Content uploaded by Laszlo Gulyás
Author content
All content in this area was uploaded by Laszlo Gulyás
Content may be subject to copyright.
Evolutionary Technology and Phenotype
Plasticity: The FATINT System
George Kampis1,3and L´aszl´o Guly´as1,2
1Collegium Budapest, Budapest, Hungary gkampis@colbud.hu
2AITIA International Inc., Budapest, Hungary lgulyas@aitia.ai
3Department of Biology, ETSU, Johnson City, USA
Summary. This is a comprehensive summary paper of works done by the authors
in 2004-7 but never published as a single narrative. Our works address various as-
pects of an Evolutionary Technology. We present the challenge, the approach (using
simulated interactions from phenotype plasticity called ’fat interactions’) and the
first results. On the road, we develop and test a platform for the general agent-based
modeling of biological populations, which serves as the evolutionary engine behind
our FATINT system.
1 Introduction: Evolutionary Technology
The direct route to artifacts is via design, i.e. by specification and subsequent re-
alization. Evolutionary Technology, as defined in [21], looks for an indirect route.
The task is to produce an abundance of forms and functions of a practically endless
variety by means of evolutionary methods. This implies the twin challenges of the
’arrow of complexity’ and ’open-ended evolution’, i.e. of producing increasingly com-
plex machineries in the course of time, and doing that in a persistent, self-supporting
process propelled by entirely endogenous causes.
Open-ended evolution is widely recognized as a difficult and unsolved problem.
John Holland, the founder of Genetic Algorithms has recently asked: ”Can we pro-
duce an existence-proof model, akin to von Neumann’s model of self-reproduction,
that exhibits open-ended evolution, with increasing diversity and complexity? ” [16].
Also, the ’arrow of complexity’ hypothesis is challenged from several directions ([2]),
for a discussion see [22]. Currently there is no accepted general evolutionary theory
for the origin of complexity or the maintenance of evolutionary change. Recent at-
tempts to overcome one or both difficulties come from Artificial Life (such as the
AVIDA system, [1][4]) and robotics such as the renowned GOLEM project [32] or
swarm-based developments [27].
In the above terms, we understand Evolutionary Technology as an attempt to
realize open-ended evolution showing complexification and to harness it for tech-
nological needs. The work reported in this paper deals with the first part of this
problem (i.e. the growing space of functionalities and attributes), leaving issues of
controllability aside.
2 George Kampis and L´aszl´o Guly´as
2 Species, Niches, and Phenotypes
Our approach extends earlier work on evolving populations, such as Tierra, Avida,
or Echo [35] [4] [17]. We maintain that in order to achieve open-ended evolution, the
understanding of the structure and dynamics of evolving populations is crucial, and
this is incompletely achieved with the current tools. This motivates our efforts to
develop a comprehensive system with automatic observers for population modeling.
Our open-ended evolution model is built on top of that.
Two fundamental concepts of biological populations that serve as the ground-
work for our models are species and niches. To put it shortly, we attempt to build
new species that recursively construct and fill new niches, giving rise to more species
and more niches in the course of time. The third component of our approach is phe-
notype plasticity arising from interaction.
In the simplest definition, a species is a set of co-reproducing organisms repro-
ductively isolated from others (’biological species’ concept, Fig. 1, [23]). Speciation,
the process by which species give rise to more species, is difficult to achieve in both
natural and model populations. A particular mechanism considered sufficient is sex-
ual selection, e.g., the dynamic change of mating preferences [10] [20]). The current
theory about the origin of the giraffe ([37], although debated, [3]) assumes that it is
male fighting (called sparring or ’necking’) that was the key to the giraffe’s long neck
and to the separation of a species, the pre-okapi, into the modern giraffe and the
okapi. Valid or not in the end, the giraffe story feeds on the power of sexual selection,
and is closely related to the idea of phenotype plasticity as we use it below.
Fig. 1. Example species in a topological (Fruchterman-Reingold) plot, left; a re-
alistic species from a simulation, right. Nodes denote individuals, edges represent
pairwise reproduction ability. Species are identified as connected components
Niches are segments of a (multi-dimensional) resource space that can be occupied
by a species ([18] [14]). Niche construction is a recently advocated theory of evolution
that builds on mutual organism-environment interactions ([33] [29]) [12]). In niche
construction, the central notion is the organism’ activity that alters the (biotic or
abiotic) environment in such a way which in turn can induce changes in the given (or
a different) organism. A simple example is ’ecological inheritance’: a new generation
of organisms inherits not just genes but also a specifically altered environmental
pressure left behind by the earlier generations.
