Dylan Evans
Walter de Back
Abstract: 174 words
Main text: 10,575 words
References: 905 words
Entire text: 11,770
Evolutionary
psychology is an approach to the study of the mind based on principles drawn
from evolutionary biology. In their
research so far, evolutionary psychologists have used many different methods,
from experimental manipulation of human behaviour in the laboratory to observation
of indigenous peoples and analysis of archaeological data.
All these methods may be called analytic, in the sense that they
collect data about already-existing systems and then analyse them.
Here we propose that evolutionary psychologists could benefit from
extending their methodological repertoire to include synthetic methods,
which involve constructing artificial systems.
Such artificial systems can provide useful models of evolved minds
and evolutionary histories that might provide evolutionary psychologists
with additional means to test their hypotheses about mental structure and
evolutionary trajectories. One kind
of synthetic method that evolutionary psychologists have so far shown little
interest in is evolutionary robotics.
We argue that, by ignoring this field, evolutionary psychologists are missing
out on a valuable research tool, and sketch out a research program involving
the use of robots to test evolutionary psychological hypotheses.
Keywords:
artificial
life
biorobotics
computer
modelling
evolutionary
psychology
evolutionary
robotics
methodology
simulation
synthetic
methods
Evolutionary psychology views the human mind as a product of evolution.
It is concerned with
identifying
the adaptive problems
tha
t
our ancestors faced and the
adaptations
that evolved to enable our ancestors to solve those problems
(Barkow, Cosmides
et al. 1992)
. These problems range
from evading predators and getting food to finding mates and rearing
children. The adaptations that evolved
to enable our ancestors to solve these problems include drives such
as hunger and thirst, emotions such as fear and love, and modules for
reasoning about social exchange
(Cosmides and Tooby 1992; Evans and Zarate 1999; Evans 2001)
.
Evolutionary
psychology pursues this goal with a variety of methodologies ranging from
experimental manipulations of human behaviour in the laboratory to observation
of indigenous peoples in pre-industrial societies.
All these methods may be called analytic, in the sense that they
collect data about already-existing systems and then analyse them.
In this article, we argue that evolutionary psychologists could
benefit from extending their methodological repertoire to include synthetic
methods, which involve constructing artificial systems.
We
begin by providing an overview of the epistemic goals of evolutionary psychology,
and of the methods currently used by evolutionary psychologists in pursuit
of these epistemic goals (section 2).
We then introduce the distinction between analytic and synthetic methods,
and describe one kind of synthetic method that evolutionary psychologists
have made limited use of in the past computer modelling (section 3).
In the next part, we describe another kind of synthetic method autonomous
robotics and focus particularly on a recent branch of autonomous robotics
known as evolutionary robotics (section 4).
Evolutionary psychologists have so far shown little interest in
this field, and in the final section we argue that, by ignoring this field,
evolutionary psychologists are missing out on a valuable research tool.
We conclude by sketching out a research program involving the use
of robots to test evolutionary psychological hypotheses (section 5).
Despite
its relatively short history, research in evolutionary psychology has
drawn on a wide range of methods. To some extent, this diverse methodology
reflects different opinions regarding the fundamental goals of the discipline.
For some evolutionary psychologists, the goal of their research
is to produce a map of the mind, a flow-chart detailing the various
mental modules that comprise the human mind and the relationships between
them (Tooby and
Cosmides 1992)
. For others, the goal is to
produce a historical account of the various stages through which the
mind passed as it evolved on its way to becoming human
(Mithen 1998) .
Clearly, these two goals are not incompatible.
It would be perfectly possible to envision an integrated evolutionary
psychology that combined them both, in which each of the various stages
through which the mind passed in its evolutionary history was described
in terms of a flow-chart. Such
an ambitious goal has not, however, been pursued by evolutionary psychologists,
at least up to now. Rather, evolutionary
psychologists have tended to pursue one or other of the two sub-goals
just described.
Different
goals require different methods. In
this section, we examine the methods that are appropriate to each of the
two goals of evolutionary psychology.
The
methods that evolutionary psychologists have used for mapping the mind
are primarily those of experimental psychology.
For example, one of the most widely discussed hypotheses proposed
by evolutionary psychologists the existence of a cheater-detection
module is based principally on evidence drawn from the Wason selection
task (Cosmides
and Tooby 1992)
. The Wason selection task involves
asking subjects to identify which card(s) need to be turned over in
order to falsify a hypothesis. For
such an experiment, all the standard apparatus of experimental psychology
is called upon: recruitment of
subjects, a well-defined experimental procedure involving at least one
control group in addition to the experimental group, and statistical
analysis of the results.
Evolutionary
psychologists fond of such experimental procedures argue that they provide
evidence for a specialised mental module when they reveal a dissociation.
A dissociation means that subjects perform better on task X
than on task Y, despite the fact there is only a small and apparently
trivial difference between the two tasks.
For example, Cosmides and Tooby found that people do much better
on the Wason selection task when the hypothesis to be falsified is cast
in terms of a social rule that must be policed than when it is cast
in terms of a logical statement that must be disproven
(Cosmides and Tooby 1992)
. They argue that this provides
evidence for the hypothesis that our logical capacities are rooted in
a mental module that is specialised for detecting cheats rather than
in a general-purpose reasoning device.
Other
methods that are also appropriate to the goal of mapping the mind include
those of cognitive neuropsychology.
This discipline also emphasises the evidential value of dissociations, but
in contrast to the subjects in the case above, they are typically the victims
of neurological damage. A dissociation in these experiments means that the
patient is impaired on task X but normal on task Y - for example, poor at
understanding printed words but good at understanding spoken words.
