CS372: Artificial Intelligence (Fall 1998)
Week 2: Responses
Readings
- Interactive C Manual
- Chapter 2, from Nilsson's text.
Diana Applegate
What I find most intriguing about S-R agents is that they can exhibit seemingly
complex, human-like, intelligent behaviors with minimal programming and/or
knowledge of their environment. In fact, it is the complexity of the agent's
environment that often causes these behaviors to "emerge". I see
a definite parallel between S-R agents and human beings -- not to say that
our "cpu", the brain, is at all simplistic. Of course, it's quite
the contrary. But assuming a materialistic theory of mind, where the brain
is analogous to a computer, we aren't "programmed" to know how
to respond to every possible environment/situation that we might encounter
in our lifetime. Instead, we are granted a number of general rules. And
these general rules, along with their corresponding actions can get us through,
even in unfamiliar territory. If these agents were, in fact, programmed
to run only within a specific environment, they obviously wouldn't be as
versatile or as interesting.
Nilsson's discussion of T-R programs and animal behavior (namely phototropism)
reinforces the fact that S-R agents, albeit simple in nature, can nevertheless
explain a lot about the way that animals respond to certain stimuli. And,
to me, this is a very important aspect of AI. Not only are we trying to
create artificially intelligent systems. In doing so, we are constantly
learning more about the ins and outs of our own systems.
Jocelyn Arcari
Including the mini-robot in our discussion of Reactive Agents was a good
way to see this agent in action. Incorporating an actual robot into the
discussion where we drew a graph with obstacles and discussed how the robot
would follow the walls helped us see one of the possibilities for a very
simple agent. The robot made me think of people's pets. I could just as
easily see someone buying the "cute little robot" which avoided
walls on the table in our CS class as I could see them buy a gerbil. It
was actually quite entertaining to watch it move about and this was a very
simple example. A more complex example with goals would seem even more life-like
and to me, suggests that it would, at the very least, replace a child's
toy (if not an actual pet). With this simple example, though, I see the
possibilities of 'intelligent' computers.
Another interesting point came in the discussion of long and short-term
memory. I was surprised and skeptical (at first) to hear that humans only
store about 7 or 8 things in their short-term memory at one time. The example
of trying to remember a chain of numbers and not being able to repeat them
after a certain amount was a good example to show this. The rules which
we live and act by, however, are stored in our long term memory. These things
are reminiscent of ROM and RAM in a computer.
The stimulus-response agents reminded me a bit of something that was taught
at basketball camp - the idea that basketball is a game of "action
and reaction." The game is composed of players each doing certain things
and it requires a good deal of practice to learn to Anticipate these actions
so the player is not simply reacting, but is acting before their opponent
catches them off-guard. I imagine that in the 'learing processes' of computers
they could be programmed to act before something gets in their way - to
sense it is coming before it does. Is this true? Are there examples?
Finally, I think "subsumption architecture" is very close to how
humans make choices. The higher level of behavior subsumes lower behavior
when we make decisions A human would probably weigh all the factors of his/her
environment before making choices, but if the environment one is in is very
simple, the factors that need to be weighed would be fewer and one's behavior
would be more easily determined. The more complex the environment, the more
the upper-level action would override the lower levels.
David Costello
The second chapter of the textbook and the last two classes bought up many
questions which I hope you can answer in the coming weeks. I, for one, did
not understand why we can't achieve true artificial intelligence if we are
now able to model the chemistry of the brain in computer science. Shouldn't
the ability to make artificial neural
networks render the problem of intelligent machines solved? If we can simulate
the parts of the brain why can't we simulate the brain itself? Perhaps the
answer lies in the lack of knowledge of what a neuron's threshold actually
is. Or perhaps the sum of the parts of the brain do not add up to true intelligence.
Whatever the case maybe, I hope this issue can be investigated more thoroughly
in future classes.
Sonia Dubielzig
I apologize for my late response. One of the advantages of AI machines is
that they, unlike humans, don't forget to do things once they've been programmed.
I may have many more capabilities than robots, but I won't ever have their
memory capacity.
