Exploring Biological Intelligence through Artificial Intelligence and Radical Reimplementation Joel Lehman

Document technical information

Format pdf
Size 1.1 MB
First found May 22, 2018

Document content analysis

Category Also themed
Language
English
Type
not defined
Concepts
no text concepts found

Persons

Organizations

Places

Transcript

How Should Intelligence Be Abstracted in AI Research ...
AAAI Technical Report FS-13-02
Exploring Biological Intelligence through Artificial
Intelligence and Radical Reimplementation
Kenneth O. Stanley
Joel Lehman
Department of Electrical Engineering and Computer Science
Department of Computer Sciences
The University of Central Florida, Orlando, FL
The University of Texas at Austin, Austin, TX
[email protected]
[email protected]
Abstract
both in biology and AI attempt to decipher the design principles underlying human intelligence.
Such efforts are motivated by how studying prior precedent can facilitate deeper understanding. As a result, observation, analysis, and inference from nature are common
techniques in biological fields such as ethology and cognitive psychology (Bekoff and Allen 1997). In contrast, computational approaches often probe questions of biological
intelligence through synthesis. For example, biologicallyinspired AI abstractly emulates biological processes such as
information processing in neural networks or natural evolution. The idea is that such biologically-inspired algorithms
may help isolate the key elements of biological intelligence
and its origins. That is, through simulations with varied features, it may be possible to separate the essential features
of biological processes from those that are only incidental.
However, a danger inherent in any biologically-inspired approach is that inferring generalities from a limited or biased
distribution can often be deceptive.
For example, a digital multi-gigahertz computer is not
found among the forms crafted by natural evolution, though
such computation may have enhanced biological intelligence. Yet its absence is not surprising: Biological evolution is not an engineer, but a tinkerer constrained to making incremental tweaks to previous designs instead of holistic revisions (Jacob and others 1977). As a result, the attractors for biological design differ from those of more directed design methodologies such as human engineering.
Additionally, historical contingency may lead natural evolution to converge to particular solutions to particular biological problems, leaving other plausible possibilities unrealized
(Blount, Borland, and Lenski 2008). In total, the effect of
these constraints (i.e. being restricted to tinkering and historical contingency) biases natural evolution’s solutions to a
systematically reduced subset of all that are possible.
Importantly, because of this systematic bias in how nature generates solutions to biologically-relevant problems,
features that are only incidental to a deeper phenomenon
may be amplified to become ubiquitous. The result is that inferences based on observing only biological examples may
be misleading (Sterelny 1997). This is particularly true in
fields like biologically-inspired AI where it is often assumed
that particular mechanisms such as biological development
or natural selection are essential rather than only particular
An important goal in artificial intelligence and biology is to uncover general principles that underlie intelligence. While artificial intelligence algorithms need
not relate to biology, they might provide a synthetic
means to investigate biological intelligence in particular. Importantly, a more complete understanding of such
biological intelligence would profoundly impact society. Thus, to explore biological hypotheses some AI researchers take direct inspiration from biology. However,
nature’s implementations of intelligence may present
only one facet of its deeper principles, complicating the
search for general hypotheses. This complication motivates the approach in this paper, called radical reimplementation, whereby biological insight can result from
purposefully unnatural experiments. The main idea is
that biological hypotheses about intelligence can be investigated by reimplementing their main principles intentionally to explicitly and maximally diverge from
existing natural examples. If such a reimplementation
successfully exhibits properties similar to those seen in
biology it may better isolate the underlying hypothesis than an example implemented more directly in nature’s spirit. Two examples of applying radical reimplementation are reviewed, yielding potential insights into
biological intelligence despite including purposefully
unnatural underlying mechanisms. In this way, radical
reimplementation provides a principled methodology
for intentionally artificial investigations to nonetheless
achieve biological relevance.
Introduction
Artificial intelligence (AI) researchers seek to develop intelligent software and machines. Interestingly, such AI research need not relate to the specific biological implementation of human intelligence. Thus through more broadly
exploring intelligence, AI research may enable discovering
the general laws underlying intelligence and intelligencegenerating processes. However, understanding biological intelligence in particular remains important and could profoundly impact humanity’s self-perception. Furthermore,
any general theory of intelligence should explain both biological and artificial instances. For these reasons, researchers
c 2013, Association for the Advancement of Artificial
Copyright Intelligence (www.aaai.org). All rights reserved.