Cognitive neuropsychologists also look for double dissociations,
in which at least two brain-damaged subjects are typically required, such
that subject A is impaired on task X but normal on task Y , while subject
B is normal on task X but impaired on task Y.
Sometimes in instances of double dissociation no task is normally
performed; rather, the double dissociation takes the form:
subject A is worse on task X than on task Y , and subject B is better
on task X than on task Y. Sometimes
double dissociations can be found within a single subject: for example, a
subject might do worse with verbs than with nouns when producing them in
speech, but worse with nouns than verbs when producing them in print.
As
with dissociations in normal subjects, single and double dissociations in
brain-damaged subjects may be taken as evidence that there are separate cognitive
modules responsible for the tasks involved.
For example, if subject A understands speech perfectly but cannot
read any longer, we might infer that there are separate modules for spoken
word comprehension and written word comprehension, and the latter module
is damaged.
Since
these methods are not unique to evolutionary psychology, it may be asked
what the evolutionary psychologist brings to such methods that experimental
psychologists and cognitive neuropsychologists do not.
The answer lies not with any methodological innovation, but
with the nature of the hypotheses that are tested and generated in the
course of the research. The evolutionary
psychologist is distinguished by the fact that his hypotheses are generated
by constant reference to evolutionary theory in addition to whatever
other psychological theories he is working with.
Thus the reason why Cosmides and Tooby inferred from their
version of the Wason selection task that our
logical capacities are rooted in a mental module that is specialised
for detecting cheats rather than in a general-purpose reasoning device
was because they reasoned that such a module would have made good sense
in the social environment in which our hominid ancestors evolved
(Cosmides and Tooby 1992)
.
Given a time-machine,
evolutionary psychologists could use the same methods to map the structure
of our ancestors minds as those used to map the structure of the human mind
today. In the absence of such a machine,
however, evolutionary psychologists must use rather different methods for
tracing the various stages through which the human mind has passed in its
evolutionary history.
Since
the goal here is to find out about the past rather than the present, experimental
approaches strictu sensu are
not possible. Rather, the methods
of historical sciences such as archaeology and palaeontology are required.
These include the analysis of artefacts and fossils, tracing
settlement patterns and identifying changes in flora and fauna that
may have caused, or been caused by, the behaviour of our ancestors.
For example, Steven Mithen has produced a plausible story about
the mental architectures of successive hominid species on the basis
of archaeological data such as hand axes and other tools
(Mithen 1998) .
Other
data that is relevant to such historical reconstructions can be gleaned
from research in biogeography, which might reveal influences of past
climate and local ecology on hominid evolution.
For example, r
esearchers examining deep-sea sediments off the coast of Namibia, West
Africa, have found evidence that rapid global cooling occurred about
two million years ago, at the time when the first ancestors of modern
humans emerged in sub-tropical southern Africa, a finding that adds
weight to the theory that climate change played a significant part in
the evolution of early humans
(Marlow, Lange et al. 2000)
. Precisely how such data bears
on hypotheses about the mental architecture of our hominid ancestors
is, however, a thorny question.
Non-historical
sciences can also be used to throw light on the historical question
of how the various hominid minds were organised by providing useful
comparative data. For example, if
we are interested in finding out how the first
Homo sapiens lived some 150,000 years ago, and by inference how
their minds might have been structured, we can look for common themes
in the data recorded by anthropologists in their studies of hunter-gatherers
alive today. It is reasonable to assume
that patterns of behaviour that are common to most of the hunter-gatherer
societies that still exist were present in their last common ancestor
i.e. in the first humans who lived in
All
the methods discussed in section two may be described as analytic, in the
sense that they start with a real system (the minds of modern humans and/or
of various ancestral species) and attempt to collect data about this system
that might permit us to infer the internal structure of this system.
There is nothing inherently wrong with analytic methods indeed,
they are the backbone of most modern science but researchers are increasingly
aware of their drawbacks. One important
drawback is that the difficulty of analysing a system grows exponentially
as the complexity of the system increases.
Since minds especially the minds of advanced primates such as our
recent ancestors and ourselves are notoriously complex systems, it follows
that all the analytic methods described above are very hard to apply in a
way that all researchers agree upon.
One has only to look at the voluminous polemical literature that has grown
up in response to the version of the Wason selection task originally described
by Cosmides and Tooby to realise how difficult it is to derive uncontroversial
conclusions from analytic methods in evolutionary psychology.
Valentino
Braitenberg has proposed, in respect of such difficulties, that when it comes
to complex systems it is often easier to discover their internal structure
by synthetic methods than by analytic ones.
In other words, if we wish to discover how some system works, it
is often easier to do so by building successively more complex models, rather
than by attempting to infer the mechanism from mere observation:
It is pleasurable and easy to create little machines that do certain tricks.
It is also quite easy to observe the full repertoire and behaviour
of these machines even if it goes beyond what we had originally planned,
as it often does. But it is much more
difficult to start from the outside and to try to guess internal structure
just from the observation of behaviour.
(Braitenberg 1984)
Braitenberg
refers to this generalisation as the law of uphill analysis and downhill
synthesis
(Braitenberg 1984)
. But this way of putting things
is misleading to the extent that it suggests that the researcher is
faced with a choice between analytic and synthetic methods.
In reality, analytic and synthetic methods are not alternatives,
but complementary aspects of a dialectic that involves moving back and
forth between the analysis of empirical data and the construction of
simple models of underlying mechanisms.
