A Bottom-Up Investigation of Emotional
Modulation in Competitive Scenarios
Lola Ca˜namero and Orlando Avila-Garc´ıa
Adaptive Systems Research Group
School of Computer Science, University of Hertfordshire
College Lane, Hatfield, Herts AL10 9AB, UK
L.Canamero@herts.ac.uk, newoavila@hotmail.com
Abstract. In this paper, we take an incremental, bottom-up approach
to investigate plausible mechanisms underlying emotional modulation of
behavior selection and their adaptive value in autonomous robots. We
focus in particular on achieving adaptive behavior selection in competitive
robotic scenarios through modulation of perception, drawing on
the notion of biological hormones. We discuss results from testing our
architectures in two different competitive robotic scenarios.
1 Introduction
One of the main problems for autonomous robots is behavior selection or “what
to do next” [12]. Motivation-based architectures [12,4,17,2] integrate a combination
of internal and external factors to select the appropriate behavior and
satisfy the robot’s needs in real time. However, these architectures are not always
sufficiently adaptive to rapid environmental changes. Previous work [5]
postulated the use of second-order mechanisms, akin to some of the functions
of emotions in biological systems, that act on other elements in the architecture
for improved performance in dynamic, unpredictable, and dangerous environments.
In that architecture and others that have followed a similar approach,
the adaptive functions of emotions are predefined by the designer. While this
nowadays widespread design practice can produce efficient behavior selection, it
leaves unanswered the question of which are the underlying mechanisms and how
they integrate and interact with other elements to achieve adaptive behavior. In
the work presented here, we take an incremental approach to investigate plausible
mechanisms underlying emotional modulation of behavior selection and
their adaptive value. We are particularly interested in how such modulation can
achieve different functionalities from the same architecture by interacting with
other elements, rather than including emotions as additional components. In
this paper, we focus on discussing how behavior selection can be made adaptive
(i.e., its output biased) to different environmental situations (two different competitive
robotic scenarios) by modulating different sensory channels—perception
of external and internal stimuli. Drawing on the notion of biological hormones,
 Currently with OpenCanarias S.L.
A. Paiva, R. Prada, and R.W. Picard (Eds.): ACII 2007, LNCS 4738, pp. 401–412, 2007.
c Springer-Verlag Berlin Heidelberg 2007
402 L. Ca˜namero and O. Avila-Garc´ıa
Fig. 1. Experimental setups used to carry out the studies: a Competitive Two-Resource
Problem scenario (left), and a “prey-predator” scenario showing the prey robot inside
the nest and the predator outside (right)
we have modeled two of the functionalities ascribed to them in order to improve
the adaptation of motivation-based architectures to different problems.
To achieve different functionalities from the same architecture, we have taken
inspiration from neuroscience models of hormonal control [9,10], in particular regarding
the following ideas: (a) Sensory inputs enhance the release of hormones
that act at different levels of the nervous system; (b) they act as gain-setting
sensitization processes that bias the output of the organism in particular directions;
and (c) after modulation, the organism responds to particular sensory
stimuli with an altered output appropriate to the new situation. We have tested
our “hormone-like” mechanisms in two dynamic and unpredictable competitive
robotic scenarios depicted in Figure 1, and show how they improve adaptation
and performance using quantitative indicators based on the notion of viability.
Finally, we analyze the results in terms of interesting behavioral phenomena that
emerge from the interaction of these artificial hormones with the rest of architectural
elements and the environment, and that resemble “emotional” behavior
in biological systems confronted to similar situations.
2 Behavior Selection Architecture
Following [5], in our architecture behavior selection results from the interactions
of a number of elements integrated through an artificial physiology and in
interaction with the environment.
The physiology consists of (1) survival-related, homeostatically controlled
essential variables and (2) hormones. Essential variables are abstractions representing
the level of internal resources that the robot needs in order to survive.
They must be kept within a range of permissible values for the robot to remain
viable or “alive,” thus defining a physiological space [14] or viability zone [1,13]
within which survival (continued existence) is guaranteed, whereas transgression
A Bottom-Up Investigation of Emotional Modulation 403
of these boundaries leads to “death.” Hormones can be seen as second-order control
mechanisms that affect the behavior of other elements of the arhitecture.