Looking back on the reading and the class lecture last week, all I can think
of is the "x1x2" and "x1 +x4" figures scattered throughout
the textbook and covering the chalkboard. "This boolean stuff is a
language?" I asked myself. Obviously, it has syntax and grammar. There
is logical order to commands, too, just as there would be if you were giving
instructions in any language. And these cryptic x's and addition signs and
multiplication symbols CAN be translated into english, although it took
me some time before I could figure out how.
Because I am new to the language of boolean algebra, I don't know if I can
make a very detailed observation of its features. However, what struck me
is how efficient the language is. A long sentence like "move north
if all adjacent cells are empty or if the cells to the west are blocked
and the cells to the north are empty" can be shortened to a simple
two-term disjunctive normal form expression, which I can't write out properly
in e-mail. Once a programmer has mastered the language and developed the
ability to "think" in those x1 and x2 terms, it is very easy to
translate commands from english into boolean forms.
However, for those of us who still lack the ability to translate quickly,
the boolean system is frustrating, and it highlilghts the amazing skills
needed to form and understand human language. Why do I have to program 3
steps to say "run the motor for one second"? (start motor, do
nothing for one second, turn off motor). Every small detail has to be included
in the program, because computers can't simply function intuitively or assume
anything.
To understand human language, much depends on intuition. Words rarely have
obvious, direct definitions--language is much more subjective than simple
algebra, because of the feelings and associations made for each word. Perhaps
"intuitive thinking" could also be considered an "AI-complete"
task, because once a machine has intuition, it has the ability to make associations,
learn, and do things without requiring programming.
Ben Flynn
This week my thoughts are somewhat limited in their nature. This is largely
due to my leaving of the photocopy of chapter 2 at Camp Darkwater, the place
of our Chamber Singer retreat. A tragic irony indeed, that I was studious
enough to read a few pages while at the camp, only to leave the rest behind
with my return to Haverford. My thoughts, therefore, will focus on the two
previous lectures.
The material of last week seemed preparatory, a calm before the storm, if
you will. The boolean rules seem much as I left them in Theory of Computation.
I think they are rather fun. As far as agents, I thought a bit about what
sort of agents human intelligence is composed. Certainly we have reactive
agents, perceptions that lead directly to actions, such as when we jerk
our hand back when we touch something hot. We seem to have model-based agents
that allow us to follow instructions; if I were driving to your house I
could follow an iconic map, or a descriptive set of directions. We have
goal directed agents. Some of my goals are to do well and learn a lot in
AI. Perhaps these are the same goal. Perhaps if we took goals some logical
extreme we would find all goals are essentially equivalent. Perhaps not.
As far as our look at production systems and the wall following model, I
think some interesting points about intelligence are being made. We perceive
that the machine is following the wall. We do so by maintaining a history
of the machines previous movements. We have a concept of following a wall
and can perceive this concept being played out in another being. Yet we
know that the machine itself has no sense of history. It is composed entirely
of reactive agents, and follows the wall simply by creating a symbolic map
of its present surroundings and moving in a precedence directed way towards
the most available open space. The machine has no concept of the intelligent
task of wall following that it is performing, and has no way of conceiving
of it. We discussed in class that some of the actions our machines would
perform would seem intelligent to us, but not to observers outside the class.
Intelligence thus seems to have curious failings in its definition for an
object may be intelligent and not intelligent simultaneously.
The object, the machine following the wall appears intelligent because it
is correctly accomplishing the task of following the wall. Without the wall,
or with an unsuitable wall, the object would not appear intelligent. Since
the intelligence of the object depends directly on the presence or lack
thereof of the wall, we might say that the wall makes the machine intelligent.
This would support the theory that environment accounts for intelligence,
at least to some degree.
If the intelligence is a property bestowed onto the object by the presence
of the wall, the object is not intrinsicly intelligent. What the object
has is a potential to be intelligent, triggered on and off by its environment.
Perhaps this points to some other entity whose "on" state is intelligence
and whose "off" state is not intelligence.