Synthetic
methods involve building models of the system under investigation and then
observing their behaviour. The more
closely the behaviour of the model corresponds to the behaviour of the target
system, the more confident we can be that the internal structure of the model
corresponds to the internal structure of the target.
Because we have built the model ourselves, its internal structure
is transparent, and need not be inferred by analysis.
In the next section, we describe the use of some synthetic methods
in evolutionary psychology up to now, and discuss the ways in which they
can be used to contribute to each of the two goals of evolutionary psychology.
The
most commonly used synthetic method in evolutionary biology and evolutionary
psychology is computer modelling based on game theory. Originally developed
as a tool to predict rational human economic behaviour, it has been successfully
applied to many evolutionary problems (Maynard-Smith 1982). Game theory is
useful in understanding situations where the individual's behaviour depends
in part on the types and frequencies of behaviours exhibited by other animals
in the population. When there is no or little change in relative frequency
of strategies over time, this situation is called an evolutionary stable
strategy.
The
best known example of the use of game theory in evolutionary psychology
is the study by Axelrod and Hamilton of the evolution of cooperation
(Axelrod and Hamilton
1981) .
Axelrod and Hamilton organised a tournament in which hand-coded
computer programs played the iterated prisoners dilemma against each
other. The fact that the tournament
was won by a very simple program called Tit-for-tat, which merely acted
in whatever way its opponent had acted in the previous move, was taken
by Axelrod and
The
synthetic agents in the original study by Axelrod and Hamilton were
very simple, and were embedded in a very simple environment.
Later studies have progressively added more features to this
early model, making both the agents and the environment more complex.
For example, the agents have been made capable of reproducing and allowed
to evolve, and the environment has been endowed with a two-dimensional
spatial structure that allows simultaneous interaction of multiple agents
(Sigmund
1995) .
However, in contrast to the influence exerted upon evolutionary
psychology by the original study by Axelrod and Hamilton, the many follow-up
studies have been almost entirely neglected. This is consistent with
the more general tendency to neglect computer modelling of any kind
by most evolutionary psychologists. To an impartial observer, this neglect
must appear extremely puzzling, since there would seem to be a good
prima facie case that certain
kinds of computer modelling, especially the wide range of approaches
to modelling that go by the name of artificial life (A-life for short),
would make ideal intellectual tools for evolutionary psychology
(Miller and Todd 1994)
.
There are one or two notable exceptions to the general tendency among
practitioners and fans of evolutionary psychology to neglect computer
models. For example, apart from Axelrod
and Hamilton, Daniel Dennett, Peter Todd and Geoffrey Miller have also
seen the value of such models for research in evolutionary psychology
(e.g. Miller, 1994). More recently,
Tatsuya Kameda and Daisuke Nakanishi have conducted an evolutionary
simulation of socio-cultural learning, building on earlier work by Boyd
and Richerson
(Kameda and Nakanishi 2002)
. Such exceptions only bring more into relief the widespread neglect
of computer modelling in evolutionary psychology as a whole.
The
potential benefits offered by computer modelling and other synthetic methods
to evolutionary psychology include:
an additional means of testing the hypotheses generated by evolutionary psychologists;
forcing evolutionary psychologists to make the details of their
hypotheses more explicit; and an additional
means for generating novel hypotheses.
The first of these benefits is discussed in greater detail in section 3.2,
so here we will just say a few words about the second and third of these
benefits.
Computer
modelling and other synthetic methods can benefit evolutionary psychology
by forcing researchers to be more explicit about the details of their
hypotheses. Vagueness and ambiguity
are inherent properties of natural languages, so when hypotheses about
mental structure are couched in such languages, they are often ambiguous.
By contrast, the formal languages employed in computer programming
are designed to eliminate vagueness and ambiguity, with the result that
when researchers formulate hypotheses in terms of computational models,
they are forced to be clearer and to be more explicit.
As Daniel Dennett has commented, the demands of program writing
force into the open any incoherencies, gaps, or unanswered questions
in a theory; it keeps the theoretician
honest (Dennett
1974) .
While clarity is certainly not everything, and a certain degree
of ambiguity may at times be necessary while we lack a sufficiently
detailed knowledge of the phenomena under investigation, scientific
progress is more often facilitated by clear hypotheses than by ambiguous
ones. Even a false hypothesis is better
than one that is so vague that it is not even wrong.
Computer
modelling and other synthetic methods can also benefit evolutionary psychology
by helping researchers to generate novel hypotheses.
Even simple models can manifest emergent behaviour that surprises
the model designer, and this emergent behaviour can prompt the researcher
to consider ideas that had not occurred to him before.
Perhaps
the most important way in which computer modelling and other synthetic methods
can benefit evolutionary psychologists is by providing them with an additional
way to test their hypotheses. This
applies equally to the two kinds of hypotheses generated by evolutionary psychologists:
those that specify something about the mental architecture of modern
humans, and those that postulate something about the evolutionary history
of the human mind.
When
using models, computational or otherwise, to test hypotheses about our current
mental architecture, the procedure is relatively straightforward.
A model is constructed, and its behaviour in a given domain is observed
and recorded. The behaviour of the
model is then compared to human behaviour in the same domain.
If the two behaviours are sufficiently similar, this provides (defeasible)
evidence that the internal structure of the human mind is similar to that
of the model.
When
it comes to using models to test hypotheses about the evolutionary history
of the human mind, the situation is more complex.
This complexity arises from the fact that historical hypotheses
may be pitched at different grains of analysis.