Evolutionary Technology and Phenotype Plasticity: The FATINT System 3
Known theoretical models for niche construction tend to study consequences
of organism-environment feedback on genetic composition (i.e. they are ’replicator
level’ models, in Dawkins’ [6]terms). Laland et al. [28] considers a fixed genome
two locus model where one locus operates controls on environmental states and the
other suffers consequences via environmental selection. Suzuki and Arita [38] use
a modified Kauffman NKES model [26] to analyze alternative strategies of genetic
adaptation and the induction of environmental change. As was first pointed out,
however, by Lewontin [30][31] and later others (e.g. Dusek [11]), a niche construction
process is the product of an underlying physical interaction between the organism
and its environment. Lewontin’s own example is a bug that lives on the upper half
of a leaf. He discusses how the bug can alter the evolutionary environment for its
descendants by a variety of selected actions, as simple as moving to the other side of
the leaf. A flexible phenotype expression (that adheres to a reaction norm) has the
upper hand in such a process. This places the interactor, i.e. the phenotype, into
the focus of interest.
The third source, phenotype plasticity, is one of the oldest ideas in evolution
[25]. Yet until recently, little attention was paid to it. Phenotype plasticity can take
various forms that includes, but is not limited to, function change ([5] [36]), exapta-
tion [13], or ’tinkering’ (bricolage, [19]). In our works the concept is used in a specific
sense, for which the giraffe example offers a case. Here, unused traits (such as the
length of a neck) become significant for interaction. This bears upon the very pheno-
type definition. Narrowly taken, the phenotype denotes those organismic properties
that are relevant for an ecological-evolutionary context at a given moment [9]. We
will consider organisms with changing phenotype definition. We treat phenotypes as
variable entities we call fat phenotypes [22]. A fat phenotype is essentially a multi-
ple phenotype that permits different, ecologically defined interactions. We note that
every natural phenotype is ’fat’ in the sense that the number of potential phenotype
traits is usually very large compared to the focal/functional (’active’) phenotype
properties at a moment (Fig. 2)
Linking together Evolutionary Technology (i.e. open-ended evolution) with the
species concept, niche construction and plastic phenotypes, we advanced a general
Phenotype Hypothesis, PH, [22]. It predicts that no evolution model and no evo-
lutionary technology/ALife design will be able to produce a sustained evolutionary
process or support the increase of functional complexity without admitting a chang-
ing phenotype space as a natural part of the interaction potential of the organism.
In this paper we deal with testable consequences of PH. We will consider emer-
gent species that occupy dynamically created niches based on phenotype interaction
and examine two Hypotheses.
Hypothesis1. An evolutionary engine with phenotype-based emergent species and
without phenotype plasticity remains stable and non-productive of species.
Hypothesis2. The introduction of phenotype changes in the above systems leads
to a sustained production of new emergent species.
4 George Kampis and L´aszl´o Guly´as
Fig. 2. Fat phenotype (P). A focal/functional phenotype P0is constructed by genes
(Gi) and environment (Ej) together
3 The FATINT System
To address the above crucial issues of phenotype representation and to study their
effects in maintaining open ended progress, we perform an extended study using
digital organisms. We are developing and using the FATINT system (the name
derives from Fat Interaction and hence from fat phenotype), which is part of the
EvoTech project run at Collegium Budapest and E¨otv¨os University. We reported
early work on this system in [20] [22] [23] and elsewhere. The EvoTech home page
is found at www.evotech.hu, with downloadables, papers, and technical reports.
The FATINT system was written in a consistent fully encapsulated agent-based
approach [15], using the Java based REPAST environment developed by the Univer-
sity of Chicago [39]. The FATINT system can also be compiled into a stand-alone
application. It offers a graphical user interface (Fig. 3) allowing for a wide selection of
models with a broad range of interactive experiments using different parameter set-
tings. The system is equipped with a high number of automatic (runtime) observers
that include a metric distance plot, various histograms, population plots (such as
number and age distributions, and programmable 2D projections), species graphs
(using Fruchterman-Reingold and Kamada-Kawaii graph algorithms) and others.
For simplicity we fix the representation of the domain as follows. Each agent has
its phenotype represented as a vector viof integers from the interval [Vmin, Vmax].
