arti cial evolution in the physical world

25

Upload: haduong

Post on 22-Jan-2017

222 views

Category:

Documents


2 download

TRANSCRIPT

Page 1: Arti cial Evolution in the Physical World

Evolutionary Robotics: From Intelligent Robots to Arti�cial Life (ER'97),T.Gomi (Ed.), pp101{125. AAI Books, 1997.Arti�cial Evolution in the PhysicalWorldADRIAN THOMPSONCCNR, COGSUniversity of SussexBrighton BN1 9QH, UKTel: +44 1273 678754, Fax: +44 1273 671320Email: [email protected] IntroductionWhen arti�cial evolution is used to automatically design a structure, thatstructure usually exists in a software simulation, to make it easily manipula-ble. When evolving control systems for autonomous mobile robots [1], evenwhen the real robot is used instead of a simulation, the actual structure un-dergoing evolution | often an arti�cial neural network (ANN) | is usuallysimulated in software.Recently, however, technology has become available which allows arti�cialevolution to manipulate the con�guration of a silicon chip directly: electroniccircuits can be evolved without the use of simulation, with every �tness mea-surement being the evaluation of a physically real electronic circuit's perfor-mance at the desired task. But why should one be interested in this? Afterall, we can easily simulate ANNs on a standard desktop PC that are largerthan the current capabilities of arti�cial evolution, so we do not need to re-sort to hardware implementations because of software being too slow (pacede Garis [2]). The answer is that evolution of recon�gurable hardware neednot be just a high speed implementation of what could easily be done insoftware: evolution is crafting a physical object that exists in real time and

Page 2: Arti cial Evolution in the Physical World

space, and behaves according to semiconductor physics. This raises a set ofopportunities for science and engineering that are not normally addressed bysimulation work:1. Evolution can exploit real-world physics that is di�cult to analyse ormodel in simulation or theoretical studies. Once the simplifying con-straints of conventional design methodologies have been dropped, thiscan allow highly e�cient circuits to be evolved, which exploit the naturalbehaviour of the electronic medium.2. The physical components have a size, shape and location, and these arecrucial in determining the interactions between them. This can make theinteractions richer, but in some ways more constrained, than the per-fectly controllable point-to-point topological interconnections normallyused when evolving in simulation.3. The characteristics of the components and their interactions are notexactly predictable or constant over time. Evolution must �nd ways ofcoping with this.The �rst point above provides the engineering motivation: extremely ef-�cient (small, low-power) circuits can be produced. The penalty for theengineer is that to do this, the second two points must also be considered.For the scientist, all three are of great interest, as they apply as much toevolution in nature | and attempts to draw inspiration from it | as toelectronics. As we shall see, they have implications for the organisation ofa physical `nervous system', whether it be natural or arti�cial. This papersummarises some results from the author's work on the evolutionary engi-neering of electronics in general, with the intention of showing its relevanceto the Evolutionary Robotics (ER) enterprise.In the next section, I describe the technology making the direct evolutionof electronics possible. The later sections then consider the three points abovein turn, showing experimental results. Only an overview is given | see thereferences for full details. Finally, the implications for ER are summarised.

