ArticlePDF Available

Goal-Seeking Behavior in a Connectionist Model

Authors:

Abstract and Figures

Goal-seeking behavior in a connectionist modelis demonstrated using the examples of foragingby a simulated ant and cooperativenest-building by a pair of simulated birds. Themodel, a control neural network, translatesneeds into responses. The purpose of this workis to produce lifelike behavior with agoal-seeking artificial neural network. Theforaging ant example illustrates theintermediation of neurons to guide the ant to agoal in a semi-predictable environment. In thenest-building example, both birds, executinggender-specific networks, exhibit socialnesting and feeding behavior directed towardmultiple goals.
Content may be subject to copyright.
Goal-Seeking Behavior in a Connectionist Model
Thomas E. Portegys, Lucent Technologies, portegys@lucent.com
Abstract
Goal-seeking behavior in a connectionist model is demonstrated using the examples of foraging
by a simulated ant and cooperative nest-building by a pair of simulated birds. The model, a con-
trol neural network, translates needs into responses. The purpose of this work is to produce life-
like behavior with a goal-seeking artificial neural network. The foraging ant example illustrates
the intermediation of neurons to guide the ant to a goal in a semi-predictable environment. In the
nest-building example, both birds, executing gender-specific networks, exhibit social nesting and
feeding behavior directed toward multiple goals.
1 Introduction
1.1 Purpose
The purpose of this work is to produce lifelike behavior with a goal-seeking artificial neural net-
work. This is demonstrated using the examples of foraging by a simulated ant and cooperative
nest-building by a pair of simulated birds. The foraging ant example illustrates the intermediation
of neurons to guide the ant to a goal in a semi-predictable environment. In the nest-building exam-
ple, both birds, executing gender-specific networks, exhibit social nesting and feeding behavior
directed toward multiple goals.
1.2 Background
The natural world poses many challenges to animals for survival and reproduction which have
fostered the evolution of capabilities beyond those of current machines. For example, computers
may excel at monitoring the global stock market, but have difficulty harvesting fruit in an orchard.
The human animal has also evolved intelligence, and according to James Albus (1979), plausibly
from prior neural mechanisms:
The rarity and late arrival of the ability to plan suggests that a highly developed precursor,
or substrate was required from which planning capabilities evolved... The implication is
that a sensory-interactive, goal-directed motor system is not simply an appendage to the
intellect, but is rather the substrate in which intelligence evolved.
As a corollary, in his AAAI-98 Presidential Address, David Waltz (1999) offers this conjecture:
...What is intelligence for? That one I think we have some chance of answering. And the
answer to it is that any kind of intelligent phenomenon that we see, whether physical or
behavioral, is there because it really serves the organism’s survivability purposes.
Reductionistic and somewhat behavioristic (Skinner 1957) views such as these suggest that intel-
ligence emerges from basic goal-seeking brain mechanisms.
The goal of artificial intelligence is to create synthetic brains, and modeling neurological sys-
tems is a common course of action. But in what manner should machines model brains? At the
fine-grained level of neurons? By simulating brain structures? This project takes the abstract, nor-
mative approach that “classic” artificial neural networks do. Steven Hampson (1990) offers this
rationale for such abstractions:
Connectionism, like artificial intelligence (AI), has no necessary commitment to biologi-
cal relevance, but it is generally assumed that a better understanding of biological intelli-
gence has something to offer to the study of artificial intelligence.
While artificial models may not be of direct relevance to biology (Dudai 1989), there is no conclu-
sive evidence that the brain is either the only or the even best possible intelligent mechanism. To
use an aeronautical analogy, the study of bird flight was useful, but not constraining, in the devel-
opment of the airplane.
The model selected for this project, Mona, was introduced by Portegys (1999) as a connec-
tionist model of motivation. A connectionist model bears a functional resemblance to a natural
neurological system in that many local interactions between units produce an emergent system
effect. A type of control neural network (Fu 1994), Mona produces goal-seeking responses based
on environmental input aimed at reducing homeostatic needs (Parten 1990). In living organisms
these are inherent needs such as thirst, hunger, sex, etc. In the context of Mona, motivation, often
a loosely defined term (Pfaff, 1982; McClelland 1987), is interpreted as a function which trans-
lates needs into responses.
2 The Mona Model
2. 1 Neurological Abstraction
Human and animal intelligence is performed by a network of neurons which operate by mutual
excitation and inhibition (Thompson, Berger, and Berry, 1980; Holmes and Rall, 1992). Mona is
an abstraction of a neurological system consisting of a network of computational units, each of
which is capable of receiving and expressing mutually mediating influences. Consider a func-
tional view of the nervous system of a simple organism which controls feeding behavior, shown in
Figure 1. Feeding consists of the sequence of catching, killing and eating prey.
“Feed”, “Catch”, “Kill”, and “Eat” are neurons which fire when their namesake events occur. The
solid arrows are enabling (or disabling) signals directed from one neuron to another; these signals
are analogous to the excitatory and inhibitory influences of living neurons. The dotted arrows are
signals derived from the organism’s need for food. The above is read as follows. When the organ-
ism becomes hungry, the goal associated with the need of hunger-reduction, “Eat”, becomes a
source of need signals propagating to antecedent neurons in a kind of “bucket brigade” from pri-
mary to secondary goals, causing them to become capable of responding. Once the prey is caught,
the killing neuron is enabled, and once that is done, the eating neuron is enabled. The enabling of
a neuron means that a context has been established in which it may successfully fire. A neuron’s
state thus consists of the 3-tuple {need, enablement, firing}. The events with which neurons are
associated can be drawn from sensors, responses, or, as in the case of the mediating “Feed” neu-
ron, the states of component neurons.
Neurons maintain a base level of enablement, analogous to long term memory, which is
dynamically modified, as part of short term memory, to accomplish a concerted operation. For
example, the “Feed” and “Catch” neurons might be enabled by default, while the “Kill” and “Eat”
neurons rest in a disabled state awaiting the catching of prey. This would prevent an attempt to kill
an object being caught by the organism for the purpose of mating or building a nest.
The ability of neurons to disable other neurons allows further opportunity for context-depen-
dent cooperation. For example, it would be sensible for the “Catch” neuron to disable itself upon
firing to prevent the seizing of prey while a catch is being eaten.
Figure 1 - Feeding control
Feed
Kill
Catch Eat
enable enable
need need
mediate
2.2 Description
Mona has a simple interface with the environment, shown in Figure 2, similar to that utilized by
Portegys (1986) and Nolfi and Parisi (1996). All knowledge of the state of the environment is
absorbed through “senses”; there are no special modalities or channels by which instructions or
meta-information are given. Responses are expressed to the environment with the goal of eliciting
sensory inputs which are internally associated with the reduction of needs.
