CS372: Artificial Intelligence (Fall 1998)
Week 4: Responses
Readings
- The Robots: Dogged, Insecure, and Driven
- Chapter 5, from Nilsson's text.
- Multitasking in Interactive C
- Chapter 1 from Moravec's Mind Children
Diana Applegate
After reading chapter 1 of Moravec's Mind Children, I was very surprised
at how optimistic it was. Especially of interest to me was Moravec's discussion
of a general-purpose, household robot. In his opinion, this creation (perhaps
similar to Rosie from the Jetson's) would appear "in time for the new
millenium". As we know, robots are still extremely specialized...I
don't think we'll have a Rosie anytime soon. In any scientific field of
research, it seems that as more progress is made, it becomes very difficult
to keep up with a fast-paced rate of discovery. Advances in AI have definitely
slowed down within the past ten years, but understandably so. So much progress
was made within a few short decades Moravec's book did a great job at catching
me up to speed about what has been happening in AI and robotics. His discourse
was very readable and wasn't at all boring. Before I end this discussion,
I'd also like to comment on Moravec's ideas about "bottom-up"
AI. I agree with him in that this seems to be the only way to go. If we
start from ground zero with perception and simple sensory behaviors, it
seems that there's a good chance that more complex behaviors will emerge
along the way.
This week, I was finally able to overcome my uneasiness about the math involved
with neural nets. In doing so, I was able to appreciate what they were all
about. I especially enjoyed the computer demos, and am looking forward to
running the bp program myself. I'm hoping to do a little experimentation
to figure out what significance the number of middle layers has on the accuracy
and timing associated with the output.
The lab, as always, was fun. I'm enjoying the programming aspects more than
the lego design aspects, but both are manageable tasks. All of my friends
are envious of the fact that I'm fufilling a lab science requirement this
way...they seem to think that robotics is so mysterious and complex, even
on the lego level! I guess this stems from all of those crazy science fiction
movies and novels. It seems that most people take for granted that AI is
a part of our daily lives.
Jocelyn Arcari
Reading the first chapter of Mind Children by Moravec gave a nice over-view
of how AI has developed and touched on some examples we had not covered
(like electronic turtles from the 50's) as well as things we mentioned in
class (Stanford's Shakey-- a robot created from reasoning programs). It
was amazing to think that people in the 1960's felt that within a decade,
we would have produced a genuinely intelligent machine. As Moravec mentions,
"This was an understandable miscalculation." Many of the comments
made by Moravec about the relationship to the survival of mutations in organisms
versus the successes in programs for robots in the real world point to the
fact that it will take Many, Many years for things to develop. If it took
billions of years for humans to evolve to thestate we are now in, it will
take a good number of years before we understand all the intricacies involved
in building the most advanced robots. Moravec likened the nervous system
and complexity of behavior of insects to that of today's robots. It was
also interesting that Moravec expected a general-purpose robot to be "usable
in the home within ten years" of 1988. Is there anything today that
would be considered a "general-purpose robot" used in the home?
It will be interesting to see what another 10 years down the road brings
us. Elman nets seem to be a good way of training robots more quickly since
previously sensed data can be located quickly and used in the next round
of training.
Will we be discussing "Blackboard systems" anymore? I understand
the basic idea but might like a little more explanation. This reminded me
of the similar system of "decoding" which I think is used in CD's
when they are played. It is my understanding that scratches on the CD or
pieces of dust can be "overlooked" and the gaps filled with a
1 or a 0 by the computer... So, the computer uses the information it has
nearby the string of 1's and 0's and fills the gap with what it "thinks"
is the appropriate value. Am I on the right track? Finally, a general question
about whether you will discuss some more with us regarding our first exam
which is approaching rather rapidly?? Thanks.
David Costello
After reading the chapter ìMind in Motionî I have a few questions.
It seems to me that the difference between cybernetics and the field of
artificial intelligence is that the former uses analog computers while the
latter uses digital computers. Is that the only difference? Also, do you
think that quantum theory will one day be used in artificial intelligence?
