Proc. 4th Eur. Conf. on Arti cial Life (ECAL97), Husbands, P. and Harvey, I. (eds.

), 388-397, MIT Press 1997.

Temperature in Natural and Articial Systems

Adrian Thompson
Centre for Computational Neuroscience and Robotics,
University of Sussex, Brighton BN1 9QH, UK.
adrianth@sussex.ac.uk

Abstract are strong correspondences between the natural and elec-

tronic cases at every stage of the discussion, with highly
Recent experiments in evolutionary electronics suggestive consequences.
have shown how articial evolution can craft ex- The motivation behind this research is to improve evo-
tremely ecient electronic circuits by manipulat- lutionary electronics as an engineering technique. How-
ing a real physical silicon medium. Each individ- ever, there are broader implications for ALife researchers,
ual circuit is physically instantiated in a recon- and these will be identied in the penultimate section.
gurable chip (FPGA) for its tness evaluation,
so evolution can exploit all of the natural physi-
cal properties exhibited by the electronic medium,
resulting in circuits well tailored to it. This can
only be done properly by rigorously rejecting con-
There is a lot of biology in this paper, but I am not a
biologist. The biological information is taken from [1]
except where indicated otherwise by citations, but any
errors are my own responsibility.
The next section summarises a recent experiment in
ventional design methods. Articial evolution is evolutionary electronics to illustrate the motivation for
then faced with a similar problem to that encoun- this research. We then consider the fundamentals of
tered in nature: how to construct a system from how temperature in
uences biological and silicon sys-
processes which all vary with temperature, such tems. The subsequent sections consider rst how a sys-
that the system can perform adequately over a tem may cope with temperature change (`Temperature
wide range of temperatures? It is benecial to
do this in a more natural way than simply for-
bidding all analogue continuous-time dynamics,
as conventional digital design does. Engineering
proposals are formulated by analysing the corre-
spondences between nature and evolutionary elec-
tronics | some of these are promising and sur-
prising. There are wider implications for ALife,
in that thermal considerations cannot be as eas-
ily ignored as `implementation details' as might
have been thought.
1 Introduction
Compensation') and then how a system's internal tem-
perature can be stabilised (`Thermal Regulation'). Both
are eective methods of sustaining operationality in the
face of changing external temperature, and they can be
used together. Finally, some implications for ALife mod-
elling are noted.
2 Motivation: Evolutionary Electronics
A Field-Programmable Gate Array (FPGA) is a Very-
Large Scale Integration (VLSI) silicon chip. It consists
of an undedicated array of components interspersed with
wires of various lengths. Within the components are elec-
tronic switches which control their behaviour, and elec-
It is now possible to allow articial evolution to manip- tronic switches also control how the inputs and outputs
ulate directly a real physical semiconductor medium to of the components connect to the wires. A particular set-
construct automatically electronic circuits that satisfy ting of the switches denes which out of the vast range
an engineering specication. Once preconceptions from of possible electronic circuits is physically instantiated
conventional design methods are rigorously rejected, the on the chip at any one time. Recent advances in FPGA
silicon electronic medium can be exploited in a way that technology make it possible to place the settings of the
is truly natural to its properties, resulting in highly ef- switches (and hence the physical circuit) directly under
cient circuits. Freely exploiting the physical properties the control of an evolutionary algorithm | previously
of a medium has its pitfalls, and one is that the evolved there were technical diculties due to limitations of the
circuit may be unable to operate over a sucient range chips [2]. With the newly introduced XC6216 device
of temperatures to be of wide applicability. Natural evo- from Xilinx [3], the switch settings can be partially or
lution must have faced the same problem, and this pa- completely recongured in a negligible amount of time,
per aims to learn lessons from it. We shall see that there there is no limit on the number of recongurations, and
the chip's architecture is such that it can never be inter- 1kHz 10kHz
nally damaged by any conguration. In this experiment,
only a subset of the chip's capabilities was enabled: each IN
component performed a single function of three inputs,
and only nearest-neighbour connections were used.
The settings of the switches are determined by the
contents of bits of RAM spread throughout the chip.

0
This RAM can be written to as easily as conventional
computer memory. Thus, any appropriate avour of
evolutionary algorithm can easily be implemented on a

3500 1400
standard desktop computer and allowed to manipulate
the physical circuit instantiated in silicon on the FPGA.
Fitness evaluations can be the measurement of the per-
formance of an evolved conguration as a real physical
electronic circuit, behaving in real-time according to the
laws of semiconductor physics. This means that evolu-
tion can be free to explore all possible congurations,
and hence all possible behaviours in the repertoire of
the FPGA. There is no need to constrain evolution to
work only within the relatively small subset of circuits
we know how to design, analyse or feasibly simulate. In-
stead, evolution can be allowed to explore beyond the
scope of conventional design methodologies, exploiting
richer architectures and dynamical behaviours, and us-
ing complex aspects of semiconductor physics.
Figure 1: The voltage waveforms of the two dierent
tones presented to the input (top row), and of the cor-
responding output produced by the best of the population
after 0, 1400 and 3500 generations (the three rows be-
low). These are photographs of the oscilloscope screen,
with the probes directly attached to the input pin and the
output pin of the FPGA. See text for interpretation.
In this spirit, the conguration bits of a 10  10 corner ponents | intended to be used to perform operations of
of the XC6216 FPGA were directly encoded as a linear Boolean logic | are really just high-gain congurations
bit-string genotype of 1800 bits, and allowed to evolve of transistors, and in the absence of the constraints of
without constraint using a fairly conventional genetic al- a digital design methodology, this is just how evolution
gorithm. The individuals were evaluated one-at-a-time treats them. Finally, the best circuit of generation 3500
as congurations of a real FPGA chip. The task was a gives the desired behaviour, showing the `digital-looking'
simple rst-step towards signal-processing and pattern- output demanded by the tness function, though it is
recognition applications. A single input was driven with certain that complex dynamics are still used internally.
a sequence of audio tones of 1kHz and 10kHz frequency,
presented in a random order. The target behaviour was
for the single output immediately to go to a steady 5V
level as soon as one frequency was present, and to go
to 0V for the other one. Such a circuit might be used
to demodulate binary data sent over a telephone line.
