Advertisement
English transcript
PROFESSOR: [INAUDIBLE].
"Thus Spake Zarathustra" was made famous
and popular by 2001.
And that is music played at this magic moment when some
primate suddenly gets an idea,
presumably one of our ancestors.
So how do we explain all that?
We've got all of the ingredients on the table.
And today I want to talk about various ways of putting those
ingredients together.
So we talked about representations, we've talked
about methods, and today we're going to talk about
architectures.
And by the end of the class you'll know how to put one of
those things together.
Actually, no one knows how to put one of
those things together.
But what you will know is about some alternatives for
putting those things together so as to make something that
is arguably intelligent in the same way we are.
So that is our agenda for today.
We'll also talk a little bit more about stories.
I think it was in 2007 when the Estonians moved a war
memorial from the center of Tallinn off to
a Russian war cemetery.
Prior to that time the Estonians had been building up
their national computer networks because they thought
that computation was the wave of the future-- networks and
all of that.
Shortly after the movement of that war memorial, someone
brought the Estonian national network down--
a cyber attack.
It was widely believed to be the Russians.
There's a large Russian ethnic population in
Estonia to start with.
And the movement of that war memorial
irritated the Russians.
And so everybody everybody thinks that they did it.
But you know what?
No computer can understand the story I just told.
They can revel through all of the worldwide web finding
information that's relevant to that, but no computer can
understand the story I just told, except one.
You'll see a demonstration of that later on today.
So by the way, if you're interested in understanding
the nature of intelligence, this is, of course, the most
important lecture of the semester.
And I should tell you a little bit about what we're going to
do on Wednesday.
Because for some reason the day before Thanksgiving tends
to be a lecture that's lightly populated,
except in this class.
Because I'm going to talk about the artificial
intelligence business and what can be learned from it about
how to avoid going broke when you start your company.
So for many of you that will be the most important lecture
of the semester.
It all started back in the dawn age of artificial
intelligence.
And really, it all started at Carnegie Mellon, sad to say.
Because the people at Carnegie Mellon, notably Newell and
Simon, were the first to think about sort of a general
purpose way of putting things together so as to build a
structure or architecture in which particular intelligent
systems could be built.
So their idea was called the general problem solver.
A long name for a simple idea.
And the simple idea is that you start your life out in a
current state, call it C. And you want to get
to some goal state.
Call it S. And a way you do that is you measure somehow
the symbolic difference between where you are and
where you want to be.
So that's the difference.
We'll call that difference D. And when you observe that
difference that's enough, they say, in this general approach
to problem solving.
For you to select some operation that will move you
from your current state to some new state, an
intermediate state.
Call it I. So I, or that operator, O, is determined by
the difference, D.
And then, of course, the next thing to do is to measure the
difference between that intermediate state and the
state you want to be in, and choose some operator that's
relevant to reducing that state.
So we'll call that D2, and we'll call this O2.
And D2 is what leads you to 02, and so it goes.
So that's the idea.
And that's often called means-ends analysis.
Why?
Because the end that you want to achieve is being in that
final state, S. And the means is that operator, O. So you
have some notion of where you want to be and the difference
of where you are and where you want to be.
And you pick an operator so as to reduce that difference.
So this is all very abstract.
Let's exercise it in solving a problem that you will all be
faced with here in a day or two.
That is, for many of you--
most Of you, I hope-- the problem of going home.
So here you are.
You're at MIT.
And where you want to be is over here, at home.
So you measure the difference between MIT and home.
And for many of you it's further than you can go by car
and not so far that you can't go at all.
So what you do is, you say, well, the right operator is
taking an airplane.
So there is the operator, take an airplane.
And this is the difference, D. And the difference, D, being
sufficiently large, you take the plane.
Trouble is, if you happen to be sitting here in
[? 10-250 ?]
there's no way you can take an airplane, because they don't
fit in here.
So you've got another problem, and that is
to get to the airplane.
So the distance between here and Logan is such that the
right way to do that is to take the MBTA.