89
realizations of a deeper principle.
As an analogy, consider trying to infer the principles of
flight (instead of intelligence) from natural examples. If a
man isolated on an island observed only birds, he might
conclude that flappable wings with feathers are essential to
flight. However, after seeing a bat flying, that same man
could conclude that because bats do not have feathers, feathers must not be essential after all. Yet it might still seem to
him that flappable wings are; in other words, ornithopters
are a powerful attractor when extrapolating flight from natural examples. If the man on the island, having already seen a
bird and a bat, then observed a helicopter, he could correctly
conclude that the action of flapping – and even wings in the
traditional sense – are unnecessary for flight.
This example illustrates the value of divergent implementations for isolating and validating basic principles. Through
its vast deviation from prior natural examples, the helicopter
helps to illuminate that the general principles of flight run
deeper than wings and feathers. Importantly, inferring the
deeper principles of aerodynamics, which is easier given the
helicopter, helps to unify understanding of both artificial and
biological instances of flight. At the same time, the helicopter validates aerodynamics by demonstrating what engineering with such knowledge enables: agile powered flight
for human transportation. In other words, the value in abstracting a phenomenon a particular way can in part be estimated by the divergent implementations it enables, i.e. what
it makes possible, and the degree to which it elegantly unifies all examples of a phenomenon.
These insights motivate a principled approach to investigating biological principles such as intelligence through
biologically-inspired AI that can be called radical reimplementation. The main idea in radical reimplementation is to
craft an abstract hypothesis about the crux of a given biological principle, and then to test this hypothesis by reimplementing the abstraction in a way that is maximally different
from how it exists in nature, yet that still exhibits biological similarity in its effects. Just as a helicopter’s significant
divergence from a bird yields valuable perspective on the
nature of flight, a successful radical reimplementation of biological intelligence may also more clearly expose the core
of the underlying phenomenon than would considering additional biological examples.
Thus AI, which seeks to explore the abstract possibilities for intelligence in all forms, can enable exploring radical reimplementations of biological intelligence. Note that
AI researchers otherwise seeking to be relevant to biology face the significant challenge of demonstrating that
their models relate convincingly to biological truth (Sterelny
1997). This challenge is especially relevant to researchers
in biologically-inspired AI where a direct connection to biology is often expected or necessary. As a result, such researchers often must argue that the abstractions made by
their models are principled, i.e. that the biological details
filtered out through abstraction into a compact algorithm are
non-essential.
Thus one way to view radical reimplementation is as a
principled alternative to the tradition within biologicallyinspired AI of minimizing deviations from nature. While
the differences in models derived from nature must usually
be well-motivated, explicitly maximizing divergence can be
equally as principled as minimizing and justifying it.
In this way, the advantage of the radical reimplementation
approach is that it escapes the underlying restrictiveness of
attempting to mimic nature directly. There is no attempt to
claim that a model is intrinsically biologically plausible or
even that it directly relates to biology; instead the idea is to
provide radically novel examples of intelligence. In short,
the approach challenges researchers to maximize the divergence from the workings of terrestrial biology while instantiating qualitatively similar output. When successful, such
reimplementations indirectly provide evidence for the underlying principles guiding their creation.
In this paper the radical reimplementation approach is illustrated by reviewing two examples of existing research in
the spirit of radical reimplementation that aim to shed light
on the principles underlying biological intelligence and its
origins.
Novelty Search: Questioning Optimization
Natural evolution is a profoundly creative process responsible for crafting the diverse animal intelligences found on
Earth. While much is understood about natural evolution,
one question still debated is the relative importance of evolutionary forces to evolution’s creativity (Orzack and Sober
1994; Pigliucci and Kaplan 2000). In particular, this section
focuses on investigating the importance of optimizing biological fitness relative to non-adaptive evolutionary forces
like genetic drift or exaptation. The overall idea is that evolution’s creativity is what facilitated the evolution of complex
intelligence and thus better understanding such creativity
may lead to algorithms able to realize similar intelligence.