Page 3: Arti cial Evolution in the Physical World

2 Technology: Evolvable HardwareA Field-Programmable Gate Array (FPGA) is a Very Large Scale Integra-tion (VLSI) silicon chip containing a large array of components and wires.Switches distributed throughout the chip determine how each component be-haves, and how they connect to the wires. By con�guring these switches, anFPGA's behaviour is determined by the user `in the Field' rather than at thechip factory. In RAM-based FPGAs, the switches are electronic, and havetheir settings determined by bits of memory onboard the chip. The XilinxXC6200 [21] is the �rst such device ideally suited to evolutionary work [19],and a simpli�ed view of it is given in Figure 1. It can be interfaced to a hostcomputer so that its con�guration memory can be written to by softwarejust like normal computer memory. An Evolutionary Algorithm (EA) run-ning on the computer can write to the FPGA's con�guration memory, settingthe electronic switches, and thus creating a physically real electronic circuit.This circuit can be evaluated according to its real-world performance at atask, and successively modi�ed by the EA of choice (eg. a genetic algorithm(GA) [7, 4, 5], evolutionary programming [3], evolution strategies [13] or ge-netic programming [8, 9]) until satisfactory performance is achieved. Figure 2depicts the operation of a simple GA applied in this way.Note that in this evolutionary process of automatic circuit design, thereis no simulation, modelling, or analysis of the circuit. The FPGA is notprogrammed to follow a sequence of instructions, it is con�gured and thenallowed to behave in real-time according to semiconductor physics: evolutionis manipulating a physical medium.The electronic equipment needed is not necessarily bulky or complicated.Figure 3 shows the tiny `Khepera' robot (a common tool in ER [11]) equippedwith an XC6216 FPGA onboard. With the FPGA controlling the real robot,behaving in the real world, a simple wall-avoiding behaviour has been evolved.Evolution was by a GA running on a PC connected to the robot by a serialcable, but there was no simulation of the robot or the control circuits. Thissimple demonstration shows that evolution of FPGA circuits is not necessarilyany more complicated or di�cult than evolving a software structure such asa simulated ANN.

Page 4: Arti cial Evolution in the Physical World

E

F

S

ES

EN

N

FS

W

N

F

W

W F

S EN W

E

NS

W

SN

EW

F

N

E

S

W

Figure 1: A simpli�ed view of the XC6216 FPGA. Only those features used later in theexperiments are shown. Top: A 10 � 10 corner of the 64 � 64 array of blocks; Below:the internals of an individual cell, showing the function unit at its centre. The symbolrepresents a multiplexer | which of its four inputs is connected to the output (viaan inversion) is controlled by the con�guration memory. Similar multiplexers are used toimplement the user-con�gurable function F.

Page 5: Arti cial Evolution in the Physical World

001 1 0 0 0 1 0

0 1 110 0 010

0 1 10 0 0 110

00 1 0

1

1 1 0 00

1 1 0 00 0 1 11

A newpopulation isformed, madeof the offspring

(on average)of the fitter

members of theold one.

A population of

bit-string

maintained,each individual coding for apossible FPGAconfiguration.

genotypes is

how well it performs

4.851

9.001

0.000

3.942

0.030

1 01 10

1 0 1 1 1 0

0 1 1 1 0 1 1

1 0 1

1 0 0

0 0

0

Population Next GenerationScores

REPEAT UNTIL SATISFACTORY

Fitness

Higher scoring individuals are more likely

randomly inverting a few bits (mutation).from each parent (crossover), and by

(initially random)

formed by stochastically combining segmentsto parent offspring (selection). Offspring are

Each individual istaken in turn andused to configurea real FPGA, whichis then scored at

the desired task.

Fitness Evaluation:

Figure2:EvolvinganFPGAcon�gurationusingasimplegeneticalgorithm.

Page 6: Arti cial Evolution in the Physical World

Figure 3: The miniature Khepera robot. The top two layers are an FPGA extension turretallowing onboard evolution of electronic control systems. They were designed by the author,and constructed in collaboration with the Xilinx Development Corp.

Page 7: Arti cial Evolution in the Physical World

3 Evolutionary Exploitation of a Physical MediumThe process of direct hardware evolution just described works by taking ac-count of changes in the real-world performance of a circuit as variations aremade to its structure. This is very di�erent from the design methods fol-lowed by humans: these always take place at a more abstract level, so thatthe designer does not have to consider the detailed behaviour of every compo-nent and their interactions. Figure 4 gives a sketch of this crucial di�erencebetween conventional design and evolution.Design, analysis, or simulation of any but the smallest circuits is infeasibleunless some of the details of the semiconductor physics are `abstracted away'to form a simpler model. If such designs, analyses, or simulations are to saysomething useful about the behaviour of the real hardware, the circuits un-der consideration must be constrained: the details that have been suppressedin forming the abstract model must not be allowed to in uence the overallbehaviour of the system at the level of description of interest. This meansthat circuits that can be designed (by humans), analysed, or simulated, cannot put to use all of the natural behaviour of the silicon medium: some ofit must be discarded for the sake of simplicity of modelling. The standardways of doing this are embedded into all design methodologies | the waythe system is broken down (perhaps hierarchically) into parts, and the inter-actions between these parts restricted so that their collective behaviour canbe readily understood from a knowledge of their individual properties.Evolution needs none of this (at least, not for the same reasons). Thereis no analysis, simulation, or modelling, so no constraints need to be placedon the circuits to facilitate these. Evolution proceeds by taking account ofthe changes in the overall behaviour as variations (usually small) are madeto the circuit's structure: this means that the collective behaviour of thecomponents can be freely exploited without having to be able to predict itfrom a knowledge of their individual properties. Evolution can be set freeto exploit the rich structures and dynamical behaviours that are natural tothe silicon medium, exploring beyond the scope of conventional design. Thedetailed properties of the components and their interactions can be used incomposing this system-level behaviour. It takes considerable imagination toenvisage what these evolved circuits could be like: the kinds of systems we

