swarm-intelligent foraging in honeybees: benefits and costs of...
TRANSCRIPT
SWAM INTELLIGENCE
Swarm-intelligent foraging in honeybees: benefits and costsof task-partitioning and environmental fluctuations
Thomas Schmickl • Ronald Thenius •
Karl Crailsheim
Received: 30 May 2009 / Accepted: 18 March 2010
� Springer-Verlag London Limited 2010
Abstract For honeybee colonies, it is crucial to collect
nectar in an efficient way. Empiric experiments showed
that the process of decision making, which allows the
colony to select the optimal nectar source, is based on
individual decisions. These decisions are made by return-
ing nectar foragers, which alter their dancing behaviours
based on the nectar source’s quality and based on the
experienced search time for a receiver bee. Nectar receiv-
ers, which represent a shared limited resource for foragers,
can modulate the foraging decisions performed by the
colony. We investigated the interplay between foragers and
receivers by using a multi-agent simulation. Therefore, we
implemented agents which are capable of a limited set of
behaviours and which spend energy according to their
behaviour. In simulation experiments, we tested colonies
with various receiver-to-forager ratios and measured col-
ony-level results like the emerging foraging patterns and
the colony’s net honey gain. We show that the number of
receivers prominently regulates the foraging workforce.
All tested environmental fluctuations are predicted to cause
energetic costs for the colony. Task-partitioning addition-
ally influences the colony’s decision-making concerning
the question whether or not the colony sticks to a nectar
source after environmental fluctuations.
Keywords Swarm intelligence � Honey bees �Task partitioning � Foraging � Equal foraging distribution �Cross inhibition � Choice � Nectar economics
1 Introduction
Honeybee colonies have to collect huge amounts of nectar
in their environment to accumulate enough honey for
periods of dearth. Colonies that perform the nectar col-
lection process in a less efficient way will have a lowered
chance to survive (e.g., throughout winter), thus natural
selection favours those colonies that efficiently forage for
nectar. The collective foraging of social insects (ants, bees,
wasps, termites) was studied intensively, revealing many
insights into the proximate mechanisms that are used by
animals to decide upon their foraging behaviour. Inside of
a colony, these individual foraging behaviours (recruit-
ment, abandonment, navigation) sum up to colony-level
foraging strategies, which can differ significantly between
species or even within one species in differing environ-
mental situations.
Mathematical models are used to analyse and interpret
these foraging decisions of social insect colonies. Colony-
level decisions that arise from known mechanisms of
individual decision-making were studied with top-down
models, which predict dynamics of forager and receiver
group sizes with sets of linked differential equations.
A comparative study [1] showed the similarities and dif-
ferences between ant foraging and honeybee foraging.
Detailed studies of recruitment mechanisms and their effect
on forager group sizes, foraging decisions, and resource
throughput were performed by means of mathematical top-
down modeling [2–4]. In addition, bottom-up (individual-
based, agent-based) models were frequently used to predict
T. Schmickl (&) � R. Thenius � K. Crailsheim
Artificial Life Lab of the Department of Zoology,
University of Graz, Universitatsplatz 2, 8010 Graz, Austria
e-mail: [email protected]
R. Thenius
e-mail: [email protected]
K. Crailsheim
e-mail: [email protected]
123
Neural Comput & Applic
DOI 10.1007/s00521-010-0357-9
how well local cues can serve as a source of information
for individual foragers [5–8]. In honeybees, foraging
decisions are based on such local cues, which are exploited
by individual forager and receiver bees. Based on these
cues, and on communication transferred through bee dan-
ces (see for eample [9, 10]), one worker can affect the
behaviour of other nearby workers. This way ‘behavioural
feedback loops’ emerge, which lead to a self-organised
regulation of workforce and foraging decisions [11].
Honeybee foraging is a very good example of natural
‘swarm intelligence’, because the group of foraging bees is
not controlled by a central decision-making unit and
because the collective decisions of colonies in varying
environments were found to act in an intelligent way.
1.1 Honeybees select foraging targets by swarm
intelligence
At first sight, foraging of honeybees looks similar to the
foraging of other mass-recruiting social insects, which
were shown to forage for food (and also for new nests) in a
swarm-intelligent way. When looking closer onto details,
there are significant differences between the foraging per-
formed by ants (or termites) and honeybees:
1. Honeybees perform waggle dances, which communi-
cate vectors (distance, direction) pointing to the
foraging targets. For further details, see [9, 12]. The
decision-making upon which foraging target is
exploited the most is based on the likelihood of naive
bees to encounter such waggle dances, and not on
direct quality information passed from one bee to
another [13].
2. Honeybees do not simply select the nearest nectar
source; they collectively take also into consideration
the quality of the source, what is the sugar concentra-
tion of the offered nectar and/or the flow of nectar at
the source [13]. This multi-factorial information pro-
cessing is an impressive ability, because it is achieved
without a central unit of decision-making, and it was
found to be very robust and flexible [3].
3. The foraging of honeybees involves task partitioning
between forager bees and receiver bees. Foragers
collect the nectar in the environment and bring it back
to the hive. There, receiver bees take over the nectar
load from the forager bees, transport it to the storage
area of the hive and store it there. As this sequential
task partitioning involves nectar transfer from forager
bees to receiver bees, this generates the need for
swarm-intelligent coordination of these two groups of
co-workers, which is regulated by the ‘tremble dance’
behaviour. This dance is (sometimes) performed by
forager bees and was found to trigger unemployed bees
to engage in the nectar-receiving task [10].
These characteristics of the honeybees’ way to forage
for food in a swarm-intelligent way is of high significance
also for the community of engineers and mathematicians
who develop new bio-inspired optimisation algorithms or
novel swarm robotic systems. The foraging decision-mak-
ing is performed collectively by the foraging bees and it
involves also individual information processing that mod-
ulates the behaviour of single bees. The decisions made by
individual bees can be modelled by rather (mathematically)
simple equations as their behavioural modulations are
usually not very complex (stimulus–response thresholds
and correlations). Frequent interactions among bees and the
exploitation of behavioural modulations of other bees as
cues for their own behavioural decisions make this forag-
ing a clear example of swarm intelligence: it was shown
that individual bees assess the quality of a visited food
source (=net energy efficiency) in a way that is expressed
best by
qualitysource ¼gain� cost
cost; ð1Þ
as was shown in Seeley [13, 14]. In Eq. 1, the variable gain
represents the energetic value of the collected nectar and
the variable costs represents the metabolic (energetic) costs
of the foraging trip, both variables are measured in Joules.
Based on this dimensionless quality index, the foragers
decide how many dancing rounds they perform. As the
ultimate effect, this leads to the ability to collectively select
the optimal nectar source. For further details, see [3, 13].
The sequential task partitioning between forager bees
and receiver bees adds interesting features to the honeybee
foraging system: It was shown that nectar receivers play an
important role, because they represent a limited resource
shared by the whole cohort of foragers. If a forager sear-
ches too long for a receiver bee that unloads the collected
nectar, it does not perform a waggle dance, thus this for-
ager does not recruit additional foragers. It was pointed out
in Johnson [15] that nectar receivers work also in pro-
cessing the food, thus they need to have a specific physi-
ological constitution. This physiological status is usually
found in ’middle-aged bees’ (11–20 days old), which are
usually present in high numbers (several thousands) in a
colony. However, this age-class performs not only nectar-
receiving, it also performs a high variety of other tasks:
e.g., guarding, wax-building, or undertaking [16, 17]. In
recent studies, we showed that the age-structure of a hon-
eybee colony is affected by external (e.g., weather) and
internal (e.g., cannibalism) factors, thus the size of age-
groups can change quickly in a colony [18, 19]. In con-
sequence a limitation in the number of available receiver
Neural Comput & Applic
123
bees, which are mid-aged bees located at the nectar transfer
places and that actually engage in the food handling task,
can significantly decrease the recruitment rate of new for-
agers bees. It was shown by Seeley [20] that honeybees use
a dynamic system to regulate the number of active receiver
bees: when the searching (queuing) delay of a loaded for-
ager exceeds a certain period, this forager starts to perform
a tremble dance, which recruits additional nectar receivers
instead of performing a waggle dance. This way, a
decentralised auto-balancing mechanism is established that
regulates the size of the two groups of workers (foragers
and receivers) in a way that keeps them at an optimal ratio.
Remembering the fact that the two groups split their task
sequentially, such a regulation system could be interpreted
as an adaptation to optimise the overall performance of the
colony, which is the unit of natural selection in honeybees.
For further details on this regulation regime, see [21].