Furthermore, this biological question itself may be of interest to AI researchers given the ubiquity of optimization in
AI and machine learning. That is, if optimization is not the
key driver of evolution’s creativity it may also point to potential advantages for driving AI search processes by other
mechanisms.
The motivation for exploring this question through radical reimplementation is that natural evolution as a whole is
subject to the N = 1 problem: There is only one example
of life evolving through natural evolution that we are aware
of, which complicates inferring statistically-valid principles
about evolution in all its possible instantiations (Sterelny
1997). Thus experimental methods facilitating indirect investigation, like radical reimplementation, may be important tools for isolating the effect and importance of particular
evolutionary forces.
In particular, evolution can be abstracted in different ways
and reimplemented to stress a particular abstraction in a
manner explicitly unlike nature. For example, a coarse way
of abstracting evolution (in a selection-centric interpretation) is to view evolution as an optimizer, driving relentlessly
towards higher fitness. An alternate non-adaptive abstraction
is instead to conceive natural evolution as a process driven
to continually create novelty without any direct pressure towards adaptation. Of course, both of these abstractions focus
90
on only one aspect of natural evolution, but radical reimplementation allows for exploring the impact of those aspects
in isolation.
Interestingly, the abstraction of evolution as an optimizer is prevalent in evolutionary computation (EC; Holland
1975). In most evolutionary algorithms (EAs), biological fitness is abstracted as a fitness function. Individuals from a
population of computational genomes are selected for their
ability to maximize the fitness function and are mated algorithmically to form the next generation. Because the fitness function is central in such an abstraction, these models can be seen as selection-centric radical reimplementations of natural evolution. In other words, traditional EAs
abstract natural evolution’s key driving force as optimization
and reimplement such optimization much differently from
natural selection.
However, although they have proven successful in many
cases, such objective-driven EAs have never generated artifacts exhibiting complexity on the level of biological organisms. Furthermore, the high-level properties of such EAs
often contrast starkly with those of natural evolution. For example, EAs are nearly always convergent, while a characteristic feature of natural evolution is its open-ended march towards greater diversity. These problems put into question the
central abstraction of evolution as an optimizer (Lehman and
Stanley 2011a), which abstracts to some degree the adaptationist view of natural evolution.
However, such a negative result of radical reimplementation is not as informative as a positive result. It could be that
a particular radical reimplementation is misguided, and that
another more well-designed implementation of the guiding
abstraction would yield better results. On the other hand, a
positive result for a radical reimplementation more assuredly
provides evidence for its underlying abstraction, because it
is improbable that reimplementing an incorrect abstraction
would yield desirable results. For example, demonstrating in
reality a new form of flight provides evidence for the principles the engineer applied; and because a new form of flight is
so difficult to achieve, it would be unlikely that relying on a
fundamentally incorrect abstraction would prove productive
merely by chance.
A more positive example of a radical reimplementation of natural evolution is based on an abstraction other
than optimization: evolution as a generator of novelty. This
optimization-free abstraction is plausible because a signature of evolution is its tendency to diverge and fill reachable
niches. Such an abstraction can be radically reimplemented
as an algorithm that searches only for novelty. This idea motivates novelty search, an EA that focuses on novelty instead
of natural selection (Lehman and Stanley 2011a).
Novelty search is a non-adaptive EA driven only to diverge, continually finding forms different from what has
been encountered by the search in the past. While in nature the accumulation of novelty is mainly passive, novelty
search explicitly seeks it. In this way, novelty search allows
investigating the properties of a search without the pressure
for organisms to adapt to their environment.
In particular, novelty search replaces the fitness function
that characterizes the optimization-based EA abstraction of
(a) Optimization-based search
(b) Novelty search
Figure 1: Contrast between optimization-based search and
novelty search. (a) Traditional EAs guided by a fitness function tend to converge towards a prescripted objective. (b)
Novelty search is instead driven to diverge from previously
encountered artifacts.
natural evolution with a novelty metric, which is a userdefined measure of distance between evolved artifacts in a
particular domain. In this way, novelty search can be driven
to find only what is different from what it has previously
encountered in a user-defined space of artifacts. Figure 1 illustrates the main difference between novelty search and a
traditional objective-based EA. While novelty might at first
seem an uninformative gradient of information to follow, often performing something novel requires exploiting regularities in a domain. For example, for a robot to behave in a
novel way may require learning about how to avoid walls or
navigate a corridor. In this way, a search for novelty can lead
to functional and interesting results.