The events which neurons represent can be drawn from sensors, responses, or the states of compo-
nent neurons, calling for three types of neurons. Neurons attuned to sensors are “receptors”, those
associated with responses are “motors”, and those mediating other neurons are “mediators”. A
receptor neuron can process a logical combination of multiple sensor data. A mediator neuron
controls the transmission of need and enablement through the sequence of its component neurons.
To elucidate by example, consider this task: Mona must get into her home from somewhere
out in the world, a locked door barring the way inside, thus necessitating the use of a key to
unlock the door. She needs to know several things, such as how to get to the door, how to unlock
the door, and how to enter her home through the unlocked door. Mona must produce a sequence of
responses to proceed from an initial keyless condition in the world to her home.
Environment Mona
responses
sensory data
Figure 2 - The Mona/environment interface
Figure 3 depicts the portion of Mona’s neural network which manages the entering of home
through an unlocked door. Let the house-shaped objects be receptor neurons, such as the one
marked “Door”; the inverted houses be motor neurons, such as “Move”; and the diamonds be
mediator neurons, such as “Enter home”. The numbers in parentheses indicate need levels, which
will be discussed presently; suffice it to say for now that the “Home” receptor has been associated
with the reduction of a need, and is thus a goal for Mona. The numbered arrows proceeding from
a mediator indicate a sequence of neurons mediated by it, known to the mediator as its “events”.
In this case, “Enter home” mediates a sequence of events associated with the receptor “Door”, the
motor “Move”, and the receptor “Home”. This mediator thus governs the process of entering
home by moving through a door. The type of mediation exerted by “Enter home” is an enabling
one, meaning that it allows firing events to propagate enabling influences. Although not depicted
in this example, a disabling mediator has dotted arrows instead of solid.
Initially the door is locked, thus the “Enter home” mediator is disabled, meaning that it cannot
function until preconditions establish an enabling context for it. This is represented by the dotted
outline of the mediator. In order to enable “Enter home”, another mediator must come into play:
“Enable enter home”. This mediator will enable the “Enter home” neuron when the “Unlock
door” neuron fires.
Enable
enter
home
(0)
Unlock
door
(0.97)
0
Enter
home
(0)
1
Door
(0)
0
Move
(0)
1
Home
(1)
2
Mediator
Disabled
mediator
Receptor
Motor Goal
(need=1)
Figure 3 - Enable enter home/Enter home
However, the “Unlock door” neuron is also in a disabled state, requiring “Get key”, shown in Fig-
ure 4, to fire as a precondition - the door cannot be unlocked without the key.
Figure 4 - Enable unlock door/Unlock door
Door
(0)
0 2
Use
key
(0)
1
Enable
unlock
door
(0)
Get
key
(0)
0
Unlock
door
(0.97)
1
The final two pieces are supplied in Figure 5: how to get a key (“Get key”),and how to get to the
door from the world (“Go to door”).
Figure 5 - Get key/Go to door
Get
key
(0.95)
0 1 2
Go
to
door
(0)
World
(0)
0
Move
(0)
1
Door
(0)
2
No
key
(0)
Take
key
(0.94) Key
(0)
Since these diagrams show the initial state of network, the “World” and “No key” receptors are
firing, denoted by the double outlines on their graphical symbols. Figure 6 shows the entire net-
work, with each neuron type segregated into its own “cortex”.
Figure 6- Getting home network
Receptors
Motors
Mediators
Key
(0)
No
key
(0)
Home
(1)
Door
(0)
World
(0)
Use
key
(0)
Take
key
(0.94)
Move
(0)
Enable
unlock
door
(0)
Unlock
door
(0.97)
1
Get
key
(0.95)
0
Enable
enter
home
(0)
Enter
home
(0)
10
2 01 01
Go
to
door
(0)
20
1
2 01
Neurons use a simple firing threshold function. Receptor and motor neurons fire when their asso-
ciated sensory/response events occur. A mediator neuron contains an eventFiring() function,
shown in Figure 7, which fires the mediator when each event in its sequence fires within the max-
imum delay imposed by the mediator’s maxEventDelay value.
// Mediator event firing.
Mediator::eventFiring(eventNumber)
{
// Discard unexpected events.
if (eventNumber != expectedEvent) return;
if (eventNumber == finalEvent)
{
// Final event firing: fire mediator.
firing = TRUE;
// Fire event in this neuron’s mediators.
for (all superMediator)
{
mediator = superMediator->mediator;
event = superMediator->eventNumber;
mediator->eventFiring(event);
}
// Reset event counter.
expectedEvent = 0;
} else { // More events expected.
// Expect next event.
expectedEvent++;
eventTimer = maxEventDelay;
// If enabled, propagate enablement to expected event.
if (enabled == TRUE)
{
neuron = components[expectedEvent];
neuron->enabler += eventEnabler;
}
}
}
Figure 7 - eventFiring() function
When an event fires, if the mediator is enabled, a mediator-specific eventEnabler value is accu-
mulated in the next event in the sequence, as shown in Figure 8.
Enabler propagation occurs when neurons fire; enablement represents a more persistent quantity.
In the example, getting the key, a transient event, enables the ability to unlock the door, a persis-
tent state. In addition, a neuron may be enabled by multiple mediators, each exerting enabling
influences. These semantically different variables can be related in the following way. Let an
enabler value be represented as a real number, positive denoting an enabling influence, negative a
disabling one. Let enablement be boolean valued. Enablement can then be expressed by a hyster-
esis function of the sum of its enablers, as shown in Figure 9. A possible drawback to this function
is that it may be less amenable to optimization techniques, e.g., an error-reduction algorithm.
Enabled
mediator
01
Firing
Enabler
propagated
Figure 8 - Enabler propagation
False
True
Enablement
Enabler 1.0
0
-1.0
Figure 9 - Enabler/enablement hysteresis
Mona’s raison d’être is need-reduction. For this purpose, some receptors are associated with the
reduction of needs and are thereby defined to be goals. For example, a warmth receptor would be
associated with a reduction of feeling cold.The drive() function, shown in Figure 10, propagates
need from goal sources to other neurons in the network, attenuating to preferably drive “closer”
neurons and to prevent endless propagation.
// Neuron drive.
