Artificial Immune, Neural and Endocrine Systems
DOI: http://dx.doi.org/10.7551/978-0-262-31709-2-ch122
EMANN - a model of emotions in an artificial neural network
Ronald Thenius, Payam Zahadat and Thomas Schmickl
Artificial Life Laboratory, University of Graz, Austria.
ronald.thenius@uni-graz.at
Abstract
With this paper we propose an advanced concept of Artificial Neural Networks (ANN, described in detail in the following), using biological observations as guideline. Please
note, that the proposed model was developed for the purpose
to simulate emotions, and to investigate features of the interplay of biochemical, neuromodulatory processes with neural
system. If and to what extend this is advantageous for technical solutions is topic of ongoing research (as mentioned
below).
With the work at hand we want to present a model of a neural system, that is influenced by emotions. This model is
based on the state of the art concept of artificial neural networks (ANN) which we improve by adding ‘artificial emotions’. The way of the implementation of emotions is based
on the research results regarding the biological and biochemical processes modulating neural cells in animals and humans.
The described modulation takes place not only on the level
of synapses, but also on the level of calculations happening
on the membrane of the cell, or the node of the ANN, respectively. The suggested model also includes the biological fact, that neuro modulatory glands (e.g., hypophysis) are
mainly controlled by the neuronal system itself. The resulting
proposed system is named EMotional Artificial Neural Network (EMANN). It shows that EMANN has different abilities, compared to ANNs without emotions.
Short overview over the state of the art of ANNs
Introduction
This work is mainly inspired by the work of Fellous (Fellous, 1999; Fellous et al., 2002; Fellous, 2009; Arbib and
Fellous, 2004), who discussed the necessity of emotions in
artificial systems. Fellous and his co-authors argue, that
models of emotions can not only be used in artificial systems as a tool (e.g., to organise activity levels of different
tasks with a single robot), but also serve as a model for emotions in animals and humans. Fellous described his ideas
and requirements about the implementation of emotions in
artificial systems as follows:
• Emotions should not be implemented as a separate, specialized module in charge of computing an emotional
value on some dimensions.
• Emotions should not simply be the result of cognitive
evaluations.
• Emotions are not linear (or non linear) combinations of
some pre-specified basic emotions.
• Emotions should be allowed to have their own temporal
dynamics, and should be allowed to interact with one another.
ECAL 2013
830
The concept of ANNs was developed in the first half of
20th the century (McCulloch and Pitts, 1943; Rosenblatt,
1958). Since then several variations of ANNs where developed (Jain et al., 1996), be it differential equation based
ANNs (CTRNN, by Beer (1995)) or models of spiking
ANNs (Hindmarsh and Rose, 1984).
In all this mentioned methods, information is filtered by
the ANN without any feedback of the information on the
method of processing: The activity of one node influences
the activity of another node via the connection of the two
nodes. Via this signal transduction the signal is modulated
by the weight associated with the regarding link. Manipulations of the process of filtering usually takes place on the
level of synapses, and is done during the learning (or evolving) process. Only a few works exist, which try to modulate
the process of data processing itself, based on input or output of the network (Neal and Timmis, 2003; Timmis et al.,
2009). To our knowledge, no concept was suggested, in
which the way of calculation is modulated and controlled by
the network itself. From a biological point of view this ability of self-modulation is highly relevant in biological neural
networks, and its implementation will add many new features to the concept of ANNs.
Emotions and neurons in biological brains
On a biological level, moods and emotion can be understood as the condition an animal is in, based on an internal
chemical (mostly hormonal) situation. This internal chemical situation modulates mostly all physiological and neural processes, what comes along with a modulation of as-
Artificial Immune, Neural and Endocrine Systems
sociated behavioural phenomenons. Examples for this are
love (Carter, 1998; Young et al., 2001), fear (LeDoux, 2000;
Derntl et al., 2009) and stress (Vermetten and Bremner,
2002). This way the moods influence as well “low-level” social behaviours (e.g., aggression (Pope et al., 2000)) as well
as “high level” social behaviours (e.g, trust) (Kosfeld et al.,
2005; Donaldson and Young, 2008; Guastella et al., 2008)
etc. On a cellular/chemical level the emotional status is the
result of a feedback-loop between the neural system and the
hormonal system: On one hand the hormones influence the
activity of the biological neural system (Derntl et al., 2009;
Joëls, 1997), on the other hand the highest unit in the hierarchical cascade of interacting hormone glands (hypophysis,
glandula pituitaria) is influenced by the central nervous system via the hypothalamus. Please note, that not all emotions
emerge from such a feedback loop in the first place, but can
also be triggered by other sources, e.g., by the genome directly.