The length of the vector is Land is always identical for all agents but variable in
time. Formally, the phenotype of agent iis:
vi∈[Vmin, Vmax ]L
Fat phenotypes are represented in various ways as associated with either growing
dimensions (i.e. variable length vectors) or as dimension switches (using silent and
active values of vector entries). We experimented with both solutions (and some
others [22]).
To accommodate phenotypic plasticity, note that the dynamic notion of pheno-
type applied here is global. Our phenotype definition is global in the sense that,
if a given phenotype trait in organism Aenters a particular interaction, then the
Evolutionary Technology and Phenotype Plasticity: The FATINT System 5
Fig. 3. Snapshot of FATINT’s graphical user interface.
same trait typically also appears for organism B. (If one giraffe fights with a neck,
others should defend themselves using theirs.) The notion is also dynamic in that it
permits the change of the phenotype during the lifetime of an organism.
Table 1 presents different phenotype transformations [20], their causes (inter-
nal or external) and extension (local or global), corresponding to different types of
genotype-phenotype maps (GPMs). In the current version of the system, we tested
two families of GMPs: a type-dependent and a type-independent one [20] [22]. In
the first, phenotype change occurs in an identical way for all organisms that share
a given genotype, whereas in the second, this is not the case (the two situations
approximately correspond to the first and the last rows of Table 1. We return to the
issue in Section 6.
Experiments were conducted using a basic evolution engine, the task of which
was to maintain a stable population before developing new species, as well as its
extension with a phenotype change support, which is an independent system com-
ponent.
4 The Evolution Engine
In FATINT we currently use sexually reproducing gender-less (i.e. ’snail-like’) agents,
which represent individuals. Unless stated differently, genotypes and phenotypes
are always assumed to coincide; however, the model’s behavior is based on the
phenotype interpretation. Reproductive success depends on a phenotype similarity
metric which introduces a sexual selection. The resulting selection force is variable
6 George Kampis and L´aszl´o Guly´as
Table 1. Mechanisms of phenotype change
Form Cause Type
Point mutation endogenous local
Phenocopies exogenous partly global
Epigenetic change both partly global
Horizontal adaptation both global
Behavior change exogenous global
(any individual can become the ’center’ of a species) yet well-defined (i.e. once a
center is formed, ’distant’ individuals are selected against). In the FATINT system,
a species emerges as a dynamically maintained cluster of interbreeding individuals,
around a dynamically defined center characterized by the predominant types.
At the base level the system uses an evolution engine that consists of a popula-
tion of agents using a single non-replicating resource (energy), as well as usual evo-
lutionary operators for crossing over, mutation, resource uptake and ageing/death.
In addition to having a phenotype, agents only have the minimum of properties: age
and accumulated energy.
Each organism has an equal chance to ’eat’ at every time-step. Note that this
nevertheless introduces an implicit competition for energy, which leads to density-
dependent effects (genotypes in higher numbers obtain a larger share of energy).
A consequence of the above implicit competition is competitive exclusion [23] [8],
that is, in the long run only one single species can survive. This puts an important
control on our open-ended experiment in the present form and serves as a kind of
negative feedback.
The evolution engine uses a quasi-parallel activation regime. Each agent acts
once per every time step in a dynamically randomized order. The agents’ activity
consists of the steps of energy intake, energy consumption, and reproduction. The
agent first seeks Ein units of energy in the shared environment, and, depending
on the amount available, it receives ein units (possibly 0). The efficiency of energy
intake decays with age:
eaccumulated =eaccumulated +ein ×(Ediscounting)ag e
where 0 < Ediscounting <1. Next, the agent consumes Econsumption units of
its accumulated energy. If it does not have this amount available, the agent dies.
Surviving agents attempt to reproduce with probability Pencounter (i.e. every time
step produces part of an entire new generation). Updating the shared environment
completes the iteration. This means the addition of Eincrease units to the energy
pool.
In reproduction an active agent picks a random mate from the list of potential
partners, which are individuals similar enough to bear an offspring with the given
agent. Similarity is measured by the Euclidean distance between the agents’ phe-
notype vectors. An advantage of this metric is that it is dimension independent. It
allows for sexual selection to occur in the same way between any selected phenotype
pairs of arbitrary dimensionality. (This choice has also known side effects. To test
them, in [24] we discussed three different families of similarity operators, but no
fundamental difference in speciation experiments were found).