Neuron::drive(need) {
{
// Save maximum propagated need.
if (need <= currentNeed) return;
currentNeed = need;
// Attenuate need.
if ((need -= ATTENUATION) <= 0) return;
// If this neuron is an enabled mediator,
// drive its expected event.
if (type == MEDIATOR && enabled == TRUE)
{
neuron = components[expectedEvent];
neuron->drive(need);
return;
} else {
// Drive this neuron’s mediators.
for (all superMediator)
{
mediator = superMediator->mediator;
event = superMediator->eventNumber;
mediator->eventDrive(event, need);
}
}
}
Figure 10 - drive() function
Need causes a mediator neuron to perform a check: if it is enabled, the eventDrive() function,
shown in Figure 11, will pass the need into its expected event neuron in order to motivate it to
occur, as shown in Figure 12.
// Mediator event drive.
Mediator::eventDrive(eventNumber, need)
{
// Attenuate need.
if ((need -= ATTENUATION) <= 0) return;
// If this mediator is disabled, drive its mediators.
if (enabled == FALSE)
{
for (all superMediator)
{
mediator = superMediator->mediator;
event = superMediator->eventNumber;
mediator->eventDrive(event, need);
}
} else {
// Drive expected event.
neuron = components[expectedEvent];
neuron->drive(need);
}
}
Figure 11 - eventDrive() function
Expected
event
mediator
Enabled Need
Figure 12 - Driving expected event
Otherwise, a disabled mediator (or a receptor or motor) drives its mediating neurons to cause
them to enable it, as shown in Figure 13.
Upon completion of drive, the need resident in motor neurons is translated into potentials of the
responses associated with those neurons. The system response is that associated with the maxi-
mum potential value:
The complete algorithms, written in C++, are available on the World Wide Web (see Conclusion).
Need
Receptor/
Mediators:
Motor/
Disabled
mediator
Figure 13 - Driving mediators
response motormax Σpropagated need()
()
Table 1 contains a trace of firing neurons obtained when the “Getting Home” example is run on a
software implementation:
Table 1: Firing neuron trace for “Getting home” network
***time=0***
Receptor firing: No key
Receptor firing: World
Motor firing: Take key
***time=1***
Receptor firing: Key
Mediator firing: Get key
Receptor firing: World
Motor firing: Move
***time=2***
Receptor firing: Key
Receptor firing: Door
Mediator firing: Go to door
Motor firing: Use key
***time=3***
Receptor firing: Key
Receptor firing: Door
Mediator firing: Unlock door
Mediator firing: Enable unlock door
Motor firing: Move
***time=4***
Receptor firing: Key
Receptor firing: Home
Mediator firing: Enter home
Mediator firing: Enable enter home
2.3 The Network vs. a State-space Model
The network must embody a model of the environment which, for comparison purposes, can be
sized against the well-known state-space model. In the network, the enabling and disabling opera-
tions allow the “topology” of the network to be modified by the act of operating it. The network
may be considered to be a hybridization of a logic engine and a state-space search engine having
two key properties: (1) more than one state (neuron) can be current (firing) at a particular time,
and (2) current states can “prove” (enable) or “disprove” (disable) the reachability of states. These
properties allow a network to assume a large number of states relative to the number of neurons
which comprise it.
As an illustration, consider the personal financial state-space shown in Figure 14, in which
money is earned at a job (pocket), spent at a store (broke), and deposited/withdrawn at a bank
(saved). For the sake of simplicity, let there be a single quantum of money in the economy, e.g., if
the money is in the bank, more money cannot be earned. The space contains (3 places) X (3
money situations) = (9 states).
broke
saved
pocket
broke
pocket
saved
broke pocket saved
job bank
store
Figure 14 - Financial state-space
Figure 15 shows a network representation of the problem in abbreviated graphical form, for clar-
ity. The ‘+’/’-’ indicate enabling/disabling influences. It can be seen that moving between places
(the top portion of the figure) is independent of the transactions which transpire at those places
since the enablement states (for earn, spend, deposit, and withdraw) store the transaction possibil-
ities. The addition of places not involved in monetary dealings, a reasonable real-world supposi-
tion, alters only the place transition portion of the network, avoiding the combinatorial expansion
of the state-space model. For example, if “home”, “park”, and “museum” are added as places, the
state-space expands by 9 states, while the network expands by 3 neurons.
saved
pocket broke
pocket
job
store
bank
job
store
bank
+
++
Figure 15 - Financial network
++
bank ++
-
+
+
+
+-
--
earn spend
deposit withdraw
3 Demonstration
3.1 The Foraging Ant
This problem demonstrates how the neural network can be used to simulate a foraging ant. Forag-
ing is a social enterprise among ants, which are known to follow trails of chemical markers left by
each other to efficiently gather food (Bonabeau and Théraulaz, 2000). In this problem, the behav-
ior of an individual ant performing one portion of the foraging process is simulated. The artificial
ant must follow a meandering trail of marks from its nest to a piece of food, which it must then
carry back to the nest. The problem illustrates the interplay of mediator neurons, using mutual
enablement and disablement, to guide the ant to a goal in a semi-predictable environment.
A sample trail is shown in Figure 16. The ant starts at its nest and follows the trail marks to the
cake. The trail is randomly generated in such a way that it never crosses itself. Generally a trail
leads straight on, so the most efficient strategy is to plunge ahead and orient upon leaving the trail.
An initial positive need is associated with the receptor which detects the presence of food at the
nest.
Sensory Capabilities.
Presence of mark at current location.
Presence of food at current location.
Presence of nest at current location.
Response Capabilities.
Move forward and backward, Grab and drop food, Orient itself in the direction of the trail.
Need.
Food to be present at nest.
Figure 16 - Meandering ant trail
Mediators:
Format:
<mediator>:<“enabled”|”disabled”>/<“enabling”|”disabling”>(<event sequence>)
Top-level control: “Forage” is to “Get food” then enable “Bring food” to bring it back to the nest:
“Forage”: enabled/enabling (“Get food”,”Bring food”)
“Get food”: enabled/enabling (“Grab food”,”Orient”,”Mark”)
“Grab food”: enabled/enabling (“Food”,”Grab”,”No food”)
“Bring food”: disabled/enabling (“Nest”,”Drop”,”Food at nest”)
After food obtained, reset “Bring food” to disabled state for next forage:
“Disable bring food”: enabled/disabling (“Food at nest”,”Bring food”)
“Travel trail” is the normal way to move. “To food” and “To nest” associate “Travel trail” with
getting somewhere:
“Travel trail”: enabled/enabling (“Mark”,”Forward”,”Mark”)
“To food”: enabled/enabling (“Travel trail”,”Food”)
“To nest”: enabled/enabling (“Travel trail”,”Nest”)
If the ant steps off the trail (“No mark”), these control getting it back on:
1. Disable normal traveling (“Travel trail”).
2. Initiate “Trail search” to back-up and orient.
3. Once oriented, enable normal traveling.
“Disable travel trail”: enabled/disabling (“No mark”,”Travel trail”)
“Trail search”: enabled/enabling (“No mark”,”Backward”,”Mark”,”Orient”,”Mark”)
“Enable travel trail”: enabled/enabling (“Trail search”,”Travel trail”)
Figure 17 shows the entire network:
Figure 17 - Foraging ant network
Motors
Mediators
Food
at
nest
(0)
No
food
(0)
Food
(0)
No
mark
(0)
Mark
(0)
Nest
(0) Drop
(0)
Grab
(0) Orient
(0)
Backward
(0) Forward
(0)
Enable
travel
trail
(0)
Trail
search
(0)
0
Travel
trail
(0)
1
02 31
Disable
travel
trail
(0)
0
1
To
nest
(0)
1
0
To
food
(0)
1
0
0 1
Forage
(0)
Get
food
(0)
0
Bring
food
(0)
1
2 1
Grab
food
(0)
0
Disable
bring
food
(0)
0
1
20 1
20 1
Receptors
The initial response of the ant is to move forward, driven by the need to fetch food to the nest, as
shown in Figure 18:
Figure 18 - Driving initial foraging response
Nest
(0)
Mark
(9.91)
Food
(9.95) No
food
(0)
Food
at
nest
(10)
Forward
(9.92)
Grab
(0)
Drop
(0)
Grab
food
(9.97)
021
Bring
food
(0)
0 2
1
Forage
(0)
0 1
Travel
trail
(9.93)
201
To
food
(0)
1
0
Response
Goal
A
B
C
D
E
F
GPath of need (need=10)
A: Goal “Food at Nest” to mediator “Bring food”.
B: “Bring food” disabled, so pass to “Forage”.
C: “Forage” to expected event “Grab food”.
D: “Grab food” to expected event “Food”.
E: “Food” to mediator “To food”.
F: “To food” to expected event “Travel trail”.
G: “Travel trail” to expected event “Forward”
(“Mark” (event 0) firing so is not expected).
Running the program on the sample trail results in the responses listed in Table 2 to fetch the food
to the nest:
Table 2: Foraging responses
Forward X 5
Backward
Orient
Forward X 2
Backward
Orient
Forward X 8
Backward
Orient
Forward X 6
Backward
Orient
Forward X 3
Grab
Orient
Forward X 4
Backward
Orient
Forward X 6
Backward
Orient
Forward X 8
Backward
Orient
Forward X 2
Backward
Orient
Forward X 4
Drop
Of particular interest is the interplay of neurons involved in orienting the ant after stepping off the
trail, shown in Table 3:
Table 3: Firing neuron trace for trail orientation
***time=n***
Receptor firing: Mark
Receptor firing: No food
Mediator firing: Travel trail
Motor firing: Forward
Ant on trail.
***time=n+1***
Receptor firing: No mark
Receptor firing: No food
Motor firing: Backward
Ant steps off trail:
“Disable travel trail” disables “Travel
trail”.
“Trail search” mediates moving backward.
***time=n+2***
Receptor firing: Mark
Receptor firing: No food
Motor firing: Orient
Ant now orients to trail:
“Trail search” mediates orient response.
***time=n+3***
Receptor firing: Mark
Receptor firing: No food
Mediator firing: Trail search
Motor firing: Forward
“Trail search fires”, allowing “Enable travel
trail” to re-enable “Travel trail”.
Ant moves forward in correct direction.
3.2 The Nesting Birds
In this task, a pair of nesting birds are simulated. The birds are the ground-nesting sort, and live in
a cellular world having three terrain regions: grassland, forest, and desert. Beginning together in
the grassland, the birds build their rectangular nest using stones found in the desert, and once con-
structed, the female lays an egg in it. They periodically require food in the form of mice which are
found in the forest.
Gender-specific Roles.
Male roles:
Find food and feed self when hungry.
Find and fetch stones for female when requested by her.
Find and fetch food for female when she is hungry.
Stay by female when not busy.
Female roles:
Signal need for food to mate when hungry and feed when he provides food.
Build nest: (1) request and accept stones from mate, and (2) place stones in rectangular con-
figuration.
Repair nest: (1) replacing missing stones, and (2) removing extraneous stones.
Lay egg in completed nest.
Sensory Capabilities.
Terrain at current location.
Object at current location.
Condition: wanting food, wanting stone, and object being held.
Condition of mate if co-located.
Response Capabilities.
Do nothing, Eat, Get (object), Go to desert, Go to forest, Go to mate, Lay egg, Look for
food, Look for stone, Put (object), Receive (object), Step, Toss (object), Turn, Want stone,
Do not want stone.
The “Go to” responses cause the bird to step toward the nearest cell with the indicated terrain or
mate. An innate ability to determine the correct direction is assumed. The “Look for” responses
likewise step the bird to the indicated object, but only if the bird is in the correct terrain. The
“Want stone” and “Do not want stone” responses change the condition of the bird. The “Toss”
response causes any held object to be discarded to a random nearby location.
Gender-specific Needs.
Needs are listed in order of high to low precedence.
Male needs:
1. Food:
Initiated periodically.
Satisfied by eating.
2. Female needs food:
Initiated by presence of female wanting food.
Satisfied by presence of female not wanting food.
3. Female needs stone:
Initiated by presence of female wanting stone.
Satisfied by presence of female not wanting stone.
4. Stay by female: a constant need.
Female needs:
1. Food:
Initiated periodically.
Satisfied by eating.
2. Egg in nest:
A constant need, prompting not only building of initial nest and laying of egg in it, but also
vigilance in the repairing nest and replacing a missing egg.
Figure 19 shows an abbreviated snapshot of the end of an animation of the nesting birds program.
The male is bringing a mouse back to the hungry female, who minds the nest and egg. The com-
plete animation can be seen in its entirety at http://www.corecomm.net/portegys/NestViewer.html
(JAVA applet).
Figure 19 - Nesting birds snapshot
egg
food for mate
forest
desert
male bringing food
Gender-specific Mediators.