Based on this phenomenon we developed the EMANN
system (EMotional Artificial Neural Network), which improves the classical ANN paradigm (Jain et al., 1996) by the
ability to develop hormone glands, that influence the abilities of the neural nodes. In this context the “mood” of an
EMANN can be understood as the current hormonal status
of the network. How much this “mood” influences the current behaviour of the controlled agent is coded into the individual responsiveness of the nodes of the EMANN.
Method
The concept
The suggested model is based on standard ANNs (Jain et al.,
1996) which are improved by the features of a simulated
neuromodulatory hormonal system. The simulated hormonal glands emit a number of virtual hormones, which
represent all types of neuromodulators in a biological brain.
These virtual hormones influence the behaviour of the individual nodes (which represent the individual cells within a
biological neural system), including the synapses, the functions which sum up the inputs of the node, and the output
function (which is analogous to the function of the axon hill
of the neural cell). In return, the hormone glands are controlled by the neural network: each cell within the described
network is able to emit hormones, which turns each node
in the network into a possible hormone gland. To enable a
virtual cell to react to a given hormonal level, hormonal receptors for every part of the virtual cell are simulated. These
hormonal receptors are analogous to surface hormonal receptors in the membrane of biological neurons. A schematic
drawing of the concept of ANN and EMANN is depicted in
the Figures 1 and 2.
Throughout this paper, ‘node’ and ‘cell’, and also ‘edge’
and ‘synapse’ are used as synonyms. The main difference between the nodes of an ANN and an EMANN is that
EMANN does not only map information from an input to an
831
Figure 1: Schematic drawing of a single node in a stateof-the-art artificial neural network (ANN). The input of a
node (I - III), which represents cells in a biological neural
network, comes from other nodes of the ANN (IV ) and/or
from the input of the ANN. The weights of the edges between the nodes (I), representing synapses in the biological
counterpart, are usually adapted to a given task by learning
algorithms or by using artificial evolutionary methods. The
net-function (II), modelling the processes in the dendritic
part of a biological neuron, sums up the weighted inputs of
the node. The out-function (III), from a functional point of
view the analogon to the axon-hill in the biological neural
cell, calculate the new status of the node. For more details
about the state of the art in ANNs see (Jain et al., 1996). The
solid arrows indicate the flow of information within the system, round edged triangles indicate the weights (synapses)
of the node.
output level, but at the same time elements of a feedback system, that change the behaviour of the cells based on dynamic
hormone levels. The “out-function” of the ANN differs from
the EMANN analogon which is called “Hill-function” since
the later is strongly influenced by hormones during runtime,
but the former is usually static during runtime.
From the biological concept to a mathematical
representation
As represented in Figure 2, inputs to a cell of an EMANN
system first pass through synapses where each synapse contains a weight that can be influenced by hormones (Figure 2:I). The value from the synapse then passes a netfunction (represented by g(X) in Figure 2:II) and the total value of the net-outputs from all the input synapses is
calculated. Finally, this sum passes a hill-function (represented by f (X) in Figure 2:III). The parameters of both
net-functions and hill-functions are influenced by hormones.
In the following formal representations, Wi,j represents
the weight of jth input synapse of cell i. Its value is specified
as:
Wi,j (H) = θi,j +
φi,j,h Hh
(1)
h
where H is the set of hormone levels in the system and
Hh is the level of hormone h. θi,j is a constant weight for
jth synapse of cell i and φi,j,h is the responsiveness of the
synapse to hormone h.
ECAL 2013
Artificial Immune, Neural and Endocrine Systems
An input to cell i from its jth synapse is represented by
Xi,j . The input passes through the synapses and the netfunction N eti,j that is described by:
N eti,j (X, H) = N etki (H)×f (Wi,j (H)Xi,j )+N etdi (H),
(2)
N etki (H) = βi +
ηi,h Hh ,
(3)
h
N etdi (H) = αi +
κi,h Hh ,
(4)
h


x < −1
1
f (x) = −1 x > 1


x
else
(5)
where βi and αi are constant parameters of cell i and
ηi,h and κi,h represent influences of hormone h on the netfunction of the cell.