The task might sound trivially easy until it is remem-
bered that there are no external timing components:
evolution must craft a continuous-time arbitrarily recur-
rent network of 100 components (each of which has an
input)output time delay of only a few nanoseconds) to
distinguish between input periods that are on a timescale
The nal circuit is shown in Fig. 2. Note the highly
recurrent structure. It behaves in continuous-time as an
unfolding of the laws of semiconductor physics: the chip
is a dynamical system, not a computational one. The
components shaded gray are part of the circuit's func-
tion even though they appear to be disconnected from
it: they must be interacting by some subtle property of
semiconductor physics such as electromagnetic coupling
or interaction through the power-supply wiring. As a
result of this rich unconstrained architecture, dynamics,
and use of physical resources, the circuit is incredibly
ecient: using just 32 cells it is one or two orders of
ve orders of magnitude slower. If achieved, this would magnitude smaller than one would expect from conven-
be a signicant feat, because this corner of the chip is
only 1mm2 in area, and requires no external components.
Fig. 1 shows the progress of evolution. In the initial
random population of 50 individuals (generation 0), all
the circuits happened to be as bad as possible, for exam-
ple maintaining a constant output voltage irrespective
of the input. After 1400 generations, a promising par-
tial solution is shown: the rich continuous-time analogue
dynamics of the circuit are evident. The FPGA's com-
tional methods. See [4] for full details of the experiment
and the theory behind it. To many people, the experi-
ment seems somewhat akin to the evolution of physical
nervous systems in nature.
The circuit operates perfectly over the 5 C range
of temperatures that the population experienced dur-
ing evolution. However, the more the temperature de-
viates from this range, the more the circuit malfunc-
tions. Conventional digital design, in forbidding the use
 Out In practice, the interaction of several potentially rate-
limiting processes (physical as well as chemical) can lead
to gradual, rather than sharp, changes in  and Q10 with
temperature. Even for a single biochemical reaction, the
rate increase with temperature falls o as temperature
increases, presumably because of the destruction of en-
In zymes on which they depend [5]. Nevertheless, the Ar-
rhenius relationship holds for many biological phenom-
ena under temperature ranges of interest, and is even re-

ected in behaviours such as the rate of creeping of ants,
the chirping of crickets, the
ashing of re
ies, the beat-
ing of cilia, and some respiratory and cardiac rhythms.
`The slope of the linear relationship between the log of
the rate of most biological reactions and the reciprocal of
absolute temperature is the Arrhenius  divided by ap-
Figure 2: The functional part of the nal evolved circuit.
The large boxes represent the components, the inputs of
which are marked with small squares.
of continuous-time analogue dynamics, can produce cir-
cuits with much greater ranges of operating temperature
| at the expense of less ecient utilisation of the physi-
cal resources. How can evolution be allowed to naturally
proximately 4:6, with  dened by the limiting step.' [6,
Chap. 37]. For thermochemical (enzymatic) reactions,
Q10 in typically somewhere between 2 and 3 : they often
go about twice as fast for every 10C rise in temperature.
The temperature dependencies of neural systems also
arise from physical, as well as chemical, origins; in semi-
conductors the processes are entirely physical. In gen-
eral, the Q10 values associated with physical processes
(such as for diusion or conductivity), and also of those
exploit the physical medium, and thus reap the benets associated with photochemical reactions, are less than
seen above, but yet produce circuits able to operate in 1.5. The operation of both neurons and semiconductor
a wide range of temperatures? Natural evolution must devices is to a large extent based upon the movement
have faced the same problem: can we draw some inspi- of charge-carriers in an electric eld (at a speed propor-
ration from it? That is the purpose of this paper. To tional to their mobility) and the interplay between this
begin, the next section considers the roots of tempera- and the diusion of those particles (at a speed propor-
ture's in
uence on biological and silicon systems. tional to their diusion constant ) in the concentration-
gradient which is in
uenced by that movement. In neu-
3 The eect of Temperature upon Rate rons and semiconductors, the charge-carriers are dier-
The rate of chemical reactions increases with tempera- ent, and the processes establishing electric elds and con-
ture. This is described by the Arrhenius relationship, centration gradients are dierent, but the link between
which can be written in the form:
k 

1 1
 The Einstein relationship is crucial to this link: the
2
ln k = R T T (1)
1 1 2 plied by the absolute temperature, so temperature ap-
where T1 and T2 are the absolute temperatures corre-
sponding to reaction velocities k1 and k2, R is the gas
constant, and  is the critical thermal increment : a con-
stant characterising the particular reaction. When this
relationship holds, a more friendly measure can be cal-
culated: Q10 = kt+10=kt, where kt is the rate at tem-
perature t, and kt+10 is the rate at 10 C higher. So Q10
simply tells us by what factor a rate increases for a 10 C
rise in temperature.
In biochemical processes, often composed of a com-
plex pathway of many intermediate reactions,  is only a
constant over a limited temperature range, which might
be smaller than that of the phenomenon under study.
A change in temperature can change which of the steps
in the pathway is the rate-limiting one, resulting in a
sharp change in  (and Q10) at particular temperatures.
diusion and the electric eld is fundamental to both.
diusion constant is proportional to the mobility multi-
pears in many of the fundamental equations of neurons
and semiconductors (eg. the Nernst equation for equi-
librium potentials at neuron membranes, and | via the
Boltzmann distribution | basic equations in semicon-
ductor physics) [7, 8]. The rest of this section will now
give some examples of how these temperature dependen-
cies are manifested in neural and electronic systems.
