The Origin of Mimicry

Deception or merely coincidence

Bram Wiggers, s2485834, [email protected],

Supervisor: Dr. Harmen de Weerd

Abstract: One of the most remarkable phenomena in nature is mimicry, the ability of one
species (mimics) to imitate the phenotype of another species (models). Several reasons for the
origin of mimicry have been proposed, but no definitive conclusion has been found yet. In this
paper, different hypotheses will be tested by an agent based co-evolution model. We consider
an adaptive plane with a camouflage and toxic gradient. Animal populations can move over this
adaptive plane by evolving. Two options will be compared: (1) Deception, which assumes that
mimics evolve to take advantage of the warning color of the models. (2) Coincidence, where
models evolve a warning color that coincidentally benefits the mimics. We found that both the
hypotheses are a plausible reason for the origin of mimicry, but that deception takes place in the
coincidence environment as well. When the model loses its camouflage, the mimics do not move
toward the model, but stay in their own place, which supports the idea of coincidence.

1 Introduction (Rabosky, Cox, Rabosky, Title, Holmes, Feldman,

and McGuire, 2016) and other animals (Maan and
One of the most remarkable phenomena in nature Cummings, 2011). Various research has been done
is mimicry, the ability of animals (mimics) to mimic (Mallet and Joron, 1999), but no clear answer has
the phenotype of other animals (models). Typically, been found for the origin of mimicry.
the effect is called mimicry when the mimic imitates
a phenotype of a model which can be dangerous We distinguish Batesian mimicry and Müllerian
for a predator. For the mimic, mimicry has a pos- mimicry. In Batesian mimicry the mimic is not
itive effect on the chance of survival, since it gets toxic, which is disadvantageous for the model. The
confused for the dangerous model. For the models, mimic will be eaten or tasted by the predator and
it might be beneficial (Müllerian, Müller (1879)) found to be harmless, giving a positive feedback
or disadvantageous (Batesian, Bates (1862)). Sev- stimulus to the predator since the mimic is not poi-
eral reasons have been proposed for the origin sonous. This results in models being eaten more,
of mimicry, but no definitive conclusion has been since the predators have found positive feedback
found as yet. on eating these colored animals. Hence, the more
The first one to discover mimicry was Bates mimics exist in the habitat of the model, the lower
(1862). Bates found there are animals that are poi- the survival chance of the model. This results in a
sonous with very bright colors, and camouflaged an- negative or parasitical effect on the model.
imals which were not poisonous. The effect of hav-
ing a bright color to scare off predators is known as Müllerian mimicry on the other hand, considers
aposematism. This effect became more remarkable two animals that are both toxic to a certain de-
when Bates found animals that had similar colors gree, and therefore enhance each others aposema-
and shapes as the toxic animals, but were not toxic tism. This will be explained by taking animals A
at all. and B, which both are toxic and have a similar phe-
This type of mimicry has been called Batesian notype. Each animal which starts to look more like
mimicry, the effect where non-toxic animals imi- the other one will be better remembered, and there-
tate the toxic animals. This effect can be found fore have a better chance of survival. As a result an
in butterflies (Bates, 1862; Müller, 1879), snakes evolutionary co-operation mechanism is created.

1
1.1 Camouflage and toxicity in their evolutionary progress. Not all animals get
eaten when predators learn of their toxicity. Preda-
In mimicry we can distinguish between preys and
tors tend to taste their prey first, before consum-
predators, where the predators are hunting the
ing them whole. Therefore preys do not have to die
preys. The preys have two defense mechanisms
when they are tasted, but it is not beneficial for
against the predators: (1) camouflage and (2) tox-
their health either. The evolution of toxicity is also
icity.
explained by kin selection. When a species is more
toxic, the group as a whole will be eaten less and
1.1.1 Camouflage have a benefit. As an individual it is less beneficial
The camouflage of a prey can be beneficial to avoid to be toxic, but since the group has a benefit when
detection of the prey. To stay undetected the an- developing toxic, toxicity still evolves.
imal needs a good camouflage color, which makes
them undetectable for the predators. This causes 1.2 Hypotheses: Deception & Coin-
animals with a higher camouflage to be eaten less cidence
(Stevens and Merilaita, 2009). However, when an
aposematism is established, a prey that is camou- In our research we want to get a closer understand-
flaged will signal the predators that the prey is safe ing of the origin of mimicry and the course of evo-
to eat. Hence, when the camouflaged animals are lution in the preys and predators. A common as-
found, the chance that a predator decides to eat sumption in the literature is that the preys delib-
them is very big. erately deceive the predators (Holmgren and En-
The warning color, or low camouflage, can be quist, 1999; Rabosky et al., 2016). However, this
beneficial since it scares off predators. However, a assumption can be argued by the idea of coinci-
warning signal can be hard to establish since the dence, which argues that mimics had the phenotype
prey has to deviate from its camouflaged color, all along, and the toxic animals coincidentally got
which makes it more vulnerable to predator at- the same phenotype. These two hypotheses will be
tacks. When aposematism is created, the warning tested by a co-evolution model. We will elaborate
colors can be useful. However, aposematism is hard on the hypotheses in the next two sections. This
feature to establish. In our research, we represent will be done by two defenitions: speciation and the
camouflage as an optical characteristic, but it can adaptive plane
be applied to odor, sound, and taste as well. Speciation is the term we use for populations
which evolve in different sub-populations. For ex-
1.1.2 Toxicity ample, an animal population which is all camou-
flaged can split into one group which stays the
Similarly, synthesizing toxic compounds can be same, and another which gets bright colors and be-
beneficial when they serve their purpose, namely, comes a mimic of another population.
scaring away predators. However, the first preys to The adaptive plane is used to explain the evolu-
explore this peak in the adaptive plane will still be tion of the preys. The plane has a camouflage and
eaten by the predators since these have not learned toxic dimension over which the populations move.
which preys are toxic. A move is described as a change in features for a
The synthesis of toxic has a small negative in- population. When the mean toxicity and camou-
fluence on the fitness. The reason for this decrease flage changes, we say the population moves over
in fitness is that these toxic acids or proteins have the adaptive plane.
to be synthesized in the animal, which takes an
adaptive cost. However, examples of animals can be
1.3 Deception by the preys
found that feed on toxic flora, and use the residues
to keep their toxicity intact. An example of this In deception, the preys have a dominant role in the
is the Zygaena (Holzkamp and Nahrstedt, 1994), evolution of mimicry. As seen from the perspective
which feeds on toxic plants and uses the residues of the preys, the mimicry group will need to be as
for its own protection. Therefore, the content of similar to the phenotype of the toxic group as pos-
toxic in the animals can be a very low cost adaption sible, to make sure the predators do not mistake