Another question I have has to do with the authorís statement that
within 50 years there will be machines with human intelligence. I read elsewhere
that such goals of high level artificial intelligence are impossible. Which
one do you think is correct? My last question has to do with artificial
emotions. The author says that a robot can be programmed to feel. Isnít
there a difference, however, between simulating emotions and feeling emotions?
Can a robot actually experience happiness or sadness or are these machines
simply mimicking emotional states?
Sonia Dubielzig
While reading Moravec's Mind Children, I started to wonder what makes human
thought special. The complex thoughts that my mind was processing as I wondered--are
these unique? And if other animals experience the intelligent behavior of
wondering, where does it stop? Are mammals the only animals capable of the
conscious, silent articulation of ideas that we call "thinking",
or can birds be included in the list? And what about the octupus, which,
as Moravec mentioned, has a highly evolved nervous system of its own. It
seems that bees can communicate, argue, weigh the advantages and disadvantages,
and consense on a new place to move a hive. However, the awareness of "self"
that bees experience MUST be completely different from the "self"
we experience. Can we really refer to both of them as "self-awareness"?
And if so, then what about the "self" that a robot experiences
when it is able to place its location on a mental map--no, I mean data representation--of
the world around it? I am intrigued by the proposals Moravec mentions for
ways to program learning by association of "pain" and "pleasure";
or program against repetitiveness by making it "painful". Perhaps
this is only how our brains work, and we are simply making associations--albeit
on a much more complex scale--we think we are "bored" or "interested".
It seems like a very mechanistic way of thinking about thinking. Somehow
there must be more to our thoughts than simple associations combining to
create a whole entity even with "emergent behavior". However,
the only other explanation seems unscientific. Is the existence of a "soul",
an intellect separate from the humdrum of bodily functions, really out of
the question?
Ben Flynn
Well, right off the bat I'd like to say that I learned a new word right
at the beginning of Chapter 4 -- instantiation. It was not in my American
Heritage Dictionary, but was on Merriam Webstar's Online Dictionary. It's
the noun form of instantiate which means "to represent (an abstraction)
by a concrete instance <heroes instantiate ideals -- W. J. Bennett>"
It will be celebrating its 50th birthday next year. (Webster's Dictionary
is at: http://www.m-w.com).
In general, nothing too earth shattering has come up. The robots are a bit
finicky, but that's not a big problem. If I were to be concerned about anything,
it would be what's going to be on the exam that seems to be fast approaching.
How deeply should I be concerned with memorizing and applying the fuctions
we have learned? Could you give us a sample problem or two to get us oriented?
That would be great.
I have to say that found the business in the book on "genetic programming"
quite interesting, and was disappointed that it did not go into greater
detail.
Emily Greenfest
I'm not sure what to say this week -- I've a million thoughts running, through
my head right now, and unfortunately, very few have to do with AI, save
that, again, so far all is very familiar and I feel as I have discussed
much of it to death in previous classes, or at least, that I have discussed
much of the issues brought up in Mind Children (the lecture material is
new, but I am never sure what to comment on lecture material svae that it
was interesting and I will do my best to recall it for the exam and maybe
for some weird usage 10 years from now) so often the most familiar arguments
now come second hand.
That is not to say that I did not enjoy the book. In fact, I found it to
be a fun and easy read. As well as an interesting one. I bring up the point
that the discussion in Mind Children brings back memories from the last
two semesters in that while much of what we study today seems "new
and exciting" it was in fact discussed, discovered, or at least consider
way back when fifty odd years ago at the advent of the science. Mind Children
was written in 1988 and it is now ten years later, but our outlook on the
future (i.e. the ideal basic robot and its construction/behavior) has changed
little. In some ways, I suppose, you could say that we farther along the
path to devloping that ideal AI being, but from the other direction you
could say that we have merely perfected what we had to begin with -- made
horizontal but no vertical change along the "evolutionary pathway"
of the robot. Today's industrial robots have enormously sesitive gripping
hands, the Mars Rover travels (fairly well) across the uncertain terrain
of the Martian terrestrial landscape, and advances in programming techniques/technology
have allowed for the better represetnation, interpretation, and reception
of sensory input than described by the author. But the missing link is still
maybe missing (there still is the question of: do we have it and just can't
recognize it?) -- the robots just don't seem to be "intelligent,"
although they have an increasingly greater semblence of being so. And the
arguments for or against the observation of intelligence in robots are interestingly
also the same (issues discussed in Mind Children are the same issues argued
continuously in classes I've had here at Bryn Mawr -- if an observer considers
a robot to be behaving intelligently than is it, regardless of the programming
or the view of the creator?).