Format:
<mediator>:<“enabled”|”disabled”>/<“enabling”|”disabling”>(<event sequence>)
Male mediators:
Food:
“Eat food”:enabled/enabling(”Got food”,”Eat”,”Not hungry”)
“Get food”:enabled/enabling(”Food on ground”,”Get”,”Got food”)
“Find food”:disabled/enabling(”Any terrain”,”Look for food”,”Food on ground”)
“Find forest”:enabled/enabling(”Any terrain”,”Go to forest”,”Forest”)
“Enable find food”:enabled/enabling(”Find forest”,”Find food”) # Once in forest, can find
food
“Disable find food”:enabled/disabling(”Find food”,”Find food”) # To force re-finding of
forest first
Stones:
“Get stone”disabled/enabling(”Stone on ground”,”Get”,”Got stone”)
“Find stone”:enabled/enabling(”Any terrain”,”Look for stone”,”Stone on ground”)
“Find desert”:enabled/enabling(”Any terrain”,”Go to desert”,”Desert”)
“Enable get stone”:enabled/enabling(”Find desert”,”Find stone”) # Once in desert, can
find stone
“Disable find stone”:enabled/disabling(”Find stone”,”Find stone”) # To force re-finding of
desert first
Coordinators:
“Get food disables get stone”:enabled/disabling(”Get food”,”Get stone”) # Can hold only
one object
“Get stone disables get food”:enabled/disabling(”Get stone”,”Get food”) # Can hold only
one object
“Toss object”:enabled/enabling(”Any terrain”,”Toss”,”Got no object”) # Have no object
after toss
“Toss enables get food”:enabled/enabling(”Toss object”,”Get food”) # Can hold food after
toss
“Toss enables get stone”:enabled/enabling(”Toss object”,”Get stone”) # Can hold stone
after toss
“Toss disables find stone”:enabled/disabling(”Toss object”,”Find stone”) # Must find
desert first
“Toss disables find food”:enabled/disabling(”Toss object”,”Find food”) # Must find forest
first
Mate:
“Feed female”:enabled/enabling(”Got food”,”Go to mate”,”Mate has food”,”Do noth-
ing”,”Mate does not want food”)
“Give stone to female”:enabled/enabling(”Got stone”,”Go to mate”,”Mate has stone”,”Do
nothing”,”Mate does not want stone”)
“Check on mate”:enabled/enabling(”Any terrain”,”Go to mate”,”Mate present”)
Female mediators:
Food:
“Eat food”:enabled/enabling(”Got food”,”Eat”,”Not hungry”)
“Receive food”:enabled/enabling(”Mate has food”,”Receive”,”Got food”)
Nest building:
“Egg in nest”:enabled/enabling(”Stone on ground”,”Step”,
”Stone on ground”,”Turn”,”Stone on ground”,”Step”,”Stone on ground”,”Step”,
”Stone on ground”,”Turn”,”Stone on ground”,”Step”,”Stone on ground”,”Step”,
”Stone on ground”,”Turn”,”Stone on ground”,”Step”,”Stone on ground”,”Step”,
”Stone on ground”,”Turn”,”Stone on ground”,”Step”,”Stone on ground”,”Turn”,
”Stone on ground”,”Step”,”Egg on ground”) # Sequence for constructed nest
“Build nest”:enabled/enabling(”Empty ground”,”Want stone”,
”Mate has stone”,”Receive”,”Have stone”,”Do not want stone”,”Have stone”,”Put”,
”Stone on ground”) # To request and place a missing stone
“Clear nest”:enabled/enabling(”Stone on ground”,”Get”,”Have stone”,”Toss”,”Ready to
lay egg”) # To remove extraneous stone
“Begin nest check”:enabled/enabling(”Any terrain”,”Step”,”Any terrain”,”Turn”,”Any ter-
rain”) # Must step out of nest to restart egg in nest sequence
Egg laying:
“Lay egg in nest”:enabled/enabling(”Ready to lay egg”,”Lay egg”,”Egg on ground”)
“Disable egg in nest”:enabled/disabling(”Egg in nest”,”Egg in nest”) # Forces bird out for
nest checking
“Enable egg in nest”:enabled/enabling(”Begin nest check”,”Egg in nest”) # Can re-lay egg
if necessary after checking nest
Graphical views of the entire networks are omitted for space reasons.
Analysis.
Several scenarios in the nesting process are cited to illustrate the workings of the neural networks.
1. Female requests food, male goes to forest, fetches food and gives to female to eat.
Context: Female needs food, raising want food condition. Male, co-located with female,
senses want food condition of female which raises need to feed her.
Responses: See Table 4.
Table 4: Food fetching responses
Male Female
Go to forest Want stone
For male, “Feed female” drives “Get food”
which drives “Find food” which drives “Find
forest”. Female also wants stone for nest, but
food takes precedence.
Go to forest Want stone
Look for food Want stone
“Find forest” fires, enabling “Find food”.
Get Want stone
Go to mate Want stone
Go to mate Receive
Do nothing Eat
“Mate does not want food” fires, lowering
need to feed female.
2. While building nest, female requests stone, male goes to desert, fetches stone and gives to
female, who places it in nest.
Context: The “Egg in nest” mediator drives the female to step around the perimeter of the
nest, observing stones as she goes. “Egg in nest” itself does not mediate the placement of
stones, but employs another mediator, “Build nest”, to do this by transferring need to it.
Responses: See Table 5.
3. Male becomes hungry while returning with stone, tosses stone and goes for food.
Context: The need to food overrides the need to fetch a stone. The “Toss enables get food”
mediator enables “Get food” by freeing the bird to pick up food. After eating, “Give stone
to female” reasserts itself, and the male proceeds to fetch another stone from the desert.
Responses: See animation.
Table 5: Stone fetching responses
Male Female
Go to mate Want stone
Male has obtained stone and is returning.
Go to mate Receive
Do nothing Do not want stone
Do nothing Put
“Build nest” fires, “Egg in nest” resumes.
Do nothing Step
Go to mate Want stone
“Build nest” repeated; male instilled with need
to fetch another stone.
Go to desert Want stone
4 Related Work
In contrast with Hopfield and backpropagating ANNs (Hopfield and Tank, 1986; Munakata 1998),
which are primarily stateless pattern classifiers, Mona employs a short term memory capability to
navigate the environment toward goals. Short term memory is implemented by the retention of
neural firing sequences and by the enabling and disabling operations, which modify the state of
the network based on sensory events and responses motivated by needs. The incorporation of state
memory into ANNs is a topic of continuing investigation (Roy 1997; Kodjabachian and Meyer,
1998)
A large body of cross-disciplinary work on the subject of animal and animat social behavior
exists (Goss and Deneubourg, 1992; Drogoul and Ferber, 1993; Anderson, Blackwell, and Can-
nings, 1997; Bonabeau, Dorigo, and Théraulaz, 1999; etc.). Reynolds’ (1987) much-publicized
simulation of bird flocking, Boids, showed how several simple rules applied by individual birds
combined to produce emergent group behavior. In the multiagent system developed by Murciano
and Millán (1996), agents cooperate and specialize to perform a collective gathering task by com-
municating relevant location data using light signals. Behavioral parameters were subject to post-
trial learning to improve performance. Mataric (1995) produced various emergent group behav-
iors, such as herding, by combining more elementary forms. For example, “herding” behavior
consists of a combination of “flocking” and “surrounding”.