2
Figure 1.1: In deception, the mimic and camou- Figure 1.2: In coincidence, the toxic and camou-
flaged animals will start with the same charac- flaged animals will start with the same charac-
teristics. The hypothesis states that the mimicry teristics. The mimicry group is already in place.
group will move toward the place without toxic The toxic population now gets more toxic, and
and without camouflage: the mimicry place, consequently moves toward the lesser camou-
since this place is best if the animals do not get flaged site. Our hypothesis is that the mimicry
eaten. group will move towards the toxic group first,
after which it moves back with the toxic group.

them for a camouflaged animal. Therefore, when

the toxic group is less camouflaged, the mimicry
group will have to evolve to have a similar pheno-
type to maintain the mimicry, as can be seen in
group moving to the less camouflaged part of the
Figure 1.1.
adaptive plane, to create a better distinction be-
Deception can also be explained in terms of spe-
tween them and camouflaged animals. The mimicry
ciation. When all the preys are starting as a big
group can stay at the brightly colored place, since
population, this is first divided in an camouflage
the predators take them for toxic animals.
non-toxic population and a non-camouflaged toxic
population. This is the starting position for this
hypothesis. After this the hypothesis is that the How does mimicry evolve out of this? One ex-
mimicry population is a sub-population of the cam- planation is the recognition of poisonous colors in
ouflaged population, by lowering their camouflage. animals. This is explained by the learning of the
Since the warning color protects them from the predators. They learn to recognize the preys that
predators, we expect that this is beneficial for the are not camouflaged as toxic and will stop eating
mimicry population. them. The mimicry preys can benefit from this by
having the same phenotype as the models.

1.4 Coincidence
Coincidence can be explained in terms of specia-
The coincidence hypothesis assumes that the tion as well. The big population is first divided in
mimicry animals do not change their phenotype, two non-toxic sub-populations, a camouflaged and
but that the toxic animals evolve a distinct look, non-camouflaged (mimic) population. The mimics
which happens to be the same as the mimics, as can already have a distinctive color, which can be ex-
be seen in Figure 1.2. The knowledge of the preda- plained by pre-adaption(Section 5). After this, a
tors is important in this hypothesis. In the case of sub-population of the camouflaged group becomes
coincidence, the predators create a discrimination toxic, by changing their diet, and starts to lose
line between the toxic population and the camou- camouflage to distinguish themselves from the cam-
flaged population, regardless of the mimic group. ouflaged population. Now, the mimics and models
This idea is known as: Pattern Enhancement and start to look similar by accident, not because of the
Peak Shift, explained in 5. This results in the toxic mimics, but because of the models.