Page 8: Arti cial Evolution in the Physical World

a b = a + b ?

Design by humans: Design through evolution:

proceeds by taking account of theoverall behavioural effect of variations made to the structure.

takes place at an abstract level.

a a = 0

Figure4:Acaricaturecomparisonofthedi�erencebetweendesignbyconventionalmethods,andthrougharti�cialevolution.

Page 9: Arti cial Evolution in the Physical World

are familiar with (eg. digital, discrete-time, computational, or hierarchicallydecomposed circuits) are but a subset of what is possible. (See [20] for thefull details of this argument.)As an example of an application of these ideas in the �eld of ER, con-sider the robot shown in Figure 5. This two-wheeled autonomous mobilerobot has a diameter of 46cm, a height of 63cm, and was required to dis-play simple wall-avoiding/room-centering behaviour in an empty 2.9m�4.2mrectangular arena. For this scenario, the d.c. motors were not allowed to runin reverse and the robot's only sensors were a pair of time-of- ight sonarsrigidly mounted on the robot, one pointing left and the other right. Thesonars �re simultaneously �ve times a second; when a sonar �res, its outputchanges from logic 0 to logic 1 and stays there until the �rst echo is sensedat its transducer, at which time its output returns to 0.This experiment was the �rst work designed to explore the possibilities ofdirectly evolving real hardware [14], and at that time suitable FPGAs werenot available. For this reason, an evolvable hardware architecture dubbed the`Dynamic State Machine' (DSM) was developed, to be built out of severalreadily available chips assembled onto a circuit-board. It is based upon astandard electronic implementation of a �nite-state machine (a common sim-ple computational architecture) using a RAM memory chip to hold a look-uptable de�ning the machine's behaviour. The contents of this RAM chip wereplaced under evolutionary control. Conventionally, the dynamics of the sys-tem would be given by a `clock' which causes the machine to change from onestate to the next at regular intervals (`clock ticks'), in a way easily describedby Boolean (binary) logic. This is an example of a constraint introduced onthe circuit's dynamics in order to allow it to be modelled in an abstract frame-work (in this case, Boolean logic). Thus, for the evolutionary experiment, theclocking constraint was removed. It was placed under evolutionary controlwhether each signal in the circuit was allowed to run freely in continuoustime, or whether it would be synchronised to the clock in the usual way. Forthe clocked signals, the clock frequency itself was placed under evolutionarycontrol. The clock, which used to be a constraint on the system's dynamics| forcing it to behave synchronously in discrete time | has been turned intoa resource which can be used to further enrich the continuous-time dynamicsof the circuit.

Page 10: Arti cial Evolution in the Physical World

RotationSensors

Wheels

Sonars

EvolvableHardware

SonarEmulator

VirtualWorldSimulator

Figure 5: The robot known as \Mr Chips."