Very crudely, they may be fine-grained or coarse-grained, or,
in the words of Sterelny and Griffiths, they may involve actual sequence
explanations or robust process explanations
(Sterelny and Griffiths 1999)
.
Actual sequence explanations
seek to explain the nuances of the causal history of the real world.
Such explanations postulate that a particular evolutionary
outcome is sensitive to various historical contingencies.
An example of a computer model designed to test an actual sequence
explanation is that constructed by Gumerman and Gell-Mann to examine
the reasons for the disappearance of the Anasazi people from Long House
Valley in northern Arizona in A.D. 1300
(Gumerman and Gell-Mann 1994)
. Gumerman and Gell-Mann created
a computerized replica of the
This suggests two things. First, environmental conditions alone can indeed
explain much of what is known about Anasazi demography and settlement patterns.
Second, environmental changes do not explain their complete disappearance.
The model suggested that such changes would have caused a steep decline in
the Anasazi population, but a small population could have stayed.
This in turn suggests that there are other factors, besides environmental
conditions, that influenced the settlement patterns of the Anasazi people
in this area.
Robust process explanations
are quite different to actual sequence explanations.
They reveal the insensitivity
of a particular outcome to some feature of its actual history.
In other words, robust process explanations aim to capture
general regularities that may crop up in many different historical and
evolutionary trajectories. An
example of a computer model designed to test various robust process
explanations is Sugarscape
(Epstein and Axtell 1996)
. Sugarscape simulates the behaviour
of artificial people located on a landscape of a generalized resource
(sugar). Initially, the people
are scattered about a twin-peaked landscape; over time, there is self-organization
into spatially segregated and culturally distinct "tribes" centred
on the peaks of the Sugarscape. Population growth forces each tribe to
disperse into the sugar lowlands between the mountains. There, the two
tribes interact, engaging in combat and competing for cultural dominance,
to produce complex social histories with violent expansionist phases,
peaceful periods, and so on. The proto-history combines a number of
ingredients, each of which generates insights of its own. One of these
ingredients is sexual reproduction. In some runs, the population becomes
thin, birth rates fall, and the population can crash. Alternatively,
the agents may over-populate their environment, driving it into ecological
collapse.
It will be clear from this brief description that Sugarscape is not meant
to model any particular human society in detail.
Rather, the aim is to test hypotheses about more general processes,
such as the influence of resource-distribution on population growth
and conflict, that occur in many different times and places.
As interest in this more abstract kind of
computer modelling grows, more software tools are
becoming available that allow researchers to construct such models of
their own. Two popular examples
of such software tools are StarLogo, which has been developed at MITs
Media Lab
(Colella, Klopfer et al. 2001)
, and NetLogo, developed at the Center for Connected Learning and Computer-Based
Modeling at Northwestern University
(Wilensky 1999)
.
If
there is no particular historical reality against which Sugarscape and
other such robust-process models can be measured, how can such models
be used to test hypotheses? This
is an important and difficult theoretical question that deserves a longer
answer than there is space for here.
We will therefore limit ourselves to making a few brief remarks
on this point. Part of the justification
for developing generic models such as Sugarscape lies in the capacity
of computer simulations to augment our ability to infer consequences
from the premises expressed in our hypotheses.
Indeed, as Dennett has commented, a simulation program is
ultimately a high speed generator of consequences that some theory assigns
to various antecedent conditions
(Dennett 1974)
. Simulations can therefore be
regarded as inference-machines which can help us derive the consequences
of hypotheses that might otherwise be too complex for our limited minds
to grapple with. If those consequences
turn out to be inconsistent with the data that the hypothesis was originally
designed to explain, we can reject the hypothesis straight away, even
though it is couched at a high level of generality.
We might even find that our initial understanding of the general
phenomenon was internally inconsistent, since simulations also function
as a consistency checker. In
this way, the construction of convincing, consistent alternative worlds
is a good way to test our understanding of evolutionary trends and mechanisms
(Cohen and
Stewart 2002) .
Another
important part of the justification for developing robust-process models
is that they allow us to explore the role of contingency in historical
and evolutionary processes. The
important role played by historical accident in evolution has led Stephen
Jay Gould to remark that if we could rewind the tape of biological history
and start it again, the outcome would be very different. Not only might
there not be humans, Gould suggests; there might not even be anything
like mammals. On the other hand, evolution
may not be quite as contingent as Gould suggests. If we could rewind
the tape of biological history and start it again, perhaps we would
find similar kinds of outcomes. Since life on earth has only evolved
once, it seems that there is no way of arbitrating between these different
possibilities; we are left trading intuitions. Computer simulations
offer the hope that we can bring more to bear on this question than
mere intuition. By running programs
like Tom Rays Tierra over
and over again, perhaps varying the initial parameters occasionally, we might
discern various constants in evolution, and thereby refute Gould's claim
that each case of evolution must lead to radically different outcomes
(Ray 1992)
. What counts as similarity
and difference depends, of course, on your frame of reference. If we
are concerned with details, such as the number of digits on a limb,
then perhaps we will find a different outcome each time we run our computer
simulation of evolution. However, if we use a less fine-grained taxonomy,
we may find the same broad classes of organism turning up every time
we let our virtual world evolve. This line of thought is what prompted
Ray to note that he found virtual viruses evolving in Tierra. These
viruses did not use RNA, and were not encased in a protein shell; they
were simply strings of digits on the computer's hard disk. However,
they had certain important properties in common with natural viruses. They
could not, for example, replicate in isolated culture, but only when
cultured with normal (self-replicating) creatures. Like natural viruses,
the artificial parasites executed some parts of the code of their hosts.