Evolutionary Technology and Phenotype Plasticity: The FATINT System 7
Given two parents, and similarity ddefined as above, the number of offsprings
is Mconst + (Mlimit −d)Mslope (for d > Mlimit). New agents inherit the active par-
ents phenotype, except for mutation and crossing over that occur with probabilities
Pmutation and Pcrossing , respectively, per gene. Mutation shifts the value of a gene
by a random value in [−Vmutation,+Vmutation ]. If the mutated value falls outside
the interval [Vmin, Vmax ], the offspring is dropped.
Table 2. Default parameter settings
Vmin = 0 Mlimit = 15
Vmax = 100 Mslope = 0
Pencounter = 0.1Mconst = 1
Pcrossing = 0.2Econsumption = 5
Pmutation = 0.1Ein = 10
Vmutation = 2 Ediscounting = 0.9
Eincrease = 1000
5 Behavior of the Basic Evolution Engine
First we ask what happens if we initialize the basic system with a randomly cho-
sen set of initial conditions and let it run with the default parameters or different
parameter settings. Our results are summarized in Fig. 4.
On the Figure we see that the evolution engine, left alone, converges to a single
stable species and maintains that indefinitely under a very wide range of parameters
and initial conditions [20]. Initial conditions were defined as randomized populations,
i.e. where each individual had a randomly defined genotype and phenotype. Fig. 4
(above left) summarizes 10 different runs with different random seeds. The solid line
show the average of ten runs, grey indicates the envelope of areas covered by any run.
It is seen that the convergence speed is a function of the initial conditions but other-
wise convergence and the subsequent maintenance of a single species is unequivocal.
From other representations, e.g. distance plots or histograms (not shown here) we
can conclude that each random initial condition yields a different surviving species
with a different emergent center in phenotype space. This typical behavior shows
the development and stability of biological species based on phenotype-to-phenotype
interaction alone.
After demonstrating this typical behavior we want to test its limits. To that end
first we varied encounter frequency, a parameter with a strong effect on reproduction
and hence long-term survival (Fig. 4, upper right). Each solid line shows an average
of 10 runs with different initial conditions. As expected, low values of Pencounter
make reproduction difficult or impossible; for values higher than 0.07 the typical
behavior is found; no speciation events occurred at any value. Varying Pmutation
(Fig. 4, lower left) had little effect. This counter-intuitive outcome is due to the fact
that speciation in the model population is not based on external fitness values but
relational (i.e. pairwise) reproduction events among organisms. Such events are not
8 George Kampis and L´aszl´o Guly´as
easily facilitated by single distant mutants, and as the next experiment shows, news
species must typically emerge from within the hyper-sphere containing the currently
predominant or ’old’ species. Fig. 4 (lower right) shows the effect of varying crossing
over. At high frequencies of crossing over, a high internal variability is produced
within the diameter of the existing species, and this leads to several new pairwise
matches among reproduction candidates, resulting in a temporary instability of the
surviving species.
Fig. 4. Number of connected components (vertical axis) against time. From top-left:
basic engine behavior with default parameters; effects of varying Pencounter(in the
range 0.05 −0.095), Pmutation (0.10.5), and Pcrossing (0.10.8). 10 different random
seeds; 6,000 steps
To answer questions about the ’reality’ of species found by the method of con-
nected components above, in a study involving three different automatic species
observers [23] we demonstrated the robustness and realistic behavior of the evolu-
tion engine. Fig. 5 (left) shows convergence to a single cluster in phenotype space
(’phenetic species’), using a standard hierarchical clustering method, characterized
with a clustering constant dcluster. We express dcluster here as a ratio to Mlimit: at
dcluster = 1 we ask about clusters of size Mlimit .
Fig. 5 (left) demonstrates that even at small values (e.g. dcluster = 0.9) the
same stable convergent behavior as in Fig. 4 is found. In intuitive terms, this means
that the ’diameter’ of the species in n-space is about as small as Mlimit in the long
run or smaller (or otherwise, smaller values of dcluster should split the population
into different phenetic species). Using a cladistic species concept [23] that asks for
lineages we obtain a similar picture (Fig. 5, right): all individuals in the long run
stem from one single lineage. A similar coincidence of the various species concepts
will be demonstrated (by calculating the Rand index [34] of various classifications)
during the speciation experiments, except at the time of the very speciation events.
The FATINT evolution engine also permits further studies such as drift, Hardy-
Weinberg equilibrium, density-dependent effects and other phenomena which are
well understood at population level (i.e. mean field) models but require further
study in the present individual-based framework (which we began in [8] [7]).