The net-outputs are then summed up and pass the hillfunction and then produce output of the cell that is described
as follows:
Figure 2: Schematic drawing of a single cell (analogon to
the nodes in a state-of-the-art ANN) in an Emotional Artificial Neural Network (EMANN). In contrast to a stateof-the-art ANN (figure 1) the cell of the EMANN (I-III)
is able not only to receive information from the EMANN
(IV ), or send information to it, but also to produce hormones (V III). These hormones are represented by global
variables, that sum up the hormone output of one time step
of all cells of the EMANN (V II, equ. 12). Each cell is influenced by these hormones (IX), be it on the synaptic level
(XII, see equ. 1) on the level of the net-function (XI, see
equ. 2) or on the level of the hill function (X , see equ. 6).
The degree, in which a cell functions as a neural cell or as a
gland for a given hormone is determined by weights (V , V I,
see equ. 13 and equ 11). Due to the interaction of cellular
activity and hormone levels, a feedback arises, that allows
to change the method of information processing, based on
the information itself. Solid arrows indicate the flow of information within the EMANN (by which the input of the
neural net is mapped to the neural output), the dotted lines
indicate hormonal modulatory pathways.
Hilli (X, H) = Hillki (H)×f (
N eti,j (X, H))+Hilldi (H),
j
Hillki (H) = λi +
(6)
σi,h Hh ,
(7)
ρi,h Hh ,
(8)
h
Hilldi (H) = δi +
h
Yi (X) = g(Hilli (X, H))
(9)
where g(x) is a unit step function. λi and δi are constant
parameters of cell i and σi,h and ρi,h represent influences of
hormone h on the hill-function of the cell.
The output of a hill-function that is produced based on
inputs to the cell can be summarized as:
Yi = g((λi +
[(βi +
h
+ (αi +
h
σi,h Hh ) × f (
j
ηi,h Hh ) × f ((θi,j +
κi,h Hh )]) + (δi +
h
φi,j,k Hh )Xi,j )
(10)
h
ρi,h Hh ))
h
The output of the cell then is calculated as a factor of the
output of the hill-function as:
outputi = neuralityi × Yi
(11)
The overall amount of hormone h in the system is calculated as follows:
Hh =
Hi,h ,
(12)
i
ECAL 2013
832
Artificial Immune, Neural and Endocrine Systems
Hi,h = glandityi,h × Yi
(13)
where Hi,h is the share of cell i in production of hormone h
and glandityi,h is a parameter representing the production
factor of hormone h in the cell.
By deactivating the parameters of different modules of
this system, e.g. net-function, we get simplified versions
of the system.
A
B
A
B
0
1
0
1
2
3
2
3
The task
To test, if the proposed model has features, that differ from
a “non-emotional” neural network, we developed a task, to
test the very basic features of the system.
The implemented task is to solve the following equation:
A
if A + B > 0.5
output(A, B) =
(14)
0
otherwise
A simplified version of EMANN which uses the hillfunction is implemented. A population of randomly generated ANNs and a population of randomly generated
EMANNs are evolved for the task. The experiment is repeated for 100 independent runs in each case. The structures
of the implemented ANN and EMANN are represented in
Figure 3. The genomes are a sequence of parameters for
each network and astandard genetic algorithm is used as the
evolutionary algorithm. Experimental settings are summarized in Table 1.
Table 1: Experimental settings
number of cells
4
number of synapses
4
number of hormones
1
active-modules
hill-function
population size
250
number of generations
25000
mutation rate
0.7
mutation probability
0.2
Output
Hormone
EMANN
Output
ANN
Figure 3: Network structures for the given task. The ANN,
as well as the EMANN, have two inputs, and one neural output (as described in the section ). Besides this, the
EMANN has a hormonal output, that influences the hillfunction (equ. 7) of all cells in the network.
Results
The test of the EMANN concept in the task described above
showed that the EMANN is able to generate higher fitness
values in an evolutionary run comparing an ANN. The experiment was performed several times. An exemplary evolutionary run is depicted in figure 4. It shows, that both ANN
and EMANN have a very fast increase of the fitness level
reached within the first view generations. ANN reaches its
maximum fitness of about 0.6 with the first few generations,
while EMANN increases its reached fitness throughout the
whole experiment.