Motivations are abstractions representing tendencies to act in particular
ways as a function of internal and external factors [18]. Internal factors are
mainly (but not only) physiological deficits (0 ≤ di ≤ 1) or bodily needs—
traditionally known as “drives”—that set urges to action to maintain the state
of the controlled physiological variables within the viability zone. External factors
are environmental stimuli, commonly termed “incentive cues” in Ethology,
(0 ≤ ci ≤ 1) that allow to satisfy bodily needs through behavior execution. In
our implementation, each motivation performs homeostatic control of one physiological
variable. We have used the equation proposed in [2] to combine cue and
physiological deficit when computing motivational intensities:
mi = di + (di × αci) (1)
In addition to physiological deficits (di) and incentive cues (ci), this equation
introduces a weighting factor (0 ≤ α ≤ 1) that affects the relevance given to the
external cue.
Behaviors are coarse-grained subsystems (embedding simpler actions) that
implement behavioral competencies similar to [12,5]. Following a classical distinction
in ethology [15], motivated behavior can be consummatory—“goalachieving”
and needing the presence of an incentive stimulus to be executed—or
appetitive—“goal-directed” search for a particular incentive stimulus. In addition
to modifying the external environment, the execution of a behavior has an
impact on (increases or decreases) the level of specific physiological variables.
Therefore, behaviors take part in the homeostatic control to maintain the state
of the physiological variables within the viability zone.
Behavior Selection is performed in a continuous loop consisting of three
main steps: (1) The deficit of the physiological variables (internal needs) and
the intensity of the external stimuli are calculated; (2) motivational intensities
are computed combining (perception of) deficits and external stimuli ponderated
by the weight α, following equation 1; (3) the behavior that (best) satisfies the
motivation with the highest intensity is executed, modifying the physiology and
possibly the position of the robot relative to external stimuli in the environment.
3 Competition for Resources
In previous work [2] we analyzed different motivation-based behavior selection
architectures within a static Two-Resource Problem (TRP), in which a single
robot must maintain appropriate levels of two internal variables by consuming
two resources available in the external environment. The TRP constitutes the
minimal scenario to test behavior selection mechanisms, and it has become a
standard testbed for behavior selection both in animals—see e.g., [17]—and autonomous
agents and robots—e.g., [4,7,3]. Its simplicity, although not devoid of
problems, favors a systematic analysis of results. The particular implementation
of the TRP in [2] used a Lego Mindstorms robot (see Figure 1, left, for a similar
404 L. Ca˜namero and O. Avila-Garc´ıa
arena, although the TRP uses only one robot), with the need to maintain temperature
and energy levels by consuming heat (white gradients on the floor of the
arena) and food (black gradients), respectively. The robot had two motivations:
mcold to increase temperature, which can be satisfied by executing the consummatory
behavior bwarmup, and mfatigue to increase energy, which can be achieved
by executing the consummatory behavior bfeed. In addition, the robot had a reflex
obstacle avoidance behavior bavoid, and the appetitive behavior bsearch. The
execution of all behaviors affects both essential variables1.
To measure results in TRP, we used different performance indicators based
on the notion of viability, in particular: Life Span, defined as the time that the
robot survived in each run (LS = tlife/trun); Overall Comfort, the average level
of satisfaction of the physiological variables during a run (OvC = tlife
i=1 (1 −
di)/tlife); and Physiological Balance, the homogeneity with which physiological
needs are satisfied during a run (PhB = tlife
i=1 (1−σ2(di))/tlife). We also noted
that, when doing behavior selection in TRP, the robot executed regular cycles of
activities rather than isolated behaviors, and those activity cycles were reflected
in the physiological space of the robot, as shown in Figure 2: from the initial
state, the robot would start looking for a given resource, e.g. heat (arrow noted
as A in the figure), then consume it until satiated (B), then start looking for the
other resource (C), consume it until satiated (D), and start all over again. The
position of the cycles in the physiological can be changed: the same cycle (i.e.,
with the same shape and duration of each activity) would be executed closer
to the ideal state, therefore preserving viability “better”, or father away from it
(and therefore in a “less viable” way) depending on the value of α, the parameter
that weighed the significance of external stimuli in equation 1, as depicted on
Figure 2 (right). The regular shape of those activity cycles reflects the fact that
behavior selection in TRP was static and highly predictable.