Figure 1 shows how colony saturation (amount of stored
honey in the hive) and the distance to the exploited nectar
source influence the ratio of receivers to foragers in the
colony: upon interaction, foragers get unloaded and
receivers get loaded. The fuller the colony is, the longer it
takes for receivers to get unloaded again, as they have to
search longer for empty storage cells. The further away the
foraging target is, the lower is the frequency of returning
loaded foragers.
1.2 Hypotheses and motivation
The task-partitioning between receivers and foragers is a
remarkable characteristic of honeybees. Our goal was to
study the ‘swarm intelligence’ of honeybee foraging and
how the group size of receivers affects the collective
decision-making process. We hypothesise that the receiver-
to-forager ratio is ‘shaping’ the collective foraging in the
following ways:
H1: Changing the receiver-to-forager ratio affects the
collective foraging decision made by the simulated hon-
eybee colony significantly.
H2: How these foraging decisions are affected by the
receiver-to-forager ratio additionally depends on the envi-
ronmental situation or the intensity of an environmental
fluctuation.
H3: Every foraging decision induced by our simulated
environmental fluctuations—characterised by a sudden
change in location of the most profitable food source
without change in mean quality of all available nectar
sources—causes an energetic loss for the foraging honey-
bee colony due to the (physiological) costs of the decision-
making process that is involved. The receiver-to-forager
ratio also affects the nectar economy of the colony sig-
nificantly in these cases.
To allow such an analysis, we extended an existing
multi-agent simulation platform (HoFoSim, honeybee for-
aging simulator), which is described in detail in Schmickl
et al. [22–24]. As we wanted to investigate the role of
receiver bees, we extended our model by implementing
also receiver bees as autonomous agents, which make their
own behavioural decisions and which have their own
metabolic expenditures. This new model is called
HoFoReSim (honeybee forager and receiver simulator) and
has the following key characteristics:
1. We model the foraging purely based on individual
behaviours, no global communication, no global
supervising unit is implemented. For example, waggle
and tremble dances affect only local neighbouring
agents.
2. The platform HoFoReSim allows a variety of interac-
tive simulation experiments to address many questions
that arise during the study of honeybee foraging.
3. HoFoReSim implements all relevant factors that
represent a realistic honeybee foraging scenario, e.g.,
errors in communication and navigation, metabolic
costs of behaviours, differences among individuals
(dance–response curves, weights, speeds, crop sizes,
etc.).
Using this multi-agent model, we performed a series of
simulation experiments in which we predicted the number
of foragers on different nectar sources, as well as the col-
ony’s predicted honey accumulation. We exposed the
modeled colony to a variety of environmental fluctuations
and observed how the colony was able to react to these
disturbances. By evaluating the honey accumulation, we
were able to assess the energetic costs induced by these
Fig. 1 Regulation of the number of active receiver bees and of active
forager bees according to the distance of the nectar source, according
to the quality of the nectar source and according to the colony’s nectar
need
Neural Comput & Applic
123
environmental fluctuations. By varying the receiver-to-
forager ratio, we could investigate the resulting changes in
foraging strategies and also how this ratio affects the col-
ony’s energetic efficiency.
2 Methods
Our multi-agent simulation platform is able to simulate up
to 5,000 foraging bees and up to 5,000 receiver bees
simultaneously. In contrast to the model described in
Schmickl et al. [22, 23], we extended our model in a way
that also receiver bees are now simulated individually.
They can be loaded by returning foragers and they are then
occupied for the time it takes to store the nectar. For
details, see [24]. Our model focuses on the tasks ‘nectar
foraging’, ‘nectar receiving’, ‘idle’ (=work reserve) and
‘scouting’. We modelled the three signals that affect this
process prominently that are ‘waggle dancing’, ‘tremble
dancing’, and ‘nectar offering’. In addition, we modelled
physiological costs of all tasks in much detail, to allow us
to predict colony-wide economics. Other signals (e.g. the
‘shaking signal’) were omitted so far, to keep the com-
plexity of the model in reasonable range.
2.1 Environment
Our simulation platform is able to perform a variety of
experiments focusing on the foraging of honeybees. The
bees (foragers and receivers) move randomly around in the
hive area and they interact (e.g., nectar transfer) when they
meet on the same patch, which represents an area of
7.5 mm 9 7.5 mm of comb space. The foragers leave the
hive through a special ‘entrance patch’. In the outside
environment, one patch represents 2.7 m 9 2.7 m of
space. The outside environment holds up to three nectar
sources, which represent artificial feeders that can be freely
positioned in the environment. The sugar concentration of
the offered nectar (sucrose) is adjustable during runtime
between 0.0 Mol/l (feeder removed) and 2.5 Mol/l. For the
spatial layout of our model, see Fig. 2.
2.2 Agent behaviours
We implemented two types of agents: ‘foragers’ and
‘receivers’. All agents are modeled as finite state automata
(Fig. 3). They are modeled as ‘reflex agents with inner
status’ (see [25]). Each agent is situated in one behavioural
state, which determines its behaviour. This behaviour is
performed every time step unless a behavioural transition is
triggered. Such transitions can be caused by one of the
following reasons: a fixed time delay has expired, an
external event has happened, an internal event has
happened, a communication event has happened or an
event was triggered with a fixed probability. For details,
see [22–24]. The ‘inner state’ of the agent is characterised
not only by its ‘behavioural state’ as described above, but
also by its crop load (nectar load) and sometimes by its
experienced queuing delay when waiting for unloading. In
addition, it possesses a memory to recognise one foraging
target, which is the vector towards a food source and the
distance of this source from the hive.
In our model, most of the transitions from one behav-
ioural state to another are implemented on a threshold
based system following the suggestions of Bonabeau et al.
[26], which we already compared to empiric data in
Schmickl and Crailsheim [23]. The probability to change
behavioural states depends on the intensity of local stimuli
(cues) in a sigmoid-shaped relationship. For example, the
time period tsearch(i, t) of a forager i in time step t that
searches for a receiver bee is exploited as a cue (stimulus).
It influences the chance to perform a waggle dance or a
tremble dance (see Fig. 4) according to the following two
equations:
To model an agent’s probability to dance a tremble
dance, we modeled
ptrembleði; tÞ ¼tsearchði; tÞn
tsearchði; tÞn þHn; ð2Þ
where we found best agreement to empiric literature data
(see [23]) with n = 5 and H ¼ 40. The probability of an
agent to perform a waggle dance was modeled by
pwaggleði; tÞ ¼ 1� tsearchði; tÞn
tsearchði; tÞn þHn; ð3Þ
where we set n = 2.5 and H ¼ 10 according to literature
(see [23]).
Bees within a distance of rhearing (6 patches = 4.5 cm)
are able to perceive a dance and are able to react according
Fig. 2 Screenshot of the simulation. a The hive: inside the hive
foragers and receivers interact. Waggle dances and tremble dances
take place there. b The hive entrance: foragers leave and enter the
hive here. Most interactions between foragers and receivers take place
near the hive’s entrance. c, d The outside: in the environment outside
the hive nectar sources are placed. e, f Experienced foragers fly to the
nectar sources. g Some scout bees do not collect nectar from the
known nectar sources but fly randomly in the environment to find new
nectar sources
Neural Comput & Applic
123
to Fig. 3. The duration of the waggle dance is modeled
according to Seeley [13], see also Fig. 5. For further
details, please see [27].
For modelling the tremble dance, we followed a similar
scheme, but had to consider the topology of a bee hive and
the spatial distributions of dance-searching naive forager
bees and of inactive potential receiver bees. In our
simulation, we did not model all combs of a full-sized bee
hive. The modelled colony area is 58 9 58 patches wide,
representing 1892 cm2 of comb surface. This area repre-
sents approximately the two comb sides of the ‘dance floor’
area on which foragers perform waggle dances and on
which the majority of nectar transfers take place. As
waggle dancers and potential foraging recruits preferen-
tially locate in this area, we could take over the dance
length and probabilities of dances as they are reported in
empiric studies. In contrast to foraging recruits, inactive
mid-aged bees, which can be activated by tremble dances,
are located in all remote areas of the hive. Based on the
data given in Seeley [10], we calculated that a tremble
Fig. 3 Two linked finite state automata, that we used to model the
behaviour of bees inside and outside the colony. The two behaviours
‘foraging’ and ‘scouting’ are associated with high metabolic costs
(Eq. 11), all other behavioural states are associated with lower
metabolic costs (Eq. 12). a Forager–forager interaction: one forager
can recruit another forager to a nectar source by performing a waggle
dance. b Forager–receiver interaction: a forager can recruit an idle
receiver to start receiving nectar by performing a tremble dance.
c Nectar transfer: during the unloading-process, nectar is transferred
from a forager to a receiver bee
0 50 100 1500
0.2
0.4
0.6
0.8
1
searching time for receiver bee [s]
prob
abili
tyof
danc
e
waggle dancetremble danceno dance
Fig. 4 Sigmoid-shaped stimulus–response curves that model the
dancing behaviour of returning forager bees depending on the
forager’s searching-time for an available nectar receiver. For details,
please see Eq. 2 and equation. These stimulus–response curves are
based on empiric observations reported in Seeley [10]
20 30 40 50 60 70 800
10
20
30
40
50
60
source quality
danc
ero
unds
perf
orm
edbee #1bee #2bee #3bee #4bee #5bee #6bee #7
Fig. 5 Duration of dances performed by foragers depending on the
source quality, as it is described by Eq. 1. Redrawn from [13]
Neural Comput & Applic
123
dance can cover 2–5% of hive space. We assumed the same
communication distance for waggle dances and tremble
dances. We parametrized our model that one tremble dance
can reach 4% of all receiver bees in the modelled hive,
which corresponds to a tremble dance length of 1 time step.