I do not mean to say that the thoughts and ideas discussed in this (and
other) classroom(s) are outdated, just that perhaps, in considering the
"future" described in Mind Childern, we must also consider that
today is not yet that future.
Ayishih Hakim
Mind Children by Hans Moravec is an interesting read. I like the personal
touch he adds to the study of AI. First thing i noticed in his analysis
of cybernetics, AI and robotics is his prediction that in 50 years we will
be able to design computers that would have minds resembling those of humans.
This book copywrited in 1988 is making this prediction for the year 2038.
To me this doesnt seem like that unreasonable of a thought. However,comparing
Moravec's notes to what we have been studying so far in class, it seems
as if computer science hasnt really made many strides from then, 1988 to
now, 1998.
Some questions that generated from the reading is what is the difference
between cybernetics and AI? I recall in an earlier class we said that cybernetics
was an early term for AI but Moravec seems to be making a distinction between
the two. Also I didnt understand Moravec's definition of capacitors. I automatically
associated capacitors with our state machines, is this a correct connection.
I made this connection because a capacitor's purpose is to intergrate memory
into the learning process.
Also what proved very interesting to me was Moravec's reasoning for the
dichotomy between computer capabilities and human capabilities. Where computers
can exhibit adult level and beyond performance on intellegence tests but
they can not do what a child can when it comes to perception and mobility.
Moravec accounts this phenomenon to the fact that a human's perception and
mobility skills have had billions of years to perfect itself (due to evolution)
but human reasoning skills are a relatively new concept that hasnt fine
tuned itself yet. Therefore he is concluding that in a million or so years
human minds would develop itself to a much higher level. Taking this a step
farther, I would say that we the children are smarter than our parents.
When I say smarter-I mean in ability and not in knowledge.
Im not sure I believe this, but then Im not a believer in evolution either.
Ada Hogan
Question: while we do a gradient descent, we try to reach the "ideal
weight vector". Is this i.w.v. simply the value of 0 or 1 ? Also, is
the momentum term remembered from one pass to another? Or could it change
so randomly (we talked about an uneven 'surface') that this isn' t necessary?
Is the learning constant "_" stored after each set of inputs,
or does it have to be calculated each time? Does "state machines"
include all machines that store and re-supply information about a previous
action? So Recurrent nets and and Ellman nets simply fall under the title
of "state machines" because they have a memory?
I thought that some interesting points on how to model a machine were brought
up in "Mind Children" . Instead of starting with a specific "advanced"
behavior, such as picking up blocks and moving them around, it seems to
make sense to start at a more basic level; instead of aspiring to a perfect
of copy of the human brain/ or nervous system, perhaps we should concentrate
on modeling simple animal nervous systems. I understand that the evolution
of animal minds (and intelligence) would be useful as we try to "build"
intelligent behavior, but it seems that biologists' research on the evolution
of animal nervous systems would also provide a basis, physically, for some
physiological process that led to progress in "intelligence".
And as for "mobility linked to intelligence"... doesn't a robot
need to be told what a tree is, what a dining center is, what Founders Hall
is, for it to act appropriately? When he spoke of a need for 'World Knowledge",
i took that there would have to be basic information supplied by a program.
Otherwise, a robot might be able to "see" something, but how will
it understand the ways in which this object influences its environment?
The machine would have to have a very high learning ability and power of
observation. It machine might eventually collect all the information about
its environment if it were in a relatively small, fairly static area. But
how could it learn in a chaotic environment, where the conditions change
constantly, such as a household? I understand, at least, why mobility in
this case would be linked to intelligence, or at least to learning.
Peter Ingebretson
This week was interesting, however I wish we had not skipped over genetic
algorithms so quickly. I have looked at some interesting software making
pretty good use of them, and I would like to learn more about strategies
of implementation that people have used.