As in the real world, some potential hosts in Tierra evolved immunity
to the virtual viruses, and some of the viruses then evolved mechanisms
to circumvent this immunity.
Ray's analysis of evolution in Tierra supports the idea that, while the
details may change, the underlying patterns may be the same whenever evolution
occurs. Given enough time, we may find that every evolutionary process tends
to produce the same basic classes of organism, filling the same kinds of
niche. If we make our taxonomy coarse enough, this statement may become trivially
true. If we use very abstract ecological categories, such as parasite and
host or predator and prey, for example, re-running the tape of evolution
will almost certainly produce similar outcomes. The interesting questions
focus on categories that are neither too general nor too specific. If evolution
always produces parasites, this may not be very interesting since the term
parasite is so broadly defined. If evolution only rarely produces animals
with five digits on each limb, this may not be very interesting either, because
this kind of detail is of no particular consequence. If evolution always
tends to produce an animal that is well adapted to the 'cognitive niche',
however, this would be a very substantial finding. Humans would then appear,
not as an odd offshoot on the tree of life, an incredible cosmic accident,
but rather as a perfectly ordinary evolutionary product or, in Stuart Kauffman's
words, 'we the expected' (Kauffman, 1995: 8).
Computer
models of the kind described in the previous section are not the only kind
of synthetic method that evolutionary psychologists can add to their methodological
armoury. In this section, we outline another synthetic method available to
evolutionary psychology: autonomous robotics. We focus on a new paradigm
in autonomous robotics that should be of particularly interest to evolutionary
psychologists: evolutionary robotics. First, however, we set out the principal
motivation for using robots in psychological research.
Game
theoretical approaches to model behaviour, as described above, can help
to illuminate the benefits of adopting certain strategies in particular
(social) situations. However, it is not very effective in pointing out
how these strategies are implemented in the situated, embodied minds
of natural animals. Despite, for example, the many attempts to confirm
the fact that animals play Tit-for-Tat, hardly any evidence for it
has been accumulated. This is caused by the inherent simplicity of the
games. Generally, in game theory, the benefits and costs for actions
are precisely defined, whereas in nature these values are in the eye
of the beholder and immeasurable. Furthermore, these games view just one
isolated behaviour observed in nature and studied it as an all-or-nothing
concept (e.g. to co-operate or not), whereas behaviour in real animals
is heavily related to other behavioural activities and the environment.
Investigating the origins of behaviour from a game theoretical
perspective is much like studying chess in order to investigate intelligence,
as is done in traditional AI
(Hemelrijk 1997)
. Both explore cognitive reasoning abilities without providing the behavioural
fundaments on which it is based.
Since
the late 1980s, a growing number of cognitive scientists have become
critical of the traditional focus of their discipline on abstract reasoning,
arguing that minds are always situated in bodies and worlds, and cannot
be understood apart from them
(Clark 1997) .
The body, mind and environment are continuously involved in
intricate complex dynamics, which adds to the complexity of understanding
minds by analytic reverse engineering. Autonomous robotics offers the
unique possibility of building an artificial mind inside an artificial
body that operates inside a real environment. This provides psychology
and cognitive science with a powerful tool for testing hypotheses through
synthesis of artificial autonomous agents
(Pfeifer and Scheier 1999)
.
It
is noteworthy that although this approach has only recently gained recognition
among cognitive scientists, pioneering in this field was conducted as
early as 1948 by Grey Walter
(Walter 1951) .
His work with analogue autonomous robots that displayed phototropic
behaviour remained largely neglected until recently due to the dominance
of computational approaches in cognitive psychology and artificial intelligence.
The
concept, advocated by roboticists, that complex behaviour can often
be explained by looking at the complexity of the environment, instead
of the internal mechanisms
(Simon 1969/1996)
, has inspired opposite reactions among psychologists. In particular
the claim that intelligence does not require
any internal representations
(Brooks 1991)
and the apparent resemblance to psychological behaviourism has caused
many psychologists to deny the power of this methodology. However, the
situated and embodied approach to cognition does not in principle deny
the existence of internal representations. The crucial point is, rather,
that internal and external processes are often more closely coupled
than has traditionally been thought.
Autonomous
robotics has been used to model different approaches to the mind, although
often implicitly. Building a robot controller that somehow converts sensation
into behaviour implies making assumptions of a philosophical, biological
and psychological nature. Running a robot inside an environment and analyzing
its performance provides insight into the underlying assumptions. In this
way, constructing controllers for autonomous robots can be used for investigating
models of the mind and for testing psychological hypotheses.
In
the relative short history of autonomous robotics, several quite different
robot control architectures have been proposed, all of which are grounded
in a different view to cognition. For the purposes of this brief overview,
it is sufficient to divide these control architectures into three classes:
symbolic, non-symbolic, and sub-symbolic.
Symbolic
(or deliberative, cognitive) robot controllers are based on the information
processing metaphor. The rise of computers inspired people in psychology
and artificial intelligence to model the human mind in terms of functional
modules that process information computationally
(Newell and Simon 1976)
. Early symbolic robot controllers typically consisted of several modules
for sensing, modelling, reasoning, planning, and task execution that
are processed in sequence. Moving inside an environment involves mapping
sensory data onto an internal map of the world, planning a trajectory
based on this map, and executing a movement, after that the process
is run again. This is known as the sense-plan-act cycle
(Pfeifer and Scheier 1999)
.
