasks to the individual processes com-
prising the overall system.
Many of these processes are either difficult or way be-
yond current computational abilities for modelling. It is
likely given a supercomputer and the set of differential
equations and boundary conditions that describe some of
these processes we would find that our computed results are
only an approximation of the real-world system. We find
that the world is a better model of itself than the models we
can induce from our data. Many of these problems are not
only computationally intractable, but also computationally
undecidable [15] [11].
As an example, consider an array of magnetic spins in
which each site takes on only one of two spin states. At
a high temperature the spins will be randomized, but as
we cool the array down to much lower temperatures we
will find spatial correlations among the spins. This sys-
tem is computationally tractable only if we make certain
simplifying assumptions. But even then, we cannot com-
pute the exact spatial correlations, only the general picture.
This example is a particularly interesting problem because
we can compute correlations either using detailed quantum
mechanics and differential equations, or we can utilize au-
tomata theory and obtain essentially the same result (more
on this later). Of course the automata theory approach is
computationally much faster then the differential equation
approach, and the real-world process is even faster. Feyn-
man has implied that the automata theory approach is a po-
tentially more realistic description of the dynamics at the
meso- micro- and nano-scale, than systems of differential
equations. The tiny ”computational agents” at those scales
do not compute differential equations. They simply interact
with their nearest neighboursand swap information,Garzon
cites a Feynman quotation[4]:
”It always bothers me that according to the laws
[of physics] as we understand them today, it takes
a computing machine an infinite number of log-
ical operations to figure out what goes on in no
matter how tiny a region of space and no matter
how tiny a region of time. How can all that be
going on in that tiny space? Why should it take
an infinite amount of logic to figure out what one
tiny piece of space-time is going to do?” (Richard
Feynman)
If the only process taking place is information being
Figure 1. Schematic of proposed computa-
tional system with bulk matter
swapped by nearest neighbours, then as Stephen Wolfram
proposes, there may be a universal rule set that governs
nearly all the dynamics observed in the universeat allscales
[16].
As Yashihito [17] and many others, have pointed out
that as the device sizes shrink and the level of integration
in microcircuits increases we will more closely approach
the nanoscale. What was clearly articulated by Yashihito is
that we should be able to use matter itself for our computa-
tions. We should be able to exploit the molecular dynamics
and meso-scale physics for computations. Yashihito did not
make explicit suggestions on how to undertake this task. It
is amusing to note that in the 1950s an eccentric scientist
was conducting experiments to grow neural structures us-
ing electrochemical assemblages[12, 13, 1]. Gordon Pask’s
goal was to create a device sensitive to either sound or mag-
netic fields that could perform some form of signal process-
ing - a kind of ear. Using electric currents, wires can be
made to self-assemble in an acidic aqueous metal-salt so-
lution. Changing the electric currents can alter the struc-
ture of these wires and their positions - the behaviour of the
system can be modified through external influence. Pask
used an array of electrodes suspended in a dish containing
the metal-salt solution, and by applying current (either tran-
siently or a slowly changing source) was able to build iron
wires that responded different to two different frequencies
of sound- 50Hz and 100Hz.
Recently, Miller suggested a variety of physical systems
that might be configured to carry out computation[9]. One
of the suggestions was liquid crystal. This has recently
be shown to be possible; Harding and Miller [6, 5] have
demonstrated for the first time that liquid crystal can be
evolved to do analogue filtering. This will be discussed in
the following section.
Figure 1 shows a schematic of the proposed technology.
Basically we will utilize a block of matter (solid, liquid, or
gas) for which we can change its properties/behavior by ex-
ternal forces. The external forces induce property/behavior
changes, which we will call the ”computer program.” So
there is a direct link between the external forces that we
have control over and the induced changes in the block of
matter. Now by measuring the behavior of the altered block
of matter we can essentially submit input data to the sample
and receive output data. In this way we have performed a
type of computation.
Of course we cannot directly program the molecular dy-
namics and we do not have control of the molecules, at least
not directly. The molecules will interact with their near-
est neighbor and we can exploit this phenomena along with
the state changes in regions of the block of matter, in or-
der to perform computations. Figure 1 shows a nanoscale
schematic. In this work, we use liquid crystal as the bulk
matter, and demonstrate a technique that can be used to
”program” liquid crystal to perform computation in the
form of signal processing.
2 Evolution In Materio
2.1 Introduction
Miller and Downning[10] argued that the lesson that
should be drawn from the work of Thompson[14] is that
evolution may be used to exploit the properties of a wider
range of materials than silicon. They refer to this as evolu-
tion in materio. Thompson had found that an evolutionary
algorithm used some subtle physical properties of an FPGA
to solve a problem[14]. It is not fully understood what
properties of the FPGA were used. This lack of knowledge
of how the system works prevents humans from designing
systems that are intended to exploit these subtle and com-
plex physical characteristics. However it does not prevent
exploitation through artificial evolution. Miller suggested
that a good candidate for evolution in materio was liquid
crystal[10]. Recently this suggestion has been vindicated
by recent work by Harding and Miller[5] who showed that
it is relatively easy to configure (using computer controlled
evolution) liquid crystal to perform various forms of com-
putation.
