Chapter 3 The
Louse and the Mars Explorer
The antithesis of Logical
Extremism is Extreme Behaviourism, which denies any Life of the Mind and views
Life instead entirely in behavioural terms. Extreme Behaviourism, in turn, is
easily confused with the condition-action rule model of thinking.
Behaviourism
If you were analysing the behaviour of a thermostat, which regulates the temperature
of a room by turning on the heat when it is too cold and turning it off when it
is too hot, you might describe the thermostat’s
input-output behaviour in condition-action terms:
If current
temperature is T degrees and target temperature is T’ degrees and T < T’ - 2°
then the
thermostat turns on the heat.
If current
temperature is T degrees and target temperature is T’ degrees and T > T’ +
2°
then the
thermostat turns off the heat.
But you wouldn’t attribute to the thermostat a mind that
manipulates such descriptions to generate its behaviour.
In the same way that you would
view the thermostat, the behaviourist views agents in general.
Thus, in the story of the fox
and the crow, a behaviourist,
unable to examine the fox’s internal, mental state, would view the behaviour of the fox in the same way that the fox views the
behaviour of the crow:
If the fox sees that
the crow has cheese, then the fox praises the
crow.
If the fox is near the
cheese, then the fox picks up the cheese.
The
behaviourist’s description of the fox begins and
ends with the fox’s externally observable behaviour.
The behaviourist justifies her refusal to attribute
any internal, mental activity to the fox, by the fact that it is impossible to
verify such attributions by the scientific method of observation and
experimentation.
According
to the behaviourist, the fox is a purely reactive
agent, simply responding to changes in the world around her. If, in the course
of reacting to these changes, the fox gets the cheese, then this result is
merely an indirect, emergent effect, rather than one that the fox deliberately
brings about by proactive, goal-oriented reasoning.
The behaviourist
also sees no reason to distinguish between the behaviour
of a thermostat and that of a human. The behaviourist
might use an implication:
If
a passenger observes an emergency on the underground,
then
the passenger presses the alarm signal button.
to describe
the behaviour of a passenger on the underground. But the use
of such an implication says nothing about how the passenger actually generates that behaviour. As far as the behavourist
is concerned, pressing the alarm signal button whenever there is an emergency
might be only an instinct, of whose purpose the passenger is entirely unaware.
Behaviourism
is indirectly supported by Darwinism, which holds that organisms evolve by
adapting to the environment, rather than by a goal-oriented process of
self-improvement.
Behaviourism
also shares with condition-action rules its focus on modelling
behaviour as reactions to changes in the environment.
However, whereas behaviourism restricts its attention
to descriptions
of behaviour, condition-action rules are used
in production systems to generate behaviour.
Production
Systems
Few psychologists subscribe
today even to moderate versions of behaviourism. Most adhere instead to the cognitive science view that intelligent agents engage
in some form of thinking that can usefully be understood as the application of
computational procedures to mental representations of the world.
Paul
Thagard states in his book, Mind: Introduction to Cognitive Science, that, among the various models of thinking investigated in
cognitive science, production systems have “the most psychological
applications” (page 51). Steven Pinker in How the
Mind Works also uses production systems as his main example of a
computational model of the mind (page 69).
A production system is
a collection of condition-action rules incorporated in the thinking component
of an agent’s observation-thought-decision-action cycle.
Condition-action rules (also called production rules) are similar to the behaviourist’s
descriptions of behaviour. However, because they are
used by an agent internally to generate its behaviour,
their conclusions can be expressed in the imperative, rather than in the
declarative mood:
If conditions then do actions.
Production
systems were invented in the 1930’s by the logician, Emil Post, but were
proposed as a computational model of human intelligence by Alan Newell.
The
Production System Cycle
Production systems embed condition-action rules in an
observation-thought-decision-action agent cycle:
To cycle,
observe the world,
think,
decide what actions to perform,
act,
cycle again.
Thinking is a form of forward reasoning, initiated by an
observation matching one of the conditions of a condition-action rule. In such
a case, the observation is said to trigger
the condition-action rule. As in logic, the remaining conditions of the rule
are verified and the conclusion is derived.
In logic, the conclusion is an inescapable
consequence of the conditions. However, in production systems, it is only a recommendation to perform the
actions that are the conclusion. If only one rule is triggered by the
observations, then the recommendation is, in effect, an unequivocal command. However, if more than
one is triggered, then the agent needs to choose between the different
recommendations, to decide which actions to perform. This decision is called conflict
resolution, because the different recommendations may conflict with one
another.
For example:
If someone attacks me, then attack them back.
If someone attacks me,
then get help.
If someone attacks me,
then try to escape.
Deciding what to do, when there is a conflict between
different recommendations, can be harder than generating the recommendations in
the first place. We will
come back to the decision problem later.
Production
Systems without any representation of the world
In the
simplest case, an agent’s mental state
might consist entirely of production rules alone, without any mental
representation of the world. In such a case, the conditions of a rule are
verified simply by matching them against the agent’s current observations. In
this case, it can be said (and has been said) that the world serves as its own
representation: If you want to find out about the world, don’t think about it,
just look and see!