And that's determined because you're working on this
difference reduction right here, the difference from
being at MIT and being at the airport.
So that difference dictates that you take the MBTA.
So you see, you're re-cursing.
But you know there are no MBTA cars in here either.
So there's still a difference like so.
And that difference dictates that you walk.
So you've got D1, D2, and D3.
And by the time you've excised the operators relevant to
those three differences, you're at Logan.
Then you take the airplane, you get over to your hometown,
and your faced with the smaller difference of getting
from that airport to where you actually want to go.
So that's the general problem- solver idea.
It was such an exciting idea at the time.
It was such an exciting idea at the time because people
would say to themselves, ah!
This is a general purpose problem solver, so we can set
it onto the problem of making itself smarter.
And so there was a kind of imagined chain reaction that
would take place.
And the developers of this architecture warned the public
that within 10 years-- that is to say, by about 1970--
computers would be generally as smart as people.
And a lot of people made fun of them for that prediction.
But it was actually scientists attempting to be responsible.
Because they thought something, a quite serious
dislocation was coming along, and that people should know
that it was coming.
And so they felt it was their responsibility in that age of
scientific responsibility to warn the public.
It didn't turn out that way, because the problem of
collecting the differences and finding the operators, that's
outside the scope of the architecture.
So this is the problem that has to be solved by a human
before this architecture can be used.
You have to have identified the differences that you might
encounter and the operators that you might use, and build
this table which relates the two together.
So maybe that one, that one, some off-diagonal
elements, and so on.
But building that table turned out to be a hard job.
So not surprisingly, the idea evolved.
And eventually the folks at Carnegie who developed the
general problem solver--
most notably Newell and his students--
developed a newer, fresher, more elaborate architecture
called SOAR.
And here's how SOAR works.
First of all, what does SOAR mean?
It doesn't mean anything.
It used to mean State Operator And Result.
But for some reason the proponents of the SOAR
architecture decided they don't like that acronym, and
have asserted that SOAR is merely a label that shouldn't
be thought of as an acronym.
In any event, SOAR consists of various parts.
It has a long-term memory.
It has a short-term memory.
And it has connections to the outside world, maybe a vision
system and an action system.
But most of the activity of the SOAR problem-solving
architecture takes place in a short-term memory.
So you can view the contents of the long-term memory as
shuttling in and out of short-term memory.
So you can see right away that this mechanism, this
architecture, is heavily influenced by certain
cognitive psychology experiments having to do with
how much you can hold in your short-term memory--
nonsense syllables and all that sort of thing that was
popular back in those days.
So this was an architecture devised primarily by
psychologists.
And it had amongst its features a short-term memory
and a long-term memory.
So that's part 1 of this architecture.
So what's in the long-term memory?
Well, assertions and rules, AKA productions.
A production being the Carnegie vernacular for rule.
It's just the rule-based stuff like you saw on almost the
first day of class.
So the whole thing is a gigantic rule-based system
with assertions and rules the shuttle back and forth from
long-term memory into short-term memory where
processing takes place.
The third thing that comes to mind when you think of SOAR
architecture is they had an elaborate preference system.
You recall that when we talked about rule-based systems
there's always a question of what do you do when more than
one rule would work?
You have to have some way of breaking those ties.
The SOAR architecture has an elaborate
subsystem for doing that.
But I said that these are the first three things you think,
and maybe that's not right.
Because the next thing you think about is perhaps a
better thing to identify with the SOAR architecture.
And that's the idea of problem spaces.
And that's the idea that if you're going to solve a
problem you have to develop a space and do a search through
that space.
Just like we did when we talked about how we can get
from here to home.
There's a space of places, that's our problem space.
We can do a search through that space to find a way to
get from one place to another.
That's the sort of thing that SOAR is focused on.
Finally, the fifth element that you tend to think about
when you think about SOAR is the idea of universal
subgoaling.