Compared to a comparable recent model [5], where every
tremble dance recruits always one additional receiver, our
modelling approach is closer to empiric data, as more
inactive mid-aged bees can be recruited by one dance.
In contrast to the sigmoid-shaped curves that model the
chances of dancing, the duration of waggle dances is lin-
early scaling with the quality of the nectar source visited by
the dancing forager bee, as it is expressed by Eq. 1. These
stimulus–response curves for the calculation of the dance
duration are implemented according to Seeley [13]. The
duration dwaggle of a dance depends on the quality index
qualitysource of a source (see Fig. 5) and is calculated
according to the Eqs. 4, 5 and 6:
kwaggle ¼ 0:51 � gð2Þ: ð4Þ
owaggle ¼ 6 � cnnþ 22
1:032e8:798kwaggle� 25 ð5Þ
dwaggle ¼ qualitysource � kwaggle þ owaggle ð6Þ
In these equations, g(x) indicates a uniformly distributed
floating-point random number, which is:
0 � gðxÞ\x: ð7Þ
The variable cnn indicates the actual ‘colony nectar
need’ according to Seeley [13], who showed that the global
status of honey stores (=accumulated recent honey income
rate) modulates the offset of the linear dance–response
curves of individual foragers. The more nectar gets
accumulated in the colony, the less they dance. The
variable kwaggle indicates the slope of the individual
stimulus–response curve, and owaggle indicates the curve’s
offset from the origin. qualitysource is modeled according to
Eq. 1 and expresses a dimensionless quality index (net
energetic gain) of the foraging trip.
The quality index qualitysource also influences the
probability of a bee i in time step t to abandon a source, as
is expressed by the following equation
pabandonði; tÞ ¼ðc� qualitysourceÞ
n
ðc� qualitysourceÞn þ cn
; ð8Þ
which was parametrised with c = 100 and n = 4 to reflect
empiric data of abandonment given in Seeley [11].
All modeled behaviours of forager bees and receiver
bees are associated with metabolic costs, which also
depend on the current weight of the bee. This metabolism
forces the agents to consume a portion of their own crop-
load each time step. How much nectar an agent actually
consumes depends on the sugar concentration of its nectar
load. After the crop-load falls below a certain threshold, an
agent refills by decreasing the honey stores of the colony.
An agent that runs out of nectar outside of the colony dies
immediately in our model, what just happens with extreme
settings (parametrisations) that are not shown in this article.
After a successful foraging flight, nectar is transferred to
receiver bees. Receiver bees increase the colony’s honey
stores when they deposit the nectar into the stores. Thus,
we can interpret the optimality of all emerging foraging
patterns by observing the dynamics of the colony’s honey
stores. Finally, we state that the nectar sources in the
environment, the nectar in the bees’ crops and the honey
stores differ significantly concerning their sugar concen-
trations. Thus, the concentrations of nectar in the crop of a
bee could vary due to the possibility of mixture of nectar of
different sources. Due to this, we had to model all dilution
and concentration processes in detail to retrieve robust
results from our model. Nectar sources offer sucrose in the
empiric experiments described in Seeley [28]. In contrast,
concentrations of sugars in the bees’ crops are tracked in
energetic equivalents of glucose. Bees have different nectar
concentrations in their crops resulting from their past
activities.
2.3 Metabolism
The metabolic rate of a honeybee in our model depends on
the actual weight of an agent i, which is expressed by
wagent(i, t). The metabolic rate depends also on the activity
level of the bee (for details see [24]). The empty weight of
a bee is modeled as wempty(i) = 83.32 ± 6.89 mg [29].
The current crop-load volume is expressed as Vload(i, t).
The maximum crop volume a bee can load with nectar
is modeled as Vmaxcrop ¼ 45:9� 8:7 ll [29]. The weight
wload(i, t) of the actual nectar load having a sugar con-
centration of cload(i, t) is calculated by Eq. 9, based on own
empiric measurements (data not shown), we modelled
wloadði; tÞ ¼ Vloadði; tÞ � ð/ � cloadði; tÞ þ nÞ: ð9Þ
We found best agreement to our own empiric
measurements (data not shown) by setting / =
0.065 mg/Mol and n = 1.0026 mg/ll. According to this
equation, the actual weight of a bee wagent(i, t) is calculated
by
wagentði; tÞ ¼ wemptyðiÞ þ wloadði; tÞ: ð10Þ
Based on data presented in Seeley [13], the metabolic
rate of a bee l(i, t) [ {lhigh(i, t), llow(i, t)} is calculated
either by
lhighði; tÞ ¼�high � wagentði; tÞqhigh
r � cloadði; tÞ� Dt; ð11Þ
Neural Comput & Applic
123
or by
llowði; tÞ ¼�low � wagentði; tÞqlow
r � cloadði; tÞ� Dt; ð12Þ
whereby ehigh = 0.00287, elow = 0.00248, qhigh = 0.629,
qlow = 0.492, r = 2.827 (according to Seeley [13]),
depending on the activity level of the agent i in time step t.
The metabolic rate is the volume of nectar which an agent
consumes per time step.
Receiver bees add the nectar that was transferred to them
to the honey stores after a period of search. The nectar
deposition process is modeled as follows: receivers stay in
state ‘storing’ for a period of time dstore, which depends on
the current colony’s nectar need [21], which we express by
the parameter cnn (=‘colony nectar need’) in our model. In
Seeley [21], the author refers to an ‘empty’ hive and to a
‘full’ hive, which we parametrise in our model as cnn = 0.0
and 1.0, respectively. The author notes that the receivers
spend 10 min (smin = 600 s) to store the nectar in an empty
hive and 28 min (smax = 1,680 s) in a full hive. We assumed
a linear negative regression of the storing duration between
the two extremes of cnn, thus we modeled the storage cycle
time (in time steps of Dt ¼ 0:5 s) as
dstore ¼smin þ ðsmax � sminÞ � ð1� cnnÞ
Dt: ð13Þ
Receiver bees consume energy (and thus nectar)
according to their current weight, also during their storage
trips. The remaining transported nectar is then added to the
colony’s honey stores at the point in time a receiver switches
from the ‘storing’ state to the ‘receiving’ state again, see
Fig. 3.
2.4 Parametrisation and individuality of bees
Our simulation experiments shown in this article were
performed with time step size Dt = 0.5 s. Other important
parameters of the Simulation are the spatial distances in
and outside the hive in the simulation environment: the
edge length of one patch inside the colony corresponds to
0.75 cm, outside the colony 2.7 m. The speed s of a bee
was implemented with s = 5 m/s outside the colony (flight
speed), within the colony s = 1.5 cm/s (running speed).
The sugar concentration of honey in the honey stores is
set to 6.25 Mol/l in our model (own measurements). To
produce honey, nectar receivers have to evaporate water
from the nectar and to alter sugars in an enzymatic way.
The water evaporation leads to a decrease of volume when
nectar is stored as honey in the colony’s honey stores, what
is also part of the material flow, which is modelled in detail
in our model. The sugar concentration of honey is of big
importance for the ‘refilling’ processes of honeybees inside
the colony. If bees run out of nectar inside the colony they
refill their crop directly from the honey stores of the col-
ony. Foragers upload 1 ± 0.2 ll of honey, receivers upload
9 ± 1.8 ll of honey. This data is estimated, based on
Huang and Seeley [29]. At the beginning of a simulation
run 15 foragers in the hive have the information about
heading and distance of each nectar source, and 50% of the
receivers are in the state ‘‘receiving’’.
The individuality of foragers regarding the duration of
the waggle dance (see Fig. 5) is implemented by a indi-
vidual value, represented by g(2) in Eq. 4, as described in
detail in Thenius et al. [27]. Other parameters, like the
probability to leave the colony for scouting, the probability
to forget a source or the probability to start foraging were
(uniformly) distributed randomly around ±30% around the
mean values. Other parameters of the simulation were
parametrized using empiric data from literature as it is
mentioned in the previous sections.
