<! OPEN THIS DOCUMENT WITH YOUR WEB BROWSER!>
<b><font size=8>The Neural Servo Model</font><br>
<font size=5>Mechanism, Strategy and Tactics for Survival</font></b>
In this paper it is suggested that simple feedback control:
a) is an effective means of increasing chances of survival for any organism
b) carries a hidden agenda for more complex systems and organisms.
An additional mechanism that exposes this hidden agenda
has been explored and tested in simulations, providing a neural network
with a generic strategy for survival in the real world.This additional mechanism is suggested to model our system of (mental) arousal.
<b>Keywords:</b> arousal, EEG, neural network, servo control, survival</b>
<b>Introduction: Learning computers?</b>
If computers can learn, what do we need programmers for?
Well, computers can't learn, but what we <i>can</i> do,
is design programs that can learn.
Designing such a program can be very similar to programming
a computer game for a human player. This is because playing a game
is all about learning. Most games offer a world (that is to some degree
similar to the world we usually live in) and an agent, controlled by the player.
The agent can have human appearance, but it can also be a helicopter
or a space ship.
Every game allows the agent a degree of control. For example, control over the
location of the agent. However, control is useless without knowledge.
Where should the agent go next? Which locations are dangerous?
To answer these questions the player has to acquire knowledge.
The act of exerting control is simple (moving the joystick), it is
exerting useful control that has to be learned.
Both people and computer programs can acquire useful control skills by means
of feedback. For most games, this feedback takes a very simple form.
An indication of how well you are playing is often simply the length of your
game. This would be all the feedback needed to train a neural network as well.
A nice example of a neural network is a crossing with traffic lights and
induction sensors. This is the kind of network that only gives a green light
when you move your car close enough to the traffic light. The knowledge
that this network represents is very limited. The most important rule
is that two crossing roads should not get a green light at the same time.
One solution would be that one traffic light dominates. This would
be practical if this light regulates a very sparse stream of traffic. In other
cases it is more practical to alternate between the two directions, just like
a network without sensors would do.
The point of this exercise is to demonstrate that the structure of network
can represent knowledge and that the structure can exhibit useful control.
To make such a network learn, one can introduce evolution. Evolution
requires two things: variation and selection. Variation can be obtained
by introducing changes to the structure (knowledge) of the network.
Selection can be obtained by introducing a learning rule, for example
'keep those changes that reduce the number of sensed cars'.
This learning rule is quite effective, since it would acquire quite an amount
of knowledge, especially in a community without traffic regulation except
'green is go'!
First the system shall find our basic rule: don't give two green lights. Why?
The resulting traffic jam and collisions provide an increase in the number
of sensed cars so that the system will never preserve connections that do give two
green lights at the same time. It <i>will</i> preserve connections that give
green lights in such a manner that the crossing will become less crowded,
implying less resistance for the flow of traffic. So, we have a learning system
that would be useful in real life!
Even more practical is it to simulate such a network, the traffic and the
manipulation (evolution) of the network inside a computer. The traffic
would represent the world (the problem) and the network the agent. This
allows us to train and test the network without causing serious accidents.
Now that you are more familiar with the mode of thought associated with
artificial intelligence, I shall continue with the introduction
of a control mechanism that is much simpler than this traffic control
system: the servo.
<b>1 Mechanism for survival: The servo</b>
The first reason why a servo can be suggested to be a fundamental
mechanism for survival is that it is active until it is satisfied.
An example of a servo is a thermostat that controls a central heating system.
You provide a preset, a goal temperature. All the thermostat has to do, is
measure the actual temperature and activate the heater if the actual temperature
is lower than the goal temperature: <i>heat supply = reality - goal</i><br>
(negative 'heat supply' could activate a cooling system).
The question is, who controls the heater? Well, the only thing that changes
is reality itself (as a result of the heat supply). For this reason, another term
for servo control is feedback control. Reality can control itself by means of
structures that exploit feedback.
Any organism advanced enough to establish (genetically or through experience)
preferences or <i>goal states</i> (like having a full stomach) can benefit from these
control structures that automatically initiate actions (like swimming) until all
goal states are satisfied. This is the first level of explanation that suggests
that the servo could well be a fundamental mechanism for survival.
<b>2 Strategy for survival: The hidden agenda of the servo</b>
The second level of explanation is hidden. Hidden in the sense that
to provide a step by step demonstration would take to much time.
This is what the study of complexity is all about: finding out what
certain rules do if applied to large scale games and vast amounts of
moves. In other words, doing computer simulations and observing
The rule that has been tested during this study is the single rule that is
incorporated in the structure of any servo: <i>be active until satisfied</i>.