I've played around a bit with the neural net you made avaliable over the
network, but only insofar as trying new patterns with a net trained on testseq
and getting the net trained on xor by setting the momentum and learning
constants lower.
I'm looking forward to a programming assignment in any of the topics we've
covered, either in or out of the lab. I think it would be particularly interesting
to see a neural net learning as the net is reflected in the actions of a
robot.
Sarah Klaum
I found the first chapter of Hans Moravec's _Mind Children_ an interesting
read. While the history he gave and the "breakeven" concepts were
informative (if dated), what I enjoyed the most was his speculation on the
convergent evolution of emotions and consciousness. His "thought experiment"
involving the COUNT-DOORS module did indeed stimulate a great deal of thought.
With excitement I envisioned some of the rudimentary processing and planning
that would go into programming such a complex behavior, and I once again
realized that it is this process--the planning and programming of behaviors
that most interests me. I was facinated with his ideas for conditioning
software, and how easily he was able to extend certain concepts into scenarios
where his robot will be able to imagine, dream, model another's mental state,
or possibly even develop criminal behavior. I was amused by the images that
were conjured up when he described such robots completing household chores,
as such a reality seems so far removed from our present technology, in most
respects.
With regard to our recent lectures, I am particularly struck by the blackboard
method, and would be interested to explore the actual implementation of
such a system.
Maralee LaBarge
It's not hard to realize, not just image anymore, the idea sported in so
many sci-fi epics, having a robot around the house. I mean, the first thought
that comes to mind is the Jetson's smart aleck maid/chamberlain who ensured
that the family was on top of things which, as silly humans, they usually
weren't. The benevolent robot, the family protector, the appliance-replacer,
who wouldn't want one?
I don't know. I'm not as excited about "robotic progeny" as Hans
Moravec seems to be. Now I need to clarify because I love AI. There are
very few philosophies with so many practical applications and very few scientific
pursuits with such broad philosophies. I like it for what it offers the
human race both in product and in knowledge. What better way to find out
about ourselves than trying to emulate ourselves? But I still think we have
to be careful. Moravec makes it sound like synethetic human beings will
soon be far superior to biological ones and then it's out with us, as if
we're making our own natural selection and giving ourselves a bad fitness
rating. Of course, we can just gab about this now--robots so far are extremely
primitive relative to the complexity of even insects, let alone animals
or humans, but we're a smart race. We figure things out. It could take years--centuries
even. But something we persist at, we eventually do. And it's an equally
frightening and exhilarating thought.
If we really want robots to think, then can we expect them to be our servants--household
and otherwise? If we want robots to be our servants, wouldn't it be cruel
to teach them how to think? Well, maybe we'll say, we'll only teach them
so much and then they can serve us and they won't mind. Well, what if we
were talking about another human being? Using that kind of philosophy is
called "brain-washing" and it's a small minority of the world
that doesn't consider that to be unethical. If robots have the potential
to be like us, to have the same freedom of thought and freedom of expression
that we do, we can't hold them back. Maybe it's the creationist bit of all
this that gets to me. I'm a firm believer in God. He created earth; He created
us. He knew what He was doing and He did it with a purpose. Are we trying
to make our electrodes walk, talk, and think just for the sake of it? And
what happens to them then? Every sci-fi story of doomsday predictions about
a mechanical world overcoming humankind looms up pretty ugly next to the
cheerful demeanors of the Jetson's maid and Data from Star Trek:TNG.
Technology is wonderful. It shouldn't be held back. But there are some technological
powers that human beings--at least now, though maybe the future will be
different--are ill-equiped to handle. I think we have to be careful with
our toy guns; they could go off.
Frank Rusch
In chapter 1 of Mind Children, the author described a rift in the development
of intelligently-behaving machines: There are machines that attempt to reason,
and those that address more primitive problems, that is, they start with
the basic dilemmas that early life had. The author seems to favor the latter,
proposing that starting from the bottom would enable the machines to evolve
just as living things have. The author makes an interesting argument when
he juxtaposes a chess-playing computer with a robot that could accurately
simulate a one-year-old child. Which of these would be considered a greater
achievement? Certainly, the chess-playing playing computer could beat world
champions, not to mention the robot one-year-old. But evan at such a tender
age, the robot child is capable of doing immensely impressive, if not readily
obvious, things. This child has already begun to recognize patterns in a
world infinitely more complex than the 64 squares of a chessboard.