The Competitive Two-Resource Problem2 (CTRP) is an extension of this
problem that consists in the introduction of two robots in the same environment
simultaneously performing their own TRP, as depicted in Figure 1 (left). The
robots do not explicitly communicate or compete; however, the fact that they
have to use the same resources to satisfy their needs introduces competition
for those resources, as both robots might need access to the same resource at
the same time. Therefore, new forms of environmental complexity—availability
and accessibility of resources—appear due to the interaction between robots,
breaking the predictability and symmetry of TRP. The question that needs to
be examined here is to what extent the architecture used for the TRP can solve
the CTRP.
1 At each execution cycle, bwarmup increases temperature by 0.3 units while decreasing
energy by 0.1 units, bfeed increases energy by 0.3 units while decreasing temperature
by 0.1 units, and bavoid and bsearch decrease each variable by 0.2 units.
2 We refer the reader to [3] for an in-depth technical quantitative analysis of this
scenario, while here we focus on a qualitative discussion of the adaptive value of
hormonal modulation and its significance from the point of view of emotion.
A Bottom-Up Investigation of Emotional Modulation 405
Fig. 2. Activity cycles in TRP. Left: cycle as reflected in the physiological space. Right:
position of cycles in the physiological space as a function of α.
Fig. 3. Decreased performance in the CTRP, as measured by viability indicators (left)
and activity cycles (right)
Analysis showed that the new forms of complexity dramatically decrease the
performance of that behavior selection architecture, as clearly reflected by the
different viability indicators and the activity cycles depicted in Figure 3. In
particular, analysis of the activity cycles shows that the cycles easily loose
the regularity and symmetry they showed in TRP, as illustrated in Figure 3
(right), and that the robot very often dies from two problems that the behavior
selection mechanism used within the TRP presents when used in the CTRP.
First, the robot can fall in a pathological sequence of opportunistic activities—
consuming the same resource—that eventually can drive it to death due to overopportunism.
Second, when one robot is located on top of a resource—i.e., consuming
it—the other robot might bump into it and push it out of the resource.
This will result in the interruption of the ongoing consummatory activity and
to death due to goal interference.
The next step in our incremental design approach is to analyze what needs
to be added to the architecture to be able to solve those problems. A solution
to the “over-opportunism” problem requires shifting attention away from less
406 L. Ca˜namero and O. Avila-Garc´ıa
needed resources when the robot is in a high risk of death (RoD), that we define
as the inverse of the distance between physiological state (dtemper, denergy) and
lethal boundaries. A solution to the “goal interference” problem requires that
the robot in need of an occupied resource does not avoid the “intruder” as if
it were a mere obstacle. Both problems can be solved by altered perception of
external stimuli, i.e., by modulation of exteroception.
3.1 Modulation of Exteroception
Rather than adding more structural elements to our architecture, our solution
consists in trying to achieve additional functionality from the same architecture.