Page 11: Arti cial Evolution in the Physical World

Figure 6 represents the resulting evolvable DSM circuit as a mixed syn-chronous/asynchronous recurrent logic network, where the two logic functionsare implemented by the RAM chip, and are thus under evolutionary control.The `genetic latches' in the �gure allow evolution to determine independentlywhether each signal is synchronised to the clock of evolved frequency, orwhether it is free-running in continuous time. Relaxing the dynamical con-straints on the circuit has so enriched its capabilities that the sonar echosignals are directly connected to its inputs, and its outputs directly drivethe power stages for the motors: normally, pre- and post-processing of thesensorimotor signals would be required.This control system was evolved as a piece of real hardware, with thephysical circuit controlling the real motors for all �tness evaluations. Forconvenience, during evolution the sonar input waveforms were emulated inreal-time on the basis of a `Virtual Reality' simulation of the robot's sensoryenvironment, based on velocity measurements taken from the wheels (whichwere just spinning in the air). Figure 7 shows the behaviour induced in therobot by the �nal evolved hardware controller: the long-exposure photographshows the excellent performance in the real world when the real sonars wereconnected and the robot placed in the arena.Remarkably, the �nal evolved control system goes directly from sonar echosignals to pulses sent to the motors, using only 32 bits of RAM and three ip- ops (excluding clock generation). This is a truly miniscule amount ofelectronics to comprise the entire sensorimotor control structure for this ro-bust behaviour, which is able to cope with the highly misleading multiplere ections which are often picked-up by the sonars. Analysis showed thatthe circuit had very rich dynamics, exploiting a stochastic interplay betweencontinuous-time and discrete-time signals. It is not a �nite-state machine,and could not have been designed by conventional methods because the de-tailed analogue continuous-time properties of the hardware (such as time-delays and metastability constants) are important to its operation: it cannotbe modelled by Boolean logic. Control experiments showed that the stan-dard synchronous �nite-state machine could not perform this task, so we canconclude that evolution really has been able to explore a richer repertoire ofbehaviours arising from the same circuitry once the simplifying constraintsnecessary for designers have been removed. See [14] and [20] for full details.

Page 12: Arti cial Evolution in the Physical World

M M

FUNCTIONLOGIC

FUNCTIONLOGIC

RIGHTLEFT

SONARS

MOTORS

LEFT RIGHT

Figure 6: A representation of the evolvable Dynamic State Machine, as used in the experi-ment. Each is a `Genetic Latch' (see text).

Page 13: Arti cial Evolution in the Physical World

Figure 7: Room-centering in virtual reality and (bottom right) in the real world, after 35generations. The top pictures are of 90 seconds of behaviour, the bottom ones of 60.

Page 14: Arti cial Evolution in the Physical World

4 Components Interacting in Physical SpaceFor our next example, consider evolving the 10 � 10 array of FPGA cellsshown in Figure 1. Again, the task is to be a simple | but non-trivial |one, formulated to explore fundamental issues. The circuit is to have a singleinput, and a single output. The input will be a square-wave audio-tone ofeither 1kHz or 10kHz, and circuit is to discriminate between them. Ideally,the output should go to a steady +5V as soon as one of the frequencies ispresent, and 0V for the other one. The task was intended as a �rst stepinto the domains of pattern recognition and signal processing, rather thanbeing an application in itself. One could imagine, however, such a circuitbeing used to demodulate frequency-modulated binary data received over atelephone line.This FPGA is intended to perform digital logic, so would normally beused with a synchronising clock, as discussed in the previous section. Thatwould make the frequency discrimination task quite straightforward: theclock could be used to time the input period. In this experiment, however,there will be no clock | can evolution exploit the rich natural unconstraineddynamics of the silicon to achieve the task? This seems almost too much toask: all that is available is 100 FPGA cells, each intended to perform a singleBoolean logic function, and each having a delay from input to output of justa few nanoseconds (billionths of a second). How could an arbitrary structure(potentially having many recurrent | feedback | connections) of these 100simple high-speed logic gates be evolved to discriminate perfectly betweeninput periods �ve orders of magnitude longer than the delay through eachcomponent? Success would be signi�cant: as well as vindicating the `un-constrained' approach to hardware evolution, the resulting circuit (requiringno external components or clock) would be incredibly e�cient in its use ofsilicon.The experimental arrangement is shown in Figure 8. A genetic algorithmruns on a standard PC, and con�gures the real FPGA for each �tness eval-uation. The XC6216 FPGA has 64 � 64 cells, so only a 10 � 10 corner wasused. For each individual circuit, a sequence of test tones (of 1kHz and10kHz) were applied to the pin designated as the input, and the signal atthe pin chosen to be the output was monitored. The �tness function was to

Page 15: Arti cial Evolution in the Physical World

configurationPCDesktop

XC6216 FPGA

Output

Tonegenerator

Analogueintegrator(to oscilloscope)