Observing the current state of the world is a
lot easier than trying to predict it from past
observations and from assumptions about the persistence of past states of
affairs. And it is a lot more reliable, because persistence assumptions can
easily go wrong, especially when there are other agents around, changing the
world to suit themselves. It’s too early to consider
this issue further in this chapter, but it is an issue we will return to later
when we look more closely at what’s involved in reasoning about the persistence
of states over time.
What
it’s like to be a louse
To see what a production
system without any representation of the world might be like, imagine that you are a wood louse
and that your entire life’s behaviour can be summed
up in the following three rules:
If it’s
clear ahead, then move forward.
If
there’s an obstacle ahead, then turn right.
If
I am tired, then stop.
Because you are such a low form of
life, you can sense only the fragment of the world that is directly in front of
you. You can also sense when you are tired. Thus, your body is a part of the
world, external to your mind. Like other external objects, your body generates
observations, such as being tired or being hungry, which have to be attended to
by your mind.
It doesn’t matter where the rules
come from, whether they evolved by a process of evolution, or whether they were
gifted to you by some Grand Designer. The important thing is, now that you have
them, they govern and regulate your life.
Suppose, for the purpose of
illustration, that you experience the following stream of observations:
Clear
ahead.
Clear
ahead.
Obstacle ahead.
Clear
ahead and tired.
Matching the observations, in
sequence, against the conditions of your rules results in the following
interleaved sequence of observations and actions:
Observe: Clear
ahead.
Do:
Move
forward.
Observe: Clear
ahead.
Do:
Move
forward.
Observe: Obstacle
ahead.
Do: Turn right.
Observe:
Clear ahead and tired.
At this point, your current
observations trigger two different rules, and their corresponding actions are
in conflict. You can’t move forward and stop at the same time. Some method of
conflict resolution is needed, to decide what to do.
Many different conflict resolution
strategies are possible. But, in this as in many other cases, the conflict can
be resolved simply[1] by
assign different priorities to the different rules, and selecting the action
generated by the rule with the highest priority. In this case it is obvious
that the third rule should have higher priority than the second. So the appropriate
action is:
Do: Stop.
Once a louse has learned its rules,
its internal state is fixed. Observations come and go and the associated
actions are performed without being recorded or remembered. The price for this
simplicity is that a louse lives only in the here and now and has no idea of
the great wide world around it. But, for a louse, this is probably a small
price to pay for enjoying the simple life.
Production
Systems with Memory
Although the simple life may
seem attractive, most people would probably prefer something more exciting. For
this, you need at least a production system with an internal memory. The
memory can be used, not only to record observations of the current state of the
world, but also to store a historical record of past observations.
Typically, an individual observation
has the form of an atomic sentence[2], so called because it contains no proper subpart that
is also a sentence. Thus, the logical form of an observation contains none of
the logical connectives, “and”, “or”, “if” and “not”, which turn simpler
sentences into more complex ones. An atomic sentence is also called an atom.
In a production system with memory,
a rule is triggered by a current observation that matches one of the conditions
of the rule. Any remaining conditions are then verified by testing them against
records of current, past or future observations, and
the actions of the rule are derived as candidates for execution.
What
it’s like to be a Mars Explorer
To imagine what a production
system with memory might be like, suppose that your life as a louse is finished
and you have been reincarnated as a robot sent on a mission to look for life on
Mars.
Fortunately, your former life
as a louse gives you a good idea how to get started. Moreover, being a robot, you never get tired
and never have to rest. However, there are two new problems you have to solve:
How do you recognise life when you see it, and how do you avoid going around in
circles.
For the first problem, your
designers have equipped you with a life recognition module, which allows you to
recognise signs of life, and with a transmitter to inform mission control of
any discoveries. For the second problem, you need a memory to recognise when
you have been to a place before, so that you can avoid going to the same place
again.
A production system with
memory, which is a refinement of the production system of a louse, might look
something like this:
If the place
ahead is clear
and I
haven’t gone to the place before,
then go to
the place.
If the place
ahead is clear
and I have
gone to the place before,
then turn
right.
If
there’s an obstacle ahead
and
it doesn’t show signs of life,
then
turn right.
If
there’s an obstacle ahead
and
it shows signs of life,
then
report it to mission control
and
turn right.
To recognise
whether you have been to a place before, you need to make a map of the terrain.
You can do this, for example, by dividing the terrain into little squares and
naming each square by a co-ordinate, (E, N), where E is the distance of the
centre of the square East of the origin, N is its distance North of the origin,
and the origin (0, 0) is the square where you start.
For this to work, each square should
be the same size as the step you take when you move one step forward. Assuming
that you know the co-ordinates of your current location, you can then use
simple arithmetic to compute the co-ordinates of both the square ahead of you
and the square to the right of you, and therefore the co-ordinates of your next
location.