2.5 Statistics
We repeated all analysis multiple times, as we applied ran-
dom uniform noise symmetrically to the bees communica-
tion and to individual agent properties based on the
parametrised mean values. The exact repetition number is
given with every experiment in the text and in the figure
legends. To analyse differences between two datasets, we
used two-sided Mann–Whitney U-tests. To analyse the effect
of multiple factors in datasets, we used ANOVA (two-way
randomised blocks), whereby experiment repetitions with
the same settings were interpreted as blocks. In these anal-
yses, means were analysed by Student–Newman–Keuls
tests. For correlation analysis, we performed Pearson cor-
relation tests. In all analyses, the level of significance (a) was
set to a = 0.01.
3 Results
In the simulation experiments described in this article, we
focus on the role of receiver bees in the decision-making
process. To test the reliability of our model, we performed a
set of simulation experiments that reflect empiric experiments
performed with real honeybees. Then, we performed a series
of simulation runs in fluctuating environments and observed
the energetic optimality of the emerging foraging patterns.
Therefore, we investigated the nectar-economic effects that
arise from the intensity of the simulated environmental fluc-
tuations and from the receiver-to-forager ratio.
3.1 Experiment 1: equal foraging distribution
In an experiment described in Seeley [13], the dependency of
waggle dance duration on the net energy efficiency
Neural Comput & Applic
123
(see Eq. 1) was shown. Two nectar sources were present in
the environment, one in a distance of 250 m (source A) and
one in a distance of 550 m (source B) from the hive. Source
B offered always a sucrose solution of 2.5 Mol/l. By
manipulating the concentration of the near feeder (source A),
the scientists searched for the concentration which stimu-
lated the bees to perform exactly the same dance length for
both nectar sources. We modeled this set-up with our sim-
ulator by simulating 500 foraging bees and 400 receiver
bees. To demonstrate the reliability of our model, we para-
metrised our model to the details given in Seeley [13]. We
observed the number of bees carrying information about one
of the two sources. This correlates with the waggle dance
durations. Additionally, we observed the size of the forager
groups that were visiting both nectar sources.
We found an equal foraging distribution (=EFD) with a
sugar concentration of 1.34 Mol/l at the near feeder, which is
within the range of concentrations reported in Seeley [13] for
real bee experiments (1.32–1.46 Mol/l). As Fig. 6 demon-
strates, there are some difficulties in searching the point of
EFD: while the number of bees which hold information
about either one of the two sources stays almost in a ratio of
1:1 (Fig. 6a), the number of bees actually visiting the near
source is always higher than those visiting the far source
(Fig. 6b). This difference is due to the fact that our model
incorporates communication error in the communication of
nectar sources (see for example [12]). Also, the flight of the
bees was modeled to be erroneous, according to the data
given in von Frisch [9]. In addition to that the flight cycle of
those bees that forage on the more distant source is longer,
thus more bees are—on average—in the flying state, in turn
decreasing the average number of bees found drinking on the
distant source per time interval. This effect can be seen by
comparing Fig. 6a with Fig. 6b. As Fig. 6c shows, the
quality assessments of the bees for their nectar sources are
similar at EFD conditions.
We performed additional simulation runs to test how
much the ability to reach an EFD depends on the environ-
mental conditions and on the receiver-to-forager ratio. We
placed two nectar sources in equal distance to the hive
(400 m) and offered an equal sugar concentration CAsource ¼
CBsource on both sources. This experiment was performed with
two sugar concentrations (fCAsource ¼ CB
source ¼ 1:0 Mol=l;
CAsource ¼ CB
source ¼ 2:5 Mol=lg). We simulated several
ratios of receivers to foragers (foragers = 500 bees,
receivers [ {200, 400, 600, 800, 1000, 1200} bees). The
foraging distribution was calculated according to the fol-
lowing symmetry index
SA;Binfo ¼
maxðf Ainfo; f
BinfoÞ
f Ainfo þ f B
info
; ð14Þ
where a value of SA;Binfo ¼ 0:5 indicates an EFD. The vari-
ables f Ainfo and f B
info represent the number of foragers that
have information about one of the nectar sources A or B as
it is depicted for one single run in Fig. 6a. As Fig. 7 shows,
our model predicts distributions closer to EFD with higher
ratios of receiver-to-foragers. This effect is stronger with a
low quality CAsource ¼ CB
source ¼ 1:0 Mol=l of the available
nectar sources.
3.2 Experiment 2: cross-inhibition between groups
of foragers
In an empiric experiment [28], two nectar sources were
placed 400 m away from the hive. Both sources offered a
sugar concentration of 1.25 Mol/l. After some time, there
was an EFD on both feeding sites, comparable to the sit-
uation demonstrated in experiment 1. At this point in time,
the sugar concentration at one of the nectar sources was
increased suddenly to 2.5 Mol/l. The number of foragers
recruited to this feeding site increased very fast and
(a) (b) (c)
Fig. 6 One exemplary run showing equilibrium foraging on two
nectar sources. Source A 250 m, 1.34 Mol/l, source B 550 m,
2.50 Mol/l. a Number of foragers having information in their memory
about one of the two sources. b Number of individual foragers found
on each source within 30-min time slots. c Distribution of quality
assessments (according to Eq. 1) within the foragers (cumulative sum
of 5 snapshots during runtime, taken in intervals of 30 min). This
correlates with dance length of waggle dances. Upper graph shows
data for source A, lower graph for source B
Neural Comput & Applic
123
simultaneously lowered the number of foragers visiting the
other feeding site, which remained unchanged in quality.
We resembled this experimental scenario and simulated it
with several realistic receiver-to-forager ratios. In all runs,
we used 500 foragers, the number of receivers were 300,
600 and 900 bees.
To investigate the effect of the quality increase of one
nectar source onto the foraging on the other source, we
performed the following series of simulation experiments.
In the first part of the experiment (before the quality
increase), we waited for two simulated hours to establish an
EFD. Then, we stored this point in time on hard disk. One
interesting feature of our simulation platform is the fact
that we are able to save the state of all agents and of the
environment to the computer hard disk and reload this
point in time later again. This allows us to perform ‘time-
machine’-like experiments, like rewinding the tape and
running the simulation from the same starting point again.
Taking this saved state, we continued the simulations
with a sudden quality increase of source A for 3 h. This
experiment was repeated 25 times. After that, again starting
with the initial state saved on disk, we simulated a scenario
without the quality increase of source A for 25 times. By
comparing the foraging patterns at the never-changed
source B in both scenarios, we investigated whether a
cross-inhibition has occurred or not. These ‘save-and-
reload’ experiments allowed us to interpret the cross-
inhibitory effect by only looking at the number of foragers
on the unchanged source after the disturbance occurred,
while other studies (e.g., [28]) had to compare the periods
before and after the occurrence of the disturbance.
In all simulation runs we observed cross-inhibition
between the two groups of forager bees. In all three set-ups
(Fig. 8a, d, g) the quality increase of source A led to an
immediate increase of the number of bees foraging on source
A. The more receivers were present in each simulation run,
the higher was the number of foragers on source A at
t = 14:00. Between 30 min and 1 h after the quality change
of source A the difference between the number of foragers on
the two nectar sources was found to be statistically signifi-
cant. This indicates that the simulated colony made a for-
aging decision according to the sudden environmental
fluctuation (two-tailed Mann–Whitney U-test, p \ 0.01,
N1 = N2 = 25 per setting). In contrast to that, in the 25 runs
without quality changes, the number of foragers on both
sources stayed the same (except for the first half hour in two
cases, two-tailed Mann–Whitney U-test, p [ 0.01, N1 =
N2 = 25 per setting; see Fig. 8b, e, h). When comparing the
number of foragers on unchanged source B in the stable
environment and in the unstable environment, we found the
following pattern. The model predicts a significant decrease
in the number of foragers on source B for the time period
30 min-1 h after the environmental fluctuation (two-tailed
Mann–Whitney U-test, p \ 0.01, N1 = N2 = 25 per setting).
Thus, we observed clear cross-inhibition in all of these
simulation runs. The number of foragers on the unchanged
nectar source at t = 14:00 correlated positively (Pearson
correlation, r = 0.67, N1 = N2 = 25 per setting) with the
receiver-to-forager ratio. This demonstrates the importance
of receiver bees as shared limited resource in explaining the
cross-inhibitory effect between forager groups.