The author says that once prices start to come
down and demand increases, robots will become more prevalent around the
house, to help with chores, etc. I'm not sure that the subsequent increased
revenue will generate more intelligent robots-- the development will likely
follow patterns of demand, even fashion. For example, fuel-efficiency in
cars has not made much progress because it is not really demanded by the
public. It seems likely that, once robot production (and competition) increases,
manufacturers will be forced to make robots that are precisely programmed
to do specific tasks people demand, rather than develop general knowledge/intelligence.
I just want to say that I really like the "start_process" thing
we're doing in the lab. It makes the programming environment seem more realistic,
closer to modeling a real brain. Being able to store continually-monitored
information in global variables is like preserving observations in one's
consciousness-- the information is always there, though it need not be examined
unless it's necessary for the task at hand.
Edina Sarajlic
In the first chapter of Mind Children Moravec discusses characteristics
of a general purpose robot which would have a cost to utility ratio acceptable
for mass production. Most of the attributes he talks about touch on the
concepts we covered in class, such as stimulus response, learning, subsumption
architecture etc. However, Moravec uses a more practical, bottom-up approach
to artificial intelligence, so he also focuses on some practical issues
we have not yet discussed, such as machine mobility. He points to the imbalance
between the level of development of reasoning and sensorimotor skills in
robots, claiming that the latter ones have not been given enough attention
by the research teams.
I do not believe that the only way to achieve intelligence in machines is
through duplicating the evolution of human (or any other animals) physiology.
However, I agree with Moravec on the importance of motor and sensory proficiency
for the emergence of intelligent behavior. He reminds us that the mobile
organisms alone exhibit the mental characteristics associated with intelligence,
and that the basis of our intelligence developed in the interaction of our
motor and sensory facilities with the complex environment. This comes back
to the problem already mentioned in class: how does one model human thinking
if only a part of it is in the realm of conscious and describable symbolic
reasoning.
Towards the end of the chapter, Moravec talks about the possibility of a
world simulator, the program which would model complex hypothetical situations
in robots environment. If the simulated events were to be incorporated into
conditioned learning, the robot would be able to train itself for many real-world
situations. Moravec speculates that a far consequence of this built-in simulator
would be the emergence of capabilities similar to human intuition or instinct.
Regardless of the predictions for the future, I find this concept of internalized
learning very interesting. I wonder if such simulators are actually included
in the software of contemporary robots, as Moravec predicted it would. If
they are, is such method of learning more advantageous than standard conditioning?
Jim Speer
This past week in class I found myself particularly interested in blackboard
systems. I can picture this system as a method for intelligence implementation
easily fitting into an artificial construct. The lab assignment helped me
this past week, since they required the use of a small blackboard system
themselves. After last week's confusing presentation on a number of topics,
I was encouraged to see that the robot lab hands-on assignments would be
integrated into the more abstractly presented topics from the book. The
use of global variables directly corresponds to the housing of blackboard
system information, and the functions that modify those variables are like
little knowledge systems.
In general, I found the introduction to the concept of state machines a
step towards a better understanding of how AI relates to intelligence in
humans. Whereas the simulation of insect-like behavior is impressive and
recognizable, it is only so with respect to an insect's stimulus-response
dominated behavior. A human's learning and memory system would seem to be
unable to be mimicked by any S-R system, no matter how complex. The introduction
of an internal state to a system gives that system's behavior a context
for it's evaluation of it's next actions. The distinction is to be made
between "reactions" and "judgments."
Still, the robots we programmed this week can't really be credit for making
decisions. We were the one's who instructed the robot to turn in a certain
direction when it detected certain things about it's environment. At best,
the robot is accurately carrying out our own decisions, or the decisions
we would like to be made under the detected circumstances. We've transferred
our will into an otherwise will-less being. This seems to me more like proxy-
intelligence than a distinct intelligence of an individual. I think that
AI enthusiasts secretly wish for robots to one day do things in an unpredictable
manner, so that the humans can be confused and impressed by the illusion
of an independent individual.