Figure 8: The arrangement for the tone discriminator experiment. The 10 � 10 corner ofcells used is shown to scale with respect to the whole FPGA. The single input to the circuitwas applied as the east-going input to a particular cell on the west edge, as shown. Thesingle output was designated to be the north-going output of a particular cell on the northedge.maximise the di�erence in the average output voltage between the case whenthe 1kHz input was present, and the case when the 10kHz was present (see[19, 17, 6] for full details). This average output voltage was measured by theanalogue integrator shown in the �gure: the circuit must be evaluated as acontinuous-time analogue system, now we have abandoned all of the digitaldesign principles with which the FPGA was intended to be used. A photo-graph of the circuit-board carrying the FPGA and the circuitry used as partof the �tness measurement is shown in Figure 9: it plugs directly into thePC, and is simple and easily built.Throughout the experiment, an oscilloscope was directly attached to theoutput pin of the FPGA (see Figure 8), so that the behaviour of the evolvingcircuits could be visually inspected. Figure 10 shows photographs of theoscilloscope screen, illustrating the improving behaviour of the best individualin the population at various times over the course of evolution.The individual in the initial random population of 50 that happened to getthe highest score produced a constant +5V output at all times, irrespectiveof the input. It received a �tness of slightly above zero just because of noise.Thus, there was no individual in the initial population that demonstrated

Page 16: Arti cial Evolution in the Physical World

Figure 9: The circuitry to evolve the tone discriminator.any ability whatsoever to perform the task.After 220 generations, the best circuit was basically copying the input tothe output. However, on what would have been the high part of the squarewave, a high frequency component was also present, visible as a blurredthickening of the line in the photograph. This high-frequency componentexceeds the maximum rate at which the FPGA can make logic transitions, sothe output makes small oscillations about a voltage slightly below the normallogic-high output voltage for the high part of the square wave. After another100 generations, the behaviour was much the same, with the addition ofoccasional glitches to 0V when the output would otherwise have been high.Once 650 generations had elapsed, de�nite progress had been made. Forthe 1kHz input, the output stayed high (with a small component of the inputwave still present) only occasionally pulsing to a low voltage. For the 10kHzinput, the input was still basically being copied to the output. By generation1100, this behaviour had been re�ned, so that the output stayed almostperfectly at +5V only when the 1kHz input was present.By generation 1400, the neat behaviour for the 1kHz input had been aban-doned, but now the output was mostly high for the 1kHz input, and mostlylow for the 10kHz input. . . with very strange looking waveforms. This be-haviour was then gradually improved. Notice the waveforms at generation2550 | they would seem utterly absurd to a digital designer. Even thoughthis is a digital FPGA, and we are evolving a recurrent network of logic gates,the gates are not being used to `do' logic. Logic gates are in fact high-gain ar-rangements of a few transistors, so that the transistors are usually saturated

Page 17: Arti cial Evolution in the Physical World

3500

2800

2550

1400

2100

1100

650

320

220

0IN

10kHz1kHz

Figure 10: Photographs of the oscilloscope screen. Top: the 1kHz and 10kHz input wave-forms. Below: the corresponding output of the best individual in the population after thenumber of generations marked down the side.

Page 18: Arti cial Evolution in the Physical World

| corresponding to logic 0 and 1. Evolution does not `know' that this wasthe intention of the designers of the FPGA, so just uses whatever behaviourthese high-gain groups of transistors happen to exhibit when connected inarbitrary ways (many of which a digital designer must avoid in order to makedigital logic a valid model of the system's behaviour). This is not a digitalsystem, but a continuous-time, continuous valued dynamical system madefrom a recurrent arrangement of high-gain groups of transistors | hence theunusual waveforms.By generation 2800, the only defect in the behaviour was rapid glitchingpresent on the output for the 10kHz input. Here, the output polarity haschanged over: it is now low for the 1kHz input and high for 10kHz. Fitnesseswere measured such that this swap would have no e�ect; in general it is agood idea to allow evolution to solve the problem in as many ways as possible| the more solutions there are, the easier they are to �nd.In the �nal photograph at generation 3500, we see the perfect desiredbehaviour. In fact, there were infrequent unwanted spikes in the output (notvisible in the photograph); these were �nally eliminated at around generation4100. The GA was run for a further 1000 generations without any observablechange in the behaviour of the best individual. The �nal circuit (which Iwill arbitrarily take to be the best individual of generation 5000) appears tobe perfect when observed by eye on the oscilloscope. If the input is changedfrom 1kHz to 10kHz (or vice-versa), then the output changes cleanly betweena steady +5V and a steady 0V without any perceptible delay.The �nal circuit is shown in Figure 11; observe the many feedback paths.No constraining preconceptions were imposed on the circuit, so evolution wasgiven the freedom to explore the full space of possible designs.