The refractory period for nerve impulses increases for
lower temperatures [5]. Mammals (which have nerves
capable of propagating impulses at speeds ranging all
the way from 0:3ms 1 to 100ms 1 ) have a Q10 for the
speed of nerve impulse propagation of around 1.7, with a
lower bound of 5C below which propagation ceases. In
fact, the Q10 is higher at low temperatures: at high tem-
peratures cell-membrane permeability changes become
so fast that discharging of the cell membrane capaci-
tance by fully activated permeability mechanisms begins
to be rate-limiting. Conduction velocity in cold-blooded
animals is less temperature-sensitive than in mammals.
Conduction velocity can also change with acclimation1
and seasons [7, Chap. 4] | see the next section.
As well as aecting the rate of action potential propa-
gation, temperature in
cal processes display adaptation to temperature, coping
with seasonal changes and dierent latitudes. How?
As temperature goes down, the primary thing to avoid
or cope with is ice crystal formation | the crystals can
cause mechanical damage to the cell membranes, and
can have drastic eects on the vital osmotic and liquid
uences the rate of neuron ring: balances as liquid water is removed. There are many
`The changes of action potential frequencies with temper-
ature are associated, although not in a simple manner,
with changes in resting potentials. Cooling reduces the
resting potential (depolarization) and this leads to a rise
in action potential frequencies; but certain nerve cells
show a frequency increase when temperature is raised.'
[9]
A similarly changeable situation prevails in VLSI
chips such as the FPGAs used in evolutionary elec-
tronics. These devices are made in complementary
metal-oxide semiconductor (CMOS) technology, where
wonderful ways in which this is achieved (eg. through
supercooling [6, Chap. 21 ]). Of particular interest is the
use of antifreeze substances; it appears that these can
also prevent low-temperature changes in protein struc-
ture. In this way, important enzymes may be maintained
in the active state even at low temperatures, partly es-
caping from the Arrhenius equation.
`Acclimation depends on exploitation of the accelera-
tions and maintenance of the independence of the limita-
tions of the Arrhenius equation.' [6, Chap. 37]. Inspec-
tion of the change in the rate-temperature curve before
the rate-limiting factor is the time taken for a eld-eect
transistor's `ON' resistance to charge or discharge the
gate capacitances of the transistors connected to it and
the parasitic capacitance of the interconnections. The
overall outcome is that delays increase by about 0.3%
per  C, which translates to a Q10 of 1.03. This sounds
very good compared to the biological case, until we re-
member that electronic circuits are commonly expected
to work with no observable change in performance over
a very wide range of ambient temperatures | typically
40 C to 85 C or even 125 C.
Clearly, both biological evolution and evolutionary
electronics has a challenging task in producing systems
with adequate thermal stability. How does nature do it?
The rest of the paper addresses this question, attempt-
and after acclimation indicates that one, or most com-
monly both, of the following is responsible: (1) `altered
enzyme activity due to changes in concentration, pH, wa-
ter activity, or relation among enzymes'; (2) `a change in
activation energy due to alteration of enzyme protein, a
cofactor, or shift to alternate pathways.' [6, Chap. 37].
Thus, in response to a temperature change, the bio-
chemical reaction pathway responsible for a particular
function may be altered in mechanism to use reactions
which are suited to the new conditions. It is conceivable
that this strategy could be applied to silicon: the system
could be composed of many alternative low-level mech-
anisms for each sub-function which automatically come
into play as appropriate for the current temperature.
Antifreeze (and potentially all the above mechanisms)
ing to apply the ndings to evolutionary electronics along can be regulated by (possibly seasonal) patterns of neu-
the way, before nally commenting on the implications roendocrine activity, as well as being directly in
uenced
for ALife modelling studies. There are two complemen- by temperature. Particular physiological processes can
tary possibilities: the rst is to produce a system that be indicated as being concerned with compensation by
works even when the temperature changes, which I will identifying temperature-dependent changes in the spe-
call compensation. The second is to produce a system
that regulates its internal temperature to be within lim-
its it can cope with. Many animals do both.
4 Temperature Compensation
4.1 Cellular and Biochemical Compensation
exes such as cardiac and respiratory rhythm. Other
cic enzymes associated with them. Acetylcholinesterase
activity points to nervous tissues or processes as the
site of compensation or lethal collapse. It has long
been known that heat-death in some animals is due to
nervous-system failure with a loss of indispensable re-
If one were to go out in summer and measure the Q10 of
some property of a cold-blooded animal one might infer
that in winter that process will come to a complete stand-
still. Returning in winter to repeat the measurement,
however, it may be found that the process is proceeding
uence
studies (eg. in sh) have indicated that the nervous sys-
tem is also the locus most sensitive to cold.
These observations suggest that there is a useful role
for the nervous system in partially controlling low-level
biochemical adaptation. In the silicon scheme, then,
perhaps some high-level controller should also in
at the same rate as it did in summer. Many biologi- the choice of ne-grained mechanisms to be invoked by
1 The word `acclimation' refers to the laboratory-equivalent of
the current temperature. However, that controller it-
acclimatisation: temperature is manipulated, but other environ- self might demand particular thermal precautions, as in
mental factors which might simultaneously change in signicant the case of the nervous system. Note that, using an
ways in the natural environment are held constant. FPGA capable of rapid partial reconguration, the al-
ternative mechanisms in the above scheme need not all
be present on the silicon simultaneously: sub-circuits can
be `swapped' in and out of the chip as appropriate for
the current temperature. Indeed, several entire systems
could be evolved, each for a dierent temperature range,
the entire circuit being swapped to cope with a tempera-
ture change. There is a biological analogy: `The seasonal
development of arthropods, particularly those with one
generation per year, has usually evolved in such a way
mutants lack temperature compensation in a particular
circadian rhythm that is otherwise normal [12].