Repeating the experiment several times showed that the
better results from EMANN paradigm in comparison to the
used state-of-the-art ANN (see figure 5 for more details) is
statistically significant.
Discussion
In his work Fellous (Fellous, 2004) argues that the implementation of emotions in an artificial system may not only
be useful as a model for emotional processes in real-world
lifeforms, but also may have advantages for purely artificial
systems. As depicted in figure 5 we can show these assumptions are valid for the test described in using an EMANN
and comparing it to a state-of-the-art ANN.
Regarding the biological counterparts, it is interesting
from a biological point of view that state of the art ANN
paradigms in most instances omit the fact, that in biological
neural control systems are massively modulated, not to say
‘controlled‘ by emotions. This phenomenon can be found
in both, ‘regular‘ emotions (as already mentioned above) or
emotions based on diseases (e.g., depression, as described
by Harding et al. (2004)). These modulations of the nervous
system and the associated behaviour can vary from a biasing of the behavioural pattern, to a complete change of be-
833
ECAL 2013
0.8
Fitness
0.6
0.7
0.7
0.6
0.5
0.4
Fitness
*
0.9
0.8
1.0
Artificial Immune, Neural and Endocrine Systems
0
5000
10000
15000
20000
0.5
0.3
EMANN
ANN
EMANN
25000
ANN
Generation
Figure 4: Exemplary evolutionary run of EMANN and
ANN. It shows that the EMANN is able to solve the task
better then the ANN. The task to solve is described in section . Curves indicate the maximum fitness value of every
generation. Fitness is calculated according to equ. 14
haviour. Due to this, emotions can be understood as a kind
of biological task-selection mechanism, from an engineering point of view. In literature the concept of emotions has
been tested before, but always with a strict separation of hormonal system (e.g. hormonal glands) and the neural system:
Gadanho and Hallam (2002) proposed a hormonal-neural
system, which is described as a complex multilevel model,
consisting of different compartments for cellular structures,
neuromodulators and emotional activities. Neal and Timmis
(2003) and Timmis et al. (2009) showed a system, in which
the hormones interact with the neural system on synaptic
level, but the hormone glands are a unit on its own on the
input level of the system, with no ability to be influenced directly from the ANN. In the proposed EMANN paradigm,
all actions and interactions, be it neurological, hormonal,
hormone-gland related, neuromodulatory or emotional, take
place on the level of cells and cellular interactions. Especially the ability of a cell to communicate with other cells on
a local level (via neurotransmitters/synapses) or on a global
level (via hormones) allows more complex activity patterns
than a state of the art ANN. Further the cells in an EMANN
have the ability not only to exchange information between
cells (equ. 9) but also to send commands regarding changes
in the information processing within the network (equ. 13)
to groups of other cells within the EMANN. This results in
a control system, which has the ability to modulate parts of
the information processing system by itself, depending on
the input of the EMANN.
Regarding the suggestions for the design of robotemotions of Fellous, described in (Fellous, 2004), are re-
ECAL 2013
834
Figure 5: Maximum fitness values of last generation of 100
independent runs with 25000 generations. It shows that
the EMANN performs significantly better in the given task
compared to am ANN (*:p < 0.05; Wilcoxon signed-rank
test, unpaired date, ”two.sided”-hypothesis, program used
for statistical analysis: “R”). Box-plots indicate median and
quartiles, whiskers indicate minimum and maximum, circles
indicate outliers. Fitness is calculated according to equ. 14
flected in our work:
• Emotions should not be implemented as a separate, specialized module in charge of computing an emotional
value on some dimension. In the EMANN paradigm hormone glands and neural cells are interlinked on a cellular
level. As displayed in figure 2, a cell can act as a neural
cell (equ. 9) as well as a hormone gland (equ. 13).
• Emotions should not simply be the result of cognitive evaluations. In an EMANN the emotions are the result of
the interplay of neural activity and the interaction of hormones (by the modulation of neural cells and glands). The
actual level of hormones are the outcome of a complex
self-organising feedback system.
• Emotions are not linear (or non linear) combinations of
some pre-specified basic emotions. The hormones in the
EMANN paradigm do not represent any predefined emotions. Hormones act as neuromodulators, which change
the functionality of the controller (in this case a neural
network). This modulation can be understood as emotion.