Because novelty search realizes a search without an overarching goal (just as natural evolution is not driven overall towards any one specific type of organism), it can act as
a tool for understanding the potential of non-adaptive processes. In effect, novelty search allows isolating the potential of a raw search for novelty entirely separated from any
optimization pressure, which would be a challenging experiment to perform with real organisms.
In this way, novelty search can both test the abstraction
of evolution as a search for novelty and potentially provide
evidence and insight for the importance of fitness optimization in evolution. To review such evidence, in EC there have
been a series of experiments that compare novelty search
with a more traditional objective-driven algorithm (Lehman
and Stanley 2011b; 2011a; 2010; Risi, Hughes, and Stanley
2010). The basic idea is to investigate which type of search
can better evolve an artificial neural network that can control a robot to perform a target task (e.g. navigating a maze
or walking bipedally). That is, an artificial neural network
is connected to a simulated robot, receiving sensory information and outputting motor commands; fitness is measured
based on how well the robot controlled by the neural network performs, whereas novelty is measured by how different the robot’s behavior is from previous robots. While the
expectation might be that optimizing directly for the target
task would be more successful, in practice novelty search
has often performed better (Lehman and Stanley 2011b;
2011a; 2010; Risi, Hughes, and Stanley 2010; Krcah 2010;
Mouret 2009), more effectively evolving controllers for
maze-navigating robots (Lehman and Stanley 2011b; 2011a;
2010; Mouret 2009), artificial ants (Lehman and Stanley
2010; 2011b), bipedal robots (Lehman and Stanley 2011a),
91
determine its identity and further differentiate (Raff 1996).
From this traditional perspective, the most intuitive inference is that temporal unfolding and local interaction may be
the key principles that explain the expressive power of developmental processes. Yet it is interesting to consider whether
development in the traditional biological sense (and neural
development in particular) might be only a particular implementation detail that masks a deeper principle.
Problematically, investigating this possibility through biological inference is challenging because life has largely converged to a single overarching means of development and
cellular replication. Thus inferring principles about development’s necessity or purpose from biology, like when considering natural evolution, is also subject to the N = 1 problem. However, an intriguing possibility is that this problem
can be circumvented by radically reimplementing development such that it holistically differs from biological development on Earth. While much work has explored alternative
developmental schemes (Stanley and Miikkulainen 2003),
nearly all of these schemes operate fundamentally in the
spirit of natural development, i.e. one cell-like entity gives
rise to many more through an iterative process of splitting
and differentiation, or at least growth (Stanley and Miikkulainen 2003). Such approaches are thought-provoking and
can illuminate some features of development, but because
of their similarity to terrestrial biology they may not well
isolate the deeper principles behind neural development.
For this reason it may be more informative to investigate models through the radical reimplementation approach.
Thus, if an interesting hypothesis can be derived for the root
mechanism of development, it can be empirically investigated by reimplementing it in a way purposefully divergent
from biological development.
One such high-level hypothesis for the root mechanism
of development is that it is a means for expressing an organism as a function of geometry. In other words, the fullydeveloped form of an organism can be understood as the
product of an abstract function that maps from points in
three-dimensional space to the type of cell (if any) that
should occupy each point in the space. The main motivation
for this more broad abstraction is that the most important
aspect of development is the patterns that it generates, and
that development over time is only one way among many to
realize such patterns.
While this perspective departs from the more common
(and intuitive) view of development as a temporal unfolding process that relies upon local interaction, there are reasons to pursue such an abstraction. For example, that such a
functional relationship is possible at all is supported by the
universal approximation theorem (Cybenko 1989), which
establishes that a series of enough nested functions can
approximate any pattern. However, though they thus can
clearly approximate phenotypic form, whether nested functions of geometry are a productive way to view developmental patterns is a separate question. Interestingly, it has been
demonstrated that many heavily conserved genes active in
development work to establish chemical gradients that act
as nested coordinate frames that provide context to a particular cell about its role within the organism (Raff 1996).
and robots that learn from experience (Risi, Hughes, and
Stanley 2010). Thus while bipedal walking may normally
be viewed as an adaptation, the explanation for its discovery by novelty search (Lehman and Stanley 2011a) cannot
be adaptationist because novelty search does not explicitly
favor walking over anything else.