On the other hand, we could argue that the machine is not intelligent on
the grounds that it is not aware of its intelligence. The machine cannot
see that it performing actions that are accomplishing a goal. If the potential
for an object that performs seemingly intelligent actions to not be intelligent
because it is achieving a goal it is not aware of exists could it not be
possible that humans are fundamentally not intelligent? After all if the
object is _incapable_ of perceiving the goal it is attaining, we too might
be achieving a goal that, no matter how hard we try, we cannot see. But
if we are not intelligent, what is intelligence?
Emily Greenfest
There are just a couple of things that struck my mind from this past week's
discussion/reading/activities. The first was a statement Nilsson made in
chapter 2 of the textbook to the effect that the separation of a stimulus-response
agent's processing into "perception" and "action" was
a purely arbitrary designation. It seems interesting that an action can
be described, as he indicated, in terms of perception -- i.e. as the observation
that such a thing may occur in the given environment -- and that, likewise,
a perception can be described as an action. This is intriguing because,
depending on your perspective, it either simplifies or complicates the modeling
process. In the one case, designating all of the robot's behaviors as a
phase of "action" (or "perception") may be seen as a
simplification of the process -- i.e. the agent performs one kind of task
instead of two. On the other hand, it reinforces how complicated that what
we call an action (or perception is) -- if the two terms are in some ways
interchangeable, than it is likely that 1) we don't really fully understand
what we are modeling (which in someways is very true) and 2) the models
are extreme simplifications that don't truly represent all that should occur
in a stimulus-response scenario (which is a fallacy but a necessary reality
of most models).
The other thing that struck me in class and about the reading is the duality
of the processes and procedure that we examined in this past week. It seems
to me that these algorithms are interesting because they both allow for
an agent to appear to be intelligent in the absence of what most observes
would actually call "intelligence" and they allow for the occurrence
of an "intelligent" process: learning. Networks, subsumption architecture,
and rule-based algorithms all allow the agent to perform behaviors that
appear intelligent by processing the agent's "perceptions" and
matching them (in one form or another -- specific rules, thresholds based
upon strength or number of inputs, or depth in a heirarchy) against a list
of possible actions. More importantly, they all facilitate learninig --
an agent could possibly be programmed to develop the set of rules, threshold
values, or heirarchical order (actually I'm not sure on this last one) necessary
to navigate a given situation. It is this potential, I believe that makes
these relative simple algorithms very interesting.
Ayishih Hakim
Monday's lab astonished me most. I was greatly intimidated by the fact that
it was expected of us to build a robot after 2 classes. Being this was my
first exposure to the creation of robots I was dumbfounded. Getting over
this initial reaction, that I undoubtedly felt other people shared, I immediatly
was intrigued. We are to build robots from Lego pieces! After hearing and
seeing the materials I assured my self that this could not be as hard as
it seems, since our materials are used by 6 year olds. Then half way through
the description of the Handiboard it was evident to me that we are not expected
to program and control a robot from 2 motors and lots of wires drawn through
the structure of the robot. Instead the Handiboard already provides a means
for us to control and program the robot.
Another shock that came to me was that not only are the building materials
made by Lego, but Lego also makes a sophisticated version of the Handiboard
that is currently being sold in toy stores!
Learning about the Handiboard, and the fact that 7th grade kids are constructing
robots at home has significantly diminished my anxiety. I am ready to take
this lab head on. :?)
Ada Hogan
I can see the advantage of using subsumption architecture to build intelligent
behavior by layers- but whatever the number of TLU's, how can a human programmer
forsee and evaluate all possible behavior? Unless the environment the machine
is in is completely closed, it seems that there could be an infinite number
of possible changes in that area. Even if a neural network is able to generalize,
to act on the information it processes from input vectors, I'm not sure
if that which seems like deductive 'reasoning' is really analagous to human
behavior. Take task specific algorithms for example: it seems terribly short
sighted in a way. It moves in little steps according to immediate surroundings,
if...then; if....then. What if the path is leading the machine to a huge
hole? It will probably avoid the hole, and reach the final destination,
but couldn't there have been a more efficient way? I find it hard to imagine
the number of sensory imputs that must be required, and how many different
ways the information could be interpreted before arriving at one final action.