• Emotions should be allowed to have their own temporal
dynamics, and should be allowed to interact with one another. In an EMANN the emotions are an effect of the hormone levels, which are a result of the collective gland activity of all EMANN cells. This way the temporal dynamics of the emotion can be at least of the same dimension
Artificial Immune, Neural and Endocrine Systems
of the neural cells, or much lower, depending from the
degree and kind of neural linking of the hormone gland
cells. Please note, that in the EMANN paradigm, the temporal dynamic of the hormones can not be higher than the
temporal dynamics of the neural cells, but it can be equal
or lower. The interaction of hormones takes place via the
cells, which are receptors of hormones as well as emitters
of hormones.
The results, depicted above show, that the EMANN for
the given task is able to outperform an ANN significantly
(figure 5). The task selected for this test was, from the authors point of view, the most simple possible, to show the
abilities of the EMANN in a well defined small scale experiment.
Conclusion and Outlook
Our work presents the first results of investigations in a
model of emotions in an neural network, the Emotional Artificial Neural Network (EMANN). The EMANN is a development based on state of the art artificial neural networks
(ANN, for an overview see (Jain et al., 1996)), improved by
the ability to get modulated in its functionality by emotions.
The emotions are implemented according to biological examples, inspired by the ideas presented in (Fellous, 2004).
It showed, that an EMANN is able to outperform a state-ofthe-art ANN in the given test. This shows, that the simulated
emotions can have a positive effect on the performance of
the ANN.
This paper is the first one in a row of investigations about
the features of emotions in artificial neural networks. In the
next step we will investigate, how learning algorithms based
on emotions (inspired from the biological counterparts) can
be used with an EMANN controller. We plan to investigate, how this system of neural, and hormonal feedbacks
interacts under conditions of a changing environment, and
how given evolutionary constraints (e.g., changing environments on an evolutionary time-scale) change the behaviour
of the evolved EMANN controlled agents in complex environments. We further want to investigate, if the effect of
genetically fixation of behaviour (“Baldwin effect” (Baldwin, 1896)) is influenced by the presence of emotions in the
control system.
To investigate the evolutionary development of neural tissues and brain structures in an artificial system we plan
to combine the EMANN paradigm with algorithms, that
are able to shape morphological structures and controller
structures in an evolutionary manner (Thenius et al., 2009b;
Schmickl et al., 2011b; Kernbach et al., 2009; Thenius et al.,
2010; Dauschan et al., 2011; Thenius et al., 2009a). To what
extend the interplay of abiotic environment, social environment and behaviour can change the shape of brains in an
artificial system (as described for biological systems, e.g.,
by Breedlove and Arnold (1980)) will be investigated in the
835
next step. We also plan to investigate the influence of emotions on the self-healing abilities (Thenius et al., 2011) of a
controlling tissue in an artificial structured controller of an
EMANN system.
One goal of the research in EMANN will be to investigate how the concept of emotions can be used in single
autonomous robots or swarms of robots in complex environment (e.g., for (Schmickl et al., 2011a; Kernbach et al.,
2008)). We plan, on a robotic swarm level, to share artificial hormones between spatially close agents (e.g., via shortrange communication), what can result in a spatial organised task allocation or even spatially organised calculation
processes based on interacting hormone levels inside a (big)
swarm. The conditions, under which hormones are emitted
within the swarm, and how the controllers of the swarm react to the given hormone level, can be as well engineered by
hand, as well as shaped by an artificial evolutionary process.
Another goal is to investigate the role of emotions in the
development of complex behaviours in an eusocial system
, e.g., behaviours comparable to the BEECLUST algorithm
(Schmickl et al., 2008; Bodi et al., 2012; Hereford, 2011).
To what extend the proposed system can be used as a
model for biological and psychological processes (as suggested by Fellous in (Fellous, 2004)) will also be investigated in the future.
Acknowledgements
This work was supported by: EU-IST-FET project SYMBRION, no. 216342; EU-ICT project REPLICATOR, no.
216240; EU-ICT ‘CoCoRo’, no. 270382; EU-ICT ‘ASSISIbf’, no. 601074; Austrian Science Fund (FWF) research
grants: P19478-B16 and P23943-N13 (REBODIMENT)
References
Arbib, M. A. and Fellous, J.-M. (2004). Emotions: from brain to
robot. Trends in Cognitive Sciences, 8(12):554–561.
Baldwin, M. J. (1896). A new factor in evolution. The American
Naturalist, 30(354):441–451.