3.3 Experiment 3: optimal source selection
Empiric experiments showed that a honeybee colony col-
lectively selects the optimal nectar source and that it is able to
reverse this decision after a severe change (=fluctuation) in
the environment [3]. Two nectar sources were placed at the
same distance from the hive (400 m), one offering a high
sugar concentration (2.5 Mol/l), the other one offering a low
sugar concentration (1.0 Mol/l). After 4 h, the concentra-
tions were altered, the former high-quality source offered
now a very poor sugar concentration (0.75 Mol/l), while the
former poor-quality source now offered a high sugar con-
centration (2.5 Mol/l). This configuration was kept for
additional 4 h. We took over this experimental set-up but
extended it by adding the following variations and
modifications:
1. We performed symmetrical fluctuations so that the
quality (=sugar concentration) of the poor source was
the same in the first and in the second half of the
Fig. 7 Distribution of foragers on two equal-quality and equidistant
nectar sources with varying ratios of receivers to foragers (SA;Binfo , see
text). Two different sugar concentrations were tested on both nectar
sources. N = 6 per setting. Bars indicate medians, whiskers indicate
quartiles
Neural Comput & Applic
123
experiment. This is different to the study described in
Seeley et al. [3], in which the experimenters offered a
lower sugar concentration on the poor source in the
afternoon compared to the poor source in the morning.
2. We simulated different amplitudes of environmental
fluctuations. The following pairs of sugar concentra-
tions were tested (given in Mol/l): 2.5:2.0, 2.5:1.5,
2.5:1.0.
3. We performed simulations with different receiver-to-
forager ratios. We kept the number of foragers
constant at 500 bees and used the following sizes for
the receiver group: 250, 500, 1,000 and 1,500 bees. In
total, each of these combinations (number of receivers
in combination with fluctuation amplitudes) was
repeated six times.
We found both, an impact of the fluctuation amplitude
and of the receiver-to-forager ratio onto the emerging
foraging patterns. Figure 9 shows that higher numbers of
receivers lead to higher numbers of active foraging bees.
We calculated the mean number of active foragers on both
nectar sources (f). The values for f varied between f = 74.2
active foragers found with 250 receiver bees and f = 127.9
active foragers found with 1,500 receiver bees (also
depicted in Fig. 13b). The differences observed in the
number of active foragers were significantly affected by all
tested numbers of receivers (ANOVA, N = 24 per setting,
p \ 0.01 between all settings) but it was not affected sig-
nificantly by the strength of the tested environmental
fluctuations (ANOVA, N = 24 per setting, p [ 0.01).
The most important question in this experiment is:
does the receiver-to-forager ratio influence the ability of
the colony to perform a second decision within one day,
that is, to select the better feeder in the afternoon? To
investigate this, we made the following definition: if the
number of foragers on source B exceeds the number of
foragers on source A at t = 16:00, we say that the colony
has performed a second foraging decision in the
afternoon.
Figure 9 shows that an increasing amplitude of the
environmental fluctuation favours such a second foraging
decision in the afternoon (depicted in sub-figures 9i–l). The
fewer receivers were available in a simulation run, the earlier
this decision was made. Such sharp foraging decisions are
performed also with lower amplitudes of the environmental
fluctuation (e.g., see Fig. 9e), if the number of receivers is
low (250 bees). With high numbers of receivers and with low
amplitudes of the environmental fluctuation, the emerging
foraging pattern is significantly different. The colony keeps
on to exploit source A massively (which was not worsened
very much in the afternoon in these runs) and additionally
recruits to source B in the afternoon, which was improved
in quality then. Using ANOVA analysis, we analysed
the dependence of the observables foragerssource Aðt ¼16 :00Þ; foragerssource Bðt ¼ 16 :00Þ and Dforagers B�Aðt ¼16 :00Þ , which represent the number of foragers on the
sources A and B, as well as the difference between these two
group sizes at t = 16:00. Statistical analysis showed that all
three observables were significantly affected (ANOVA,
(a)
(b)
(c)
(d)
(e)
(f)
(g)
(h)
(i)
Fig. 8 Cross-inhibition
between two groups of foragers
in scenarios with different
receiver-to-forager ratios. All
runs: 500 foragers. Left column:
300 receivers. Middle column:
600 receivers. Right column:
900 receivers. Both sources are
of same quality for the first 2 h.
For the following 3 h, two
different scenarios were
performed. First row: the quality
of source A increased
spontaneously for the last 3 h of
the experiment. Second row: the
quality of source A stayed
unchanged during the last 3 h of
the experiment. Last row:
comparison of the number of
foragers on source B in both
scenarios. Stars: significant
differences (U-test, p \ 0.01,
N1 = N2 = 25), thus they
indicate presence of a cross-
inhibitory effect. Graph shows
mean ± SD
Neural Comput & Applic
123
N = 24 per setting, p \ 0.01) by the hight of the environ-
mental fluctuation as well as by the number of available
receivers. However, there was no statistical difference of
means between the runs with 1,000 and 1,500 receivers, as
well as between the runs with the two strongest
environmental fluctuations (ANOVA, N = 24 per setting,
p [ 0.01). In addition, ANOVA analysis showed that also
the interaction of the type of environmental fluctuation and
the number of receivers jointly affected the means
(ANOVA, N = 24 per setting, p \ 0.01).
(a)
(b)
(c)
(d)
(e)
(f)
(g)
(h)
(i)
(j)
(k)
(l)
Fig. 9 Source selection of the simulated honeybee colony in an
environment which always contains one high-quality foraging target.
For the first 4 h, source A offered the high nectar quality, for the
second 4 h period, source B offered the higher nectar quality. Three
different amplitudes of such environmental fluctuations were per-
formed, see column headings for exact values. These scenarios were
performed with different receiver-to-forager ratios; see row headings
for exact values. An intersection of the two datasets on the right half
in each figure (afternoon) indicates that the colony performed a
second foraging decision in response to the environmental fluctuation
at noon. The values of f indicate the mean number of foragers found
on both sources, thus they indicate the emerging colony’s foraging
activity. All runs: 500 foragers were simulated with different numbers
of receiver bees. N = 6 per set-up. Graph shows mean ± SD
Neural Comput & Applic
123
If both options in the decision-making are of poor
quality, but still one option is better than the other, a
similar pattern arises (see Fig. 10). The lower the ratio of
receivers to foragers is, the quicker a second foraging
decision is made in the afternoon.
3.4 Analysis of experiment 3: how does this colony
behaviour affect the net honey gain of the colony?
Figure 11 exemplary shows for two selected simulation
experiments of experiment 3 that these environmental fluc-
tuations significantly affect the colony’s predicted net honey
gain. During the first 2 h of the simulated 8-h period, the
colony has almost no net honey gain, as all bee agents con-
sume nectar and waggle dance recruitment is a process that
takes time. In the second half of the morning, the predicted
honey gain increases, but the switch of source qualities calls
for re-recruitment of foragers to the other food source.
Again, honey gain breaks down, and it takes some time until
it reaches the previous value of honey gain per time unit.
Figure 11 also shows that the effect of the environmental
fluctuation is observable on the honey gain of the colony
after a significant time delay, which is due to the duration of
foraging flights, queuing delays and the time receivers need
to store the nectar. To allow investigation of the energetic
costs of environmental fluctuations, we compared the pre-
dicted honey gain of undisturbed simulation runs (indicated
by the final net gain ’a’ in Fig. 11) to the predicted net honey
gain of runs with environmental fluctuations (indicated by
the final net gain ’b’ in Fig. 11). The environmental fluctu-
ation induced higher costs (vabscolony ¼ �a� �b) in the simulation
runs performed with 1,500 receiver bees compared to those
runs performed with 250 receiver bees. The variable �a rep-
resents the mean final honey store in all undisturbed runs of
the same colony size and the variable �b represnts the mean
final honey stores in disturbed runs with same colony size.
When we increased the size of the receiver group, we
increased the size of the simulated colony accordingly. Of
course, bigger colonies have more active foragers (see
values of f indicated in Figs. 9, 10), and thus these colonies
accumulate more honey per time unit. This can be seen by
comparing absolute values of a and b in Fig. 11. To allow
comparison of the costs induced by environmental fluctu-
ations of differently sized colonies, we expressed the costs
as a relative measure of the potential net honey gain
achieved in an undisturbed environment. Thus, we defined
vrelcolony ¼ 1�
�b
�a; ð15Þ
expressing the fraction of the net honey gain, which was
not achieved due to the environmental fluctuation
Fig. 10 Source selection of the simulated honeybee colony between
two poor nectar source options of differing quality. For the first 4 h,
source A offered a higher nectar quality than source B. For the second
4 h period, source B offered a higher nectar quality than source A.