A single “hormone-like” modulatory mechanism can alter perception in both
cases, with a twofold effect. First, by acting on the parameter α of equation 1—
i.e., by biasing the relevance given to external cues—the hormone reduces the
perception of both incentive cues, therefore reducing opportunistic activities
when there is any risk of death. Second, by cancelling the perception of obstacles
sobstacle (carried out using the bumper sensor), and hence the avoidance reflex
behavior, when the robot is facing the competitor, the hormone potentiates the
competition skills of the robot by enhancing its capacity to push the other robot
out the resources and not to be interrupted. To achieve this twofold functionality,
the concentration of hormone will be a function of the risk of death (RoD)
and the perception of the competitor, given by 0 ≤ scompetitor ≤ 1. Hormone
concentration is computed as:
cg = RoD + scompetitor (2)
The relation between hormone concentration and the cancellation of the perception
of incentive cues and obstacles is as follows. To achieve the first functionality,
the cancellation of α is directly proportional to the increment in hormone
concentration, i.e., when RoD increases, α decreases: α = min(1 − cg, 0) The
second functionality is obtained by cancelling the perception of sobstacle—i.e.,
bumpers—when the competitor is in front of the robot. For this mechanism to
be efficient, two conditions must be fulfilled to make a coherent pushing of the
other robot. First, the robot must avoid getting engaged in fights when it has
high RoD. Second, it must only bump blindly into the other robot, not against
the walls of the arena. To produce that effect the cancellation of the bumpers
must be at hormonal levels cg  1 and cg  2.
It is worth noting that the motivation-based behavior selection architecture
has suffered no modification; the only difference with respect to the TRP is the
fact that now one of its parameters (α, cfr. equation 1) is modulated by the
hormonal feedback mechanism.
3.2 Experiments and Results
We tested the robots for a total of 16 runs of 1200 steps (approximately 5 minutes)
each, one step representing a loop of the behavior selection mechanism
that takes 260ms in the 16MHz onboard microcontroller. As shown in Figure 4,
A Bottom-Up Investigation of Emotional Modulation 407
Fig. 4. Activity cycles in CTRP in unmodulated architecture (left) and with modulation
of exteroception (right)
the robot with hormone-like mechanism recovers the stability and viability of
activity cycles. We refer the reader to [3] for a detailed quantitative analysis,
while we focus here on various interesting functionalities that emerged as a result
of modulating the exteroception of the robot. The first functionality is to
stop consuming resources when the robot detects its competitor approaching.
This could be interpreted by an external observer as abandonment of a situation
(waiting for the other robot at the resource) in which competing is disadvantageous.
Instead, the robot will leave the resource and go straightforward towards
the competitor until it reaches it; at that moment, two things can happen. If
there is some level of RoD, the bumpers of the robot will not be cancelled and it
will avoid the competitor, showing a behavior that an observer could interpret
as “fear” after evaluating the competitor. On the contrary, if there is no RoD,
the hormonal system will cancel the bumpers and the robot will push the competitor
unconditionally—as if it showed some sort of “aggression” against it. If
we study the whole picture as external observers, such behavioral phenomena
could well be interpreted as some sort of “protection of resources”.
4 Prey-Predator Scenario
The previous scenario involved no active interaction between the two robots. It
therefore seems natural to ask whether an active relation between the robots
would introduce additional complexity, and how the previous behavior selection
architecture would cope with it. We thus designed a prey-predator scenario
(Figure 1, right) that we call the Hazardous 3-Resource Problem (H3RP). In
H3RP, a “predator” robot actively chases and can damage a “prey” robot by
hitting a home-made contact sensor in the form of a ring. To make this interaction
possible, we had to introduce new elements in the environment—a nest in
one of the corners of the arena, in which the prey can “hide” and recover from
damage—and in the architecture of the prey, namely: (a) a third physiological
variable, integrity, which is a metaphor of the essential need any organism has
408 L. Ca˜namero and O. Avila-Garc´ıa
to keep its tissue—the boundary between the organism and its environment—
intact and that is unpredictably reduced by the attacks of the predator; (b) a
new motivation mdamage to decrease the integrity deficit; and (c) an appropriate
consummatory behavior brecover to satisfy the new need.
Initial experiments showed very quickly that a purely motivation-based behavior
selection mechanism does not perform well within the new framework,
since the prey invariably died as a consequence of predator attack (see the right
graph of Figure 6 for quantitative results of additional experiments). The main
cause seemed to be the inability of the prey to react timely to the attack of
the predator, which was perceived in close proximity only. In other words, the
behavior selection mechanism paid low attention to the new motivation to recover
integrity, even when the predator is in sight. The probability to lose integrity
rises when the predator is around, therefore it would be advantageous for the
prey robot to “anticipate” that loss and start “preparing in advance” to recover
integrity.