And that's the idea that whenever you can't think of
want to do next, that becomes your next problem that
deserves it's own problem space and its own set of
differences and operators, and rules and assertions.
So you start off on a high level, then you have to solve
problems at a lower level, just like you did up there
with a general problem solver.
So if you have these two architectures you can begin to
say, well, what are they centered on?
And this architecture, this general problem solver, is
centered on the idea that everything is
about problem solving.
Because the problem solving hypothesis--
no one gave it that name.
But that's what it was.
And this architecture did get its name.
And it was said always, by Newell, to be based on what he
called the symbol system hypothesis.
The hypothesis that what we are as humans is symbol
manipulators.
And we can uncover how that all works by giving people
crypto-arithmetic problems and having them talk out loud, by
thinking about what happens when you try to remember
nonsense syllables, by all that sort of stuff that was en
vogue in terms of psychology experiments in the day when
this architecture was first articulated.
But when you look at architectures you can sort of
see where they come from and what their antecedents are.
It has a short-term memory and a long-term memory, because
Newell and his associates were cognitive scientists.
It has assertions and rules and preferences, because
Newell and his associates were also AI people.
And it has problem spaces and universal subgoaling because
those are ideas that had been work out in a more primitive
form already in the general problem-solver architecture.
So that's a glimpse of what SOAR looked like
in its early days.
It's been very highly developed by a
lot of smart people.
So although it's symbol centered, they've attached to
it things having to do with emotion and perception, but
generally with the view that the first thing to do when
faced with this perception is to get it out of there and get
it into a symbolic form.
That's sort of the bias that the architecture comes with.
So those are two architectures that are heavily biased toward
thinking that the important part of what we
do is problem solving.
But the most important, perhaps--
at least from an MIT perspective--
of these problem-solving oriented ways of thinking
about the world, is Marvin Minsky's architecture, which
he articulates in his book "The Emotion Machine."
And Marvin is not just concerned with problem
solving, but also with how problem solving
might come in layers.
So let me show you an example of the sort of problem what
motivates some of Marvin's thinking.
So you can read that, it's a short little vignette.
You have no trouble understanding it, right?
No.
It's not difficult for us humans.
Awfully tough for a computer.
In part, because the thinking you need, your ability to
understand that story, requires you to think on many
levels at the same time.
First of all, there's a sort of, at the
bottom, instinctive reaction.
You see where there's instinctive reaction?
That's the part where she hears a sound
and terns her head.
That's instinct, right?
That's practically built in.
But then what she sees is a car.
So that's something that we don't have wired in.
It would be unlikely that we've evolved in the last 100
years to have an instinctive appreciation of cars
barrelling down the road.
So the next level in Marvin's
architecture is learned reaction.
So that's the part about thinking about the car.
Now spread throughout there-- well, let's see where is a
particularly good example.
She decides to sprint across the road.
So that's where she's solving a problem.
So that's the deliberative thinking level.
It doesn't stop there, because later on she reflects on her
impulsive decision.
So she thinks not only about stuff that's happening out
there in the world, but she also thinks about stuff that's
going on in here.
So that's a level which we can call reflective thinking.
Well, you know, it doesn't stop there, because she also
considers, in another part of the story, something about
being uneasy about arriving late.
So she's not only just thinking about events that are
going on in her mind right now, but events that are going
on right now relative to plans she's made.
Some Marvin calls that the self-reflecting layer.
But that isn't the whole thing either, because toward the end
of the story she starts to worry about what her friends
would think of her.
So there's a kind of reflective thinking in a more
social context.
So he calls that self-conscious thinking.
So as the Carnegie folks think, the SOAR architecture
focuses mostly on problem solving, Minsky's "Emotion
Machine" book considers not just thinking, but thinking on
many layers.
And the blocker to doing any of that can be said to be the
development of common sense, which computers, alas, have
never had much of.
So this could be said to be based on the common sense
hypothesis.
And the common sense hypothesis holds that in order
to do all of that stuff, you have to have
common sense like people.