These scenarios were performed with different receiver-to-forager
ratios; see row headings for exact values. An intersection of the two
data sets on the right half in each figure (afternoon) indicates that the
colony performed a second foraging decision in response to the
environmental fluctuation at noon. The values of f indicate the mean
number of foragers found on both sources, thus they indicate the
emerging colony’s foraging activity. All runs: 500 foragers were
simulated with different numbers of receiver bees. N = 6 per setup.
Graph shows mean ± SD
Neural Comput & Applic
123
compared to undisturbed simulation runs with the same
colony size.
We compared the predicted colony’s net honey cost
(vrelcolony) for all choice experiments depicted in the Figs. 9
and 10 to see how the strength of the environmental fluc-
tuation and the receiver-to-forager ratio affects this
important variable. As shown in Fig. 12, both factors affect
the costs significantly. On the one hand, lower receiver
numbers lead to lower relative costs of environmental
fluctuations. This suggests that the quick foraging decisions
induced by low numbers of receivers help the colony to
prevent costs caused by environmental fluctuations. On the
other hand, low numbers of receivers lead to less amounts
of accumulated honey.
As Figs. 9 and 10 show, the predicted foraging patterns
are significantly influenced by the receiver-to-forager ratio.
Low numbers of receivers mainly lead to low numbers of
foragers on the worsened source in the afternoon. This can
be explained by the cross-inhibition phenomenon depicted
in Fig. 8. We compared the accumulated net honey gain
within the 8 h of our experiments to investigate how these
different foraging patterns affect the colony’s energetic
foraging success.
Figure 13a shows that the number of available receiver
bees affects the net honey gain of the simulated colony:
colonies with medium ratios of receivers to foragers
accumulate more honey than colonies with extreme ratios.
When receivers are rather scarce, massive honey accumu-
lation is prevented. With too many receivers, energetic
costs of these excess receivers decrease the net honey gain.
In contrast to that the number of active foragers is posi-
tively correlated to the receiver-to-forager ratio (Fig. 13b).
When evaluating efficiency criteria like the net honey gain
per active forager (Fig. 13c) and the net honey gain per
worker bee (forager and receivers combined, Fig. 13d), a
clear optimum for medium receiver-to-forager ratios (500
receivers and 500 foragers) was found. When analysing
how the environmental conditions affect these measures,
we found a clear picture. The better the best nectar source
in the environment is, the better is the colony’s predicted
net honey accumulation. The least affected measure is the
number of active foragers (f), which is more affected
by receiver-to-forager ratios than by environmental
conditions.
4 Discussion
With our model ‘HoFoReSim’, we simulated collective
foraging behaviour of honeybees by a multi-agent simu-
lation. Our aim was to investigate how environmental
changes affect the nectar economics of honeybee colonies
and how foraging decisions could influence these eco-
nomics. In our model, the agents’ behaviour are modelled
as state machines. Changes of state are triggered by events,
probabilities, thresholds and time-outs. Both, the state
machine and the functions that model thresholds could
alternatively be modelled as artificial neural networks [30].
In the real bees, these thresholds and behavioural states are
controlled by the bees’ nervous systems and by their
physiological status (energy and hormones).
We formulated three novel hypotheses (see Sect. 1) and
designed several experiments to test these hypotheses. In
addition, we were able to give qualitative and also quan-
titative predictions of the benefits and gains of foraging
decisions under a variety of environmental conditions and
fluctuations. To allow this study, we modeled proximate
(a)
(b)
Fig. 11 Dynamics of honey accumulation performed by the simu-
lated colony. a Predicted honey dynamics of a colony having a low
number of receiver bees. b Predicted honey dynamics of a colony
having a high number of receiver bees. To assess the energetic costs
induced by the environmental fluctuation at t = 12:00, we compared
the final net honey gain of undisturbed runs to the final net gain of
disturbed runs. For all settings. N = 6. Graph shows mean nectar
dynamics
Neural Comput & Applic
123
(a) (b)
(c) (d)
Fig. 12 Costs of environmental
fluctuations as observed in our
simulated choice experiments.
We compared four
environmental fluctuations of
different intensity of the
disturbance, which occurred at
t = 12:00 and investigated the
resulting relative colony-level
costs compared to undisturbed
runs performed with similar
sized colonies (vrelcolony). All runs:
500 foragers were simulated
with different numbers of
receiver bees. N = 6 per set-up
(a) (b)
(c) (d)
Fig. 13 Investigations of the
net honey gain of the colony
which results from the foraging
patterns depicted in Fig. 9.
a Mean accumulated net honey
gain of the colony after 8 h
[ml]. b Mean number of
foragers active in each
simulation set-up (values of fdepicted in Fig. 9). c Mean net
honey gain per active forager in
all experimental set-ups
[ml/bee]. d Mean net honey
gain per worker bee in all
experimental set-ups [ml/bee].
All runs: 500 foragers were
simulated with different
numbers of receiver bees. N = 6
per set-up
Neural Comput & Applic
123
behavioural rules in our agents (which represent bees), and
we compared the emerging ultimate results to empiric data.
Using the same simulator and having the same parameter
settings, our results in the experiments of equilibrium
foraging, of cross-inhibition between groups of foragers
and in the source-selection experiment match published
empiric data closely [3, 4, 13, 28] (empiric data not shown
here).
Several models of honeybee foraging have been pub-
lished. Most of them took a top–down approach and
describe the rate of changes of forager groups on different
sources by a linked set of differential equations [1, 3]. Only
few individual-based simulations of the honeybee foraging
system exist. The studies of [7, 8, 31]) investigated the
forager–receiver interactions in an individual-based model,
but did not model any nectar sources or complex behav-
ioural programs of foragers and receivers, like we did it in
our model (see Fig. 3). Other studies investigated the
occurrence of multiple unloading events in forager-receiver
interactions, without dealing with other aspects of honey-
bee foraging behaviour [32].
Two studies of de Vries and Biesmeijer [5, 6] describe a
model, which has many similarities to our model presented
here, but which differs in significant aspects that are crucial
for the analyses presented here. Our model implements the
agents’ metabolic costs in more detail, e.g. by considering
the bee’s weight and task. It models the energetic expen-
ditures of all bees, also from those bees that are ‘idle’ in the
hive. In our model, receivers have to meet forager bees
locally and dances have a limited spreading area. Our
dance–response curves differ significantly form the ones
published in de Vries and Biesmeijer [5], as can be seen in
Eqs. 4 and 5, thus our curves (Fig. 5) resemble the data
reported by Seeley [13] statistically more closely. We
implemented the dynamics of the honey stores in detail and
also the focal aspects differ significantly. Our model
focuses on the economics of foraging decisions. One sig-
nificant result of our model is the fact that the second
foraging decision is not made by the colony under several
conditions (Fig. 9). Such predictions were not made by de
Vries and Biesmeijer [5]. Without implementing the bees’
metabolism and morphological features, it is impossible to
predict an energy balance of the modeled colony, like we
did it extensively in our study.
The model ‘TaskSelSim’ described in Schmickl et al.
[33–35] has also implemented the metabolic cost of for-
aging bees and receiver bees, but it is focused on emergent
division of labour and therefore does not allow choices
between different nectar sources. An individual-based
model that implements two nectar sources and metabolism
of agents is ‘FlowerSim’ described in Thenius et al. [36].
This model abstracts the waggle dances to a rough
approximation and implements the nectar sources as a field
of many small blossoms that have to be visited multiple
times by forager bees. While this model reflects the for-
aging behaviour of bees accurately, it is not comparable to
the environmental situations investigated by Seeley [28],
who used unlimited flow feeders as artificial food sources
in empiric experiments. As we wanted to compare our
predicted foraging decisions with these data provided by
Seeley [28], we had to elaborate our model ‘HoFoSim’
[22–24] to the current model ‘HoFoReSim’, which extends
the original model mainly by its additional implementation
of receiver bees and their metabolic costs.
In experiment 1, where we searched the sugar concen-
tration for the near nectar source to achieve equilibrium
foraging, we matched the published results of empiric
experiments very closely. The fact that we always found
more bees actively foraging on the nearer source in times
when the number of bees recruited to the two sources was
in a ratio of 1:1 can be explained by the communication
error associated with the waggle dancing. These waggle
dances do not communicate coordinates of the food
source, they just encode the direction and the distance to
this source. Therefore, the error in communication mis-
guides the bees on trips to the far source more than on trips
to the near source. This leads to longer searching flights
around the far nectar source and to a lower probability of
finding the source, compared to the near source. Longer
searching flights increase the costs, what leads to the small
left-shift of the lower histogram in Fig. 1c. The higher
probability of missing the source leads to the lowered bee-
counting on the far nectar source (Fig. 6b). The fact that
honeybees are able to forage simultaneously on several
foraging targets, while in parallel scaling the forager
groups on each nectar source according to its quality, is an
interesting phenomenon. It shows, in contrast to many ant
species, that the collective foraging does not quickly
converge into just one single solution, thus the swarm
intelligence of honeybees can be described to be highly
flexible. A quick reaction to environmental changes is only
possible when the colony frequently observes and com-
pares many available solutions.