Evolutionary Technology and Phenotype Plasticity: The FATINT System 9
Fig. 5. Convergence. Left: clusters against time. Summary of 60 runs: 10 for 6
random seeds each, from dcluster = 0.9 to 1.4. Time goes left to right, dcluster
increases right to left. Vertical axis shows number of clusters. Right: cladistic analysis
of the same 6 populations. Clades up to a given time. Time as before, vertical axis
shows clade numbers
6 New Species and Functions from Phenotype Change
To test the effects of the phenotype changes, in FATINT’s phenotype change mode,
at every time step a phenotype change can be initiated by hand or, alternatively,
at every reproduction event a new phenotype dimension can be introduced with
a variable probability Pchange per offspring. When such an event occurs, a new
component is added to the agents’ phenotype vector. The particular phenotype value
an agent receives depends on the used method (we call ’stretch’ method: the name
refers to the degree new phenotype values are stretched along the new dimension, a
factor that will prove to be important).
As alluded to in Section 4, two basic methods were applied (genotype-based, and
genotype independent). The type-based stretch method calculates the value from the
agent’s old geno/phenotype. For simplicity, only the value vof a single dimension
(the last one) is used:
vnew =Vmin + (v×Vstretch)mod(Vmax −Vmin + 1)
where Vstretch is a positive parameter. The type-independent method selects a
uniform random value from [Vmin, Vmax].
To understand the effect of phenotype changes discussed in Section 3, we first
study the hand-introduction of a phenotype change. The results for a typical run
are shown in Fig. 6.
An initially compact and homogeneous species becomes structured and, in the
course of time, this structure is amplified to give rise to two new species. Structuring
occurs in terms of reproductive links formed and broken: i.e. the effect of phenotype
change is to introduce new differences and also new similarities among organisms.
Due to repeated success and failure of reproduction events both of them enter a
positive feedback loop (Fig. 6, middle panels) which results in an amplification
of the changes. Note, however, that the basic engine behavior as analyzed before
acts against this phenomenon, and so the result can depend on several, usually
random factors. Hence, speciation is bound to be a contingency-driven effect for
10 George Kampis and L´aszl´o Guly´as
Fig. 6. Phenotype plasticity induced speciation in a sample run. New phenotype is
introduced at left. First a homogenous reproduction network becomes structured,
which is further amplified. After several generations, the process leads to segregation
and the emergence of new reproduction centers. Positions are illusory in this topology
plot.
which phenotype change appears to be a necessary, but itself not sufficient, condition
in our model.
We experimented both with the type-dependent and the type-independent
method (Fig. 7, right panel). We found in each case an increased heterogeneity
of the system along the new phenotype dimension (see Fig. 7 bottom left for the
type-dependent case). The new dimension is represented in the rightmost histogram.
In line with our analysis of the stability of the basic engine in Section 5 we see sev-
eral new candidate species centers arising (Fig. 7 top left), most of them within
the diameter (usually: within the Mlimit) of the original species. Later dampening
and amplification events tend to sort out the majority of these candidates, or all of
them, leaving one (the original) or two species (the old and a new one) as survivors.
We observed rare 3-speciation events as well. The new species are characterized by
new species centers and hence different functionalities (i.e. different sexual selection
preferences).
We want to understand the process of phenotype change and the speciation tran-
sition more closely. Species analysis tools presented in [23] as well as the histograms
(Fig. 7) provide a cue. Using either method, new phenotype traits turn out to be
highly variable. For example, a sample run produced the comparisons seen in Ta-
ble 3. Stable species situations before speciation are characterized by a Rand index
R= 1 for all pairwise comparisons of the species metrics ([23]). Around the time
of the introduction of a new phenotype trait (t= 450), clusters formed using phe-
netic (similarity-based) and cladistic (lineage-based) classifications strongly diverge
from each other (whereas the former correlate well with the clusters of reproduction
ability under a biological species notion), showing that phenotype properties are
Evolutionary Technology and Phenotype Plasticity: The FATINT System 11
Fig. 7. Left panel: allele histograms at the time (bottom) and after a phenotype-
driven speciation event (top). Histograms show the population distribution of alleles
for a given gene. For modeling reasons, the picture shows a 5 gene 100 allele situation
and its transition to 6 genes or phenotype traits. Right panel: cartoon fitness func-
tions(top) and empirical distributions (bottom) before and after phenotype change
for the type-independent (bottom left) and type-dependent (bottom right) model
more different than are lineages: the same lineage supports a number of different
phenotypes. The situation changes after species segregation has taken place (around
t= 550). By the time the two new species are completely formed (t= 1000) the
same conditions as before speciation (R= 1) are restored.