Beer, R. D. (1995). On the dynamics of small continuous-time
recurrent neural networks. Adaptive Behavior, 3:469–509.
Bodi, M., Thenius, R., Szopek, M., Schmickl, T., and Crailsheim,
K. (2012). Interaction of robot swarms using the honeybeeinspired control algorithm beeclust. Mathematical and Computer Modelling of Dynamical Systems, 18(1):87–100.
Breedlove, S. M. and Arnold, A. P. (1980). Hormone accumulation
in a sexually dimorphic motor nucleus of the rat spinal cord.
Science, 210(4469):564–566.
Carter, C. S. (1998). Neuroendocrine perspectives on social attachment and love. Psychoneuroendocrinology, 23(8):779–818.
Dauschan, M., Thenius, R., Schmickl, T., and Crailsheim, K.
(2011). Using virtual embryogenesis multi-robot organisms.
In International Conference on Adaptive and Intelligent Systems, ICAIS’11, Klagenfurt, AT, September 06-08, 2011. Proceedings, pages 238–247.
ECAL 2013
Artificial Immune, Neural and Endocrine Systems
Derntl, B., Windischberger, C., Robinson, S., Kryspin-Exner, I.,
Gur, R. C., Moser, E., and Habel, U. (2009). Amygdala activity to fear and anger in healthy young males is associated with
testosterone. Psychoneuroendocrinology, 34(5):687–693.
Donaldson, Z. R. and Young, L. J. (2008). Oxytocin, vasopressin,
and the neurogenetics of sociality. Science, 322:900–904.
Fellous, J.-M. (1999). The neuromodulatory basis of emotion. The
Neuroscientist, 5(5):283–294.
Fellous, J.-M. (2004). From human emotions to robot emotions.
In Hudlicka, E. and Canamero, L., editors, Architectures for
Modeling Emotion: Cross-Disciplinary Foundations, AAAI
Spring 2004 symposium, Stanford. American Association for
Artificial Intelligence.
Fellous, J.-M. (2009). Emotion: Computational modeling. In
Squire, L. R., editor, Encyclopedia of Neuroscience, volume 3, pages 909–913. Oxford: Academic Press.
Fellous, J.-M., Armony, J. L., and LeDoux, J. E. (2002). Emotional
circuits and computational neuroscience. In Arbib, M. A.,
editor, The Handbook of Brain Theory and Neural Networks.
The MIT Press, 2 edition.
Gadanho, S. C. and Hallam, J. (2002). Robot learning driven by
emotions. Adaptive Behavior, 9(1):42–64.
Guastella, A. J., Mitchell, P. B., and Dadds, M. R. (2008). Oxytocin
increases gaze to the eye region of human faces. Biological
psychiatry, 63(1):3–5.
Harding, E. J., Paul, E. S., and Mendl, M. (2004). Animal behaviour: Cognitive bias and affective state. Nature, 427:312.
Hereford, J. M. (2011). Analysis of BEECLUST swarm algorithm.
In Proc. of the IEEE Symposium on Swarm Intelligence (SIS
2011), pages 192–198. IEEE.
Hindmarsh, J. L. and Rose, R. M. (1984). A model of neuronal
bursting using three coupled first order differential equations.
Proceedings of the Royal society of London. Series B. Biological sciences, 221(1222):87–102.
Jain, A. K., Mao, J., and Mohiuddin, K. (1996). Artificial neural
networks: A tutorial. IEEE Computer, 29:31–44.
Joëls, M. (1997). Steroid hormones and excitability in the mammalian brain. Frontiers in neuroendocrinology, 18(1):2–48.
Kernbach, S., Hamann, H., Stradner, J., Thenius, R., Schmickl, T.,
Crailsheim, K., van Rossum, A., Sebag, M., Bredeche, N.,
Yao, Y., Baele, G., de Peer, Y. V., Timmis, J., Mohktar, M.,
Tyrrell, A., Eiben, A., McKibbin, S., Liu, W., and Winfield,
A. F. (2009). On adaptive self-organization in artificial robot
organisms. In The First International Conference on Adaptive
and Self-adaptive Systems and Applications (ADAPTIVE’09).
IEEE Press.
Kernbach, S., Meister, E., Schlachter, F., Jebens, K., Szymanski,
M., Liedke, J., Laneri, D., Winkler, L., Schmickl, T., Thenius, R., Corradi, P., and Ricotti, L. (2008). Symbiotic Robot
Organisms: REPLICATOR and SYMBRION Projects. In
Madhavan, R. and Messina, E., editors, Proceedings of the
Performance Metrics for Intelligent Systems Workshop (PerMIS), pages 62–69. NIST Special Publication 1090.