Behaviour-based
robot controllers
(Arkin 1998) were
introduced as a reaction to cognitive and symbolic approaches and offers
a solution that is inspired by behaviourism and ethology. Instead of
using internal symbolic representations of the environment, these controllers
interact directly with the environment. Instead of processing information
centrally and sequentially, behaviour-based robot controllers consist
of collections of simple input-output mappings that are processed in
parallel and compete or cooperate for influence on control. This approach
offers useful lessons in the sense that the animal and human mind need
not be as complex as might be suggested by its behaviour, because this
is for a large part due to the complexity of the environment.
Behaviour-based
architectures are frequently used in autonomous robotics research for the
control of low-level behaviour, sometimes in combination with high-level symbolic
planners. This approach, however, suffers from the appearance of being purely
behaviouristic, lacking the important ability to adapt, and tends to be motivated
more by practical engineering demands than by scientific curiosity.
Subsymbolic
robot controllers rely on artificial neural networks to control robot
behaviour. Artificial neural networks (ANNs, or connectionist models)
use parallel distributed processing (PDP) and can be trained using incomplete
data and are able to generalize the acquired knowledge to novel situations
(McClelland, Rumelhart
et al. 1986) .
The neural controller self-organises by gradually changing the weights
of the connections between the artificial neurons until the network
is satisfactory for the task at hand.
Subsymbolic
robot controllers have been adopted in biorobotics and synthetic neuro-ethology.
A good example of this line of research is the work on robotic experiments
in cricket phonotaxis. Through making robotics implementation of female
crickets that can find males by moving towards a species-specific song
that males produce, evidence is found for neuro-ethological hypotheses
that the control system that is required for this task is much simpler
than traditionally been thought by biologists
(Webb 1995) .
Although subsymbolic controllers allow for investigation of the first goal
of evolutionary psychology (exploring the structure of the mind) it cannot
be employed in this form for second goal (tracing the historical route of
evolved minds). In the crickets case, evolutionary psychologists would not
only be interested in the neural structures that facilitate the phonotaxic
behaviour, but also in the historical trajectory that has led to the emergence
of the singing behaviour through natural (sexual) selection.
The
idea of constructing robot controllers through self-organisation can
be taken a step further. Evolutionary
Robotics
(Nolfi and Floreano 2000)
is a relatively new approach to robotics that is based on the use of
evolutionary computation for developing robot controllers (and sometimes
robot morphology). The controllers of evolutionary robots are usually
artificial neural networks. In this approach, however, the neural controllers
are not (only) subject to development, but are also subject to artificial
evolution.
Most
experiments in evolutionary robotics follow the same basic format.
First, a population of artificial chromosomes that code for the
control system and/or the morphology of a robot are randomly generated.
Next, these chromosomes are decoded and the resulting robot (which
may be physical or simulated) is set free to act in a given environment while
its performance on various tasks is automatically evaluated according to
a predefined criterion known as the fitness function.
The fittest robots are then allowed to reproduce by generating copies
of their genotypes with the addition of changes introduced by some genetic
operators such as mutation and recombination.
This whole process is then repeated for a number of generations,
which allows natural selection to enhance the average fitness of the population
in the same way as it has done so effectively in the history of life as we
know it.
Most
current research in evolutionary robotics tends to fall into one of the following
categories:
A
capacity for navigation is one of the most basic requirements for an autonomous
mobile robot, so it is not surprising that much of the research in evolutionary
robotics continues to focus on this area.
Typical tasks that researchers in this area set their robots include
obstacle avoidance and navigating toward a target area.
Since most animal navigation is visually guided, this field of research
includes a great deal of work on visual sensing, including such areas as
object recognition and discrimination, though a lot of robots still use infra-red
sensors or sonar instead of or in addition to a camera.
Most
experiments in the evolution of robot navigation involve testing each
member of each generation one by one in a fixed physical environment.
Experiments in co-evolution, by contrast, involve a continually
changing dynamic environment consisting of other agents.
Biologists have long speculated that such an environment may
enhance the adaptive power of natural selection.
For example, Dawkins and
Krebs argued that competing populations may reciprocally drive one another
to increasing levels of behavioural complexity by producing an evolutionary
arms race (Dawkins
and Krebs 1979)
. Researchers in artificial life
began to develop virtual simulations of such arms races in the early
1990s (Miller
and Cliff 1994)
, and more recently researchers in evolutionary robotics have developed
fully-embodied physical models of predator-prey dynamics.
Most
research in evolutionary robotics takes the robot body as a fixed parameter
and focuses exclusively on evolving robot control systems.
However, a few projects have applied the evolutionary process
to robot bodies. The most striking experiment so far in this field involved
the evolution of simple locomotive systems composed of bars and actuators,
and then used rapid-prototyping technology to produce the multi-linked
structures, so that the only human input needed was the final attachment
of snap-on motors
(Lipson and Pollack 2000)
.
Another approach in
evolvable
hardware is evolving robot controllers in
reconfigurable
electronic circuit
s such as field-programmable gate arrays
(FPGA
s
)
.
These are argued to
be analogue
dynamical
continuous-time systems and thereby
avoid the constraints of
discrete
digital
design
(Thompson 1997)
.
The
explicit discussion of evolvable hardware
draws attention to an aspect of evolutionary robotics that may not be apparent
from the discussion so far the fact that experiments can be done in simulation
as well as with real physical robots.
The relative advantages and disadvantages of each approach are the subject
of some debate in evolutionary robotics.