3
2 Number-dependent theory tors will get more positive reinforcement from eat-
ing than punishment. To explain Batesian mimicry
A lot of research has been done on mimicry, but we substitute n for n0 in which n0 is updated ac-
still there are no convincing proofs of the origin cording to
of mimicry. The number-dependent theory will be amimic
n0 = n + n ∗ . (2.2)
elaborated, since it can explain both Müllerian and amodel
Batesian mimicry. In Equation 2 we can see how n0 is dependent of
Müller (1879) argued for the number-dependent the ratio number of mimics (amimic ) to number of
theory, which states that the purifying effect of models (amodel ). When there are a lot of mimics,
non-mimicry animals becomes stronger when there the predators will have more difficulty learning the
are more animals eaten. The purifying effect is the toxicity of the animals, meaning a larger n0 . As a
working of the survival of the fittest, and the se- result, the mimics have a parasitical behavior to-
lection of fit animals compared to less fit animals. wards the models. In this situation the equilibrium
With a strong purifying effect, the weak animals will be found when there are enough models to keep
have almost no chance of reproducing and only the predators scared for the color, and the number
the fittest animals will reproduce. Contrary to this, of mimics can not increase because the predators
when there is a weak purifying effect, the weaker would eat to much.
animals have a bigger chance of reproducing by ac- Müller argued that the ratio between mimic and
cident. model is important in mimicry research. When the
The number-dependent theory can be explained mimic is more abundant than the model, the preda-
by the following equations. Let ai the number of tors will take the mimics for general, and start
animals in a given sub-prey-population and let n be hunting the model again. On the other hand when
the number of animals which are eaten before the there is a rare mimic, the predators will not change
predators have learned which animals they should their behavior, leaving the mimic model ecosystem
not eat (the learning rate). Lastly, we define A as in balance. This can be explained by using Equa-
the total number of preys with a certain phenotype tion 2. When the non-toxic mimic is a very small
and Ei the number of preys eaten per population i. part of A, the effect on the learning rate of the
When all species are totally dissimilar(ai = A), the predators is minimized.
loss for all species will be n. However, if populations Holmgren and Enquist (1999) argued that that
have the same phenotype, the losses are given by models have to change their phenotype from the
the following equation: mimics because the more they resemble the mim-
ai ∗ n ics, the bigger the chance is that they are eaten.
Ei = (2.1) This is explained with the formula, where we can
A
see that if the model moves away from the mimic,
By this formula, Müllerian and Batesian mimicry amimic decreases, and therefore the number of ani-
can be explained. For Müllerian we can derive the mals that need to be eaten decreases. This explains
following. When there are a lot of animals which that the most brightly colored preys will have a big-
look similar, A gets bigger, and therefore Ei gets ger chance of survival and the population of models
smaller. In other words, fewer preys of each sub- will evolve to brighter colors. However, the mimics
prey population are needed to teach the predators will receive more benefit when they resemble the
what is toxic. Therefore it has an evolutionary ad- models more, and will therefore evolve to have a
vantage for a species to look a lot like other species. similar phenotype as the models. Mimicry can only
However, this model only holds when all species be sustained if the movement of the mimics is faster
are equally toxic, since the predators always get a than the movement of the models.
negative feedback from eating a prey. For Batesian Mallet and Joron (1999) found that selective
mimicry the formula changes, since the mimics are pressure is determining in the outcome of mimicry
not toxic and the models are. In this case, the mim- research. When a large part of the preys is attacked
ics have a negative effect on the learning rate. To by predators(large Ei ), selection will be stronger
establish a mimicry ecosystem, the ratio of mimics and therefore the ecosystem will not change much.
to models can not be higher than 1, since the preda- On the other hand, when fewer preys are attacked,

4
the preys will experience a relaxation of selection. 3.1 Model Description
Due to this decrease of selection in the population
the preys have more opportunity to experiment, In this paper, a model is presented that simulates
causing the population to shift its balance. This the co-evolution of predator and prey to investi-
way sub-populations are formed to discover other gate the evolution of mimicry. The predators and
parts of the adaptive landscape. preys will co-evolve until mimicry is established, af-
ter which the parameters and genetic development
In this paper, the origin of mimicry will be re-
of the agents will be analyzed.
searched. Two hypotheses will be tested: deception
and coincidence. Besides this, the idea of Holmgren The model involves two types of agents: preda-
and Enquist (1999) will be verified. Lastly the ideas tors and preys. The preys are further divided in
of Mallet and Joron (1999) will be discussed, which three populations: toxic, camouflaged and mimic.
are elaborated in Section 5. Predators have three abilities. Eating preys,
The remainder of this paper is set up as follows. learning from their mistakes and reproducing. The
The ideas of Mallet will be discussed in section 5. predators consist of a neural network to store the
In section 3 we will discuss the model, first general memory of the preys. This memory is evolves,
and then in more depth. We present our results so there is no adaptive memory in the individual
in section 4. Section 6 will discuss the results and predators. Each time-step, the predator encounters
compare this to the ideas presented in section 1 and a number of preys. These preys can be found or stay
section 5. disguised depending on their camouflage. When the
prey is found, the predator decides whether or not
to eat the preys, depending on the phenotype of the
prey. After this, it receives fitness if the animal is
3 Methods not toxic, and loses fitness if it is toxic. When giv-
ing birth, the best parent is chosen out of a random
There are 3 different ways to study mimicry (Mallet sub-set of predators, and copied into the child. The
and Joron, 1999). child then has a small chance on mutations, which
will be elaborated on in Section 3.2.
• The evolutionary dynamics way, which stud- Preys have one ability, giving birth. Besides this
ies the evolution of the preys but ignores the the preys have three characteristics: camouflage,
behavior of the predators. (Gavrilets and Hast- toxicity, and pattern. All of these features range
ings, 1998) from 0 to 100. To give birth, the fittest two parents
are found and the child is created out of the par-
• The receiver psychology way, which focuses on ents. The fitness of the preys is determined by the
the behavior of the predators, but tends to ig- camouflage, toxicity and times the prey is eaten by
nore the evolution of the preys. (Holmgren and a predators. The more camouflage or toxicity a prey
Enquist, 1999) has, the less fitness it has. Pattern is added to have
an independent value only for the use of signalling
• The traditional natural historical way, which the phenotype to predators. This pattern has no
analyzes the behavior of the predator and prey. influence on fitness, but can be used by predators
In these kinds of research, the co-evolution be- to recognize the prey.
tween predators and preys are studied (Bates, The preys consist of three populations which can
1862). only reproduce within populations. The first popu-
lation of toxic preys has a small genetic drift toward
In this research the natural historical will be used more toxicity. This genetic drift exists of a small
by combining the evolution of the preys and the be- tweak in the mutation. When an animal is mutated,
havior of the predators. To study this co-evolution there will be slightly higher probability of increas-
of these animals, a model will be created. First we ing toxicity than of decreasing it. The genetic drift
will give a general explanation in Section 3.1, after of the camouflage works in the same way for the
which the more technical discussion can be found camouflaged population. The mimics have neither
in Section 3.2. drift towards toxic, nor towards camouflage. In the