In the animal world, exposure to predators triggers what has been termed
“predator-induced stress” or “predator-stress” for short, characterized by high
levels of corticoids or “stress hormones” and a number of responses related to
increased attention to and avoidance of the predator. Such reactions occur not
only in the presence of a predator. Prey animals use unconditioned and conditioned
predator cues to assess risk of predation, and they even seem to be
able to perceive risks in the absence of such cues [6]. An example of the latter
is the phenomenon known as “risk of permanence”—maintained levels of vigilance
after predator’s disappearance. Risk of predation strongly influences prey
decision-making (for example, when and where to feed, vigilance, or the use of
nest), which in this circumstances can be considered as a mechanism to allow
an animal to manage predator-induced stress [11]. Risk of predation has been
proposed to increase the animal’s level of “apprehension,” i.e. the reduction in
attention to other activities (e.g. foraging) as a result of increasing the time
spent executing defense-related activities such as vigilance or refuge use [8].
4.1 Modulation of Interoception
We have again applied “hormonal” modulation to our behavior selection architecture
to achieve such “anticipatory” behavior, this time exploiting the temporal
dynamics of hormonal decay to produce long-term modulatory effects triggered
by short-term exposure to a stimulus [10].
To achieve this, a simple solution consists in using one of the existing sensors
of the prey robot to detect the predator from a distance. Given the morphology
of the robot, this sensor must be the same as that used to locate the nest.
The problem of using that sensor is that it is fixed, pointing forwards. Since the
predator does not pass in front of the prey very often and only does it for very
brief periods, the additional stimulus (spredator) will be too weak to make any
difference. However, long-term hormonal modulation acts as a mechanism for
predation risk assessment in the absence of predator cues. Hormone concerntration
makes the system more sensitive to integrity deficit after the detection of
A Bottom-Up Investigation of Emotional Modulation 409
the predator. Hormonal secretion follows the detection of the stimulus spredator
and increases the perceived integrity deficit. Due to the hormone’s temporal dynamics,
modulation will be acting in the system long time after the predator has
disappeared. Hormone concentration modifies again one of the sensory inputs of
the architecture—interoceptive in this case—biasing behavior selection.
We have modeled hormonal temporal dynamics—release and dissipation—
using an artificial endocrine system similar to that proposed in [16] and described
by equations 3 and 4. A gland g releases hormone as a function of the intensity
of the external stimulus predator (spredator) at a constant releasing rate βg:
rg = βg · spredator (3)
Hormone concentration3 suffers two opposite forces over time: it increases
with the release of hormone by the gland, and dissipates or decays over time at
a constant rate γg:
c(t + 1)g = max[(c(t)g · γg) + rg, 100] (4)
In this implementation, the hormone increases the perception of the integrity
deficit (dintegrity ), i.e., the higher the hormone concentration, the higher the
reading of the dintegrity interoceptor:
dnew
integrity = max(dintegrity + δg · cg, 1) (5)
Factor δg determines how susceptible to hormonal modulation the interoceptor
(dintegrity) is. We use δg = 0.005, which implies that the level of perceived
dintegrity is increased by 0.5 when hormonal concentration is maximum
(cg = 100). In other words, although the level of integrity is at its ideal value
(dintegrity = 0), the interoceptor will perceive a level of 0.5 if hormone concentration
is maximum. Note that there is a constraint to avoid the level of integrity
deficit to be perceived beyond the maximum possible value (dintegrity = 1).
4.2 Experiments and Results
We tested the robot for 16 runs of 1600 steps each, i.e., each architecture (nonmodulated
and modulated) was tested for almost two hours in H3RP.