Perhaps human actions are more guided by approximation of surroundings that
I think. ( i was thinking about productions systems, and acting on the most
specific condition for action). I would have thought that we judged each
situation on more of an individual basis, but I suppose that we also learn
from experience, which means that we are simply storing up information retrieving
it and reapplying it.... as we might program a computer.
Peter Ingebretson
Although stimulus response agents are very simple, this chapter was stimulating
for me. I particularly enjoyed seeing a system of rules for following walls
in a world broken down into a purely formal logic system. Of course, any
system even a few orders more complex would require an incomprehensible
number of logical prepositions, however I like to think that in principle
any system can be understood on this level.
I am also looking forward to looking more into neural networks, since they
seem to be the bridge between formal logic systems and dynamic systems.
Although I can look at any intelligence as stimulus response with additional
state variables, it is hard for me to see how one would go
about uncovering the hidden relations inherent in a complex system without
some 'organic' way of approaching it. I think, at least so far in what we
have learned, neural network-type systems are the way to go insofar as low
level artificial intelligence goes.
I would be interested in hearing more about how subsumption architectures
actually work, because to me they seemed to linear and rigid to be very
extensible.
I'm looking forward to trying out some of the theories we've looked at in
robots in lab later, in order to understand what implementing them is actually
like.
Sarah Klaum
I have found that I have fewer questions and pressing comments than I did
throughout last week's material, but a few still remain. Within the scope
of stimulus-response "machines," I began to contemplate the great
numbers of stimulus-response _agents_ that exist within our environment
and within ourselves: from phototrophic organisms to the bundles of neurons
housed within our fragile bodies. The simple wall-following robot that Nilsson
describes may appear complex, but could we categorize it as we would a neuron,
within the arena of stimulus-response, and take this one step further? If
we were to attribute intelligence to one, would we feel obligated to apply
such a term to the second instance? I believe I have stretched several definitions
here in my categorical approach, simply because I found it an interesting
proposal--what if each neuron in our body were considered an "intelligent"
agent, culminating in our "supreme" intelligence. I recall
ne response from last week that brought up the topic of programming, and
the question of assigning intelligence to individual functions. I then applied
this image to the human biological system, using such terms as teleo-reactive,
or complex subsumption architecture (at least on some small-scale cellular
level). Granted, this can be taken much too far. It has been brought up
in lecture, however, that perhaps it is not the behavior that appears complex,
but the environment that the behavior is performed and observed in. With
this in mind, then, surely many of our own behaviors can indeed be expressed
in such terms.
On a separate track, I am interested in learning a little bit more about
the history of threshold logic gates, as it was mentioned in this semester's
Computer Organization class that such technology was, at the onset, classified.
Maralee LaBarge
There really isn't much to say about the chapter two readings. The action
functions of the agents, though interesting, don't lend themselves to much
extra thinking on the part of reader, since this is more a chapter of memorization
and comprehension. The most one can really wonder is--how was all this realized?
How many years and people did it take to come up with these different models
and in what field of AI did they originate? I don't have the book yet, so
maybe it says that in a later chapter, but I doubt it. I know we're not
taking AI for a history lesson, but so many people invest their whole post-collegiate
life into developing these things, and we rarely if ever hear a name mentioned
(exception, Rodney Brooks for subsumption architecture). Certainly, many
of the architectures and ideas have been developed and polished over time
by many authors, but I'd like to know that history.
Part of what makes AI so interesting is that it
has an impact on how we think about ourselves and our own inner workings.
It forces self- examination. I don't see how anyone could work in the field
of AI and not have to consider day by day what makes them tick, or look
at a robot running around in an environment and despite knowing that they
programmed it, wondering if in its own way, it is thinking or feeling something
that we cannot recognize. We may give ourselves too much or too little credit
for all our accomplishments, but the operation of the machine called human
being is so fascinatingly complex that we may never run out of things to
study about it and we will always want to know: can we build something that
does what we do, either the way we do it, or otherwise?
So, to me, to only recognize what AI scientists
have come up with and not how they've come up with it does the field a serious
disservice and it takes away some of the wonder of it all. The best question
in AI is do/can machines think? And we should all be anxiously engaged in
trying to answer that question, even if we never do.