Ben Sprecher
This week's reading has been especially exciting. I greatly enjoyed Moravec's
synopsis of the history of robotics and AI to date, and I think he has many
good points in his discussion of the general purpose robot. I think his
predictions about the time frame are not accurate, however. I forsee the
first such robots appearing by the middle to end of the next decade, not
of this one. Also, technologies have been advancing at different paces.
Speech recognition is advancing in leaps and bounds - appearently Dragon
Systems' Naturally Speaking can recognize spoken words faster than the fastest
typist can type them, even accounting for errors. Once that recognition
can be coupled with some kind of symbolic understanding, we can actually
tell computers what we want done, and they will understand it as a tast,
not as a series of steps, and they can develop the steps themselves. Even
without the understanding layer, I am convinced that at least command-level
speach recognition will be integral to any general purpose robot. No one
will buy a complicated appliance that they have to type into to make it
do desired tasks.
The idea of a robotic general-purpose home aid is actually quite amazing.
If you think about it, all of the labor-saving devices in a typical home
are build for exactly one purpose: the dishwasher only washes dishes, the
vacuum cleaner only vacuums, the blender only slices and chops and dices.
A KitchenAid mixer is especially great because you can change heads on it
to allow it to mix, stir, blend, whip, chop, make pasta, and even knead
dough. And that's by far the most versatile home appliannce I can think
of. The true power of a robot is that it can use the afformentioned tools.
A device that can make use of labor saving devices has so much more leverage
than a single-purpose appliance - it can make use of the decades of refinement
that went into making very good dishwashers and driers and ovens and blenders
and it can operate them efficiently. That is why the general purpose robot
is destined to find its way into the home - it fills the gap between the
plethora of everyday tasks of the household and the many appliances which
handle each task the best.
I anxiosly await my own little R2D2.
Emily Sweeney-Samuelson
The prologue and first chapter of <U>Mind Children</U> was interesting.
I wonder how close we are to fulfilling some of Moravec's predictions?
I mean, obviously his time estimations are much too early. But,
how far along is the field in his predicted sequence of events? Is
AI/cybernetics research going as he predicted it would, making progress
in the same areas? It was fun to read about his expectations for
the structure and functions of the general-purpose home robot. I
wish we did have them now, or at least that we were closer to having them.
I am also curious to see if he goes any further into the negative
possibilities of this future race of robots. He seems optimistic
about the results, only mentioning a few negative possibilities, but he
hasn't speculated very much about the behavior of robots when they are almost
independent of humans, when our species will not be needed. I
suppose he feels uncertain about predicting things that far in the future,
in "a postbiological world," but he seems very free and
confident with his other ambitious predictions. I am curious about
what he addresses in the rest of the book.
As for Nilsson and the lectures, I have several ways to represent and interpret
an agent's environment floating around in my head, but I am not sure of
their individual importance or frequency of use. I understand
the general concepts and applications of state machines, but I know different
amounts about the specific methods, and I'm not sure how much I should know
about each.
Tim Waring
i read the chapter on machine evolution anyway. The idea behind machine
evolution carrying the AI effort to new realizations, and new hieghts is
a good one. It strikes the same chord as the idea of implementing neural
networks. The similarity being that the best approach is the natural approach,
the one followed by nature. I think that makes sense because it is the only
way we know that is has worked. The chapter on state machines struck me
a little differently. The IDEA of having and internal state, even a dynamic
and modifyable state, anything from a map, any iconic representation, even
the blackboard architecture, seems that there are rules involved in how
the state is arrived at, altered and how it affects the AI. These rules
cannot exist in a conscious being who is subject only to thier own whim.
I can change the rules that i use to decide what affects me, any state machine
will not be able to do that. I also read the Moravec chapter, Mind in Motion.
He paints what seems to be a realistic picture of a general purpose robot
existing in about ten years. it does seem as if each individual piece, locomotion,
vision, rationality, speach recognition (brand new developments here make
me excited, (dragon software)), and leanring. but that doesn't mean that
they are all on the verge of being integrated. He admitted that a general
purpose robot will need to have the computational power of a supercomputer.