For the three systems cited above, while interesting group behavior is exhibited on the
assigned tasks, the individual agents are programmatically defined with task-dependent knowl-
edge, and thus limited to use in narrow domains. One of the goals in this work is to investigate the
behavior of agents constructed from task-independent components. Different configurations of
such components result in agents capable of performing varied tasks. In Mona, networks are con-
structed of three types of neurons: receptors, motors, and mediators. Within a network, neurons of
a given type are distinguishable only by their internal parameters and interconnections. A future
goal is to design systems which can re-configure their components by learning or evolution, and
thereby become capable of performing new tasks.
5 Conclusions
Real birds and ants must deal with more numerous and complex needs than the simulated crea-
tures presented in this paper, of course. In addition to a more complex nest-building process, birds
must also regulate egg temperature, rear hatchlings, fend off predators, etc. What is of primary
interest here is the general question of how artificial neural networks can produce lifelike behav-
ior. A basic assumption is that behavior is motivated by needs which are translated into responses.
Using the model, the ant successfully forages for food in a class of simulated environments, and
the pair of simulated birds are able to cooperate in the construction of a nest while keeping them-
selves fed, a process involving the orchestration of sometimes conflicting needs.
The programming for the Mona neural network model and several tasks, including the nesting
birds and foraging ant, may be downloaded from http://www.corecomm.net/portegys/ai.html.
6 Future Work
6.1 Context Dependency
In Schank’s scripts (Schank and Childers, 1984), environmental cues combine to produce expec-
tations about available scenarios. For example, the presence of a table, crowd noise, and waiter
contribute to the expectation of things available in a restaurant, such as the ability to order and eat
a meal. Cues form intersections of contexts, yielding expectations. Scripts are a powerful model
of cognition at a high level of abstraction: that of natural language. In the terminology of Mona,
the abstraction becomes grounded in a neural network; the firing of events associated with the
table, crowd noise, and waiter accumulate to enable mediators of restaurant events. A unique con-
text of events can also serve to discriminate an instance of a general type. For example, I want to
relate to my dog both as my pet (discrimination) and as a canine (generalization). Categorizing the
environment in a contextual way is a vital part of cognition and clearly necessary for effective
goal-seeking.
6.2 Learning
Although Mona adapts to its environment using short term memory, it obviously lacks a learning
or long-term adaptation capability, which necessitates manual composition of the network. This
omission exists in order to initially focus on operational aspects of the model. The general appli-
cability of the model will only be realized when learning is added. This would likely involve a
“hypothesizer” function, which observes sequences of events and posits causality between them.
Mona is amenable to this type of learning: receptor neurons embody sensory events; motor neu-
rons embody response events, and mediator neurons embody event sequences. Successful hypoth-
eses would strengthen mediators, while failing ones would weaken them.
7 References
Anderson, C., Blackwell, P. G., and Cannings, C. (1997). Stochastic Simulation of Ants that For-
age by Expectation. In P. Husbands and I. Harvey (Eds.), Fourth European Conference on
Artificial Life. Cambridge, MA: MIT Press.
Albus, J. S. (1979). Mechanisms of Planning and Problem Solving in the Brain. Math. Biosci. 45,
247-293.
Bonabeau, E., Dorigo, M., and Théraulaz, G. (1999). Swarm Intelligence: From Natural to Artifi-
cial Systems. Oxford University Press.
Bonabeau, E. and Théraulaz, G. (2000). Swarm Smarts. Scientific American. 282:3, 72-79.
Drogoul, A. and Ferber, J. (1993). From Tom Thumb to the Dockers: Some Experiments with
Foraging Robots. In J-A Meyer, H. L. Roitblat, and S. W. Wilson (Eds.), From Animals to Ani-
mats II: Proceedings of the Second International Conference on Simulation of Adaptive
Behavior. Cambridge, MA: MIT Press.
Dudai, Y. (1989). The Neurobiology of Memory. New York: Oxford University Press.
Fu, L. (1994). Neural Networks in Computer Intelligence. McGraw-Hill, Inc.
Goss, S. and Deneubourg, J.L. (1992). Harvesting by a Group of Robots. In F.J. Varela and P.
Bourgine (Eds.), Toward a Practice of Autonomous Systems: Proceedings of the First Euro-
pean Conference on Artificial Life. Cambridge, MA: MIT Press.
Hampson, S. E. (1990). Connectionist Problem Solving: Computational Aspects of Biological
Learning. Birkhäuser Boston.
Holmes, W. and Rall, W. (1992) Electrotonic Models of Neuronal Dendrites and Single Neuron
Computation. In T. McKenna et al. (Eds.), Single Neuron Computation, San Diego, CA: Aca-
demic Press.
Hopfield, J. and Tank, D. (1986). Computing with Neural Circuits: A Model. Science, 233, 625-
633.
Kodjabachian, J. and Meyer, J-A. (1998). Evolution and Development of Neural Controllers for
Locomotion, Gradient-Following, and Obstacle-Avoidance in Artificial Insects. IEEE Trans-
actions on Neural Networks. 9:5, 796-812.
Mataric, M. (1995). Designing and Understanding Adaptive Group Behavior. Adaptive Behavior.
4:1, 51-80.
McClelland, D. (1987). Human Motivation. Cambridge: Cambridge University Press.
Munakata, T. (1998). Fundamentals of the New Artificial Intelligence: Beyond Traditional Para-
digms. New York: Springer-Verlag Inc.
Murciano, A. and Millán, J. (1996). Learning Signaling Behaviors and Specialization in Coopera-
tive Agents. Adaptive Behavior. 5:1, 5-28.
Nolfi, S. and Parisi, D. (1996). Learning to Adapt to Changing Environments in Evolving Neural
Networks. Adaptive Behavior. 5:1, 75-98.
Parten, C. R. (1990). Handbook of Neural Computing Applications. Academic Press, Inc.