How can such compensation be achieved? If the os-
cillation arises from the interaction of two or more pro-
cesses, each of which is aected by temperature, then
the interactions can be arranged so as to give an over-
all stability in period. For example, the period might
depend on the products of a biochemical reaction which
increases in rate with temperature; there could be a sec-
that only one specic stage is capable of hibernating suc- ond reaction which inhibits the rst one, and which also
cessfully.' [6, Chap. 21]. Here, the genotype species
several dierent structures, only one of which behaves
appropriately in low temperatures (by hibernating).
4.2 Compensation in Behavioural Timescales
Numerous circadian (daily) rhythms have been docu-
mented in a wide range of behavioural and physiological
variables.2 For example, they play a role in thermal reg-
ulation, sleep, feeding and drinking behaviours, and en-
docrine, renal and reproductive function. In some cases
the rhythm needs to be equivalent to an internal clock of
considerable precision, for example it is widely believed
that migratory birds use such a clock to correct for the
sun's movement across the sky to allow it to be used as
a compass. Rhythms on shorter timescales (eg. cardiac
increases in rate with temperature. Another possibil-
ity is to take the net eect of processes having recipro-
cal temperature coecients: this could even be applied
to the mutual entrainment of multiple oscillators [13].
Note that even one-celled animals have temperature-
independent endogenous rhythms (ibid.). Some progress
has been made towards postulating neural bases for cir-
cadian rhythms, but this is still far from understood [12].
In an fascinating duality with the solar navigation of
birds mentioned above, a prize of $20000 was oered
in 1714 for the maker of a time-piece suciently accu-
rate to enable longitude to be ascertained at sea. One of
the problems was the variation with temperature of the
lengths of pendulums and of balance-springs [14]. The
solution was to use the net eect of two dierent metals
and respiratory patterns) and on the longer timescales
of seasons and years are also important.
Some biological rhythms are directly derived from
environmental cues, but others result from some sort
of `endogenous' oscillator in the animal. Such oscilla-
tors are capable of free-running in the absence of en-
vironmental cues but can also be entrained by multi-
ple `timegivers'3 in the environment such as light-dark
cycles, food-availability cycles, temperature cycles, and
social and acoustic cues. `Entrainment' means a grad-
ual and ongoing resynchronisation to the environmental
timegiver (rather than a sudden resetting). Ongoing in-
teraction with timegivers can result in a dierent period
of oscillation than the free-running period would be in
the absence of cues. It has been shown that, apart from
entrainment, circadian rhythms are not learnt phenom-
having dierent thermal properties: this ts into the gen-
eral scheme above. Presumably compensation schemes
of an analogous nature can be built into electronics, or
could arise through evolution given appropriate primi-
tives and a selection pressure.
In evolutionary electronics, the thermal stability of in-
ternal timescales could also be achieved through inter-
action with external timegivers, as in entrainment. For
example, if the 1kHz/10kHz discriminator is given a se-
quence of inputs consisting of both frequencies, then it
is constantly being `reminded' of what the two periods
are: the inputs are themselves timegivers. The circuit
need only say whether the current input corresponds to
the higher or lower of the two frequencies it has received
in the past. One way of doing this would be to use en-
dogenous entrained oscillators, but other mechanisms are
ena, but are genetically specied.
In studies of the circadian rhythms of plants and in-
sects, the Q10 of the free-running period's alteration with
temperature has been found to be typically in the range
0.85{1.3, where something in the range 2{3 would be ex-
possible.
In applications where the inputs do not implicitly con-
tain appropriate time cues, they could be augmented by
an extra input which does. For example, an extra input
could be driven by an external crystal oscillator (cheap,
pected from consideration of the Arrhenius equation (see
also [11, pp23{27]). This is in the absence of environ-
mental timegivers, and appears to be the result of active
compensation, rather than some inherent insensitivity in
the mechanism. There is some evidence that the temper-
ature compensating mechanism may not be an inherent
feature of the oscillatory mechanism: certain Neurospora
2 The biology of circadian rhythms given here is taken from [10].
3 Sometimes referred to as zeitgebers , even in English texts.
accurate, and temperature stable). This is supercially
similar to the `clock' used by digital designers to globally
synchronise their circuits, with the subcircuits changing
in lock-step on the beating of the clock. However, there
is a fundamental dierence: now the external oscillator
is truly a timegiver, to be exploited by evolution in any
way, and is not an enforced constraint on the system's
dynamical behaviour. Evolution could totally ignore the
clock input if it chose, or it could be used as a subtle in-

uence on the circuit's dynamics. In this way, evolution
remains free to explore rich architectures and dynamics
beyond the scope of human design, but has appropriate
resources to evolve thermal stability if there is a selection
pressure for it.
I have suggested two mechanisms here (compensation
and interaction with external timegivers) which could
ectivity
allow the evolution of temperature-stable circuits. In
5.1.2 Heat Exchange
Internal temperature can be controlled by regulating the
amount of radiation, conduction and convection near the
physical interface between body and environment (the
integument). These can be altered by varying the area,
orientation with respect to the sun and wind, posture
with respect to the ground, colour, texture, re
and thermal resistance of the exposed surfaces, and the
each case, though, there needs to be a selection pressure
for evolution to favour temperature-stability. I propose
to arrange for this by evaluating each individual circuit
on several dierent FPGA chips in parallel, each being
held at a dierent temperature (using Peltier-eect heat-
uences the rate of heat transfer between body
pumps | see x5.1.2 | and hand-designed thermostatic
ow [13]), but
use of insulation (hair, feathers, and layers of supercial
tissue). Note that the integument does not necessarily
have to be at the same temperature as the core of the
body. Varying the integument's temperature not only di-
rectly in
and environment (Fourier's law of heat
control). An individual's tness score will then re
ect its also indirectly aects the amount of convection in the
ability to perform the desired behaviour in all of these
dierent conditions.