And if you have to have common sense like people, you have to
think about how much of that is there and how
can we go get it?
And so this spawned a lot of activity in the media lab
amongst people influenced by Marvin, having to do with
gathering common sense.
The open mind project, the work of Henry Lieberman and
others, having to do with the gathering of common sense from
the world wide web as a way of populating systems that would
lay the foundation for doing this kind of layered thinking.
So that is a brief survey of some mechanisms, some older
than others, but all but GPS--
GPS too.
Let's face it, it's hard to think of solving any problem
without means-ends analysis being involved.
So GPS isn't wrong, it's just not the only tool you need to
think about what to do.
So these are early, and late, and still-current.
But it's not the only thing there is, because there have
been reactions against this problem-solving way of
thinking about the development of intelligence.
And the most prominent of those counter currents, of
those alternative ideas, belongs to Rod Brooks and his
subsumption architecture.
So along about the early--
along about the years surrounding 1990, Brooks
became upset--
subsumption--
because robots couldn't do much.
They would turn them on at night, and then the next
morning they'd come in the laboratory and they would have
moved 25 feet, nicely avoiding a table perhaps.
Not doing very much and taking a long time to do it.
So he had decided that it's because people were thinking
in the wrong way.
In those days people thought that the way you build a robot
is you build a vision system, and then you build a reasoning
system, and then you build an action system.
And it can do almost nothing, but it does something.
So you improve the vision system, and improve the
reasoning system, and improve the actual system.
And now you've broken it, because all the stuff you used
to be able to do doesn't work anymore.
So what's the alternative?
Well, the alternative, as articulated by Brooks, is to
turn this idea on its side.
So instead of having an encapsulated vision system, an
encapsulated reasoning system, and an encapsulated action
system, what you have is layers that are focused not on
the sensing and the reasoning and the action, but layers
that are specialized to dealing with the world.
So in Brook's way of thinking about things, at the lowest
level you might have a system that's capable of--
well, before we get to that, avoiding objects.
And maybe the next level up is the wandering layer.
And maybe the next level up after that is explore.
And maybe the next level up after that is seek.
Now in the old days when people took 6001 I had no
trouble getting an answer the question, what does this
remind you of in 6001?
It doesn't remind you of anything in 6001 since you
haven't taken it.
But it viewed, as a generalization of a
programming idea, what is the programming idea?
There are only a few powerful ideas in program, and this is
a generalization of one of them.
What is it?
Do you have a name?
Yes, Andrew?
STUDENT: Layers of abstraction?
PROFESSOR: Layers of abstraction,
and abstraction barriers.
That nails it pretty well.
Because each of these guys can have its own vision, action,
and reasoning system.
And if you think of these as abstraction boundaries, then
when you got this thing working you don't screw with
it anymore.
You build this layer on top.
And it may reach down in here from time to time, but it
doesn't fundamentally change it.
Brooks was inspired in part by the way our brains are
constructed.
All that old stuff that we share with pigs is down in
there deep, and we put the neocortex over it.
So it looks layered in a way that would make
[? Gerry Sussman ?] proud.
So this then is the way that Brooks looks at the world, and
it's characterized by a few features just like SOAR is.
One of those features is no representation.
So this is a detail that's probably right at the level
that Brooks was operating, and very questionable when you get
above the level that Brooks was operating.
But before I go on, let me say what the hypothesis is.
The hypothesis is the creature hypothesis.
It's the hypothesis that once you can get a machine to act
as smart as an insect, then the rest will be easy.
Well, how do you get a creature to
be smart as an insect?
Maybe you don't need representation.
We focused on representation in this course, so you can see
there's a little stress---
Next thing is, what do you do if you don't have a
representation?
Let's see.
Your representation makes a model possible.
Models make it possible to predict, to understand, to
explain, and to control.
So if you don't have one what can you possibly do?
Brooks' answer is, you use the world instead of a model.
So everything you do is reactive.
You don't have anything in your head that is
a map of this room.