Pfaff, D. W. (1982). Motivational Concepts: Definitions and Distinctions. In E. Pfaff (Ed.), The
Physiological Mechanisms of Motivation. New York: Springer-Verlag Inc.
Portegys, T. (1986) GIL - an Experiment in Goal-directed Inductive Learning. Ph.D. dissertation,
Northwestern University, Evanston (Available from UMI at http://www.umi.com/).
Portegys, T. (1999). A Connectionist Model of Motivation. Proceedings of the International Joint
Conference on Neural Networks (IJCNN’99). IEEE Catalog Number: 99CH36339C.
Reynolds, C. W. (1987). Flocks, Herds, and Schools: A Distributed Behavior Model. Computer
Graphics 21:4, 25-34.
Roy, A. (1997) Panel Discussion at ICNN97 on Connectionist Learning. In D. Levine (Ed.), Neu-
ral Networks, 2:2.
Schank, R. C., with Childers, P. G. (1984). The Cognitive Computer; On Language, Learning, and
Artificial Intelligence. Addison-Wesley Publishing Company, Inc.
Skinner, B. F. (1957). Verbal Behavior. New York: Appleton-Century-Crofts.
Thompson, R., Berger, T., and Berry S. (1980). Brain Anatomy and Function. In M. Wittrock
(Ed.), The Brain and Psychology, Academic Press.
Waltz, D. (1999). The Importance of Importance. AI Magazine. 20:3, 18-35.
... For the purpose of comparison we use the standard LSTM library. To ensure that we set optimally its numerous parameters, we repeat some of the published results [117][118][119], where the performance of the LSTM model is compared to the performance of the Q-learner and Elman network. The experiments presented here are an adaptation of those proposed in literature [117][118][119]. ...
... To ensure that we set optimally its numerous parameters, we repeat some of the published results [117][118][119], where the performance of the LSTM model is compared to the performance of the Q-learner and Elman network. The experiments presented here are an adaptation of those proposed in literature [117][118][119]. The task is to learn the shortest path to the goal starting from any valid position in a discrete maze world. ...
... As is sometimes the case, this project "fell out" while organizing another project. That other project being a sort of capstone that mirrors a previous effort to model nesting bird behavior using the Mona goal-seeking ANN (Portegys, 2001), which is a type of modular reinforcement learning system (Moerman, 2009). The plan is to combine Mona with Morphognosis and leverage the strengths of both architectures to perform an enhanced nest-building task. ...
Preprint
Full-text available
Biological neural networks operate in the presence of task disruptions as they guide organisms toward goals. A familiar stream of stimulus-response causations can be disrupted by subtask streams imposed by the environment. For example, taking a familiar path to a foraging area might be disrupted by the presence of a predator, necessitating a "detour" to the area. The detour can be a known alternative path that must be dynamically composed with the original path to accomplish the overall task. In this project, overarching base paths are disrupted by independently learned path modules in the form of insertion, substitution, and deletion modifications to the base paths such that the resulting modified paths are novel to the network. The network's performance is then tested on these paths that have been learned in piecemeal fashion. In sum, the network must compose a new task on the fly. Several network architectures are tested: Time delay neural network (TDNN), Long short-term memory (LSTM), Temporal convolutional network (TCN), and Morphognosis, a hierarchical neural network. LSTM and Morphognosis perform significantly better for this task.
... Combining reinforcement learning (Francois-Lavet et al. 2018) and modularity is a strength of the Mona goal-seeking neural network (Portegys, 2001). Mona accomplishes this through cause-andeffect learning, building in the process hierarchies of networks that serve as conduits for goal values motivating responses toward goals. ...
Preprint
Full-text available
This study compares the modular learning performance of two artificial neural network architectures: a Long Short-Term Memory (LSTM) recurrent network, and Morphognosis, a neural network based on a hierarchy of spatial and temporal contexts. Mazes are used to measure performance, defined as the ability to utilize independently learned mazes to solve mazes composed of them. A maze is a sequence of rooms connected by doors. The modular task is implemented as follows: at the beginning of the maze, an initial door choice forms a context that must be retained until the end of an intervening maze, where the same door must be chosen again to reach the goal. For testing, the door-association mazes and separately trained intervening mazes are presented together for the first time. While both neural networks perform well during training, the testing performance of Morphognosis is significantly better than LSTM.
... The interest in simulating group behaviour through agentbased techniques goes back at least to the original BOIDS simulation of flock flight behaviour (Reynolds, 1987). Since then, several other researchers have used models based on behavioural results to study fast moving groups (Brogan and Hodgins, 1997), swarms of individuals working on the same task (Corner and Lamont, 2004), and even cooperative nest building (Portegys, 2001). Despite its popular place in behavioural science, little work has been done with the generalized matching equation as it pertains to simulation work. ...
Article
In this study, we present a selective overview of group and individual foraging behaviour and propose an agent-based model to unify these methodologies, enabling the exploration of plausible causal mechanisms and novel hypothesis generation. Despite its popular place in behavioural science, little work has been done with the generalized matching equation as it pertains to simulation work. Our model uses psychologically justifiable and theoretically grounded models of memory for time and decision rules. The results of preliminary simulation experiments suggest that the model is sensitive to changes in pay-off density and correctly replicates Ideal Free Distributions.
... It is important to mention that there is not one connectionist model, but there is a multitude of different connectionist models each with different assumptions, focusing on different phenomena (e.g., connectionist models of emotion regulation, Armony, Servan-Schreiber, Cohen, & LeDoux, 1997; or connectionist models of motivation, e.g., Portegys, 2001). It is also important to note that in the literature connectionism is often discussed as contrary to production-rule architectures. ...
Article
Full-text available
This article describes PSI theory, which is a formalized computational architecture of human psychological processes. In contrast to other existing theories, PSI theory not only models cognitive, but also motivational and emotional processes and their interactions. The article starts with a brief overview of the theory showing the connections between its different parts. We then discuss the theory’s components in greater detail. Key constructs and processes are the five basic human needs, the satisfaction of needs using the cognitive system, including perception, schemas in memory, planning, and action. Furthermore, emotions are defined and the role of emotions in cognitive and motivational processes is elaborated, referring to a specific example. The neural basis of the PSI theory is also highlighted referring to the “quad structure,” to specific brain areas, and to thinking as scanning in a neural network. Finally, some evidence for the validity of the theory is provided.
... Benson and Nilsson, 1995) are typically symbolic, not connectionistic systems, necessitating a novel learning solution for Mona. Mona has modeled complex behavior on a number of tasks, including foraging and cooperative nest-building (Portegys, 1999(Portegys, & 2001. See www.itk.ilstu.edu/faculty/portegys/programs/NestViewer/NestViewer.html for an exhibit of the nest-building task. ...