5 Thermal Regulation
Having dealt with ways of coping with temperature vari-
ations, I now wish to consider the complementary strat-
egy of stabilising internal temperature by some method
uttering of the mouth and throat
air surrounding the body [6, Chap. 33](which is also de-
pendent on posture).
The above applies to both the cooling and heating
of the body; for cooling only, the evaporation of liquid
(eg. water, saliva, sweat, or urine) from the body (skin,
mouth, respiratory system) or the adjacent environment
is extensively used by animals. `Forced air' cooling is
also used by panting,
of regulation. In order to do this, we rst look at the
means by which the internal temperature can be varied
with respect to the ambience: these mechanisms are the
ow of blood between parts of the body at
`actuators' of the thermal regulation control system. The
control system itself will then be considered.
ow between cold extremities and the core.
5.1 Temperature-altering Mechanisms
5.1.1 Behavioural
ow between them.
Faced with a sharp temperature change, many animals
make an immediate behavioural response (seeking a more
equitable environment). Placed in a temperature gra-
dient, most cold-blooded animals will move about and
gradually spend more time around a preferred temper-
ature, which is a function both of the animal's species
and of its thermal history. As temperature varies over
the course of the day, behaviours include selecting suit-
able micro-habitats, basking, seeking shade, burrowing
at thin slice of silicon of about 1cm2 area,
(which protects from both extremes of outside temper-
[18], or fanning with wings.
An important strategy in temperature regulation is to
control the
dierent temperatures; a typical example is the restric-
tion of blood
Some animals have a specialised arrangement of blood
vessels to perform countercurrent heat exchange, allow-
ing parts of the body to be at dierent temperatures even
when there is a considerable blood
Examples include maintenance of the core at a higher
temperature than the gills, legs, or tail; allowing partic-
ular organs to be held at a dierent temperature to the
core (eg. testes), or to have their temperature regulated
more precisely than the rest of the body (eg. brain).
The analogies with thermal management in present-
day electronics are strong. The design procedure is typi-
cally as follows. Firstly, the designer calculates the power
consumption ( ' heat generation) of the silicon chip,
which is a
stuck to the inside of a cavity inside a larger plastic or
ature), aggregation with conspecics [15]; appropriate ceramic package (which allows for the mounting of the
alternation between these can achieve a remarkably con- device on a circuit-board). For a particular packaging
stant body temperature. On a longer timescale, tem- method, a `thermal resistance from junction to ambient'
perature may alter the activity of endocrine organs and
become a part of the complex stimulus of migration. Sili-
con systems controlling an autonomous mobile robot [16]
might be at liberty to use such behaviours, but in general
there is no escape for electronics from its given environ-
ment. There also seems to be no electronic counterpart
for temperature-related activities such as nest-building,
or the rather unique strategies of Homo sapiens . .. such
as inventing air-conditioners [17].
is specied, where `junction' refers to the transistors on
the silicon. This allows the temperature of the silicon
to be calculated for a given ambient air temperature. If
the silicon would be too hot (the silicon being cold is
not normally a problem), then either a dierent type of
package is used or the heat transfer between the outside
of the package (or case) and the ambient must be im-
proved. The latter is done by mounting a heatsink onto
the case using a paste of high thermal conductivity. The
heatsink is usually a matt-black nned metal structure The heat generated in the brain is a signicant compo-
of large surface area, giving a very low thermal resistance nent of non-shivering thermogenesis, and a major part of
to the ambience. Using the sum of the junction-to-case this originates from neural activity; mostly arising from
thermal resistance (also specied for the package) and the metabolic processes needed to run the sodium-pump
the heatsink-to-ambient resistance, the silicon temper-
ature can again be calculated, and a suitable heatsink
selected.
Often forced-air cooling is used by mounting a fan on
the heatsink or nearby to the package (sometimes there is
just one fan blowing air through a large box of electron-
ics), which eectively decreases the thermal resistance
to the ambience [19]. A
uid other than air can also rate proportional to the speed of logic switching (there
be used (water or freons; the latter are electrical insu-
lators and can be allowed direct contact with the chip),
and this
uid can be refrigerated rather than at ambi- previous section.
ent temperature: such methods are currently only used
in rather exotic applications. Finally, it is possible to
mount a Peltier-eect heat-pump [20] in the form of a
wafer between the case and the heatsink. These electri-
cally powered devices are used to pump heat from the
case to the heatsink: the heatsink gets hotter with re-
spect to the surrounding
uid, accelerating its heat-loss, an interesting possibility | could a circuit evolve such
and the case becomes cooler. Note that these devices can
also operate in reverse to heat the case, and can be elec-
tronically controlled to maintain the silicon at a constant
temperature: both are currently unusual.
The analogies with nature are obvious. The main dif-
ference is that silicon chips are normally just kept be-
low a maximum temperature: we have seen that digi-
tal systems become slower with rising temperature. For
conventional digital design methodologies, faster opera-
tion at low temperatures is not a problem. Another con-
sideration is that the ageing mechanisms leading to de-
vice failure accelerate with temperature. Consequently,
to restore ion concentrations after ring [7, Chap. 5].
There is a strong analogy with electronics here. Al-
though there are research projects on reversible comput-
ing (using almost zero energy, thus generating almost no
heat), current technologies are far less ecient than this.
If used to perform digital operations, the CMOS FPGA
chips used in evolutionary electronics generate heat at a
is negligible heat generation when there is no activity in
the circuit). This heat must be dissipated, as seen in the
In the type of unconstrained evolutionary electronics
which motivates this paper, the circuit instantiated on
the FPGA is a continuous-time dynamical analogue sys-
tem, and is probably not `doing' logic. Nevertheless, it
is still true that the rate of heat generation rises with in-
creasing frequency of activity in the circuit. This raises
that the heat produced by its activity maintains it at
a preferred temperature? The circuit could potentially
even adapt its levels of activity in response to a change
in ambient temperature.