Page 19: Arti cial Evolution in the Physical World

In

Out

Figure 11: The �nal evolved circuit. The 10 � 10 array of cells is shown, along with allconnections that eventually connect an output to an input. Connections driven by a cell'sfunction output are represented by arrows originating from the cell boundary. Connectionsinto a cell which are selected as inputs to its function unit have a small square drawn onthem. The actual setting of each function unit is not indicated in this diagram.In

Out

Figure 12: The functional part of the circuit. Cells not drawn here can be clamped toconstant values without a�ecting the circuit's behaviour.

Page 20: Arti cial Evolution in the Physical World

By empirical testing, it was possible to determine which parts of the 10�10array were actually contributing to the behaviour. Figure 12 shows thisfunctional part of the circuit. Observe the cells shaded gray: they do in uencethe system's behaviour (if an attempt is made to clamp and one of them to aconstant value, then the system malfunctions), but yet they are not connectedto the main part of the circuit, and there seems to be no route of connectionsby which they could ever in uence the output pin! These components mustbe interacting with the others by some subtle unconventional means (such aselectromagnetic coupling or power-supply loading) which has been put to useby evolution in composing the overall system behaviour.By releasing the full repertoire of behaviours that the recon�gurable elec-tronic medium can manifest, evolution has been able to craft a highly e�cientcomplex dynamical system. Conventional design would require 1{2 orders ofmagnitude more silicon area to achieve the same performance with no exter-nal components or clock, and even then it would be di�cult. But we havenow stepped even further away from being able to understand the system interms of familiar models. Not only do we have the rich analogue continuous-time dynamics seen in the previous section, but now the interactions betweenthe components cannot completely be described by merely listing the wiresconnecting them. The functioning of the `gray cells' above shows that theinteractions of the components are not solely determined by the connectingwires, but also by their positions in physical space. In particular, the spatialproximity of the components is likely to be important.In general, the size, shape, and location of the components will be impor-tant, as well as the point-to-point connections (wires) between them. Theseextra means of interaction are in some ways a resource to be used, but tothe extent that they are unavoidable, they could also be viewed as a con-straint. These issues are crucial to understanding the evolution of physical`nervous systems', whether biological or electronic, which must necessarilyexist in three-dimensional space. A particularly interesting class of spatialinteractions in biology is di�use neural messengers: although it is possible toincorporate these into an ANN model [10], the very phrase `neural network'betrays the extent to which it is often assumed that a topological network ofpoint-to-point interconnections (of perfectly controllable strength) capturesall of the important aspects of neural interaction.

Page 21: Arti cial Evolution in the Physical World

I suggest that evolvable hardware, by providing a physical medium inwhich arti�cial `nervous systems' can be evolved, may provide a tool withwhich the evolution of natural nervous systems | and the engineering inspi-ration that can be drawn from them | may be investigated. Conversely, itis de�nitely the case that neuroscience is relevant to hardware evolution.5 Coping with Variations in Components and Interac-tionsUp until now in this paper, I have been guilty of a small deception. I havespoken as if the only reason that human designers work within a constrainedspace of circuits | for instance, synchronous digital logic circuits | is tomake the design process simple. In fact, precluding the detailed propertiesof the medium from contributing to the system behaviour not only supportsdesign abstractions, but also gives robustness. Those properties that havebeen excluded from the designer's model (and prevented from in uencingthe system's behaviour by means of constraining its structure and dynamics)can vary greatly without causing the system to malfunction. Such changestypically arise from process variations between nominally identical siliconchips, ageing, and temperature and power-supply uctuations.Once evolution is allowed to explore the full spectrum of possible be-haviours that the medium can support, this `automatic' robustness is lost. Atrade-o� needs to be found between exploiting the properties of the medium,and being tolerant to variations in them.1 Tolerance to variation in a prop-erty does not necessarily imply that the property is not used at all: severalaspects of the medium which vary in di�erent ways can be balanced againsteach-other to give stable overall system behaviour, or di�erent mechanismscan be called into play for di�erent conditions.As in the case of spatial interactions, the evolution of adequate robust-ness is important for all physical `nervous systems', whether electronic orbiological. Nature will be a rich source of inspiration for techniques in theevolution of robustness. Rather than precluding large swathes of the naturalbehaviour of the medium from ever being put to use (as does conventional1Note that a fault is an extreme form of variation, and there are several evolutionary mechanisms bywhich tolerance can be achieved [15, 16, 18].