But maybe I don't need one because I can get around that
table by const
…
"Thus Spake Zarathustra" was made famous
and popular by 2001.
And that is music played at this magic moment when some
primate suddenly gets an idea,
presumably one of our ancestors.
So how do we explain all that?
We've got all of the ingredients on the table.
And today I want to talk about various ways of putting those
ingredients together.
So we talked about representations, we've talked
about methods, and today we're going to talk about
architectures.
And by the end of the class you'll know how to put one of
those things together.
Actually, no one knows how to put one of
those things together.
But what you will know is about some alternatives for
putting those things together so as to make something that
is arguably intelligent in the same way we are.
So that is our agenda for today.
We'll also talk a little bit more about stories.
I think it was in 2007 when the Estonians moved a war
memorial from the center of Tallinn off to
a Russian war cemetery.
Prior to that time the Estonians had been building up
their national computer networks because they thought
that computation was the wave of the future-- networks and
all of that.
Shortly after the movement of that war memorial, someone
brought the Estonian national network down--
a cyber attack.
It was widely believed to be the Russians.
There's a large Russian ethnic population in
Estonia to start with.
And the movement of that war memorial
irritated the Russians.
And so everybody everybody thinks that they did it.
But you know what?
No computer can understand the story I just told.
They can revel through all of the worldwide web finding
information that's relevant to that, but no computer can
understand the story I just told, except one.
You'll see a demonstration of that later on today.
So by the way, if you're interested in understanding
the nature of intelligence, this is, of course, the most
important lecture of the semester.
And I should tell you a little bit about what we're going to
do on Wednesday.
Because for some reason the day before Thanksgiving tends
to be a lecture that's lightly populated,
except in this class.
Because I'm going to talk about the artificial
intelligence business and what can be learned from it about
how to avoid going broke when you start your company.
So for many of you that will be the most important lecture
of the semester.
It all started back in the dawn age of artificial
intelligence.
And really, it all started at Carnegie Mellon, sad to say.
Because the people at Carnegie Mellon, notably Newell and
Simon, were the first to think about sort of a general
purpose way of putting things together so as to build a
structure or architecture in which particular intelligent
systems could be built.
So their idea was called the general problem solver.
A long name for a simple idea.
And the simple idea is that you start your life out in a
current state, call it C. And you want to get
to some goal state.
Call it S. And a way you do that is you measure somehow
the symbolic difference between where you are and
where you want to be.
So that's the difference.
We'll call that difference D. And when you observe that
difference that's enough, they say, in this general approach
to problem solving.
For you to select some operation that will move you
from your current state to some new state, an
intermediate state.
Call it I. So I, or that operator, O, is determined by
the difference, D.
And then, of course, the next thing to do is to measure the
difference between that intermediate state and the
state you want to be in, and choose some operator that's
relevant to reducing that state.
So we'll call that D2, and we'll call this O2.
And D2 is what leads you to 02, and so it goes.
So that's the idea.
And that's often called means-ends analysis.
Why?
Because the end that you want to achieve is being in that
final state, S. And the means is that operator, O. So you
have some notion of where you want to be and the difference
of where you are and where you want to be.
And you pick an operator so as to reduce that difference.
So this is all very abstract.
Let's exercise it in solving a problem that you will all be
faced with here in a day or two.
That is, for many of you--
most Of you, I hope-- the problem of going home.
So here you are.
You're at MIT.
And where you want to be is over here, at home.
So you measure the difference between MIT and home.
And for many of you it's further than you can go by car
and not so far that you can't go at all.
So what you do is, you say, well, the right operator is
taking an airplane.
So there is the operator, take an airplane.
And this is the difference, D. And the difference, D, being
sufficiently large, you take the plane.
Trouble is, if you happen to be sitting here in
[? 10-250 ?]
there's no way you can take an airplane, because they don't
fit in here.
So you've got another problem, and that is
to get to the airplane.
So the distance between here and Logan is such that the
right way to do that is to take the MBTA.