The prey robot presented higher viability levels in terms of life span, at the
cost of overall comfort, when equipped with the modulatory mechanism, as
shown in Figure 5. Long-term hormonal modulation acts as a mechanism for
predation risk assessment in the absence of predator cues. It can be regarded as
increasing the level of “apprehension” of the prey robot after short-term predator
exposure, and this is reflected in an increment of the motivation to recover and of
the execution time of recover-related (consummatory and appetitive) activities—
the robot spends more time looking for the nest and recovering integrity in it—at
the cost of other activities, namely feed and warmup, as reflected in Figure 6
3 We constrained hormonal concentration to a maximum of cg = 100 in order to keep
more control on the hormone’s dynamics and thus facilitate the analysis of results.
410 L. Ca˜namero and O. Avila-Garc´ıa
Life Span
0.0
0.2
0.4
0.6
0.8
1.0
Overall Comfort
0.30
0.35
0.40
0.45
0.50
0.55
0.60
Physiological Balance
0.50
0.55
0.60
0.65
0.70
Fig. 5. Average performance of non-modulated (light bars) and modulated (dark bars)
architecutres in terms of LifeSpan, Physiological Balance, and Overall Comfort. Bars
show standard error of the mean.
Damage Subsystem Execution
10%
15%
20%
25%
30%
% steps
Consuming Interruption
0%
10%
20%
30%
40%
50%
% consummatory phases
Cause of Death
16
10
0
2
4
6
8
10
12
14
16
run
Integrity
Energy
Temper
Survive
Fig. 6. Comparison between non-modulated (bars on the left of each graph) and modulated
(bars on the right) architectures in terms of execution time of recover subsystem
(left), average number of interruptions of consummatory feed and warmup behaviors
(center), and causes of death in the 16 runs (right). Bars show standard mean error.
(left). This increment in the execution time of recover-related activities is statistically
highly significant. Another important phenomenon is the interruption of
ongoing consummatory feeding or warming-up activities (Figure 6, center). When
the robot is under the effect of the hormone it will abandon the resource and
go to the nest before the motivation has been satiated. The prey robot, when
equipped with the hormonal mechanism, presents statistically higher levels of
interruption of ongoing feeding or warming-up activities. Finally, analysis of the
causes of death (Figure 6, right) shows substantial differences with respect to
the non-modulated architecture.
5 Conclusion
We have discussed a bottom-up study of plausible mechanisms underlying emotional
modulation of behavior selection and their adaptive value, in particular
A Bottom-Up Investigation of Emotional Modulation 411
Motors
External Stimuli
Physiological Deficits
Motivations
Behaviors
ch
cf
so
dt de
bs
bw
ba
bf
mc
mf
A
S
M
g
Hormone
sc
Physiological Deficits
Motors
External Stimuli Behaviors
ch
cf
cn
so
dt de di
bs
bw
bf
ba
br
mc
mf
md A
S
M
sp
g
Hormone
Motivations
Fig. 7. Hormonal modulation of exteroception (left) and of interoception (right)
how such modulation applied to a motivation-based architecture can achieve different
functionalities found in biological emotions, to face different emotionallyrelevant
problems posed by different competitive scenarios.We have considered a
first scenario in which obtaining resources in competition with others is the main
survival-related problem, and a second scenario in which the attack of a predator
constitutes the main threat. Drawing on the notion of biological hormones, we
have focused on achieving adaptive behavior selection in these different competitive
robotic scenarios by modulating perception of external stimuli in the first
case, and of internal stimuli in the second, as depicted in Figure 7. In addition
to improving behavior selection performance and adaptation, modulation has
given rise to some emergent behavioral phenomena that could be interpreted
by an external observer as “emotional,” such as aggressive/defensive behavior
in the first, “fleeing” and “apprehension” in the second. We suggest that such
modulatory mechanisms provide a more principled integration of different behavior
selection elements and functions, in addition to improving the adaptation
of a robot to changing environments. The type of adaptation fostered by such
mechanisms is different from other mechanism such as learning or evolution, for
which “past solutions” are “overwritten” by new ones.
Current and future work includes the integration in the same architecture of
both types of hormonal modulation presented here, to face a more complex preypredator
problem requiring interactions among both mechanisms. We will also
continue our incremental study of plausible modulatory mechanisms underlying
emotions by changing and complexifying the environment to give rise to other
behavior selection problems.