The reason that novelty search nevertheless does well is
that a search driven to optimize fitness may converge on a
fitness peak from which there is no path to the highest-fitness
goal behavior. In other words, the stepping stones to the target behavior may not increase fitness themselves. While this
phenomenon is known in biology, it is informative to see
how pervasive it is even in a wide range of simple problems,
and how it can be overcome by searching only for novelty.
The generality of results from novelty search (Lehman
and Stanley 2011b; 2011a; 2010; Risi, Hughes, and Stanley
2010; Mouret 2009; Krcah 2010) is provocative and hints
at the importance of non-adaptive forces to the creativity of
natural evolution. It also provides a concrete example of how
non-adaptive forces can produce artifacts with the superficial appearance of adaptation. While natural selection may
be a powerful honing force, it may not be well-suited for creativity. In this way novelty search offers a controversial perspective on the key mechanism behind evolutionary search
in nature: Perhaps evolution at heart is more of a novelty accumulator than an optimizer, even if on surface their mechanisms are divergent. Interestingly, such results may also hint
at the potential of AI search algorithms driven by gradients
other than optimization towards an a priori objective.
HyperNEAT: Neural Connectivity as a
Function of Geometry
Development is a prominent feature of biological organisms that enables the large-scale nervous systems underlying
complex biological animal intelligence. Thus an important
biological question is what features of biological development enable compactly representing highly complex, functional, and evolvable neural networks? For example, the vast
complexity of the adult human nervous system is organized
over time through biological development and is efficiently
encoded by many fewer genes than there are neurons in the
final structure (Stix 2006). Furthermore, such neural development has been much modulated by natural evolution to
enable the wide diversity of complex animal lifestyles on
Earth. Identifying the most general laws underlying neural
development may facilitate human engineering of similarly
complex artificial neural networks (ANNs), which would
have profound implications for society.
However, when contemplating development, the danger is
that superficial aspects of the process might seemingly provide plausible answers to philosophical questions about the
deepest principles governing it. On its surface, development
generally begins with a single cell that then repeatedly splits
and differentiates into the multicellular adult form, e.g. a human adult develops over time from its zygote origins, with
the nervous system forming as a coupled subcomponent of
the overall developing organism. Notably, during development each cell exploits local signals from its neighbors to
92
This observation supports the idea that establishing nested
patterns are an important function of development.
Because nested coordinate frames appear to be an important feature of development, an interesting hypothesis is that
the formation of such nested frames may be the main abstract mechanism of development in biology. In other words,
development in nature may be only one way among many
to approximate an abstract mapping from geometry to form
through a series of nested functions, and such mapping may
be what enables the compact representation and significant
evolvability of organisms in nature, including the complex
brains of animals. While this hypothesis might appear tenuous in the context of biology alone, it is possible to support it through empirical investigation, i.e. through a radical
reimplementation of the hypothesized abstraction of development as a mapping from geometry to form.
One way to reimplement this abstraction that differs significantly from development in nature is explicitly to represent development as a function of geometry. This approach
is embodied by compositional pattern producing networks
(CPPNs; Stanley 2007). The CPPN is a variant of a traditional artificial neural network that composes a set of functions together in potentially complex nested ways to produce
a pattern when queried over some input geometry (e.g. a
two-dimensional coordinate frame). In this way, CPPNs can
represent complex patterns from nested coordinate frames in
a form much divergent from biological development.
Relevant to the biological hypothesis explored in this section, an approach facilitated by this alternate conception of
development, called HyperNEAT, can been seen as a radical reimplementation of neural development. In particular,
HyperNEAT evolves complex ANNs that are represented
by CPPNs. The main idea is that given an arbitrarily large
neural network embedded in a geometry (i.e. each node
is given a coordinate in space), a CPPN can represent the
strengths of connections between two nodes as a function
of the nodes’ coordinates. For example, if the nodes were
embedded in a two-dimensional geometry, then a CPPN
with four-dimensional input could encode the connectivity
among the nodes as a function of their positions. In this way,
the final complex ANN is created without local interaction
or iterative temporal unfolding. Thus the motivation is that
the underlying hypothesis about neural development’s root
mechanism is supported if over many domains HyperNEAT
facilitates computational evolution of functional ANNs with
biologically-similar regularities.