And that's determined because you're working on this
difference reduction right here, the difference from
being at MIT and being at the airport.
So that difference dictates that you take the MBTA.
So you see, you're re-cursing.
But you know there are no MBTA cars in here either.
So there's still a difference like so.
And that difference dictates that you walk.
So you've got D1, D2, and D3.
And by the time you've excised the operators relevant to
those three differences, you're at Logan.
Then you take the airplane, you get over to your hometown,
and your faced with the smaller difference of getting
from that airport to where you actually want to go.
So that's the general problem- solver idea.
It was such an exciting idea at the time.
It was such an exciting idea at the time because people
would say to themselves, ah!
This is a general purpose problem solver, so we can set
it onto the problem of making itself smarter.
And so there was a kind of imagined chain reaction that
would take place.
And the developers of this architecture warned the public
that within 10 years-- that is to say, by about 1970--
computers would be generally as smart as people.
And a lot of people made fun of them for that prediction.
But it was actually scientists attempting to be responsible.
Because they thought something, a quite serious
dislocation was coming along, and that people should know
that it was coming.
And so they felt it was their responsibility in that age of
scientific responsibility to warn the public.
It didn't turn out that way, because the problem of
collecting the differences and finding the operators, that's
outside the scope of the architecture.
So this is the problem that has to be solved by a human
before this architecture can be used.
You have to have identified the differences that you might
encounter and the operators that you might use, and build
this table which relates the two together.
So maybe that one, that one, some off-diagonal
elements, and so on.
But building that table turned out to be a hard job.
So not surprisingly, the idea evolved.
And eventually the folks at Carnegie who developed the
general problem solver--
most notably Newell and his students--
developed a newer, fresher, more elaborate architecture
called SOAR.
And here's how SOAR works.
First of all, what does SOAR mean?
It doesn't mean anything.
It used to mean State Operator And Result.
But for some reason the proponents of the SOAR
architecture decided they don't like that acronym, and
have asserted that SOAR is merely a label that shouldn't
be thought of as an acronym.
In any event, SOAR consists of various parts.
It has a long-term memory.
It has a short-term memory.
And it has connections to the outside world, maybe a vision
system and an action system.
But most of the activity of the SOAR problem-solving
architecture takes place in a short-term memory.
So you can view the contents of the long-term memory as
shuttling in and out of short-term memory.
So you can see right away that this mechanism, this
architecture, is heavily influenced by certain
cognitive psychology experiments having to do with
how much you can hold in your short-term memory--
nonsense syllables and all that sort of thing that was
popular back in those days.
So this was an architecture devised primarily by
psychologists.
And it had amongst its features a short-term memory
and a long-term memory.
So that's part 1 of this architecture.
So what's in the long-term memory?
Well, assertions and rules, AKA productions.
A production being the Carnegie vernacular for rule.
It's just the rule-based stuff like you saw on almost the
first day of class.
So the whole thing is a gigantic rule-based system
with assertions and rules the shuttle back and forth from
long-term memory into short-term memory where
processing takes place.
The third thing that comes to mind when you think of SOAR
architecture is they had an elaborate preference system.
You recall that when we talked about rule-based systems
there's always a question of what do you do when more than
one rule would work?
You have to have some way of breaking those ties.
The SOAR architecture has an elaborate
subsystem for doing that.
But I said that these are the first three things you think,
and maybe that's not right.
Because the next thing you think about is perhaps a
better thing to identify with the SOAR architecture.
And that's the idea of problem spaces.
And that's the idea that if you're going to solve a
problem you have to develop a space and do a search through
that space.
Just like we did when we talked about how we can get
from here to home.
There's a space of places, that's our problem space.
We can do a search through that space to find a way to
get from one place to another.
That's the sort of thing that SOAR is focused on.
Finally, the fifth element that you tend to think about
when you think about SOAR is the idea of universal
subgoaling.
And that's the idea that whenever you can't think of
want to do next, that becomes your next problem that
deserves it's own problem space and its own set of
differences and operators, and rules and assertions.