5
simulation the toxic population reflects the model, Algorithm 3.1 Prey Eaten
and the mimic population reflects the mimic. The let Cy camouflage of prey y, scaled [-50, 50]
camouflage population is an example for the preda- let Py pattern of prey y, scaled [-50, 50]
tors which they can eat. The camouflaged group is let NN(X,Y) neural function of predators
also used as a control group, to see if the mimic if random(101) > Cy and N N (Py , Cy ) >
population has unique behavior compared to other random f loat(1) then
populations. In this paper the term phenotype is Eat P reyy
used for the combination of camouflage and pat- end if
tern.
The model makes no use of the spatial proper-
Algorithm 3.2 Choose Parent
ties of animals and is dependent on camouflage for
finding preys, and fitness for finding parents. Terms let [P1, P2, P3 ... Pn] random animals
like movement and location of the populations re- let Parent maxFitnes([P1,P2,P3...Pn])
fer to the place on the adaptive plane. There is no
location modelled.
total number of synapses (weights) is thus 9 per
predator. After each synapse round, an activation
3.2 Model Analysis function is applied to scale the values between 0
In this section we look at the model in more detail. and 1. The activation function used is the sigmoid
The eating behavior and knowledge of the predator function. The learning of the predators is explained
is discussed in Section 3.2.1, the reproduction of in Section 3.2.2.
animals in Section 3.2.2, and the setup of different
parameters in Section 3.3. 3.2.2 Reproduction
To reproduce, the predators and preys use different
3.2.1 Eating of the predators
principles. For the selection of the parents, both the
The eating of the predators is dependent on two animals use tournament selection, which works as
steps. This can be seen in Algorithm 3.1. The first is found in Algorithm 3.2.
(1) the probability of an animal being found. Cam- Predators When the parent is chosen, the child
ouflage makes the probability to be found lower, so is made as a clone of the parent. The way of re-
that preys with a high camouflage have a bigger production is iteroparous, that is, parents can have
chance to hide from the predators. (2) The choice several children in their lifetime. For predators the
of the predator. The predator uses a simple feed- reproduction is different from the preys, in the way
forward neural network to propagate the pheno- that they are asexual. In asexual reproduction only
type. The phenotype of the prey lets the preda- one parent is needed for reproduction, which means
tor decide whether to eat a prey or not. This way, that the weights of the predator are copied and
toxic animals can distinguish themselves by having given to the child. After this, each wight has a
a very distinguishable phenotype. The network ex- chance of mutation. When a weight is mutated, a
ists of two input nodes, three hidden nodes and one value between -0.25 and 0.25 is added to it. The
output node. The nodes are connected by synapses, value is than cut to the domain between -2 and 2.
which all contain a weight. As input, camouflage Because of the survival of the fittest principle, the
and pattern of the preys are used, both scaled be- best predators evolve and anticipate on the changes
tween -50 and +50. As output a probability of eat- within the preys. This way of learning can be seen
ing is generated, between 0 and 1. The decision as a random search.
threshold is then chosen randomly, so that there is Preys The preys use sexual, iteroparous repro-
a slight chance predators will still try to eat bright duction with tournament selection. When the two
animals occasionally. Each node is connected to all best parents are chosen with the tournament selec-
nodes in the next layer. This results in 6 synapses tion, the child is made as a combination of these
from the input node to the hidden nodes and 3 parents. Different than the predators, the preys
from the hidden nodes to the output node. The have sexual reproduction, meaning that there are