It turns out that there is not a known biological prece-
dent for this: in contrast to an increase in tissue respi-
ration of liver, heart, and skeletal muscle, the O2 con-
sumption of brain tissue remains constant during the de-
velopment of non-shivering thermogenesis [6, Chap. 5].
Of course, the lack of a biological counterpart need not
necessarily discount it in engineering systems. If there
existed parts of the circuit (possibly distributed through-
the amount of cooling is kept xed at that necessary to out a large system) which played no other role than to
guarantee a particular maximumtemperature, and is not be highly active and generate heat to maintain vital sub-
adaptive. A nal dierence is that evaporative cooling
| extremely eective in nature | would be inconvenient
(though theoretically possible) in electronic systems.
5.1.3 Heat Generation
In stabilising internal temperature, many animals (eg.
circuits at a preferred temperature, then these would be
analogous to brown fat in mammals. This is not an idle
speculation, but a serious proposal: such thermogenic
circuits could be built-in by hand, could be encouraged
to evolve, or could be searched for in the analysis of a
thermally stable evolved circuit.
5.2 Homeostatic Control Systems
insects, sh, birds, mammals) use the heat generated by
all metabolizing tissues (endothermy). In many cases,
this can be adaptively controlled (perhaps via the neu-
roendocrine system), not only by shivering or otherwise
activating muscles, but also non-shivering thermogene-
sis. Some mammals even have a specialised thermogenic
tissue (brown fat) for producing bursts of heat in re-
Dierent species use dierent combinations of the
temperature-altering mechanisms studied above, and ac-
cording to dierent adaptive strategies, to maintain a
relatively constant internal temperature. Species are
often categorised as hot-blooded (homeotherms), cold-
blooded (poikilotherms), endotherms, ectotherms, reg-
sponse to cold stress: it serves no other function. `The
tissue is strategically localized in the neck and thoracic
regions in relation to major blood vessels so that its heat
is quickly transported to those organs (brain and heart)
whose continuous high temperatures are vital... ' [1].
ulators, compensators, or conformers. These classica-
tions of species are sometimes used inconsistently in the
literature, and in ways potentially misleading for this
paper, so I have avoided them until now. `Homeotherm'
means mammals and birds. The others | cold-blooded
animals or poikilotherms | dier from homeotherms
in lacking the central autonomic thermal controls (the
hypothalamus in mammals, and the spinal cord in
birds), the continuously high body temperatures, and
the emphasis on thermoregulation as a balance between
metabolic heat and insulation (in the form of feathers or
fur). However, poikilotherms can use many of the mech-
anisms detailed above to maintain a remarkably constant
body temperature in their natural environments [21, 22].
Both homeotherms and poikilotherms have biological
stabilisation in the face of changing external tempera-
ture by homeostatic feed-back control mechanisms: in
homeotherms it is the temperatures of particular parts
of the body which are the controlled variables, whereas
in poikilotherms the controlled variable might be, for ex-
ample, metabolic rate [6, Chap. 5]. This does not neces-
sarily imply a reduced involvement of the nervous system
in the thermal regulation of poikilotherms. For example,
adaptive responses in the thermal resistance of tissue in
several poikilotherms have been found to be regulated
by photoperiod (not by heat), indicating involvement of
photoreceptors and the neuroendocrine system even in
this compensatory adaptation [1].
The hypothalamic thermal control in mammals is re-
exly activated by thermoreceptors of skin and mucous
membranes, and by temperature change in the hypotha-
lamus itself, or the blood circulating through it. Eer-
ent nerves control muscles (eg. for panting or shiver-
ing), and, via other organs, the cutaneous blood vessels,
the sweat glands, and the piloerector muscles (for raising
hairs); the response via the endocrine system can also be
on a much longer timescale, as we have seen. This does
not mean that these things are solely under hypothalamic
control, for example cutaneous vessels are also directly
in
uenced by temperature. A simple feedback controller is easy to construct using
Hypothalamic control, then, can be viewed as a dis-
continuous nonlinear negative-feedback control system,
with a limited range of operation, aiming to maintain
the temperature of various parts of the body at their
particular set-points (which can be altered, eg. during
exercise or by fever). Note that not all observations are
easily explained from this viewpoint. See [23] for de-
tails of various control-theoretic models. They model
through the primitive mammals (monotremes | platy-
puses and spiny ant-eaters | and marsupials) to the Eu-
theria (placental mammals). `Increasing complexity of
organization (especially behavioural organization) makes
homeothermy a necessity; or conversely, one may argue
that complexity, both physiological and behavioural, be-
comes progressively more feasible as the internal envi-
ronmental temperature, especially that of the nervous
system, is stabilized.' [1]
These observations are tremendously suggestive for
evolutionary electronics. Should we accept that the in-
clusion of a thermostatic control system is the price to
be paid for an ecient complex `electronic nervous sys-
tem' able to operate in a wide range of ambient temper-
atures? For thermostatic regulation is indeed costly: in
nature we even see animals prepared to give up activ-
ity altogether in order to abandon homeothermy when it
becomes too costly to maintain (eg. hibernation). The
cost of regulation depends on the degree to which con-
formity with the ambient temperature must be resisted:
various trade-os are found even within mammals [25].
In electronics, there would be the cost of the thermal
sensors, the feedback controller, the actuators, and the
energy required by these. Such a system is quite simple
to build, however, and the components are only expen-
sive in relation to the incredibly low cost of electronics
currently achieved through economies of scale.
Thermal sensing might be done by measuring time-
delays on the silicon (see x3), measuring the current
through a slightly forward-biased electrostatic-discharge
(ESD) protection diode (present around the edges of the
silicon), or by mounting a thermocouple on the outside
of the case. Use of both internal and surface thermosen-
sors can enable more sophisticated control strategies [26].
conventional electronic design principles, and a suitable
actuator would be a Peltier-eect heat-pump used as de-
scribed in x5.1.2 (or even the thermogenic subcircuits
proposed in x5.1.3). All of this could be provided with-
out having to try to evolve it, once one was convinced
that thermostatic control was appropriate.