In experiment 2, we found cross-inhibition in all of our
simulation scenarios. Despite that, Fig. 8 shows that the
receiver-to-forager ratio simultaneously modulates the total
growth capacities of the two groups of foragers. The cross-
inhibitory effect is important for social insect foraging
decisions, as it is a measure of the colony’s ability to reflect
changes of environmental conditions in its internal infor-
mation processing. The parallel exploitation (shown in
experiment 1) ensures that information from various
sources flows into the colony in parallel. Only the cross-
inhibitory effect allows the foragers to react to quality
changes of food sources, which they never have visited
themselves.
Neural Comput & Applic
123
The choice experiments which we performed in exper-
iment 3 tested whether or not the phenomena demonstrated
in experiments 1 and 2 really lead to a quick and adaptive
foraging decision in a changing environment. Our experi-
ments show the huge influence of the receiver-to-forager
ratio onto the foraging patterns that emerge in fluctuating
environments (see Figs. 9, 10). With low receiver-to-for-
ager ratios, the number of foragers that exploit the wors-
ened source in the afternoon always decreases, due to the
cross-inhibition effect shown in experiment 2. The rate of
this decrease is higher, the higher the amplitude of the
environmental fluctuation is. With high receiver-to-forager
ratios, the colony reduces the number of foragers on the
worsened source only if this worsening is strong. With low
fluctuation amplitudes, the colony even increases the
number of foragers on the worsened source and addition-
ally intensifies the foraging on the improved source.
By looking at the net honey gain of the colony in the
runs of experiment 3, we conclude that a low number of
receivers (and thus a strong cross-inhibitory effect) does
not lead to a better nectar income compared to colonies
with high numbers of receivers for the same number of
forager bees. This result was expected, as bigger colonies
naturally have more active foragers and thus more nectar
income per time unit. We found that colonies with low
receiver-to-forager ratios switch quicker from one nectar
source to another after an environmental fluctuation. This
sharpening of collective foraging decisions leads to a
decrease in energetic costs caused by environmental fluc-
tuations compared to runs of a similar sized colony in
undisturbed environmental conditions. In conclusion, we
can say that colonies with low ratios of receivers to for-
agers (0.5:1) react faster and are more picky in their for-
aging decisions. This allows such colonies to make a better
deal out of the environmental situation they are confronted
with. More receiver bees allow more nectar accumulation,
but additional receiver bees increase also the metabolic
expenditures of the colony. In our simulation runs, we
found an optimum with a 1:1 ratio of receivers to foragers
when we measured the per bee efficiency of the nectar
accumulation. As was shown by Seeley [28], this optimal
ratio depends on the duration of the average foraging trip of
forager bees and on the average duration of storage trips by
receiver bees.
As shown in Fig. 14, the availability of receivers for
foragers affects the emerging foraging equilibria that arise
on both nectar sources in multiple ways. More receiver bees
per forager favour recruitment by waggle dances, as queuing
delays are short (see Fig. 4). In addition, as waggle dancers
tend to fly again to the source they danced for, abandonment
rates in forager groups decrease in parallel. The combination
of these effects favours bigger forager groups (more active
foragers, as expressed by f) but keeps foragers longer in
their foraging loop. This slows down recruitment to alter-
native foraging targets in response to environmental fluc-
tuations. This is shown in Figs. 9 and 10.
In conclusion, we can summarize the answers to our
previously formulated hypotheses:
H1: Our model predicts that the receiver-to-forager ratio
affects foraging decisions. A high ratio leads to more active
foragers, but to a slow reacting colony after the first for-
aging decision is settled. Cross-inhibition is affected by
low receiver-to-forager ratio and equal foraging distribu-
tion (EFD) is hard to find with a low number of receivers,
thus the system is more likely to show symmetry-breaking
with two food sources of equal quality.
H2: Our model predicted that stronger environmental
fluctuations enhance the speed of colony-level reactions in
the forager distribution, especially with low numbers of
receivers per forager.
H3: All simulated environmental fluctuations caused
colony-level costs (b \ a, see Fig. 11, also shown in a
summarized view in Fig. 12). Low receiver-to-forager
ratios prevent a significant fraction of these potential costs,
but keep harvests small, even in undisturbed runs. An
excess number of receivers causes an excess of metabolic
costs, thus it is lowering the colony’s harvesting efficiency.
We found an optimal receiver-to-forager ratio at 1:1.
Fig. 14 Receivers affect the emerging foraging distributions on two
nectar sources in multiple ways. More receivers per foragers favour
recruitment by waggle dances, as queuing delays are short (see
Fig. 4). In addition, as waggle dancers tend to fly again to that source,
abandonment rates in forager groups decrease. This favours bigger
forager groups (more active foragers) but keeps foragers longer in
their foraging loop, thus slows down recruitment to other foraging
targets in response to environmental fluctuations
Neural Comput & Applic
123
Our simulation study generated testable predictions for
honeybee scientists and investigated the role of nectar
receivers concerning their effect on colony-level foraging
strategies and nectar economics. Despite a better under-
standing of the biological system, this helps to derive novel
bio-inspired swarm-intelligent algorithms based on hon-
eybee foraging, as they were already proposed [37–40].
Comparable to the honeybee foraging case, ant foraging
has significantly inspired the field of swarm-intelligent
optimisation and routing algorithms [41–43] and swarm
robotics (e.g., [44–47]). Even simple aggregation behav-
iours, like aggregation of cockroaches, has influenced
swarm robotics [48]. Also in the field of swarm robotics,
foraging for energy is a topic of interest in recent years. For
honeybee colonies, nectar is a source (and also a storage
form) of energy, thus we think that the insights we got by
studying the economic gains and costs of honeybee for-
aging will be useful also in the domain of swarm intelli-
gence and swarm robotics.
Recently, we translated several behavioural patterns
found in honeybees to swarm-robotic algorithms: temper-
ature-induced aggregation of young honeybees [49, 50] and
honeybee trophallactic exchanges [51–53]. By further
investigating the key aspects and feedbacks in the regula-
tion of honeybees’ society and collective behaviours, we
plan to elaborate these algorithms and control programs.
Such investigations are performed by empiric experiments
and by model analyses, like we present it in this article.
Acknowledgments This article was supported by the following
grants: EU IST-FET-open project (IP) I-Swarm, no. 507006. EU IST-
FET project SYMBRION, no. 216342. EU ICT project REPLICA-
TOR, no. 216240. Austrian Science Fund (FWF) research grants
P15961-B06 and P19478-B16.
References
1. Sumpter DJT, Pratt SC (2003) A modeling framework for
understanding social insect foraging. Behav Ecol Sociobiol
(53):131–144
2. Bartholdi JJ, Seeley TD, Tovey C, Vate JV (1992) The pattern
and effectiveness of forager allocation among flower patches in
honey bee colonies. J Theor Biol 160:23–40
3. Seeley TD, Camazine S, Sneyd J (1991) Collective decision-
making in honey bees: how colonies choose among nectar sour-
ces. Behav Ecol Sociobiol 28(4):277–290
4. Cox MD, Myerscough MR (2003) A flexible model of foraging
by a honey bee colony: the effects of individual behaviour on
foraging success. J Theor Biol 223:179–197
5. de Vries H, Biesmeijer JC (2002) Self-organization in collective
honeybee foraging: emergence of symmetry breaking, cross
inhibition and equal harvest-rate distribution. Behav Ecol
Sociobiol 51(6):557–569
6. de Vries H, Biesmeijer JC (1998) Modelling collective foraging
by means of individual behaviour rules in honey-bees. Behav
Ecol Sociobiol 44:109–124
7. Anderson C, Ratnieks FLW (1999) Task partitioning in insect
societies. I. Effect of colony size on queueing delay and colony
ergonomic efficiency. Am Nat 154:521–535
8. Ratnieks FLW, Anderson C (1999) Task partitioning in insect
societies II: use of queueing delay information in recruitment.
Am Nat 154(5): 536–548
9. von Frisch K (1965) Tanzsprache und Orientierung der Bienen.