This high variability of new phenotype traits at the time of transition is a sim-
ple mathematical consequence in the model and a natural assumption about real
populations, where unused properties do not experience stabilizing selection. As a
consequence, the population diameter increases along these dimensions and thus
when a new dimension is activated new close matches (new types of reproduction
events) become possible.
To summarize: the typical process upon the introduction of a new phenotype
trait is the appearance and amplification of phenotype heterogeneity which often
leads to the appearance of new emergent species centers ’strong’ enough to facili-
tate the separation of the formerly stable species into new species showing different
functionalities (in the limited sense of different sexual selection forces).
Table 3. R index of biological (BSC), phenetic (PSC) and cladistic (CSC) species
at the time of a speciation event starting at t= 450
Time interval PSC vs. BSC PSC vs. CSC BSC vs. CSC
450-550 R = 0.716 R = 0.356 R = 0.640
515-550 R = 0.716 R = 0.716 R = 1.0
550-1000 R = 1.0 R = 1.0 R = 1.0
12 George Kampis and L´aszl´o Guly´as
7 Autonomous Speciation Experiments
In the autonomous phenotype change mode, as noted above, at every reproduction
event a new phenotype dimension is introduced with probability Pchange per off-
spring. To study the effect of this behavior we introduce nonzero values of Pchange
(Table 2). The findings are summarized in Figure 8.
The Figure shows that the autonomous mode of FATINT typically produces
sustained speciation events (see Figure 8, left, and [23]). We find repeated specia-
tion and subsequent relaxation. A developing species can be stable for hundreds of
generations but tend to collapse afterwards due the factors mentioned in Section
5. If, before a complete relaxation repeated phenotype transitions occur, this can
drive the system into a sustained multi-species state (Fig. 8 left) or into a runaway
condition (Fig. 8 bottom right). To the interpretation of the Figure, we note that
the typical population size using the default parameters was around a few hundred
individuals (the batch experiments are computation intensive and were performed
on a Beowulf cluster taking typically a week). Population size puts a serious limita-
tion to speciation: a long term survival of 10 species means that an average of a few
dozen individuals are allocated per species. Turning things around, we get that no
more than 10 species can meaningfully exist for a longer time. Another important
limiting factor is deliberately introduced density dependence (Section 3) which puts
an important control by making the system conservative (i.e. dominated by negative
feedback). This results in a short lifetime of the generated new species and forms a
condition that can be dropped by using different resource models in the future ([8]).
Fig. 8. Left panel: sustained speciation under various values of Pchange . Right panel:
top, decreased levels of Mlimit (5−15) disrupt species. Bottom, increasing phenotype
modification (Vstretch ,1−20) leads to species proliferation. Black line: average, grey
lines, error bars, 10 runs. Time goes to right, species numbers are on vertical axis.
Figure 8, left, delivers the main results of our simulations. It shows that pheno-
type plasticity can drive an otherwise stable system into speciation repeatedly and
push species numbers to high limits marked by population size and other factors.
We also wanted to study the behavior of the autonomous system under various
parameters. Two important parameter choices are presented in Figure 8, right panel,
otherwise using default parameters. As before, solid lines represent averages over 10
Evolutionary Technology and Phenotype Plasticity: The FATINT System 13
runs, grey cover the area spanned by all runs. (We omitted here the studies of
mutation and crossing over which produce similar results as seen before; this and
other items of sensitivity analysis are found in [20]). At the top, we see experiments
with varying Mlimit. As expected, this is a sensitive parameter and its low values
can artificially disrupt species, a method with a further potential to exploit in future
modeling studies (e.g. evolving Mlimit in the course of the simulation). Increasing
Mlimit alone does, however, not form species having well-defined new properties
(and consequently, stable high values of R). The other parameter tested is Vstretch
which is responsible for the variability of new phenotype traits. Clearly, low (or zero)
values decrease or cancel any effect from phenotype change, whereas high values lead
to a dramatic increase of variability. Correspondingly, we find runaway conditions
in speciation for high values of Vstretch (Figure 8, bottom right).
8 Conclusions and Future Works
We presented the development and results of an interactions-based evolutionary
model (in progress for more than 3 years). The model uses similarity-based sexual
selection to approach open ended evolution.