The energy consumption of thermostatic regulation
could be prohibitive in many electronics applications. It
response close to the set-points: further away, less-nely is conceivable that in some situations the circuit could go
controlled emergency responses (such as massive shiver- into dormancy during periods of extreme ambient tem-
ing independent of skin temperature) come into play. perature when the costs of cooling/heating would out-
Mammals (derived from ancient poikilothermic rep-
tiles) are often considered the most highly developed an-
imals, breaking shackles of innate habits, and being able
to adapt behaviour to changing circumstances. They are
aided in both by the constancy of their internal environ-
ments [24]. Hoar notes that `. .. there is a good general
correlation between the precision of temperature con-
trol and the complexity of behavioural organisation.' [1]
There is a graded series from the poikilothermic reptiles
weigh the benets of continued operation. As in ani-
mals, some minimal capabilities would have to remain,
to `wake up' in an emergency or when conditions are
more comfortable, and to maintain functions that are
indispensable (the equivalent of cardiac and respiratory
rhythms in animals). It seems ludicrous to think of a
hibernating circuit; however, some small-bodied animals
(which can rapidly change body temperature) abandon
homeothermy while asleep at night, and resume it dur-
ing daytime. It is easy to think of applications where
a piece of electronics could be allowed to do the same.
Note that this idea can apply to both hot and cold ex-
tremes (in nature the summer equivalent of hibernation
is estivation, and less profound forms of torpor also ex-
ist). Shutting-down of electronic sub-systems when they
are not required is already widely used when power con-
It can be concluded that there is a wide range of bio-
logical phenomena within the scope of ALife which can
only be fully understood with reference to animals' ther-
mal responses.
7 Summary and Conclusion
sumption is crucial, and indeed dormancy is also used by
animals when food or water is scarce.
In an electronic system made of many parts, ther-
moregulation of the whole could emerge from the individ-
ual actions of the parts. There is a biological precedent
for this: `Colony activity resulting in some temperature
control is common at least among some insects: avoid-
ance and preference, shivering and quieting, fasting and
feeding, varying metabolic water production, clustering
and dispersal of swarms, nest placement and structure,
opening and closing nest entrances, fetching water into
the nest and fanning to ventilate and evaporatively cool
it. Appropriate ones of these mechanisms are eective in
Articial evolution, when freely allowed to manipulate a
physical silicon medium, does indeed share many of the
same problems that natural evolution faces in crafting
systems able to operate over a wide range of ambient
temperatures. A number of engineering proposals have
been constructed by analysing these correspondences.
In analogy to biochemical compensation, it was sug-
gested that an electronic system could be composed of
many subcircuits, and that there be alternatives for each
subcircuit. For a particular subcircuit, the alternatives
would operate over dierent ranges of temperature, and
would automatically come into play as appropriate. All
the alternative subcircuits need not reside on the sili-
maintaining quite narrow ranges.' [6, Chap. 37].
6 Implications for ALife Modelling
Part of the ALife enterprise is to use computer simula-
tions as a tool in theoretical biology [27]. Animal be-
haviour, evolution, and neuroethology are studied. As
is essential in any model, some of the complexity of the
real system is left behind, and it might be thought that
thermal considerations can safely be ignored as a `mere
implementation detail' with no impact on the essence of
behaviour and evolution. It seems clear from the biolog-
ical literature summarised here that this is not the case.
Behavioural mechanisms for thermal regulation are ubiq-
con simultaneously: the fast reconguration of an FPGA
could be used to `swap' them in and out, possibly with
the intervention of a high-level controller. Indeed, several
whole systems could be evolved for dierent temperature
ranges, and the system in its entirety swapped with an-
other in response to a temperature change (there is even
a biological counterpart for that). It was also seen how
a system can stabilise its behavioural timescales by in-
teraction with external timegivers, either implicit in the
input, or explicitly supplied.
Temperature-altering mechanisms were discussed.
The behavioural responses heavily relied upon by many
animals are to a large extent not applicable in electronic
applications other than autonomous mobile robotics.
uitous and can be a signicant component of an animal's
lifestyle. This is especially the case when water economy,
not considered above, is taken into account: to use evap-
orative cooling, one must have drunk suciently or be
adjacent to an external liquid source.
We have even seen that thermal considerations can
have an in
uence on the evolution of body size: small an- there seems to be no precedent for the case where sub-
imals have a greater surface-area to volume ratio, which
aects heat transfer with the environment [6, Chap. 32 ].
They also have smaller reserves of food, water, or energy
(important for dormancy), but can easily seek thermal
shelter by behaviours like burrowing [28]. Small animals
can change their body temperature more rapidly than
larger ones, so can suspend or resume thermoregulation
relatively rapidly.
Thermal considerations can even in
Several of the most important heat-exchange mecha-
nisms found in nature are already in use in the elec-
tronics industry, with the exception of evaporative cool-
ing, which seems inappropriate. The use of circuit ac-
tivity adaptively to generate heat in analogy with ther-
mogenic tissues in animals is highly promising, though
circuits with `nervous' function also adapt their activity
for thermal considerations.
The most radical suggestion to come from the bio-
logical literature was that precise homeostatic control
of internal temperature is almost essential for a com-
plex behavioural and physiological organisation. It is
straightforward to arrange for this in electronic systems,
but in the current climate of extreme low cost and low
uence macroevo- power consumption, the benets will need to be great for
lution. Recently, physiological and morphological adap- commercial acceptance. Dormancy, and the use of collec-
tations to climatic conditions in the house sparrow tive regulation behaviours, are possibilities for reducing
(Passer domesticus) in N. America have led to the dier- the power consumption of a homeostatically thermoreg-
entiation of new sub-species living at distinct localities ulated electronic system.