Interestingly, HyperNEAT has proved successful over
many domains, including many-joint robot arm control
(Woolley and Stanley 2010), real-world Khepera robot control (Lehman et al. 2013), quadruped locomotion (Clune
et al. 2011), checkers board evaluation (Gauci and Stanley
2010), and robocup soccer (Verbancsics and Stanley 2010).
Figure 2 shows representative ANNs evolved with HyperNEAT that demonstrate regular connectivities reminiscent of
biology. Thus the conclusion from its diverse applications is
that HyperNEAT appears to be generally well-suited for representing complex ANNs with natural regularities.
In this way credence is given to the original hypothesis that inspired the radical reimplementation: Neural devel-
Figure 2: Examples of Evolved HyperNEAT ANNs. A selection of complex artificial neural networks evolved by the
HyperNEAT method (reproduced from Clune et al. 2011).
The ANNs exhibit key biological regularities like symmetry
and modulated repetition.
opment may be more at heart about realizing connectivity
patterns of nested geometrical functions than temporal unfolding or local interaction. If other mechanisms were also
physically available to evolution, then if it were “run again”
the form of development realized in this second realization
might appear as radical to us as HyperNEAT appear when
compared to biological development. Or perhaps more practically, engineering approaches can exploit abstractions of
development like CPPNs (similarly to how helicopters exploit the principles of flight) to enable computationally representing and evolving complex ANNs. At a deeper level,
conceiving development as a set of functions of nested coordinate frames may broaden our understanding of the phenomenon and hint at a fundamental mechanism independent
of physics and time.
Discussion
It is often difficult for AI researchers, even those directly inspired by biology, to convince biologists of the relevance of
their models. Part of the difficulty is that biology generally
studies intelligence as it is, centered on the concrete. In contrast, AI is centered on the abstract and aims to investigate
intelligence in all possible incarnations. As the examples of
radical reimplementation reviewed here illustrate, focusing
on the abstract and divergent possibilities for life can yield
informative and thought-provoking insights. Thus a contribution of this paper is to demonstrate that models that are
deliberately biologically implausible may still have implications for understanding biological intelligence.
The key point is that it is possible for abstractions, even
those not specifically designed to illustrate biological principles, to be biologically relevant. It may be exactly because
novelty search and CPPNs were designed as practical means
to extend the state of the art in EC, and not to speak to biology, that they are biologically relevant. That is, the practicality of novelty search and CPPNs in many different contexts provides significant evidence for their driving abstractions, and the divergence of the implementations of these
abstractions from nature isolates the concept being tested.
In this way it is possible to interpret abstractions from
non-biological fields as providing evidence for biological
theories. Traditional EAs or novelty search are not direct
models of natural evolution, nor are CPPNs a direct model
of development. Yet traditional EAs offer an indirect means
of exploring the abstraction of evolution as an optimizer, and
93
Jacob, F., et al. 1977. Evolution and tinkering. Science
196(4295):1161–1166.
Krcah, P. 2010. Solving deceptive tasks in robot body-brain
co-evolution by searching for behavioral novelty. In ISDA,
284–289. IEEE.
Lehman, J., and Stanley, K. O. 2010. Efficiently evolving
programs through the search for novelty. In Proceedings
of the Genetic and Evolutionary Computation Conference
(GECCO-2010). ACM.
Lehman, J., and Stanley, K. O. 2011a. Abandoning objectives: Evolution through the search for novelty alone. Evol.
Comp. 19(2):189–223.
Lehman, J., and Stanley, K. O. 2011b. Novelty seach and the
problem with objectives. In Genetic Programming in Theory
and Practice IX (GPTP 2011). Springer. chapter 3, 37–56.
Lehman, J.; Risi, S.; D’Ambrosio, D.; and Stanley, K. O.
2013. Encouraging reactivity to create robust machines.
Adaptive Behavior.
Mouret, J.-B. 2009. Novelty-based multiobjectivization. In
Proc. of the Workshop on Exploring New Horizons in Evol.