2.1.1 Liquid Crystal
Liquid crystal (LC) is commonly defined as a substance
that can exist in a mesomorphic state [3][7]. Mesomorphic
states have a degree of molecular order that lies between
that of a solid crystal (long-range positional and orienta-
tional) and a liquid, gas or amorphous solid (no long-range
order). In LC there is long-range orientational order but no
long-range positional order.
2
Figure 2. Equipment configuration
2.2 An Evolvable Motherboard with a FPMA
An evolvable motherboard(EM)[8] is a circuit that can
be used to investigate intrinsic evolution. The EM is a re-
configurable circuit that rewires a circuit under computer
control. Previous EMs have been used to evolve circuits
containing electronic components [8, 2] - however they can
also be used to evolve in materio by replacing the standard
components with a candidate material.
An EM is connected to an Evolvatron. This is essentially
a PC that is used to control the evolutionary processes. The
Evolvatron also has digital and analog I/O, and can be used
to provide test signals and record the response of the mate-
rial under evolution.
In the experimentspresented here, a standard liquid crys-
tal display with twisted nematic liquid crystals was used as
the medium for evolution. It is assumed that the electrodes
are indium tin oxide. Typically such a display would be
connected to a driver circuit. The driver circuit has a con-
figuration bus on which commands can be given for writing
text or individually addressing pixels so that images can be
displayed. The driver circuit has a large number of outputs
that connect to the wires on the matrix display. When dis-
playing an image appropriate connections are held high, at
a fixed voltage - the outputs are typically either fully on or
fully off.
Such a driver circuit was unsuitable for the task of in-
trinsic evolution. There is a need to be able to apply both
control signals and incident signals to the display, and also
record the response from a particular connector. Evolution
should be allowed to determine the correct voltages to ap-
ply, and may choose to apply several different values. The
evolutionary algorithm should also be able to select suitable
positions to apply and record values. A standard driver cir-
cuit would be unable to do this satisfactorily.
Hence a variation of the evolvable motherboard was de-
veloped in order to meet these requirements.
The Liquid Crystal Evolvable Motherboard (LCEM) is
circuit that uses four cross-switch matrix devicesto dynami-
cally configure circuits connecting to the liquid crystal. The
Figure 3. The LCEM
switches are used to wire the 64 connections on the LCD
to one of 8 external connections. The external connections
are: input voltages, grounding, signals and connections to
measurement devices. Each of the external connectors can
be wired to any of the connections to the LCD.
The external connections of the LCEM are connected to
the Evolvatron’s analogue inputs and outputs. One connec-
tion was assigned for the incident signal, one for measure-
ment and the other for fixed voltages. The value of the fixed
voltages is determined by the evolutionary algorithm, but is
constant throughout each evaluation.
In these experiments the liquid crystal glass sandwich
was removed from the display controller it was originally
mounted on, and placed on the LCEM. The display has a
large number of connections (in excess of 200), however
because of PCB manufacturing constraints we are limited
in the size of connection we can make, and hence the num-
ber of connections. The LCD is therefore roughly posi-
tioned over the pads on the PCB, with many of the PCB
pads touching more than 1 of the connectors on the LCD.
This means that we are applying configuration voltages to
several areas of LC at the same time.
Unfortunately neither the internal structure nor the elec-
trical characteristics of the LCD are known. This raises the
possibility that a configuration may be applied that would
damage the device. The wires inside the LCD are made of
an extremely thin material that could easily be burnt out if
too much current flows through them. To guard against this,
each connection to the LCD is made through a 4.7Kohm re-
sistor in order to provide protection against short circuits
and to help limit the current in the LCD. The current sup-
plied to the LCD is limited to 100mA. The software control-
ling the evolution is also responsible for avoiding configura-
tions that may endanger the device (such as short circuits).
It is important to note that other than the control circuitry
3
Liquid Crystal Display
8x16 Analog Switch Array
8 External Connectors
LCD contacts,
32 per side
- 64 in total.
Figure 4. Schematic of LCEM
for the switch arrays there are no other active components
on the motherboard - only analog switches, smoothing ca-
pacitors, resistors and the LCD are present.
2.3 Systems Evolved In Liquid Crystal
2.3.1 Tone Discriminator
A tone discriminator as a devicewhich when presented with
one of two signals input signals returns a different response
for the each signal. In [14], on which this experiment is
loosely based, the configuration to differentiate two differ-
ent frequencysquare waves, giving a low outputfor one and
a high output for the other. We have evolved a system in
liquid crystal that was also able to perform this task. It was
found that it was easier to evolve solution in liquid crystal
than it was to evolve a circuit in an FPGA. An example of
the evolved response in shown in figure 5.