Article
Full-text available
An important function of many organisms is the ability to learn contextual information in order to increase the probability of achieving goals. For example, a cat may watch a particular mouse hole where she has experienced success in catching mice in preference to other similar holes. Or a person will improve his chances of getting back into his house by taking his keys with him. In this paper, predisposing conditions that affect future outcomes are referred to as environmental contexts. These conditional probabilities are learned by a goal-seeking neural network. Environmental contexts of varying complexities are generated that contain conditional state-transition probabilities such that the probability of some transitions is affected by the completion of others. The neural network is capable of expressing responses that allow it to navigate the environment in order to reach a goal. The goal-seeking effectiveness of the neural network in a variety of environmental complexities is measured.
... Although a connectionist architecture, Mona is more of a state-based planning system that a conventional pattern classifying neural network. It has exhibited complex behavior on a number of tasks, including cooperative nest-building (Portegys, 2001) (www.itk.ilstu.edu/faculty/portegys/programs/NestViewer/ NestViewer.html). ...
Conference Paper
Full-text available
Instincts are a vital part of the behavioral repertoire of organisms. Even humans rely heavily on these inborn mechanisms for survival. Many creatures, for example, build elaborate nests without ever learning through experience. This paper explores this evolutionary legacy in the context of an artificial goal-seeking neural network. An instinct is defined as a simple stimulus-response sequence that is triggered by environmental and other events. The well-known "Monkey and Bananas" problem is used as a task situation. Instincts are "hard-wired" neurons in the brain of a monkey. Using a genetic algorithm, a population of monkeys evolved to successfully solve the task that none were able to solve by experience alone. The solutions were also found to be quite adaptable to variations in the task; in fact more so than a hand-crafted solution.
... Mona has modeled complex behavior on a number of tasks, including foraging and cooperative nest-building [13,14]. For an exhibit of the nest-building task, see www.itk.ilstu.edu/faculty/portegys/programs/NestViewer ...
Conference Paper
Full-text available
An important function of many organisms is the ability to use contextual information in order to increase the probability of achieving goals. For example, a street address has a particular meaning only in the context of the city it is in. In this paper, predisposing conditions that influence future outcomes are learned by a goal-seeking neural network called Mona. A maze problem is used as a context-learning exercise. At the beginning of the maze, an initial door choice forms a context that must be remembered until the end of the maze, where the same door must be chosen again in order to reach a goal. Mona must learn these door associations and the intervening path through the maze. Movement is accomplished by expressing responses to the environment. The goal-seeking effectiveness of the neural network in a variety of maze complexities is measured.
Article
Full-text available
Mona is a goal - seeking artificial neural network (ANN) that learns hierarchie s of cause and effect contexts . These contexts allow Mona to predict future events. The structure of the environment is modeled in long - term memory; the state of the environment is modeled in working memory. Mona is an active system: it uses environmental contexts to produce responses that navigate the environment toward goal events that satisfy internal needs. Goal - seeking thus also filters learning to retain more relevant information about the environment. (PDF) MONA: HIERARCHICAL CONTEXT-LEARNING IN A GOAL- SEEKING ARTIFICIAL NEURAL NETWORK. Available from: https://www.researchgate.net/publication/261698478_MONA_HIERARCHICAL_CONTEXT-LEARNING_IN_A_GOAL-_SEEKING_ARTIFICIAL_NEURAL_NETWORK [accessed Aug 31 2019].
Conference Paper
Agents that forage in groups must handle the problem of resource sharing. In this study, we present a selective overview of group and individual foraging behavior and propose an agent-based model to unify these methodologies, enabling the exploration of plausible causal mechanisms and novel hypothesis generation. Despite its place in behavioral science, little work has been done with the generalized matching equation as it pertains to simulation work. Our model uses psychologically plausible models of memory for time and decision rules. The results of preliminary simulation research suggest that the model is sensitive to changes in pay-off density and correctly replicates Ideal Free Distributions.
Thesis
Full-text available
GIL, a goal-directed inductive learning program, is presented. GIL learns without environment-dependent heuristics and without the "special" help of a teacher by modeling an abstraction of instrumental conditioning. In addition, GIL uses a secondary goal value learning mechanism as a substitute for state-space searching. GIL's learning of four problems is presented: (1) the game of tic-tac-toe, (2) a maze problem in which the ability to learn a novel maze was tested as a function of previous exposure to similar mazes, (3) a dual goal problem, and (4) a pattern recognition problem, in which "friend" and "foe" patterns were to be distinguished.
Article
Human intelligence is shaped by what is most important to us - the things that cause ecstasy, despair, pleasure, pain, and other intense emotions. The ability to separate the important from the unimportant underlies such faculties as attention, focusing, situation and outcome assessment, priority setting, judgment, taste, goal selection, credit assignment, the selection of relevant memories and precedents, and learning from experience. AI has for the most part focused on logic and reasoning in artificial situations where only relevant variables and operators are specified and has paid insufficient attention to processes of reducing the richness and disorganization of the real world to a form where logical reasoning can be applied. This article discusses the role of importance judgment in intelligence; provides some examples of research that make use of importance judgments; and offers suggestions for new mechanisms, architectures, applications, and research directions for AI.
Article
Suppose an organism, studied in a particular context, is presented with a well-defined stimulus and the well-defined response expected does not occur. A short time later (too soon for changes in maturation) the same organism in the same context at the same time of day is presented with the same stimulus, and the response does occur. What explains the response on the second trial? A change in the organism, called motivation, may be inferred. As an intervening variable, the concept of motivation identifies a logically required cause of behavioral change.
Article
This paper examines information-computation processes in time and in space and some aspects of computer intelligence using multidimensional matrix neural growing networks. In particular, issues of object-oriented {open_quotes}thinking{close_quotes} of computers are considered.
Article
In this article, we present a learning mechanism that allows a multiagent system to cooperate to achieve a gathering task efficiently in unknown and changing environments. The multiagent system is a team of autonomous behavior-based agents with limited communication capabilities. Cooperation is based on the acquisition of signaling behaviors and on the specialization of the agents into different types. Every agent has the same collection of built-in reactive behaviors. Some of the built-in behaviors are fixed, whereas others can be modified through reinforcement learning. The reinforcement signal is delayed until a trial is completed and assesses the collective performance of the team. Each agent uses this common signal to learn what individual behaviors are more suitable for the team. Simulation results, and the corresponding statistical analysis, show that the multiagent system always achieves near-optimal performances.