Another feature of a true servo controller is that its actions propel
reality towards a satisfactory state (no difference between reality and goal).
When we use a neural network to simulate a servo that gives us an important
advantage: the goal state of such a 'neural' servo is not fixed like in a servo.
A neural servo can become satisfied not only by propelling reality toward
its goal state, but also by propelling its goal state toward reality. This
probably sounds quite abstract, but it describes the most interesting device
I have ever seen! That is because when we train a neural
net to become a servo, we are not limited to just one 'control channel', but
we can build a network of hundreds of servos connected by internal circuits!
The only thing we have to do, is enable evolution to search for connections
that satisfy the system.
Initially the system has no goal state 'in mind' for reality. This means
that if the system were able to blow up its sensed environment (if there would
be a detonator switch it could connect to) it would do so with great satisfaction.
However, this is not the hidden agenda of the servo I want to discuss here.
Since a neural servo is probably not able to destroy its environment, it shall
make a trade-off between the learning and control effort. For example, if a
neural servo has to control a central heating system it would acquire connections
that generate a goal temperature that is equal to the actual temperature.
The effectiveness of this approach relies on the fact that
each time the actual temperature changes it will be much easier
to <i>control</i> the room temperature (after the structure has come to represent
a goal temperature) than to <i>change this knowledge</i> embodied by the network.
In effect, a neural servo tends to keep its environment the way it got to know it.
This means that if the neural servo is a legitimate model for human intelligence,
a child inevitably (since its lack of control over its environment) has to learn all
about its environment, while later on it will use the control it gains (by learning
to communicate and by developing senso-motorical skills) to preserve or recreate
the kind of environment it has grown up in. If its initial environment has been a
socially and physically healthy one, this tendency provides an effective strategy
for survival (for any social organism).
So far, we have explored the hidden agenda of the servo mechanism merely by means of
a thought experiment. The neural servo seems to be able to provide a master plan
for life: <i>get to know your environment until you have gained enough control to preserve (or recreate) it the way you found it.</i>
<b>3 Tactics for survival: Arousal unlocks the hidden agenda</b>
Compared to the results of the thought experiment, the results emerging from actual simulations may appear very limited. Their strength, however, lies mainly in their
modeling power for the effects of arousal, not in the absolute intelligence or control gained by the simulated neural servo. This additional mechanism, arousal, is suggested to be the key to the servo's hidden agenda.
The first neural servo that has been simulated, could exert no control at all, while it was exposed to a repetive series of patterns through an array of six input channels.
The simulated network consisted of six servo controllers ('thermostats') which could be connected internally by a maximum of 1200 connections between 36 neural nodes.
These neural nodes generated an outgoing pulse whenever their input per saldo exceeded a certain threshold.
This network, growing by means evolution (selecting for connections that satisfy the system), turned out to be unable to learn effectively. It could progress only until
it failed. After any 'bad luck', the system could not return to its old level of performance by destroying the last added connections. In order to understand this effect, we'll have to return to the basic concept of feedback control.
A servo is satisfied only if reality and goal state are equal: <i>input = reality - goal.</i><br>
So task of this neural servo was to generate goal states that absorbed reality before
it got a chance to pour into the network. If it had a lucky day, the prototype could
find a few appropriate goal states before it made a bad connection. Fortunately, these
lucky runs provided an interesting insight: Every latest improvement operated upon reverberating patterns that had already been transformed by the structure that was already in place!
Although it is impossible to imagine the effect of hundreds of feedback loops, delays
and logical operations on a reverberating pattern, it is possible to see why the prototype could not maintain its effectiveness after a failure. All the operations acting at once
simply didn't have the same effect as when improvements kicked in, one by one, being collected by evolution!
Solving this problem was easy enough. Connections already had a property
called 'physical strength', which was increased whenever the system became more satisfied. This way older connections had high strength, new ones low strength
(this allowed evolution to destroy weak connections whenever the system became less satisfied, so that the system returned to its original structural state).
So, any mechanism that could re-enable all connections in order of their physical strength, would cause the operations to transform the patterns in the right sequence!
An existing system that could do this job very well, is our arousal system.<br>
<i>One of the things our arousal system does, is creating fluctuating electric potentials
across wide areas of the brain</i>. These oscillations are often assumed to be
a mechanism for synchronizing neuronal firing. <i>I would suggest that these potentials
act as offsets that effectively disable and re-enable connections in order of their physical strength, effectively downsizing a network to a more primitive state before 'releasing' more and more recent parts of the network again</i>.