Whereas I do agree that once the market on home robots gets going it will
accelerate, i don't believe we will be starting very soon, or with such
a bang. simply because of the amonunt of intergation from different collaborators
needed to produce one highly expensive prototype.
i designed the robot for lab as well, and in so doing and using the parallel
architecture i discovered the usefullness of "organismal programming"
(as I call it) My idea is to programm in an opbject oriented manner so that
a leg becomes a leg, and not just actuators 1, 13, and 42 and sensors 229,
49 and 385. If dealing with the leg as a leg that gives output and recieves
input and moves, it is more like an organismal leg, and is easier put to
organismal tasks.
such is my hypothesis. i'm still trying to put that into code.
Sarah Waziruddin
This weeks reading contained a lot of information.
Nilsson introduced the idea of taking previous sensory information into
account when computing an action. This greatly improves the functionality
of the robot. Ellman nets are used to implement this idea. Ellman nets are
recurrent networks that learn how to compute the current feature vector
from the previous feature vectors and the current sensory information. Machines
that use this sort of network are referred to as state machines because
they take previous history into account.
Moravec is quite different from Nilsson and provides a different approach
to AI. I read the prologue of the book and this definitely emanated a post-modern
view. Advancement in AI seemed to lead to self-destruction. The prologue
much resembled the mood prevalent in science fiction movies from the 50s.
The first chapter had a different mood from the prologue. I really liked
the historical information contained in this chapter. Moravec stated that
what machines do well, we do badly and vice versa. This idea is quite appealing
to me because it suggests harmony vs. and us against them attitude. There
is no 50s paranoia in this statement. I also found it highly ironic that
Moravec suggested that as we develop advanced robots, we will have less
and less to do. I feel that we have to work harder to keep ourselves from
being outdated as these advanced machines are being created, so in reality,
all we are doing is increasing competition for ourselves. Its as if we are
in a race with ourselves.
Moravec also presented a few ideas as to why robots have not achieved a
certain level of intelligence yet. The reasoning ability of a robot depends
on what facts it is given. Therefore information such as the state of the
world has to be programmed into the robot, making it less robust and less
intelligent.
Moravec pointed out that perception of the environment and mobility are
both signs of intelligence. These two abilities prove to be the most difficult
for robots to exhibit. Moravec also made an interesting observation. He
noticed that people observing robots often attribute to them abilities which
they do not posses.
This weeks lab was also great as we learnt about concurrent programming
and what affects this has on the abilities of the robot.
Leslie Zavisca
This week's reading was very mild compared to last week's as it went very
quickly since it written mostly in laymen's terms. I enjoyed reading all
of the history that Moravec provided. At the beginning of this course I
realized how in the dark I have been when it comes to the history of AI.
I am constantly surprised at how early some of the major advances occurred.
Am I correct that his book was written ten years ago? I would be really
interested in his opinion of the last ten years and how many of his predictions
came true. For example, as far as I know, there is not a market for household
robots as he suspected there would be by now. The only thing that really
puzzled me in the first chapter had to do with the outline of the FETCH-CUP
program. Here are two of the lines of code: Step 16: If the robot is facing
a closed door, try to open it Step 17: If the door fails to open, say "knock
knock" and go to Step 16 It appears to me that the robot is just going
to get stuck in these two steps if the door does not open. The series of
Steps 6 through 9 seems to have this same problem. Not that this is that
important, I just wanted to make sure that I didn't miss anything major
in this section of reading. I was interested in the DEAL-WITH-CLIFF function
as well. Is it's priority determined by something similar to a TLU and How
is the cliff probability variable determined? Moravec also wrote about his
version of "a sensible robot". I'd like to know what similar robots
have been created since his idea and how successfully they performed.
The subjects in lecture most interesting to me were iconic representation
and ellman networks. Although I know we've only scratched the surface, these
two elements seem to have the most potential for more complex behaviors.
This week's lab assignments were fun to complete, as well as a bit frustrating.
Once we had the first function written, the second two followed quickly
and I am getting used to interactive c by now. The hard part was rebuilding
sections of our robot to accommodate these functions. I have to admit that
I am not looking forward to doing that every week.
Back
to CS372 Course Materials