Physical robots are argued to be better models since they incorporate
real problems in coping with physical forces like sensory noise, energy consumption,
damage and inertia. However, artificial evolution in physical robots is very
time-consuming. Simulated robots avoid this problem, but also circumvent
the physical problems that may well be essential to the structure of natural
minds. Another important problem with
physical robots is the inability to evolve bodily structures (evolvable hardware
as described above is a rare exception), which is essential for the modelling
of co-evolution of body and brain. This is to date only possible in simulation.
Recent
studies also combine the physical/simulated robot approaches by performing
the most time-consuming evolution in simulation, and subsequently evolve
the controllers further in physical robots. When experiments in evolutionary
robotics are conducted entirely by means of detailed computer simulations
that also incorporate physical forces, the boundary between the research
described above, in section 3.1., and the research described in this section,
becomes less clear.
As
indicated by this brief overview of research, evolutionary robotics is already
the topic of intense experimentation in the fast-growing robotics community.
In the next section we argue that this line of research is very promising
for research in evolutionary psychology as well.
The
overlap between evolutionary robotics and evolutionary psychology should
be clear by now: both disciplines are concerned with embodied, situated minds
that evolve(d) over many generations, implicitly solving adaptive problems
through natural selection. Nevertheless,
the motivations of these fields are quite different.
Most research in evolutionary robotics is primarily concerned with
(developing a methodology for) creating robot controllers that are more robust
or that perform better at a given task. This research is not directly interesting
for psychologists, since it is motivated by pragmatic or industrial goals
and is not aimed at modelling the mind. Indirectly, however, developments
in evolutionary robotics provides (and will continue to provide) wonderful
new tools for better modelling the evolution of natural and artificial minds.
With
the use of these modern techniques, we can begin to recreate the evolution
of the mind by introducing evolving robots to much the same evolutionary
forces that shaped the minds of humans and other animals long ago: synthetic
evolutionary psychology.
In
this section, we point out how the general goals of evolutionary psychology
could benefit from synthetic approaches, particularly by evolutionary robotics,
and we outline some specific issues that can be investigated by this interdisciplinary
approach. This is followed by a sketch of a framework for the design of experiments
in synthetic evolutionary psychology that differs from standard evolutionary
robotics in several important aspects.
We
propose a synthetic approach to evolutionary psychology that complements
the standard analytic methods. The use of autonomous and evolutionary robotics
in this research effort can be seen as an extension of the computer modelling
methodology, with the advantage of dealing with evolving situated, embodied
minds. We show here how this approach can benefit evolutionary psychology
and point out some specific issues for investigation. We also highlight the
relationships between our approach and some related fields.
Both
goals of evolutionary psychology (tracing the structure and history of the
mind) can be studied through the application of evolutionary robotics. A
synthetic approach to mapping the mind investigates the function of adaptive
modules. Introducing evolutionary robots to the evolutionary and environmental
forces that are thought to have shaped the human mind, and subsequently analysing
the resulting robot controllers for adaptations, provides data relevant to
both kinds of hypotheses in evolutionary psychology.
Moreover, lesioning experiments can then be performed on the evolved
robot minds to look for evidence of dissociations.
Natural
minds are the result of a history in which new evolutionary forces build adaptations
not from scratch but on the basis of existing adaptive modules. This incremental
evolution can be synthesised by increasing the complexity of the robot body
or its environment (while allowing for more complex controllers to evolve).
The scaffolding of adaptive modules that is conducted in this way provides
data relevant to hypotheses about the historical route by which the human
mind evolved, as well as insight into the structure of modern natural minds.
As
with computer modelling, evolutionary robotics can be used to test both
robust-process hypotheses and actual-sequence hypotheses in evolutionary
psychology. In a recent experiment
lying closer to the actual-sequence end of the spectrum, Den Dulk, Heerebout
and Phaf studied the evolution of dual-route dynamics for affective
processing (den
Dulk, Heerebout et al. forthcoming)
. In particular, Dulk and colleagues
used the standard methods of evolutionary robotics to examine the evolutionary
justification given by LeDoux for his dual-route model of fear-processing
(LeDoux 1998)
.
Le Doux has found evidence that, in many mammals, fear is processed
simultaneously by two neural pathways, one subcortical and the other
largely cortical. The subcortical
route is faster but generates many false positives, while the cortical
route is slower but more accurate.
Le Doux argues that this dual-route mechanism evolved by natural selection
because it allowed animals to get the best of both worlds escaping quickly
when necessary, but not wasting too much time or effort on false alarms.
Dulk and colleagues examined this argument by allowing agents
to evolve in a simple environment consisting of predators and food.
They found that agents did indeed evolve a dual-route mechanism
similar to that proposed by Le Doux, but only when certain conditions
were met: the food and the predator
had to be relatively hard to distinguish, and information must take
significantly longer to propagate via the cortical route than via the
subcortical route.
A
program of research in synthetic evolutionary psychology might use much of
the same tools as those used by mainstream research in evolutionary robotics,
but the issues for investigation would be substantially different.
Evolutionary robotics is generally concerned with navigation, co-evolution,
cooperation and evolvable hardware.
Evolutionary psychologists would instead be more interested in using robots
to explore psychological questions in particular, the history and structure
of the various mental modules that comprise the human mind.
A
representative sample of the kinds of problems that are thought by evolutionary
psychologists to have led to the evolution of mental modules include:
Experiments
in evolutionary robotics have already begun to explore the first two problems
in this list, and have touched on the problem of communication, but have
left the other problems virtually unexplored.