Acknowledgments
Support was provided partly by the European project HUMAINE (FP6-IST–
507422) and partly by a University of Hertfordshire studentship to Orlando
Avila-Garc´ıa.
412 L. Ca˜namero and O. Avila-Garc´ıa
References
1. Ashby, W.R.: Design for a Brain. Chapman & Hall, London (1952)
2. Avila-Garc´ıa, O., Ca˜namero, L., te Boekhorst, R.: Analyzing the Performance of
“Winner-Take-All” and “Voting-Based” Action Selection Policies within the Two-
Resource Problem. In: Banzhaf, W., Ziegler, J., Christaller, T., Dittrich, P., Kim,
J.T. (eds.) ECAL 2003. LNCS (LNAI), vol. 2801, pp. 733–742. Springer, Heidelberg
(2003)
3. Avila-Garc´ıa, O., Ca˜namero, L.: Using Hormonal Feedback to Modulate Action Selection
in a Competitive Scenario. In: Proc. Eight Intl. Conf. Simulation of Adaptive
Behavior (SAB04), pp. 243–252. MIT Press, Cambridge, MA (2004)
4. Blumberg, B.: Old Tricks, New Dogs: Ethology and Interactive Creatures. Unpublished
PhD Thesis. MIT Media Laboratory, Cambridge, MA (1997)
5. Ca˜namero, L.D.: Modeling Motivations and Emotions as a Basis for Intelligent
Behavior. In: Johnson, W.L. (ed.) Proc. First Intl. Conf. on Autonomous Agents,
pp. 148–155. ACM Press, New York (1997)
6. Curio, E.: Proximate and Developmental Aspects of Antipredator Behavior. Adv.
Study Behav. 22, 135–238 (1993)
7. Girard, B., Cuzin, V., Guillot, A., Gurney, K.N., Prescott, T.J.: Comparing a
Brain-Inspired Robot Action Selection Mechanism with “Winner-Takes-All”. In:
Proc. Seventh Intl. Conf.on Simulation of Adaptive Behavior, MIT Press, Cambridge
(2002)
8. Kavaliers, M., Choleris, E.: Antipredator Responses and Defensive Behavior: Ecological
and Ethological Approaches for the Neurosciences. Neuroscience and Biobehavioral
Reviews 25, 577–586 (2001)
9. Kravitz, E.A.: Hormonal Control of Behavior: Amines and the Biasing of Behavioral
Output in Lobsters. Science 241, 1175–1781 (1988)
10. Levitan, I.B., Kaczmarek, L.K.: The Neuron: Cell and Molecular Biology, 3rd edn.
Oxford University Press, Oxford, UK (2002)
11. Lima, S.L.: Stress and Decision Making under the Risk of Predation: Recent Developments
from Behavioral, Reproductive, and Ecological Perspectives. Adv. Study
Behav. 27, 215–290 (1998)
12. Maes, P.: A Bottom-Up Mechanism for Behavior Selection in an Artificial Creature.
In: Proc. First Intl.Conf. on Simulation of Adaptive Behavior (SAB90), pp. 238–
246. MIT Press, Cambridge, MA (1991)
13. Meyer, J.-A.: The Animat Approach to Cognitive Science. In: Roitblat, H.L.,
Meyer, J.-A. (eds.) Comparative Approaches to Cognitive Science, pp. 27–44. MIT
Press, Cambridge, MA (1995)
14. McFarland, D. (ed.) Motivational Control Systems Analysis. Academic Press, London
(1974)
15. McFarland, D.: Animal Behaviour, 3rd edn. Addison Wesley Longman Ltd., Redwood
City, CA, USA (1999)
16. Neal, M., Timmis, J.: Timidity: A Useful Emotional Mechanism for Robot Control?
Informatica 27, 197–204 (2003)
17. Spier, E., McFarland, D.: Possibly Optimal Decision Making under Self-Sufficiency
and Autonomy. J. Theor.Biol. 189, 317–331 (1997)
18. Toates, F.: Motivational Systems. Cambridge Univ. Press, Cambridge (1986)