6
Table 3.1: The Parameter List shows the differ- of a predator represents a day , after which the
ent parameters which were used in the model predator has learned which animal it should not eat
and which it should eat. The prey becomes older,
Parameter Predators Preys
which makes the difference in fitness between preys
Number of individuals 10 630 which are eaten and that are not eaten bigger.
(Tox. = 300, Cam. = 300, Mim. = 30) Each time-step, a predator encounters the num-
Genetic drift - 3 ber of preys divided by the number of preda-
Times Prey Found(Yencountered ) 3 - tors. This is optionally multiplied by Yencountered
Mutation rate 20 2 to make the selective pressure higher. From the
Tournament size 3 10 point of view of the prey, it has Yencountered en-
Age 3 5 counters with predators. The encounters of the
Chance-being-found 101 - predators(Enr ) can be approximated as follows
Camouflage Disadvantage - 3.0
Toxicity Disadvantage - 0.2
Enr = (#P reys/#P redators) ∗ Yencountered .
(3.1)
The encounters for each prey(Eny ) is given by
two parents needed to create a child. There is no
distinction made in gender, so that every animal
Eny = Yencountered . (3.2)
can reproduce with every other. The child is then
created by taking the mean of camouflage, toxicity The disadvantages of toxic(T) and camou-
and pattern and giving this to the child. The values flage(C) are multiplied by the toxic disadvan-
of the prey have a value between 0 and 100. The tage(TD) and camouflage disadvantage (CD) re-
child has a chance of mutation, in which case the spectively. For example, a prey with toxicity = 80
features are added by a random number between and toxicity disadvantage 0.2, will subtract 16 from
-10 and 10. If this value exceeds the borders of 0 the fitness of the prey. The fitness of a prey is up-
or 100, it is cut off at that value. In the case of ge- dated according to
netic drift, the GeneticDrif t parameter is added to
these numbers, giving more chance for an increas-
Fy = −(T ∗ T D) − (C ∗ CD) − (500 ∗ Ey ). (3.3)
ing mutation.
Generations The reproduction works in gener- Here we can see that the fitness of the preys(Fy )
ations, and in each generation the old generation is calculated by subtracting the toxicity and cam-
dies. This means that there are no animals older ouflage, multiplied by their respective disadvantage
than other animals, but all animals die at the same from 0. After that, 500 is subtracted from the fit-
moment after a predefined number of time-steps. ness, for each time it gets eaten(Ey ). The fitness is
This number of time-steps is not the same for a value below 0, with 0 being the highest value.
predators and preys, and can be found in Table The fitness of the predators is calculated in a
3.1 as the age parameter. The preys will survive an different way than the prey fitness. The fitness is
attack of a predator, and only die when their gen- determined by the eating of toxic or non-toxic ani-
eration dies. However, the fitness of a prey deterio- mals, according to the following formula
rates when it is successfully attacked by a predator,
Eny
making the chance for reproduction minimal. X
Fr = (Ey (toxic) − 60). (3.4)
y=1
3.3 Setup of the Parameters
Here we can see that the fitness of the predators
The number of predators is fixed, as well as the (Fr ) depends on the sum of the toxicity of all the
number of preys in a population. In every run of the eaten preys(Ey ). All toxicity values are subtracted
model, the number of predators is lower than the by 60, so predators get positive feedback when eat-
number of preys, corresponding to the real world. ing animals with lower toxicity than 40, and nega-
The predators have a lower age than preys, since tive when eating animals which have a higher tox-
they learn faster than the preys evolve. The death icity value.

7
Figure 4.1: The average camouflage for 100 runs Figure 4.2: The average pattern for 100 runs for
for the deception hypothesis. The starting po- the deception hypothesis. We can see that the
sition of the mimicry and camouflage are the mimics and the camouflaged animals are initial-
same, and the toxic populations starts low. The ized at the same position, but the mimics move
mimicry population moves slowly towards the towards the toxic population. Same as in Figure
toxic population. This is an average over 100 4.1 this is an average of 100 runs, so it is not an
runs, but one transition from much camouflage representative of one run.
to little takes little time. The line can be used as
the percentage of runs which already has little
camouflage at the given moment. flage as well, but due to a big population and the
genetic drift is not able to continue. We can see
that the camouflage of the toxic population is fairly
4 Results stable. In the pattern we can see that the pattern
The results are divided in 4 different sections. In of the mimics moves toward the toxic population.
Section 4.1 the deception hypothesis will be tested. These graphs support the idea of deception, since
In Section 4.2 the coincidence hypothesis will be the mimic moves from the camouflaged place in the
tested. For both of these hypotheses we will show adaptive plane towards the non-camouflaged place,
results from 100 runs, in 14000 time steps. After mimicking the toxic population. Besides this, the
this, in Section 4.3, we will discuss the peculiari- mimics use pattern to get an evolutionary advan-
ties of an individual run, and verify the theorem tage by having the same pattern as the toxic group.
of Holmgren and Enquist (1999). Lastly, in Section We can see that the mimics evolve to trick the
4.4 the difference between model and mimic will be predators into not eating them.
discussed.
4.2 Coincidence
4.1 Deception
In the coincidence hypothesis, the toxic and cam-
For the first hypothesis, the mimic and camouflaged ouflaged group start at the same location, without
population start out with high camouflage. The toxic and camouflage. The toxic group then gets a
toxic population starts at it’s final position with genetic drift toward more toxicity. The starting sit-
high toxicity and little camouflage(bright colors). uation can be found in Figure 1.2. The question was
See Figure 1.1 for the graphical view. Proof for de- whether the mimic group would move toward the
ception would be the movement of mimics toward toxic population (deception) or stay at the same
the less camouflaged side, while maintaining their place (coincidence). We can see in Figure 4.3 that,
low toxicity level. In Figure 4.1 and 4.2 we can see on average, the mimics do not move towards the
the average camouflage and pattern over 100 runs. toxic population. This is an argument for coinci-
We can see that the mimic population moves to- dence. When we look at the pattern (Figure 4.4),
wards the lower camouflage level. The camouflaged however, we can see that the mimics try to deceive
population moves in the direction of less camou- the predators by using the same phenotype as the