This is partly due to technological limitations, but such considerations
apply only to conducting experiments with real physical robots.
Computer simulations of evolving robot populations could explore
the other areas without much difficulty. There is great scope, then, for
a broad research program to explore all of these problems in a systematic
way, perhaps by a graded approach that tackles each problem in the order
in which they were faced by our ancestors.
Such an incremental approach might also throw light on the way in
which prior adaptations may be co-opted by natural selection as the basis
for solutions to later problems.
It
could be argued that synthetic evolutionary psychology does not qualify as
a new approach, since it borrows parts of its methodology from evolutionary
robotics, biorobotics and A-life. Indeed, these fields are similar in many
respects. We feel, however, that the goals and methodology of synthetic evolutionary
psychology are sufficiently distinct from the other fields to make it qualify
as an area of research in its own right.
Whereas
evolutionary robotics differs from synthetic evolutionary psychology with
respect to goals, as we have shown, biorobotics and A-life share the goal
of providing feedback to the behavioural sciences. These fields, however,
differ in their focus. Biorobotics
research focuses mainly on the plausibility and comparative performance of
artificial neural structures and behaviour, leaving the evolutionary component
virtually untouched. And A-life, lacking a clear description and goal, focuses
on many often quite different topics: from chemistry to art, and from game
theoretical accounts of human behaviour to the evolution of morphology in
virtual worlds, all of which are conducted in simulation. The topics in A-life
are too diverse to allow for a good comparison.
To
avoid confusion and relate our approach to other fields of research, synthetic
evolutionary psychology stresses that:
·
Experiments are individual-oriented
·
Individuals are situated (and embodied)
·
Populations are subject to natural selection
·
Environments are dynamic, changing and continuous
·
Environments contain simultaneous populations of individuals
·
Experiments involve real problems (i.e. those that exist in biological
contexts)
The
experimental design of synthetic evolutionary psychology may be considerably
different from mainstream evolutionary robotics. Here, we indicate where
these differences come from and argue for some methodological adjustments.
As
we have already seen, advances in autonomous robotics are often due to inspiration
from biology and cognitive science. This is obviously a good approach, since
nature provides us with the only examples of truly autonomous machines we
know today and much can be learned from them. These methods, including evolutionary
robotics, are to a growing extent biologically
inspired. The kind of synthetic
methods we propose are used for exactly the opposite purpose, however: making
advances in behavioural sciences as a result of inspiration from computer
or robotic modelling. These models should not only be inspired by biology,
but should moreover be biologically
plausible (at least in the respects relevant to the phenomenon under
investigation). Whereas mainstream
evolutionary robotics research is not, synthetic evolutionary psychology
is very much concerned with being
biologically plausible. This fact calls for adopting some methodological
changes.
Note
that this problem is certainly not new, and is already the subject of active
research in many synthetic disciplines that provide feedback to natural and
behavioural science. It is, however, of special importance to synthetic evolutionary
psychology, since it makes use of a whole range of bio-inspired techniques,
and would therefore be very sensitive to criticism from psychologists and
biologists. To avoid or overcome this,
there are two options. The first is to be very explicit about the relation
and relevance between the model and its target system, to exclude many (mis)interpretations
of possible results. The second is to make the tools and experimental design
more biologically plausible, preferably without making it more detailed and
complex.
A
good modelling methodology should be explicit about many aspects of
the relations between the model and the system(s) it is supposed to
represent (the target system). Barbara Webb has developed a framework
for describing and comparing robotic models in behavioural and biological
sciences (Webb
2001) . This framework
offers seven distinct dimensions that should be included in developing
and describing an experiment to allow meaningful comparison to be made
between behavioural data from artificial and natural creatures:
·
Relevance: does the model test and/or generates hypotheses? Are these
clearly defined?
·
Level: what are the elementary units of the model? Neurons, modules,
individuals?
·
Generality: what are the target systems? Humans, primates, or animals
at large?
·
Abstraction: how detailed or complex is the model relative to the target
system?
·
Structural accuracy: how well does the model represent the behavioural
mechanism?
·
Performance match: to what extent does the behaviour of the model and
target match?
·
Medium: what is the physical (or simulated) basis by which the model
is implemented?
A
good modelling methodology in biorobotics (for which this framework was initially
developed) as well as synthetic evolutionary psychology is one in which biological
and psychological behaviour are modelled in context of the real problems
faced by animals or our ancestors.
Besides
being explicit about the relevance of an experiment, it is equally important
that to adopt techniques that are plausible in a biological context. This
applies to all the techniques involved; from neural models to evolving morphology
and from artificial ecologies to plausible evolutionary schemes. Although
all are interesting and important, we focus on the latter, which is of special
importance to synthetic evolutionary psychology.
A
well-known and important difference between artificial and natural evolution
is that the former is goal-directed, while the latter is open-ended. Artificial
evolution is a technique originally devised to optimise parameters that is
inspired by the optimisation property observed in nature. Natural evolution,
in contrast, does not converge to a single solution and then stop: it is open-ended
and continuous. There are no optimal solutions in nature because the problems,
arising from the biological context which includes many co-evolving creatures,
are constantly changing. Evolutionary psychologists are interested in the
mechanisms that result from natural, and therefore open-ended, evolution.
A synthetic approach to evolutionary psychology should attempt to replicate
this process. Over the years, various technical procedures have been proposed
that partly overcome the major differences between artificial and natural
evolution, such as massive co-evolution, variable genotype length, and complex
genotype/phenotype mapping schemes.