Answering Hypothesis1, we have shown that a phenotype based system shows
species convergence and stability under the specified conditions. To test Hypothesis2
we demonstrated that changing phenotype-to-phenotype interactions can repeatedly
split species by the production of new selection constraints. We hypothesized that
sustained ecological evolution always proceeds by similar mechanisms. In FATINT,
the presented speciation process can produce new functions (i.e. new sexual selec-
tion processes) and can go on indefinitely using different mechanisms for phenotype
change as genetic point mutations (our type-dependent method) or behavior change
(type-independent method).
Topics that go beyond the limits of this paper include competition avoidance [22],
the analysis of the role of positive and negative feedback in open-ended evolution
[23], and our first results on emergent ecosystems using various genetic models and
genotype-phenotype maps [8][7].
Finally we ask about the relationship to the distant goal of evolutionary tech-
nology. The presented works clearly signify first steps in this direction, addressing
what appears to be a main bottleneck, namely the understanding of the role of
material structures (i.e. ’embodiment’). Our results imply that fat phenotypes may
play a key role in pushing speciation forward under specific conditions. A next step
would be to extend these results to functional differentiation (e.g. ecosystems) and
to the study of body plans (e.g. using evolvable genotype-phenotype maps).
Acknowledgments
This work was supported by the EC grant QosCosGrid IST FP6 #033883. The
authors wish to thank the hospitality of Collegium Budapest for the period of this
study. Part of the tests were carried out on the grid testbed of the Poznan Super-
computing Center (PSNC). L.G. acknowledges the partial support of the GVOP-
3.2.2−2004.07 −005/3.0 (ELTE Informatics Cooperative Research and Education
14 George Kampis and L´aszl´o Guly´as
Center) grant of the Hungarian Government. During a period of the work G.K. was
Wayne G. Basler Chair at East Tennessee State University (ETSU), TN. The above
supports are gratefully acknowledged. The authors thank Walter de Back for helpful
comments and suggestions. G.K. also wishes to thank comments and support of Dr.
Istv´an Karsai of ETSU and the ’Bunker Enterprises’ for food, music and hospitality.
References
1. Ch. Adami, Ch. Ofria, and Collier. T.C. Evolution of biological complexity.
Proc. Nat. Acad. Sci. USA, 97:4463 – 4468, 2000.
2. Mark A. Bedau. The arrow of complexity hypothesis today. In Evolution of
Complexity Workshop, AlifeX., 2006. http ://ecco.vub.ac.be/EC O/Bedau.pdf .
3. Elissa Z. Cameron and Johan T. du Toit. Winning by a neck: Tall giraffes avoid
competing with shorter browsers. The American Naturalist, 169:130135, 2007.
4. S.S. Chow, C.O. Wilke, Ch. Ofria, R.E. Lenski, and Ch. Adami. Adaptive
radiation from resource competition in digital organisms. Science, 305:84–86.,
2004.
5. Ch. Darwin. The Origin of Species. 6th edition., 1872.
6. R Dawkins. The Extended Phenotype: The Long Reach of the Gene. Oxford:
Oxford University Press., 1982.
7. Walter de Back, L´aszl´o Guly´as, and George Kampis. Niche differentiation in
various multi-resource ecosystems. In ECCS, 2007. (submitted).
8. W.N. de Back, L. Guly´as, and G. Kampis. Niche differentiation and coexistence
in a multi-resource ecosystem with competition. In European Conference on
Artificial Life 9, 2007. (submitted).
9. Th. Dobzhansky. Nothing in biology makes sense except in the light of evolution.
he American Biology Teacher, 35:125–129, 1973.
10. G. S. Van Doorn and F. J. Weissing. Ecological versus sexual selection models
of sympatric speciation: A synthesis. Selection, page 1740, 2001.
11. V. Dusek. Lewontins living legacy: Levels of selection and organismic construc-
tion of the environment. Human Nature Review, 2:367374, 2002.
12. P. Godfrey-Smith. Organism, Environment and Dialectics, volume Thinking
About Evolution: Historical, Philosophical and Political Perspectives. CUP,
Cambridge, 2001.
13. S.J. Gould. Ontogeny and Philogeny. Harvard University Press, Harvard, Mass.,
1977.
14. J. R. Griesemer. Niche: historical perspectives, volume Keywords in evolutionary
biology, pages 231–240. Harvard University Press, Cambridge, Massachusetts,
USA., 1992.