Every time you go to a square, you
record your observation of the square together with its co-ordinates. Then, to
find out whether you have gone to a place before, you just consult your memory
of past observations.
Suppose for example, that you are at
the origin, pointed in an Easterly direction. Suppose also that the following
atomic sentences describe part of the world around you:
Life
at (2, 1)
Clear
at (1, 0)
Clear
at (2, 0)
Obstacle at (3, 0)
Obstacle at (2, -1)
Obstacle at (2,
1).
Although there is life in your
vicinity, you can’t see it yet. So, when you start, the only thing you know
about the world is that it is clear at (1, 0).
Assume also that, although it is
your mission to look for life, you are the only thing that moves. So this
description of the world applies to all states of the world you will encounter
(assuming that, when you occupy a place, it is still considered clear).
With these assumptions, you have no
choice. Your behaviour is completely predetermined:
Observe: Clear
at (1, 0)
Do: Go
to (1, 0)
Observe: Clear at (2, 0)
Do: Go
to (2, 0)
Observe: Obstacle at (3, 0)
Do: Turn
right
Observe: Obstacle at (2, -1)
Do: Turn
right
Observe: Clear
at (1, 0)
Remember: Gone to (1, 0)
Do: Turn
right
Observe: Obstacle at (2, 1) and
Life at (2, 1)
Do: Report
life at (2, 1) to mission control
Do: Turn
right.[3]
Notice
that reporting your discovery of life to mission control is just another
action, like moving forward or turning right. You have no idea that, for your
designers, this is the ultimate goal of your existence.
Your
designers have endowed you with a production system that achieves the goal of discovering
life as an emergent property. Perhaps, for them, this goal is but a sub-goal of
some higher-level goal, such as satisfying their own scientific curiosity. But
for you, none of these goals or sub-goals is apparent.
The
use of production systems to simulate goal-reduction
Production systems have been
used, not only to construct computational models of intelligent agents, but
also to build computer applications, most often in the form of expert systems.
Many of these applications use condition-action rules to simulate
goal-reduction explicitly, instead of relying on emergent properties to achieve
higher-level goals implicitly.
For example, the fox’s reduction of the goal of having cheese to the sub-goals of
being near the cheese and picking it up can be simulated by the
condition-action rule[4]:
If I want to have
an object
then
add to my beliefs that I want to be near the object
and
pick up the object.
Here a goal Goal is
represented in the system’s memory as a pseudo-belief of the form:
I
want Goal.
The
reduction of Goal to Sub-goals
is simulated by a condition-action rule with a condition Goal and actions that are
either pseudo-actions of the form:
add to my beliefs
that I want Sub-goal
performed internally on the
system’s memory or genuine actions performed externally.
The main problem with the simulation
approach is that it looses the connection between goal-reduction and the belief
that justifies it, in this case with the belief:
An animal has an
object
if the animal is near the object
and the animal picks up the object.
As we have seen, the connection is that goal-reduction is the same as
reasoning backwards with the belief, which is the main idea of logic
programming.
Thagard (page 45) gives a similar
example of a condition-action rule, but uses it to illustrate his claim that
“unlike logic, rule-based systems can also easily represent strategic
information about what to do”:
If
you want to go home and you have the bus fare,
then you can catch
a bus.
Forward
reasoning with the rule reduces the goal (going home) to a sub-goal (catching a
bus), and simulates backward reasoning with the belief[5]:
You go home if you have the bus fare and you catch a bus.
Thus Thagard’s argument against logic can be viewed instead as an argument for logic programming, because it can “easily represent strategic information about what to do”.
Indeed, it seems that all of Thagard’s arguments for production rules are better
understood as arguments for computational logic instead. This is because he
confuses production
rules:
If conditions then do
actions.
with logical
implications:
If conditions then conclusions.
An unfortunate confusion
This confusion is perhaps most
apparent when Thagard writes (page 47) that “rules
can be used to reason either forward
or backward.” But this is not a true
property of production rules, but rather a characteristic feature of logical
implications.
Because conditions in
production rules come first and actions happen later, true production rules can
be used only in the forward direction, when the conditions hold to derive
candidate actions. But because conclusions in logic are always expressed in the
declarative mood, logical implications can be used to reason either forward or
backward.
Thus Thagard
mistakenly attributes to production rules a property that they don’t have, but
that logical implications do, and then he uses this attribution to argue that
“rules” are better than logic.
To be
fair, it has to be recognised that Thagard is only
reporting a generally held confusion. In this case, the rule that he uses as an
example simulates goal-reduction, which
is a special case of backwards reasoning with a belief
expressed in logical form. However, in the next chapter, we will see that true production
rules are a special case of forward reasoning with goals
expressed in logical form. We will also see how goal-reduction and production
rules can be combined in a more general framework, which uses logic for both
beliefs and goals.
Summary
The use of production systems to generate the
behaviour of an intelligent agent, as seen in this chapter, can be pictured
like this:
Observations Actions
”Forward reasoning”