So you start off on a high level, then you have to solve
problems at a lower level, just like you did up there
with a general problem solver.
So if you have these two architectures you can begin to
say, well, what are they centered on?
And this architecture, this general problem solver, is
centered on the idea that everything is
about problem solving.
Because the problem solving hypothesis--
no one gave it that name.
But that's what it was.
And this architecture did get its name.
And it was said always, by Newell, to be based on what he
called the symbol system hypothesis.
The hypothesis that what we are as humans is symbol
manipulators.
And we can uncover how that all works by giving people
crypto-arithmetic problems and having them talk out loud, by
thinking about what happens when you try to remember
nonsense syllables, by all that sort of stuff that was en
vogue in terms of psychology experiments in the day when
this architecture was first articulated.
But when you look at architectures you can sort of
see where they come from and what their antecedents are.
It has a short-term memory and a long-term memory, because
Newell and his associates were cognitive scientists.
It has assertions and rules and preferences, because
Newell and his associates were also AI people.
And it has problem spaces and universal subgoaling because
those are ideas that had been work out in a more primitive
form already in the general problem-solver architecture.
So that's a glimpse of what SOAR looked like
in its early days.
It's been very highly developed by a
lot of smart people.
So although it's symbol centered, they've attached to
it things having to do with emotion and perception, but
generally with the view that the first thing to do when
faced with this perception is to get it out of there and get
it into a symbolic form.
That's sort of the bias that the architecture comes with.
So those are two architectures that are heavily biased toward
thinking that the important part of what we
do is problem solving.
But the most important, perhaps--
at least from an MIT perspective--
of these problem-solving oriented ways of thinking
about the world, is Marvin Minsky's architecture, which
he articulates in his book "The Emotion Machine."
And Marvin is not just concerned with problem
solving, but also with how problem solving
might come in layers.
So let me show you an example of the sort of problem what
motivates some of Marvin's thinking.
So you can read that, it's a short little vignette.
You have no trouble understanding it, right?
No.
It's not difficult for us humans.
Awfully tough for a computer.
In part, because the thinking you need, your ability to
understand that story, requires you to think on many
levels at the same time.
First of all, there's a sort of, at the
bottom, instinctive reaction.
You see where there's instinctive reaction?
That's the part where she hears a sound
and terns her head.
That's instinct, right?
That's practically built in.
But then what she sees is a car.
So that's something that we don't have wired in.
It would be unlikely that we've evolved in the last 100
years to have an instinctive appreciation of cars
barrelling down the road.
So the next level in Marvin's
architecture is learned reaction.
So that's the part about thinking about the car.
Now spread throughout there-- well, let's see where is a
particularly good example.
She decides to sprint across the road.
So that's where she's solving a problem.
So that's the deliberative thinking level.
It doesn't stop there, because later on she reflects on her
impulsive decision.
So she thinks not only about stuff that's happening out
there in the world, but she also thinks about stuff that's
going on in here.
So that's a level which we can call reflective thinking.
Well, you know, it doesn't stop there, because she also
considers, in another part of the story, something about
being uneasy about arriving late.
So she's not only just thinking about events that are
going on in her mind right now, but events that are going
on right now relative to plans she's made.
Some Marvin calls that the self-reflecting layer.
But that isn't the whole thing either, because toward the end
of the story she starts to worry about what her friends
would think of her.
So there's a kind of reflective thinking in a more
social context.
So he calls that self-conscious thinking.
So as the Carnegie folks think, the SOAR architecture
focuses mostly on problem solving, Minsky's "Emotion
Machine" book considers not just thinking, but thinking on
many layers.
And the blocker to doing any of that can be said to be the
development of common sense, which computers, alas, have
never had much of.
So this could be said to be based on the common sense
hypothesis.
And the common sense hypothesis holds that in order
to do all of that stuff, you have to have
common sense like people.
And if you have to have common sense like people, you have to
think about how much of that is there and how
can we go get it?