8
 Figure 4.5: The pattern of an individual run in
the deception hypothesis. The pattern changes
with two distinctive bumps, before being sim-
ilar to the toxic population. After this we see
Figure 4.3: The average camouflage for 100 runs the toxic population moving away from the pat-
for the coincidence hypothesis. Notice that the tern of the mimics, and the mimics chasing this
starting positions of toxic and camouflage are pattern.
the same, and the mimics are already in place.
We can see the toxic population dropping fast,
decreasing the camouflage. When we take a look
models. The camouflage is an supports the idea of
at the mimics, we see that the population does
not move towards the toxic population. coincidence, but the pattern supports deception.

4.3 Individual Run

Figure 4.5 shows the pattern of an individual run
of 14000 time-steps of the simulation. Using this
figure we can evaluate the idea of Holmgren and
Enquist (1999), which said: ”For mimicry to be es-
tablished, the movement of the mimic should always
be faster than the movement of the model”. Besides
this we will take a look at Kin Selection, explained
in Section 5.
After approximately 9000 steps in the simulation,
we notice the pattern of the models moves away
from the mimics. When the mimics have a higher
pattern, the models get a lower pattern and vice
versa. This is consistent with the idea of Holmgren
and Enquist (1999). In the results we can see that
the movement of the mimics is faster, but the mod-
els try to escape this. This can be explained by the
mimic and model population. For models it is dis-
Figure 4.4: The average pattern for 100 runs advantageous to look like the mimic, because the
for the coincidence hypothesis. We can see that,
predators will get positive feedback from eating a
same as in the deception hypothesis, the mimics
mimic. Therefore do the models which look less like
move towards the toxic population.
the mimics have an evolutionary advantage, a de-
velopment we can observe in the individual run of
the model.
We can see that the the mimic changes its cam-

9
 4.4 Euclidean distance Model and
Mimic
Figure 4.6 shows the distance in phenotype between
model and mimic over time. This graph tells us that
the distance gets smaller, thus mimicry is created.
The euclidean distance is measured as the distance
between the mean of the pattern and camouflage
between the mimics and the models by the follow-
ing formula:

p
EU D = (Cmim − Cmod )2 + (Pmim − Pmod )2
Figure 4.6: The Euclidean Distance over time (4.1)
for both the hypotheses. In both hypotheses, the For the Deception hypothesis, the mimic moves
difference between model and mimic decreases towards the model, to trick the predators. Over
over time, arguing for mimicry. The coincidence time, we can see the more and more mimic popula-
hypothesis has a little less difference. Keep in
tions look similar to models. This means that a lot
mind this is an average over 100 runs, and there-
fore is not a representation of one run, but a
of the simulations evolve into mimicry. For the Co-
probability of mimic and model being the same. incidence hypothesis, the toxic group goes towards
the mimics first. After this, the mimic follows the
model over the phenotype plane. This can be seen
as the average Euclidean Distance gets smaller and
smaller. Since both distances are decreasing, the
model supports both hypotheses for mimicry.

5 The Origin of Aposematism

ouflage in several bumps. These bumps can be
explained by the models getting less toxic and
and Mimicry
the predators eating them. When we observe the The results of the model can be used to verify
model, we can see that the models start losing their certain possible hypotheses, stated by (Mallet and
toxicity when the predators do not eat them. This Joron, 1999).
is explained by the individual fitness of the preys.
However, when the toxicity drops below 40, which 1. Novelty and Recognizability can be explained
is a positive influence for the predators, the preda- by the predators fearing new colors, which is
tors will start eating again. This is seen in Figure called neophobia. When predators do not know
4.5 when the mimic moves away from the pattern a color, they fear it, and will refuse eating the
of the model. Now Kin Selection appears for the prey. The model shows that predators make a
models. It is as a group better to be toxic, and discrimination line which animals they do and
therefore the whole population starts to be toxic which they do not want to eat. New patterns or
again. When the models are not toxic, it is bet- camouflages would always fall into one of these
ter for the mimics not to look like the predators, categories. The discrimination line can be seen
but to have more camouflage. In the choice of the as a threshold to which the predator decides to
predators, we could see that a discrimination line eat an animal. With another knowledge rep-
was made throughout the phenotype plane. As a resentation this origin may be simulated, but
result, the camouflaged preys were eaten if found, not with the simple neural network we used in
the toxic animals were often skipped. The mimics this research. We hypothesize that when the
had to be on the right side of this discrimination punishment for eating toxic is very big, preda-
line to stay alive. This means there is a form of tors will not eat newly found animals. How-
pattern enhancement, explained in 5. ever, when this is not the case, predators can