2.3.2 Real-time Robot Controller
We have recently demonstrated, for the first time, that it is
possible to evolve a robot controller in liquid crystal. A con-
troller was evolved that allowed a simulated robot to move
around an enclosed environmentwith obstacles withoutcol-
liding with the walls, as shown in figure 6. The task is sig-
nificantly harder than that of our previous work with liquid
crystal (if only because the number of inputs and outputs to
the display device has been doubled). Yet we found that it
was relatively easy (in evolutionary terms) to evolve a so-
phisticated robot controller.
Figure 5. Tone discriminator response. Dark
areas indicate 5kHz input, light 100Hz
Figure 6. Path of an robot controlled using
liquid crystal
The quality of results when compared to previous work is
also high. The environment is more complex than that of
comparable work, and unlike much work on evolving GP
robot controllers or neural network controllers, we solve a
real-time control task. The results also indicated an evo-
lutionary computational effort that is comparable to other
examples of evolved controller (with simpler tasks).
2.3.3 Logic Gates
We have recently demonstrated that it is possible to evolve
logic gates exhibiting digital behaviour in liquid crystal. As
expected, they operate at much slower speeds than conven-
tional devices and the behaviour can be intermittent. We do
not feel that systems such as liquid crystal should be used
in this manner, and that they are much more suited to non-
4
classical computation.
2.4 Conclusions
We have only explored a tiny fraction of the potential of
computational matter. We have demonstrated that it is pos-
sible to program material systems, in this case using evo-
lution, to provide computation in both classical and non-
classical senses. At present our experimental set up is rather
crude. In the future we will construct field programmable
matter arrays that will allow us to control the material in
more sophisticated ways. We believe that this will enable
the development of computational devices that offer advan-
tages over conventional devices. It may be possible to build
small devices operating at a molecular level, that require
low power and may be more resistant to environmental fac-
tors.
Such devices may be more appropriate than existing tech-
nology for un-conventionaland non-Von Neumann comput-
ing. The work with liquid crystal has demonstrated the vi-
ability of programmable materials and that an appropriate
programming technique exists in the form of evolutionary
search.
References
[1] P. Cariani. To evolve an ear: epistemological implications
of gordon pask’s electrochemical devices. In Systems Re-
search, volume 3, pages 19–33, 1993.
[2] J. Crooks. Evolvable analogue hardware. Meng project re-
port, The University Of York, 2002.
[3] D. Demus, J. Goodby, G. W. Gray, H. W. Spiess, and V. Vill,
editors. Handbook of Liquid Crystals, volume 1,2A,2B,3.
July 1998.
[4] M. Garzon. Models of Massive Parallelism, Analysis of Cel-
lular Automata and Neural Networks. Springer, New York,
1995.
[5] S. Harding and J. F. Miller. Evolution in materio: A tone
discriminator in liquid crystal. In In Proceedings of the
Congress on Evolutionary Computation 2004 (CEC’2004),
volume 2, pages 1800–1807, 2004.
[6] S. Harding and J. F. Miller. Evolution in materio: Initial
experiments with liquid crystal. In Proceedings of 2004
NASA/DoD Conference on Evolvable Hardware (EH’04),
pages 298–305, 2004.
[7] I. C. Khoo. Liquid Crystals: physical properties and non-
linear optical phenomena. Wiley, 1995.
[8] P. Layzell. A new research tool for intrinsic hardware evolu-
tion. Proceedings of The Second International Conference
on Evolvable Systems: From Biology to Hardware, LNCS,
1478:47–56, 1998.
[9] J. F. Miller and K. Downing. Evolution in materio: Look-
ing beyond the silicon box. In NASA/DOD Conference
on Evolvable Hardware, pages 167–176. IEEE Comp. Soc.
Press, 2002.
[10] J. F. Miller and K. Downing. Evolution in materio: Looking
beyond the silicon box. Proceedings of NASA/DoD Evolv-
able Hardware Workshop, pages 167–176, July 2002.
[11] C. Moore. Unpredictability and undecidability in dynamical
systems. In Phys. Rev. Lett, volume 64, pages 2354–2357,
1990.
[12] G. Pask. Physical analogues to the growth of a concept. In
Mechanization of Thought Processes, Symposium 10, Na-
tional Physical Laboratory, pages 765–794, 1958.
[13] G. Pask. The natural history of networks, 1959.
[14] A. Thompson. An evolved circuit, intrinsic in silicon, en-
twined with physics. In ICES, pages 390–405, 1996.
[15] S. Wolfram. Undecidability and intractability in theoretical
physics. In Phys. Rev. Lett., volume 54 (8), pages 735–738,
1985.
[16] S. Wolfram. A New Kind of Science. Wolfram Media, 2002.
[17] A. Yashihito. Information processing using intelligent mate-
rials - information-processing architectures for material pro-
cessors. In J. of Intell. Mat. Syst. and Structures, volume 5,
pages 418–423, 1994.
5