For the simulated neural servo this additional system did the trick: it learned to absorb repetitive series of four patterns for up to 85%. The hidden agenda of the servo turned out to be within reach as well: confronted with two patterns, alternating at
random intervals, the neural servo preffered 'flipping' back to the first pattern and learn that single pattern above learning both pattern and exerting no physical control. This is about
the most basic variation of 'get to know your environment until you have gained enough control to preserve (or recreate) it the way you found it'. <i>So, the 'hidden agenda' is a real phenomenom, and arousal is all we need to unlock it.</i>
<b>3.1 Tactics for survival: Arousal in action</b>
The electric potentials in the brain (measured with an EEG) are only one observable aspect of our system of arousal. This system connects many things:
<tr><th>Level of arousal: <th>Low <th>High
<tr><td>pulse rate: <td>slow <td>fast
<tr><td>respiration: <td>slow <td>fast
<tr><td>brainwaves: <td>slow <td>fast
<tr><td>heat release: <td>slow <td>fast
<tr><td>digestion: <td>fast <td>slow
<tr><td>growth: <td>fast <td>slow
Let's put the model to the test. When we are excited, our arousal is high. We breathe
fast to acquire more oxygen, our heart beats fast in order to get the oxygen where
it is needed: to the muscles. Heat generated by the muscles is released quickly to avoid overheating. Digestion and growth are put on hold. What happens in the brain?<br>
The electric oscillations in our brain are small and fast. So, according to the neural servo model this results in having a nearly complete network ready and being able to perform many (small) operation sequences per second. A limitation of this mode of
functioning is that large operation sequences are not done. This means that context switches are very hard to follow: it is difficult to operate a telephone when you have already started to panic.
At the opposite side of the spectrum we find 'sleep'. When we sleep, our arousal is low.
We breathe slowly, we hardly consume energy. Heat release is minimal. Digestion
recombines complex molecules in order to keep our body in good shape. The electric
oscillations in our brain are slow and big. The deeper our sleep, the more primitive
is the part of our brain that still functions: we go back to our childhood. Any signal
that we pick up reverberates through this primitive brain. The lack of reality causes
more and more feedback to occur, until we fall back into an even more deep sleep.
In the morning when there is a sufficient supply of reality pouring into our senses,
the brain can finally launch a successful attack, creating a phantasy world that keeps
matching reality until we are fully 'awake': dreaming correctly or correcting reality.
A lack of sleep has the same effects as panicing. We are unable to refresh our
phantasy world, which requires a very long sequence of operations to build (slow brainwaves). We then become aware of the fact that we create reality from within:
we 'start' halucinating.
<b>4 Future work</b>
To conclude this paper I will outline, in a very condensed form, some
more issues that could be addressed by this model:
<li>One question that has remained unanswered sofar,
is whether the neural servo is a biological reality.
Many observations seem to be in favour of this model. It turns out, for
example, that people tend to go back to their early childhood during deep sleep.
This supports the suggestion that low arousal (sleep) would temporarily downsize
our neural networks to more primitive states.
<li>Another question could be the role of arousal in communication. For example, our breathing rate is controlled by our arousal, so it could well be a means for communicating our moods. Note that a rizing tone of speech relates to increasingly powerful b
reathing, rizing arousal. This may seem trivial,but why should evolution put any effort into giving people some universal protocol for communication when it can get emotional messages across for free?
<li>Consciousness, finally, could be suggested to 'stand or fall' with the way arousal
allows more or less complete networks to transform the patterns that travel within
them. Remember that new connections do not add 'new ingredients' but that they
are more like food <i>processors:</i> they transform a pattern each time it passes by!
For this reason it is likely that more advanced (recent) connections often inhibit
or correct the transformations of their precursors. This way a picture arises of a
network that can allocate each subset of itself (a more or less primitive state) time
in order to generate a certain pattern: a pattern that should match reality to the greatest possible extent. According to this model the ability to manipulate one's own arousal by any means would extent diversity of the patterns that can be grown within o
It would be a matter of experience, however, whether or not these manipulations
result in sensible patterns: patterns that match reality or have any other meaning.
This is where we might have stumbled upon a criterion for consciousness. Maybe we
lose consciousness only if our arousal makes such a shift that a very unusual sequence
of operations is performed, resulting in a pattern that is increasingly mismatching
<center><h5><b>Copyright (c) 1997 Mervyn van Kuyen -- All rights reserved.
<tr><td>Updated July 21 (1997)</td>
<td><a href="index.html">Back to MERVYN INVENTURES</a>