10
 experiment in their eating behavior, making The simulation has possibilities to research
the probability of neophobia smaller. this topic, which will be more elaborated on
in Section 6.3
2. Preadaptation Another reason for evolving
aposematismic features is found before the cre- 5. Density- and Appearance-dependent Warning
ation of toxicity. Butterflies for example, need Color This kind of aposematism assumes a
bright colors to attract mates and in territo- phenotype which can change during the life-
rial male-male interactions. This causes bright time of an animal. When many animals are
colors to evolve in preys, without the toxic- eaten it would be better to look toxic, whereas
ity having evolved yet. A change in diet can when there are not many eaten, it would be
then cause preys to synthesize toxic, result- better to maintain a camouflage. Since we look
ing in aposematism. This hypothesis is close at the evolution and genetics of mimicry, this
to the coincidence hypothesis, since the mim- topic is not simulated in the model.
ics preadapt to a bright color, after which the
toxic animals get bright colors as well. The 6. Kin Selection This way of evolution has an
model shows that, even if the mimics already altruistic approach, as it looks at groups as
had bright colors, they would not move toward a whole. When a few animals die to scare
the camouflage of the toxic animals. Thus, predators off for the rest, can this be benefi-
the model supports preadaptation. The idea of cial for the group. Other preys, carrying the
preadaptation will be elaborated on in Section same genetics, can evolve to a more aposema-
6.1. tismic look. This hypothesis is tested by creat-
ing populations which evolve together, making
3. Pattern Enhancement and Peak Shift assumes the group as a whole better instead of only sin-
the memory of the warning colors in the preda- gle animals. For each individual it is better to
tor is a symbolic representation of the actual have less toxic, since this has a negative influ-
colors. This implies that preys which resem- ence on their fitness. However, for a group as
ble the warning color more will have a bigger a whole it is better to be toxic, since preda-
opportunity to survive. In the adaptive plane tors will stop eating the group. It was found
this would mean that it would always be better that the toxic group emerged in this kind of
to have less camouflage because this is more group behavior. We observed that the toxic
similar to the general representation. This is group became less toxic as the aposematism
the case in the model, since the predators take was created, until they reached a toxic value
every non-camouflaged animal for a toxic ani- of 40 (the toxic threshold for the predators).
mal. For the mimics, this made it possible to After passing this, the predators started eat-
signal a more dangerous signal than the mod- ing the toxic group again, and the toxic group
els by having less camouflage. In the pattern used kin selection to become toxic again.
we could see that the mimics evolved towards
the most toxic location for the predator. Addi- 7. Genetic Drift & Shifting Balance The shifting
tionally, when the models changed their pheno- balance theory consists of three phases, which
type, the mimics evolved with them, according were found in the model. In phase 1, a sub-
to the idea of Holmgren. population explores a new adaptive peak due
to genetic drift. This was found as the mim-
4. Müllerian Mimicry As discussed earlier, ics moved away from the camouflaged group
Müllerian mimicry occurs when two animals of to explore another adaptive peak. In phase 2
similar toxicity start to co-evolve. This means the population gets fully adapted to this adap-
that animals will start to look more alike when tive peak. We found this in the pattern of the
combining their distinctive features. When this preys which evolved to a similar pattern as
is combined with another species later on, all that of the models. In phase 3 the populations
theses patterns are joined together create a co-evolves with other populations in the same
wide variety of aposematismic features. How- adaptive peak. This happened as the mimics
ever, Müllerian mimicry was not simulated. evolved together with the phenotype of the