ECAL 2013
836
Kosfeld, M., Heinrichs, M., Zak, P. J., Fischbacher, U., and Fehr,
E. (2005). Oxytocin increases trust in humans. Nature,
435:673–676.
LeDoux, J. E. (2000). Emotion circuits in the brain. Annual review
of neuroscience, 23:155–184.
McCulloch, W. S. and Pitts, W. (1943). A logical calculus of ideas
immanent in nervous activity. Bulletin of Mathematical Biophysics, 5(4):115–133.
Neal, M. and Timmis, J. (2003). Timidity: A useful emotional
mechanism for robot control? Informatica, 27(4):197–204.
Pope, H. G., Kouri, E. M., and Hudson, J. I. (2000). Effects of supraphysiologic doses of testosterone on mood and aggression
in normal men: A randomized controlled trial. Archives of
General Psychiatry, 57(2):133–140.
Rosenblatt, F. (1958). The perceptron: A probalistic model for
information storage and organization in the brain. Psychological Review, 65(6):386–408.
Schmickl, T., Thenius, R., Christoph, M., Timmis, J., Tyrrell,
A., Read, M., Hilder, J., Halloy, J., Campo, A., Stefanini,
C., Manfredi, L., Orofino, S., Kernbach, S., Dipper, T., and
Sutantyo, D. (2011a). Cocoro – the self-aware underwater
swarm. In Self-Adaptive and Self-Organizing Systems Workshops (SASOW), 2011 Fifth IEEE Conference on, pages 120–
126.
Schmickl, T., Thenius, R., Möslinger, C., Radspieler, G., Kernbach, S., and Crailsheim, K. (2008). Get in touch: Cooperative decision making based on robot-to-robot collisions. Autonomous Agents and Multi-Agent Systems, 18(1):133–155.
Schmickl, T., Thenius, R., Stradner, J., Hamann, H., and Crailsheim, K. (2011b). Robotic organisms: Artificial homeostatic
hormone system and virtual embryogenesis as examples for
adaptive reaction-diffusion controllers. In Kernbach, S.
and Fitch, R., editors, IROS 2011 Workshop–Reconfigurable
Modular Robotics: Challenges of Mechatronic and BioChemo-Hybrid Systems.
Thenius, R., Bodi, M., Schmickl, T., and Crailsheim, K. (2009a).
Growth of structured artificial neural networks by virtual embryogenesis. In Kampis, G., Karsai, I., and Szathmáry, E.,
editors, ECAL (2), volume 5778 of Lecture Notes in Computer Science, pages 118–125. Springer.
Thenius, R., Bodi, M., Schmickl, T., and Crailsheim, K. (2010).
Evolving artificial neural networks and artificial embryology.
In Levi, P. and Kernbach, S., editors, Symbiotic Multi-Robot
Organisms: Reliability, Adaptability, Evolution, pages 266–
268. Springer-Verlag.
Thenius, R., Dauschan, M., Schmickl, T., and Crailsheim, K.
(2011). Regenerative abilities in modular robots using virtual
embryogenesis. In International Conference on Adaptive and
Intelligent Systems, ICAIS’11, Klagenfurt, AT, September 0608, 2011. Proceedings, pages 227–237.
Thenius, R., Schmickl, T., and Crailsheim, K. (2009b). Novel
concept of modelling embryology for structuring an artificial
neural network. In Troch, I. and Breitenecker, F., editors,
Proceedings of the MATHMOD, pages 1821–1831.
Artificial Immune, Neural and Endocrine Systems
Timmis, J., Neal, M., and Thorniley, J. (2009). An adaptive
neuro-endocrine system for robotic systems. In IEEE Workshop on Robotic Intelligence in Informationally Structured
Space. Part of IEEE Workshops on Computational Intelligence, Nashville.
Vermetten, E. and Bremner, J. D. (2002). Circuits and systems
in stress. i. preclinical studies. Depression and anxiety,
15(3):126–147.
Young, L. J., Lim, M. M., Gingrich, B., and Insel, T. R. (2001).
Cellular mechanisms of social attachment. Hormones and
Behavior, 40:133–138.
837
ECAL 2013