15. L. Gulyas. Understanding Emergent Social Phenomena: Methods, Tools and
Applications for Agent-Based Modeling. PhD Thesis, Etvs University, Budapest.
http://hps.elte.hu/ gulya/Research/Thesis/index.html, 2004.
16. John Holland. http://www.nd.edu/ swarm03/keynote/keynote.html. Technical
report, 2003.
17. John. H Holland. Hidden Order. Helix Books, 1995.
18. G.E. Hutchinson. An Introduction to Population Ecology. Yale University Press,
New Haven, Connecticut, USA., 1978.
19. F. Jacob. Evolution and tinkering. Science, 196:1161–1166., 1977.
Evolutionary Technology and Phenotype Plasticity: The FATINT System 15
20. G. Kampis and L. Gulyas. Sustained evolution from changing interaction. In
Proceedings of ALife IX,, page 328333. MIT Press, Boston, 2004.
21. G. Kampis and L. Gulyas. Emergence out of interaction: A phenotype based
model of species evolution. Advanced Engineering Informatics Journal, 20:313–
320, 2005.
22. G. Kampis and L. Gulyas. Full body: The importance of the phenotype in
evolution. In Workshop on the Evolution of Complexity, ALifeX. (also, Artificial
Life, submitted)., 2006.
23. G. Kampis, L. Gulyas, and S. Sos. The species problem in artificial life. In First
IEEE Conference on Artificial Life. Honolulu, HI., 2007.
24. George Kampis and Laszlo Gulyas. Phat phenotypes for agents in niche con-
struction. In Artificial Life X: Proceedings of the Tenth International Conference
on the Simulation and Synthesis of Living Systems (Bradford Books), 2006.
25. I. Karsai and G. Kampis. Phenotype plasticity: its history, significance and
possible impact. (manuscript), 2007.
26. S. Kauffman. The Origins of Order: Self-Organization and Selection in Evolu-
tion. Oxford University Press, 1993.
27. S. Kornienko, Olga Kornienko, and Paul Levi. Generation of desired emergent
behavior in swarm of micro-robots. In ECAI 2004, pages 239–243, 2004.
28. K. N. Laland, F. J. Odling-Smee, and M. W. Feldman. The evolutionary conse-
quences of niche construction: A theoretical investigation using two-locus theory.
Journal of Evolutionary Biology, 9:293316, 1996.
29. K.N. Laland, J. Odling-Smee, and M.W. Feldman. Niche construction, biological
evolution and cultural change. Behavioral and Brain Sciences, 23:131146, 2000.
30. R. Lewontin. The organism as the subject and object of evolution. Scientia,
118:6582, 1983.
31. R. Lewontin. The Triple Helix: Gene, Organism, and Environment. Harvard
University Press, Boston., 2000.
32. H. Lipson and J. B. Pollack. Automatic design and manufacture of robotic
lifeforms. Nature, 406:974–978., 2000.
33. K.N.; Odling-Smee, J.; Laland and M.W. Feldman. Niche Construction: The
Neglected Process in Evolution. Princeton, NJ: Princeton UP., 2003.
34. W. M. Rand. Objective criteria for the evaluation of clustering methods. Journal
of the American Statistical Association, 66:846850, 1971.
35. T. S. Ray. An approach to the synthesis of life. In C. Langton, C., J. D. Taylor,
Farmer, and S. Rasmussen, editors, Artificial Life II, volume XI of Santa Fe
Institute Studies in the Sciences of Complexity, pages 371–408. Redwood City,
CA: Addison-Wesley, 1991.
36. R. Rosen. On the Generation of Metabolic Novelties in Evolution, volume Bio-
genesis, Evolution, Homeostasis, pages 113–123. Springer, Berlin., 1973.
37. Robert E. Simmons and Lue Scheepers. Winning by a neck: Sexual selection in
the evolution of giraffe. The American Naturalist, 148:771–786., 1996.
38. R. Suzuki and T. Arita. How niche construction can guide coevolution. In
Proceedings of Eighth European Conference on Artificial Life, page 373382. MIT
Press (Bradford Books), 2005.
39. E. Tatara, M.J. North, T.R. Howe, N.T. Collier, and J.R. Vos. An introduction
to repast modeling by using a simple predator-prey example. In Proceedings of
the Agent 2006 Conference on Social Agents: Results and Prospects, Argonne
National Laboratory, Argonne, IL USA., 2006.