Springer, Berlin
10. Seeley TD (1992) The tremble dance of the honey bee: message
and meanings. Behav Ecol Sociobiol 31:375–383
11. Seeley TD, Camazine S, Sneyd J (1991) Collective decision-
making in honey bees: how colonies choose among nectar sour-
ces. Behav Ecol Sociobiol 28(4):277–290
12. Gruter C, Farina WM (2009) The honeybee waggle dance: can
we follow the steps? Trends Ecol Evol 24(5):242–247
13. Seeley TD (1994) Honey bee foragers as sensory units of their
colonies. Behav Ecol Sociobiol 34:51–62
14. Schmid-Hempel P, Kacelnik A, Houston AI (1985) Honeybees
maximize efficiency by not filling their crop. Behav Ecol
Sociobiol 17:61–66
15. Johnson BR (2003) Organization of work in the honeybee: a
compromise between division of labour and behavioural flexi-
bility. Proc Royal Soc Lond B 270(1511):147–152
16. Seeley TD (1982) Adaptive significance of the age polyethism
schedule in honeybee colonies. Behav Ecol Sociobiol 11:287–293
17. Johnson BR (2002) Reallocation of labor in honeybee colonies
during heat stress: the relative roles of task switching and the
activation of reserve labor. Behav Ecol Sociobiol 51:188–196
18. Schmickl T, Crailsheim K Hopomo (2007) A model of honeybee
intracolonial population dynamics and resource management.
Ecol Model 204(1–2): 219–245
19. Schmickl T, Crailsheim K (2001) Cannibalism and early capping:
strategy of honeybee colonies in times of experimental pollen
shortages. J Comp Physiol A 187(7):541–547
20. Seeley TD (1992) The tremble dance of the honey bee: message
and meanings. Behav Ecol Sociobiol 31:375–383
21. Seeley TD (1989) Social foraging in honey bees: how nectar
foragers assess their colonys nutritional status. Behav Ecol and
Sociobiol 24:181–199
22. Schmickl T, Thenius R, Crailsheim K (2005) Simulating swarm
intelligence in honeybees: foraging in differently fluctuating
environments. In: Proceedings of the genetic and evolutionary
computation conference (GECCO) 2005, Washington, DC,
pp 273–274
23. Schmickl T, Crailsheim K (2004) Costs of environmental fluc-
tuations and benefits of dynamic decentralized foraging decisions
in honey bees. Adapt Behav Anim Anim Software Agents Rob
Adapt Syst 12:263–277
24. Thenius R, Schmickl T, Crailsheim K (2006) Economic optimi-
sation in honeybees: adaptive behaviour of a superorganism. In:
Nolfi S, Baldassarre G, Calabretta R, Hallam JCT, Marocco D,
Meyer JA, Miglino O, Parisi D (eds) From animals to animats 9:
9th international conference on simulation of adaptive behavior,
SAB 2006. Volume 4095 of Lecture Notes in Artificial Intelli-
gence (LNAI). Springer, Berlin, pp 725–737
25. Russell SJ, Norvig P (1995) Artificial intelligence: a modern
approach. Prentice Hall, Englewood Cliffs
26. Bonabeau E, Dorigo M, Theraulaz G (1999) Swarm intelligence:
from natural to artificial systems. Oxford University Press,
Oxford
27. Thenius R, Schmickl T, Crailsheim K (2005) The dance or work
problem: why do not all honeybees dance with maximum
intensity. Lect Notes Artif Intell 3690:246–255
28. Seeley TD (1995) The wisdom of the hive: the social physiology
of honey bee colonies. Havard University Press, Cambridge
Neural Comput & Applic
123
29. Huang M, Seeley TD (2003) Multiple unloadings by nectar for-
agers in honey bees: a matter of information improvement or crop
fullness?. Insectes Sociaux 50:330–339
30. Castro L (2007) Fundamentals of natural computing: an over-
view. Phys Life Rev 4(1):1–36
31. Anderson C (1998) Simulation of the feedbacks and regulation of
recruitment dancing in honey bees. Adv Compl Syst 1:267–282
32. Gregson A, Hart A, Holcombe M, Ratnieks F (2003) Partial
nectar loads as a cause of multiple nectar transfer in the honey
bee (apis mellifera): a simulation model. J Theor Biol 222(1): 1–8
33. Schmickl T, Crailsheim K (2008) Analysing honeybees’ division
of labour in broodcare by a multi-agent model. In: Bullock S,
Noble J, Watson R, Bedau MA (eds) Artificial life XI: proceed-
ings of the eleventh international conference on the simulation
and synthesis of living systems, MIT Press, Cambridge, pp 529–
536
34. Schmickl T, Crailsheim K (2008) An individual-based model of
task selection in honeybees. In: Goebel R, Siekmann J, Wahlster W
(eds) From animals to animats 10. Lecture Notes in Artificial
Intelligence, 5040, MIT Press, Cambridge, pp 383–392
35. Schmickl T, Crailsheim K (2008) Taskselsim: a model of the self-
organization of the division of labour in honeybees. Math Com-
put Model Dyn Syst 14:101–125
36. Thenius R, Schmickl T, Crailsheim K (2008) Optimisation of a
honeybee-colony’s energetics via social learning based on
queuing delays. Connect Sci 20(2):193–210
37. Wedde HF, Farooq M, Pannenbaecker T, Vogel B, Mueller C,
Meth J, Jeruschkat R (2005) Beeadhoc: an energy efficient
routing algorithm for mobile ad hoc networks inspired by bee
behavior. In: GECCO ’05: proceedings of the 2005 conference on
genetic and evolutionary computation. ACM, New York, pp 153–
160
38. Wedde HF, Farooq M, Zhang Y (2004) Beehive: An efficient
fault-tolerant routing algorithm inspired by honey bee behavior.
In: Lecture notes in computer science. Number 3172, Springer,
Berlin, pp 83–94
39. Tovey C (2004) The honey bee algorithm: a biological inspired
approach to internet server optimization. Engineering Enterprise,
Spring, pp 13–15
40. Pham D, Ghanbarzadeh A, Koc E, Otri S, Rahim S, Zaidi M
(2006) The bees algorithm, a novel tool for complex optimisation
problems. In: Proceedings of the 2nd international virtual con-
ference on intelligent production machines and systems (IPR-
OMS 2006), Elsevier, pp 454–459
41. Dorigo M, Stutzle T (2004) Ant colony optimization (Bradford
Books). The MIT Press, Cambridge
42. Dorigo M, Bonabeau E, Theraulaz G (2000) Ant algorithms and
stigmergy. Future Gener Comput Syst 16(9):851–871
43. Bonabeau E, Henaux F, Guerin S, Snyers D, Kuntz P, Theraulaz G
(January 1998) Routing in telecommunications networks with
‘‘smart’’ ant-like agents. Working papers 98-01-003, Santa Fe
Institute
44. Sugawara K, Kazama T, Watanabe T (2004) Foraging behavior
of interacting robots with virtual pheromone. In: Proceedings of
2004 IEEE/RSJ international conference on intelligent robots and
systems. IEEE Press, Los Alamitos, pp 3074–3079
45. Krieger MJB, Billeter JB (2000) The call of duty: self organised
task allocation in a population of up to twelve mobile robots. Rob
Auton Syst 30:65–84
46. Payton D, Daily M, Estowski R, Howard M, Lee C (2001)
Pheromone robotics. Auton Rob 11(3):319–324
47. Payton D, Estkowski R, Howrad M (2005) Pheromonic robotics
and the logic of virtual pheromones. Lect Notes Comput Sci
3342:45–57
48. Garnier S, Jost C, Jeanson R, Gautrais J, Asadpour M, Caprari G,
Theraulaz G (2005) Aggregation behaviour as a source of col-
lective decision in a group of cockroach-like-robots. In:
Capcarrere M (ed) Advances in artificial life: 8th European
conference, ECAL 2005. Vol 3630 of LNAI. Springer, Berlin,
pp 169–178
49. Schmickl T, Thenius R, Moslinger C, Radspieler G, Kernbach S,
Crailsheim K (2008) Get in touch: cooperative decision making
based on robot-to-robot collisions. Auton Agent Multi Agent Syst
18(1):133–155
50. Hamann H, Worn H, Crailsheim K, Schmickl T (2008) Spatial
macroscopic models of a bio-inspired robotic swarm algorithm.
In: IEEE/RSJ 2008 international conference on intelligent robots
and systems (IROS’08). IEEE Press, Los Alamitos, pp 1415–
1420
51. Schmickl T, Moslinger C, Thenius R, Crailsheim K (2007) Bio-
inspired navigation of autonomous robots in heterogenous envi-
ronments. Int J Factory Autom Rob Soft Comput 3:164–170
52. Schmickl T, Moslinger C, Thenius R, Crailsheim K (2007)
Individual adaptation allows collective path-finding in a robotic
swarm. Int J Factory Autom Rob Soft Comput 4:102–108
53. Schmickl T, Crailsheim K (2008) Trophallaxis within a robotic
swarm: bio-inspired communication among robots in a swarm.
Auton Rob 25:171–188
Neural Comput & Applic
123