Design of Robots,2009 IEEE/RSJ Intl. Conf. on Intelligent
Robots and Systems.
Orzack, S., and Sober, E. 1994. Optimality models and the
test of adaptationism. American Naturalist 361–380.
Pigliucci, M., and Kaplan, J. 2000. The fall and rise of Dr.
Pangloss: Adaptationism and the spandrels paper 20 years
later. Trends in ecology & evolution 15(2):66–70.
Raff, R. 1996. The shape of life: genes, development, and
the evolution of animal form. University of Chicago Press.
Risi, S.; Hughes, C.; and Stanley, K. 2010. Evolving plastic
neural networks with novelty search. Adaptive Behavior.
Stanley, K., and Miikkulainen, R. 2003. A taxonomy for
artificial embryogeny. Artificial Life 9(2):93–130.
Stanley, K. 2007. Compositional pattern producing networks: A novel abstraction of development. Genetic Programming and Evolvable Machines 8(2):131–162.
Sterelny, K. 1997. Universal biology. The British Journal
for the Philosophy of Science 48(4):587–601.
Stix, G. 2006. Owning the stuff of life. Scientific American
294(2):76–83.
Verbancsics, P., and Stanley, K. O. 2010. Evolving static
representations for task transfer. Journal of Machine Learning Research 1737–1763.
Woolley, B. G., and Stanley, K. O. 2010. Evolving a single scalable controller for an octopus arm with a variable
number of segments. In Schaefer, R.; Cotta, C.; Kołodziej,
J.; and Rudolph, G., eds., Proceedings of Parallel Problem
Solving from Nature, 270–279. Springer.
novelty search provides a means to investigate the power of
searches without pressure to achieve anything in particular.
In a similar way, the success of CPPNs in representing complex phenotypes in a wide range of applications provides
evidence for abstracting development as a product of nested
functions of geometry. Because abstractions are general, it
is not surprising that philosophical insight can bleed from
one field to another. However, the merit of radical reimplementation is that it suggests a principled approach to such
philosophical overlap.
Interestingly, much progress in science results from overturning assumptions widely accepted as fact. In this spirit
a new tool to investigate such assumptions about biology
may be important, especially when it is difficult to investigate them in a more direct manner. What principles underlying biological intelligence that are widely assumed essential
are in fact only incidental? What might separating these two
classes teach us about intelligence?
Conclusion
This paper defined the radical reimplementation approach
and illustrated how it can facilitate AI research achieving biological relevance. By coarsely abstracting a biological phenomenon and reimplementing it in a way maximally divergent from biological example, the validity of the abstraction
can be probed. In this way, abstractions about biology that
are divergently implemented in AI simulations can potentially be relevant to biologists. The strength of this approach
is that there is no need to defend the reimplementation itself
as biologically plausible or accurate; what is being tested is
the abstraction. In conclusion, radical reimplementation is a
new philosophical tool for investigating and isolating fundamental biological principles like animal intelligence, thereby
enabling AI researchers to ask provocative questions.
References
Bekoff, M., and Allen, C. 1997. Cognitive ethology: Slayers, skeptics, and proponents. Anthropomorphism, Anecdotes, and Animals: The Emperor’s New Clothes? 313–334.
Blount, Z.; Borland, C.; and Lenski, R. 2008. Historical
contingency and the evolution of a key innovation in an experimental population of escherichia coli. Proceedings of
the National Academy of Sciences 105(23):7899.
Clune, J.; Stanley, K.; Pennock, R.; and Ofria, C. 2011. On
the performance of indirect encoding across the continuum
of regularity. Evolutionary Computation, IEEE Transactions
on 15(3):346 –367.
Cybenko, G. 1989. Approximation by superpositions of a
sigmoidal function. Mathematics of Control, Signals, and
Systems (MCSS) 2(4):303–314.
Gauci, J., and Stanley, K. O. 2010. Autonomous Evolution
of Topographic Regularities in Artificial Neural Networks.
Neural Computation Journal 22(7):1860–1898.
Holland, J. H. 1975. Adaptation in Natural and Artificial
Systems: An Introductory Analysis with Applications to Biology, Control and Artificial Intelligence. Ann Arbor, MI:
University of Michigan Press.
94

Similar documents

×

Report this document