[15]. Such population dynamics cannot be understood In conclusion, biology provides a wealth of new ideas
without considering animals' responses to temperature. for combining the new possibilities of extreme eciency
through the unconstrained exploitation of physical re- [15] H. Pohl. Thermal adaptation in the whole animal.
sources with an adequate thermal stability. Some of In J. Bligh, J. L. Cloudsley-Thompson, and A. G.
these are suitable for immediate inclusion into research Macdonald, eds, Environmental Physiology of An-
programmes in evolutionary electronics. The fact that imals, chapter 13, pp261{286. Blackwell Scientic
thermal considerations cannot be neglected as an incon- Publications, 1976.
sequential implementation detail is also worthy of note [16] A. Thompson. Evolving electronic robot controllers
to ALife researchers interested in animal behaviour, evo- that exploit hardware resources. In F. Moran
lution, and neuroethology. et al., eds, Advances in Articial Life: Proc. 3rd
Acknowledgements: This work is supported by Xilinx Inc.,
and the Centre for Computational Neuroscience & Robotics: many
thanks to each. Thanks also to John Gray, Phil Husbands, Dave
Cli, Inman Harvey, and everyone.
References
[1] W. S. Hoar. General and Comparative Physiology.
Prentice-Hall, 3rd edition, 1983.
[2] A. Thompson. Silicon evolution. In J. R. Koza et al.,
eds, Genetic Programming 1996: Proc. 1st Annual
Conf. (GP96), pp444{452. MIT Press, 1996.
Eur. Conf. on Articial Life (ECAL95), vol. 929 of
LNAI, pp640{656. Springer-Verlag, 1995.
[17] T. H. Benzinger. The thermal homeostasis of man.
In L. L. Langley, ed, Homeostasis: Origins of the
Concept, Benchmark papers in Human Physiology,
chapter 21, pp295{327. Dowden, Hutchison & Ross,
1973.
[18] G. A. Bartholomew, J. W. Hudson, and T. R. How-
ell. Body temperature, oxygen consumption, evapo-
rative water loss, and heart rate in the Poor-will. In
R. Ruibal, ed, The Adaptation of Organisms, Dick-
[3] Xilinx, Inc. XC6200 Advanced product specication
V1.0, June 1996. In The Programmable Logic Data
Book. 1996. See http://www.xilinx.com.
[4] A. Thompson. An evolved circuit, intrinsic in sil-
icon, entwined with physics. In Higuchi & Iwata,
eds, Proc. 1st Int. Conf. on Evolvable Systems
(ICES'96), LNCS. Springer-Verlag, 1996. In press.
[5] W. B. Yapp. An Introduction to Animal Physiology.
Oxford University Press, 1970.
[6] American Physiological Society. Handbook of Phys-
iology. Section 4: Adaptation to the Environment.
Waverly Press, 1964.
[7] American Physiological Society. Handbook of Phys-
iology. Section 1: The Nervous System, Volume I:
Cellular Biology of Neurons, Part 1. Waverly Press,
1977.
[8] C. A. Mead. Analog VLSI and Neural Systems. Ad-
dison Wesley, 1989.
[9] E. M. Pantelouris. Introduction to Animal Physiol-
ogy and Physiological Genetics, vol. 32 of Int. Series
of monographs on pure and applied Biology. Divi-
sion: zoology. Permagon Press, 1967.
[10] M. C. Moore-Ede, F. M. Sulzman, and C. A. Fuller.
The Clocks That Time Us: Physiology of the Cir-
cadian Timing System. Harvard University Press,
1982.
[11] J. E. Harker. The Physiology of Diurnal Rhythms.
No. 13 in Cambridge monographs in experimental
Biology. Cambridge University Press, 1964.
enson series on contemporary thought in Biological
Science, pp15{25. Dickenson Publishing Company,
1967.
[19] R. W. Keyes. The Physics of VLSI Systems.
Addison-Wesley, 1987.
[20] R. S. Components Ltd., PO Box 99, Corby,
Northants NN17 9RS, UK. Peltier eect heat
pumps, March 1988. Datasheet 7562.
[21] V. A. Pegel and V. A. Remorov. The heat-regulating
mechanism and its development in poikilothermic
animals. In J. W. S. Pringle, ed, Essays on Physio-
logical Evolution. Permagon Press, 1965.
[22] D. W. Wood. Principles of Animal Physiology. Ed-
ward Arnold Ltd., 1983.
[23] S. Armstrong Talbot and Urs Gessner. Systems
Physiology. John Wiley and Sons, 1973.
[24] Collins encyclopedia of animals. William Collins
Sons, 1975.
[25] K. Schmidt-Nielsen, B. Schmidt-Nielsen, S. A. Jar-
num, et al. Body temperature of the camel and its
relation to water economy. In R. Ruibal, ed, The
Adaptation of Organisms, Dickenson series on con-
temporary thought in Biological Science, pp26{42.
Dickenson Publishing Company, Inc., 1967.
[26] G. Strom. Central nervous regulation of body tem-
perature. In Handbook of Physiology, Section 1:
Neurophysiology, Vol. II, chapter XLVI, pp1173{
1196. Waverly Press, 1960. American Physiology
Society.
[12] C. Ladd Prosser, ed. Neural and Integrative Animal
Physiology. Wiley-Liss, 1991.
[13] G. E. Folk Jr. Textbook of Environmental Physiol-
ogy. Lea & Febiger, 2nd edition, 1974.
[14] H. A. Lloyd. Mechanical timekeepers. In A His-
tory of Technology, vol. III, chapter 24, pp648{675.
Oxford University Press, 1957.
[27] G. E. Miller. Articial life as theoretical biology:
how to do real science with computer simulation.
CSRP 378, School of Cognitive and Computing Sci-
ences, University of Sussex, 1995.
[28] M. S. Gordon, G. A. Bartholomew, A. D. Grinnell,
et al. Animal Physiology: Principles and adapta-
tions. Macmillan Publishing Co., 1982.