Page 22: Arti cial Evolution in the Physical World

design), the natural approach in an evolutionary framework is to provide aselection pressure for robustness, and to allow evolution to build robustnessinto the overall system behaviour using the full set of resources available.This selection pressure may be provided by evaluating the circuits in thepresence of those variations with which they are required to cope, so that tobe �t they must operate well under a wide set of conditions.An especially promising idea from biology is the notion of an external`timegiver' which can stabilise the timescales of the system's internal dynam-ics [12]. This timegiver could be inherent in the system's ongoing interactionwith the environment: for example, in the tone decoder example of the previ-ous section, the fact that the input waveform is always either 1kHz or 10kHzcould be used as a time-reference. In addition, the circuit could interactwith an external timegiver more explicitly. In analogy to the daily light/darkcycles that entrain circadian rhythms in animals, a stable oscillation couldbe applied to the evolving circuits as an extra input, at the same time asa selection pressure towards robustness is maintained. This `clock' is not aconstraint on the system's dynamics | evolution could choose to ignore italtogether | but instead enriches the spectrum of possible dynamical be-haviours with stability, which can be incorporated in subtle ways.Preliminary experiments on this `unconstrained' approach to robustnessfor the tone-discrimination task are encouraging, but not yet conclusive.6 Summary: Implications for ERWe have seen how evolution, when manipulating a real physical electronicmedium, can exploit it with orders of magnitude more e�ciency than con-ventional design. This is possible because evolution can utilise the emergentbehaviour of a collection of components without having to be able to pre-dict or analyse it: the simplifying constraints of traditional methods can beremoved, releasing the full capabilities of the physical hardware.Evolvable hardware provides the �rst opportunity for the evolution ofsynthetic physical `nervous systems.' The components of a physical system|whether it be biological or electronic | have a size, shape, and location, andthese are important. Robustness cannot be taken for granted. These issuesneed to be faced in order to reap the full engineering bene�ts of unconstrained

Page 23: Arti cial Evolution in the Physical World

hardware evolution: the potentially small, low-power, fault-tolerant circuitsproduced have obvious applications in Evolutionary Robotics and elsewhere.Unconstrained direct hardware evolution may thus also be a useful routeto greater realism in ER models aiming to address questions in biology orwishing to take inspiration from it.AcknowledgementsGratitude to the School of Cognitive & Computing Sciences, and to Xilinx,Inc. for funding this work. Personal thanks to Phil Husbands, Dave Cli�,Inman Harvey, and John Gray.References[1] D. Cli�, I. Harvey, and P. Husbands. Explorations in evolutionaryrobotics. Adaptive Behaviour, 2(1):73{110, 1993.[2] Hugo de Garis. CAM-BRAIN: The evolutionary engineering of a bil-lion neuron arti�cial brain by 2001 which grows/evolves at electronicspeeds inside a Cellular Automaton Machine (CAM). In E. Sanchezand M. Tomassini, editors, Towards Evolvable Hardware: The evolution-ary engineering approach, volume 1062 of LNCS, pages 76{98. Springer-Verlag, 1996.[3] L. J. Fogel, A. J. Owens, and M. J. Walsh. Arti�cial Intelligence ThroughSimulated Evolution. John Wiley & Sons, Inc., 1966.[4] David E. Goldberg. Genetic Algorithms in Search, Optimisation & Ma-chine Learning. Addison Wesley, 1989.[5] Inman Harvey. Species Adaptation Genetic Algorithms: A basis for acontinuing SAGA. In F. J. Varela and P Bourgine, editors, Towards aPractice of Autonomous Systems: Proc. 1st Eur. Conf. on Arti�cial Life,pages 346{354. MIT Press, 1992.[6] Inman Harvey and Adrian Thompson. Through the labyrinth evolution�nds a way: A silicon ridge. In Tetsuya Higuchi and Masaya Iwata,