And so this spawned a lot of activity in the media lab
amongst people influenced by Marvin, having to do with
gathering common sense.
The open mind project, the work of Henry Lieberman and
others, having to do with the gathering of common sense from
the world wide web as a way of populating systems that would
lay the foundation for doing this kind of layered thinking.
So that is a brief survey of some mechanisms, some older
than others, but all but GPS--
GPS too.
Let's face it, it's hard to think of solving any problem
without means-ends analysis being involved.
So GPS isn't wrong, it's just not the only tool you need to
think about what to do.
So these are early, and late, and still-current.
But it's not the only thing there is, because there have
been reactions against this problem-solving way of
thinking about the development of intelligence.
And the most prominent of those counter currents, of
those alternative ideas, belongs to Rod Brooks and his
subsumption architecture.
So along about the early--
along about the years surrounding 1990, Brooks
became upset--
subsumption--
because robots couldn't do much.
They would turn them on at night, and then the next
morning they'd come in the laboratory and they would have
moved 25 feet, nicely avoiding a table perhaps.
Not doing very much and taking a long time to do it.
So he had decided that it's because people were thinking
in the wrong way.
In those days people thought that the way you build a robot
is you build a vision system, and then you build a reasoning
system, and then you build an action system.
And it can do almost nothing, but it does something.
So you improve the vision system, and improve the
reasoning system, and improve the actual system.
And now you've broken it, because all the stuff you used
to be able to do doesn't work anymore.
So what's the alternative?
Well, the alternative, as articulated by Brooks, is to
turn this idea on its side.
So instead of having an encapsulated vision system, an
encapsulated reasoning system, and an encapsulated action
system, what you have is layers that are focused not on
the sensing and the reasoning and the action, but layers
that are specialized to dealing with the world.
So in Brook's way of thinking about things, at the lowest
level you might have a system that's capable of--
well, before we get to that, avoiding objects.
And maybe the next level up is the wandering layer.
And maybe the next level up after that is explore.
And maybe the next level up after that is seek.
Now in the old days when people took 6001 I had no
trouble getting an answer the question, what does this
remind you of in 6001?
It doesn't remind you of anything in 6001 since you
haven't taken it.
But it viewed, as a generalization of a
programming idea, what is the programming idea?
There are only a few powerful ideas in program, and this is
a generalization of one of them.
What is it?
Do you have a name?
Yes, Andrew?
STUDENT: Layers of abstraction?
PROFESSOR: Layers of abstraction,
and abstraction barriers.
That nails it pretty well.
Because each of these guys can have its own vision, action,
and reasoning system.
And if you think of these as abstraction boundaries, then
when you got this thing working you don't screw with
it anymore.
You build this layer on top.
And it may reach down in here from time to time, but it
doesn't fundamentally change it.
Brooks was inspired in part by the way our brains are
constructed.
All that old stuff that we share with pigs is down in
there deep, and we put the neocortex over it.
So it looks layered in a way that would make
[? Gerry Sussman ?] proud.
So this then is the way that Brooks looks at the world, and
it's characterized by a few features just like SOAR is.
One of those features is no representation.
So this is a detail that's probably right at the level
that Brooks was operating, and very questionable when you get
above the level that Brooks was operating.
But before I go on, let me say what the hypothesis is.
The hypothesis is the creature hypothesis.
It's the hypothesis that once you can get a machine to act
as smart as an insect, then the rest will be easy.
Well, how do you get a creature to
be smart as an insect?
Maybe you don't need representation.
We focused on representation in this course, so you can see
there's a little stress---
Next thing is, what do you do if you don't have a
representation?
Let's see.
Your representation makes a model possible.
Models make it possible to predict, to understand, to
explain, and to control.
So if you don't have one what can you possibly do?
Brooks' answer is, you use the world instead of a model.
So everything you do is reactive.
You don't have anything in your head that is
a map of this room.
But maybe I don't need one because I can get around that
table by const
…