11
 models, even if the models evolved to another 6.2 The idea of Holmgren and En-
phenotype. quist (1999)
Holmgren argued that the models always move
away from the mimics, according to the number de-
pendent theory which was discussed in Section 2.
6 Discussion By creating distance from the mimics, the models
experience less confusion from the predators. How-
During this last section we will discuss the achieved ever, since the mimics follow the phenotype of the
results, and compare them to our hypotheses. In models and evolve faster, this is an endless cat and
Section 6.1 we will discuss the deception and coinci- mouse game. In the model this can be observed
dence hypotheses. We answer the question whether in the pattern, where the mimics circle around the
Deception and Coincidence are plausible origins of models. However, no quantitative measurement has
mimicry. In Section 6.2 the idea of Holmgren and been done so far, which leaves space for future re-
Enquist (1999) is discussed. We show that their search.
idea is found in our model. After this we will dis-
cuss the new implementations and usages that can 6.3 Future experiments and imple-
be used by this model in Section 6.3. mentations
The model can be used to do more research on theo-
ries of mimicry. Since the parameters can be easily
6.1 Deception and Coincidence adjusted, more experiments can be done. Firstly,
more experiments can be done with different start-
For both the origins of mimicry we can say that ing positions of the camouflage, toxicity and pat-
they are plausible. The deception has very clear tern values, starting with a control run. In this
evidence in the deception set-up, where we can see case one population without genetic drift would be
that the mimics always move toward the models. created to see how preys evolve without other an-
Similarly, we can see that in the pattern the mim- imals. Another example is the toxicity and mimic
ics evolve toward the models to deceive the preda- group starting in the toxic position, and the camou-
tors. For a small population of mimics, it is possi- flage on the camouflage position. This would make
ble to explore new peaks in the adaptive plane, and for another coincidence set-up, which assumes the
change their phenotype to get into the mimic place. speciation of the mimicry being a sub-population
For the camouflage population, we can see that this from the toxic population. Alternatively, the pat-
population is too big to explore new adaptive peaks. tern can be altered. The pattern of the camouflage
and mimic start the same in this paper, but it can
The coincidence has evidence in the change of be altered to a situation where the toxic and cam-
camouflage, since we can see that the mimics do ouflage population start the same, and see whether
not move towards the models in their appearance, the toxic population moves away. Besides this we
but stay in their own place. We can assume that the can see whether the mimics move toward the mod-
mimics do not try to mimic the models, when the els, which supports the Coincidence hypothesis.
camouflage has an influence on their fitness. How- The dynamics of the model can be altered as well.
ever, when this influence is removed, as can be seen One possibility is adding more dimensions of recog-
in pattern, the mimics will deceive the predators, nition. This would mean that instead of 1 pattern,
and choose the looks of the models. the model would have 50 patterns, which all can be
Both of the hypotheses set-ups concluded in the mutated and inherited individually. In these recog-
mimicry ecosystem. However, as soon as the ecosys- nition dimensions the scale between 0 and 100 can
tem is created, it is not possible to see which set-up be removed, so the models and mimics can move
the system came from. Therefore, we can support through the adaptive space with more freedom.
both hypotheses, but can not conclude which one This way, neophobia and the idea of Holmgren can
matches best with biological data. be researched in more detail. When the dimensions

12
are implemented we hypothesize that the models ics of mimicry evolution. Biological Journal of
will keep evading the phenotype of the mimics and the Linnean Society, 66(2):145–158, 1999.
the mimics chasing this phenotype. If the domains
are removed, we expect very high and low values in Guido Holzkamp and Adolf Nahrstedt. Biosynthe-
the dimensions, arguing for the very bright colors of sis of cyanogenic glucosides in the lepidoptera. in-
the animals. To research Müllerian mimicry, more corporation of [u-14c]-2-methylpropanealdoxime,
populations can be added which have intermediate 2s-[u-14c]-nethylbutanealdoxime and d, l-[u-14c]-
values of genetic drift toward toxic. With two toxic n-hydroxyisoleucine into linamarin and lotaus-
populations, a research can be conducted whether tralin by the larvae of zygaena trifolli. Insect bio-
the animals imitate each others phenotype or keep chemistry and molecular biology, 24(2):161–165,
their own phenotype. The number dependent the- 1994.
ory can be tested in the same way. We hypothe- Martine E Maan and Molly E Cummings. Poison
size that when more toxic populations are imple- frog colors are honest signals of toxicity, particu-
mented, there will be one center where all the ani- larly for bird predators. The American Natural-
mals converge to, to make one clear aposematism. ist, 179(1):E1–E14, 2011.
The representation of knowledge can be differenti-
ated. At the moment, the predators have a line in James Mallet and Mathieu Joron. Evolution of di-
their choice to eat camouflage or not. If a more versity in warning color and mimicry: polymor-
curved line, or other methods are implemented, phisms, shifting balance, and speciation. Annual
the idea of Novelty and Recognizability can be re- Review of Ecology and Systematics, 30(1):201–
searched in more depth. If this is combined with a 233, 1999.
variation of punishment for toxicity, we expect that
neophobia emerge from the simulation. Lastly, a Fritz Müller. Ituna and thyridia: a remarkable case
spatial model can be created, in which agents have of mimicry in butterflies. Trans. Entomol. Soc.
a x- and y-coordinate. This way mimicry rings can Lond, 1879:20–29, 1879.
be researched, which are discussed in great depth Alison R Davis Rabosky, Christian L Cox, Daniel L
by Mallet and Joron (1999), and found by Bates Rabosky, Pascal O Title, Iris A Holmes, Anat
(1862). When the spatial model is implemented, all Feldman, and Jimmy A McGuire. Coral snakes
the aforementioned can be combined in one simu- predict the evolution of mimicry across new
lation, since every place can evolve something else. world snakes. Nature communications, 7, 2016.
Especially the borders of different mimicry systems
will be interesting to research. Using a bigger adap- Martin Stevens and Sami Merilaita. Animal camou-
tion space, better knowledge of the predators and flage: current issues and new perspectives. Philo-
a spatial dimension in the model, we aim to have sophical Transactions of the Royal Society of
a better understanding of the origin of mimicry in London B: Biological Sciences, 364(1516):423–
the future. 427, 2009.

References
Henry Walter Bates. Xxxii. contributions to an
insect fauna of the amazon valley. lepidoptera:
Heliconidae. Transactions of the Linnean Society
of London, 23(3):495–566, 1862.

Sergey Gavrilets and Alan Hastings. Coevolution-

ary chase in two-species systems with applica-
tions to mimicry. Journal of Theoretical Biology,
191(4):415–427, 1998.

Noél MA Holmgren and Magnus Enquist. Dynam-

13