Page 24: Arti cial Evolution in the Physical World

editors, Proc. 1st Int. Conf. on Evolvable Systems: From Biology toHardware (ICES`96), LNCS. Springer-Verlag, 1996. In press.[7] J. H. Holland. Adaptation in Natural and Arti�cial Systems. Ann Arbor:University of Michigan Press, 1975.[8] J R Koza. Genetic Programming: On the programming of computers bymeans of natural selection. MIT Press, Cambridge, Mass., 1992.[9] J. R. Koza. Genetic Programming II: Automatic Discovery of ReusablePrograms. MIT Press, 1994.[10] Bart Krekelberg and John G. Taylor. Nitric oxide: what can it compute?Network, 8(1):1{16, 1997.[11] Francesco Mondada and Dario Floreano. Evolution and mobile au-tonomous robotics. In E. Sanchez and M. Tomassini, editors, TowardsEvolvable Hardware: The evolutionary engineering approach, volume1062 of LNCS, pages 221{249. Springer-Verlag, 1996.[12] M. C. Moore-Ede, F. M. Sulzman, and C. A. Fuller. The Clocks ThatTime Us: Physiology of the Circadian Timing System. Harvard Univer-sity Press, 1982.[13] Hans-Paul Schwefel and G�unter Rudolph. Contemporary evolutionstrategies. In F. Mor�an et al., editors, Advances in Arti�cial Life: Proc.3rd Eur. Conf. on Arti�cial Life, volume 929 of LNAI, pages 893{907.Springer-Verlag, 1995.[14] Adrian Thompson. Evolving electronic robot controllers that exploithardware resources. In F. Mor�an et al., editors, Advances in Arti�cialLife: Proc. 3rd Eur. Conf. on Arti�cial Life (ECAL95), volume 929 ofLNAI, pages 640{656. Springer-Verlag, 1995.[15] Adrian Thompson. Evolving fault tolerant systems. In Proc. 1stIEE/IEEE Int. Conf. on Genetic Algorithms in Engineering Systems:Innovations and Applications (GALESIA'95), pages 524{529. IEE Conf.Publication No. 414, 1995.

Page 25: Arti cial Evolution in the Physical World

[16] Adrian Thompson. Evolutionary techniques for fault tolerance. In Proc.UKACC Int. Conf. on Control 1996 (CONTROL'96), pages 693{698.IEE Conference Publication No. 427, 1996.[17] Adrian Thompson. An evolved circuit, intrinsic in silicon, entwined withphysics. In Tetsuya Higuchi and Masaya Iwata, editors, Proc. 1st Int.Conf. on Evolvable Systems (ICES'96), LNCS. Springer-Verlag, 1996. Inpress.[18] Adrian Thompson. Evolving inherently fault-tolerant systems. Proc.of the Institute of Mechanical Engineers, Part I: Journal of Systemsand Control Engineering, 1996. Special issue on `Genetic Algorithms inEngineering Systems.' In press.[19] Adrian Thompson. Silicon evolution. In J. R. Koza et al., editors,Genetic Programming 1996: Proc. 1st Annual Conf. (GP96), pages 444{452. Cambridge, MA: MIT Press, 1996.[20] Adrian Thompson, Inman Harvey, and Philip Husbands. Unconstrainedevolution and hard consequences. In E. Sanchez and M. Tomassini,editors, Towards Evolvable Hardware: The evolutionary engineering ap-proach, volume 1062 of LNCS, pages 136{165. Springer-Verlag, 1996.[21] Xilinx, Inc. XC6200 Advanced product speci�cation V1.0, June 1996. InThe Programmable Logic Data Book. 1996. See http://www.xilinx.com.