David Rothstein
Getting to work with the robots in lab last week was an interesting experience.
I began to understand the statement you made on the first day of class that
by the end of the semester we would be building robots more complex than
the Mars Pathfinder. Such a robot no longer seems particularly complicated
to me -- all it needs is a computer chip with several ports capable of supporting
different types of sensors, which is exactly what we have in the lab. Each
sensor can act independently as a stimulus-response agent, with its own
program to control that particular sensor's functions and a master program
to make decisions about prioritizing the various activities the robot is
capable of doing. The discussion of Threshold Logic Units gave me a good
understanding of how such a master program might work and how it could "learn"
by making adjustments to the threshold values based on input from the outside
world. However, there are still issues related to computers learning which
are difficult for me to understand. For example, in order to make adjustments
that correspond to "learning," a computer must first know that
it has made a mistake. But how do computers know that they have made a mistake?
In the movie we watched two weeks ago, with the computer program that was
being taught to recognize letters of the alphabet, the computer learned
of its mistakes when its programmers specifically told it that the correct
identity of the letter it had just been shown. The human mind, however,
has the ability to learn its mistakes through various methods, besides just
being told about them. For example, when we see a contradiction arise as
a result of something we have done, we have the ability to reason backwards
and decide which action of ours was most likely in error. For more complicated
artificial intelligence tasks than simply recognizing letters, it seems
as though it is vital for a computer to be able to recognize its own mistakes
(and then learn from them). Such an ability would probably require more
complicated programming techniques than we have discussed so far.
Frank Rusch
The robot described at the beginning of chapter 2 was primitive, since it
could not really function in a indeterminant space: The robot was made to
move from one spot in a grid to another, with no consideration of the interstitial
space in between. The little robot Deepak brought into class was more free
moving than the one in the book, though it was a little bit jerky at turns,
showing that it was indeed considering its motions as separate calculations
in time, rather than moving fluidly, as living things do. Animals are obviously
not merely S-R agents, because we can anticipate and choose what we perceive.
If you look at the algorithm we wrote in class for the 8-sensor robot's
action, you can see that certain circumstances will always result in certain
action, even though many choices may be available. This is due to the serial
manner in the robot processes information. This can be also observed in
many simple game-playing programs on computers. If a computer that plays
tic-tac-toe had two ways to win in one move (i.e. two places on the board
it could put an 'X' and win), it would always pick the same spot, based
on the order in which it scans the tic-tac-toe board. Though the is inconsequential
in the tic-tac-toe example, it makes a robot predictable. Suppose the 8-senosr
robot has no walls around it anywhere-- it will always go in the same direction
as long as no walls appear. This particular direction is always its "first
choice". People don't have a list of 4 prioritized directions from
which they pick the first available, they reason which direction to go.
However, a person stranded in a desert might act more like this robot, since
they would not have many visual details with which to reason.
The subsumption architecture prioritizes action based on its importance.
This seems very useful, and when implemented in robots, works better than
in humans because robots (unlike people) don't forget what they originally
set out to do. However, robots may lack the judgement to not continually
supersede previous actions with more important ones-- and thus take a long
time to get back to the original tasks. Though we may encounter "higher
priority" dilemmas in our everyday lives, it is not always necessary
to address the matters immediately, and sometimes good to leave those things
on the back burner when more resources are available, better circumstances,
etc...
Edina Sarajlic
I found the robot simulation game from Tuesday's class very interesting.
This exercise offered some insight into the principles of operation of a
simple goal oriented reactive agent. Two arms, eyes and a brain, each far
more sophisticated than any AI creation, were not immediately successful
in accomplishing the seemingly simple task of picking up a cardboard box.
This example shows that the performance of a reactive agent is not exclusively
dependent on the quality of its constituent modules. As mentioned in class,
the efficiency of information exchange between sensors, action mapping module(s)
and effectors is another very important factor influencing the agents performance.
Designing a reactive agent as an assembly of specialized modules is helpful,
because it divides the building process into an array of smaller, more manageable
tasks. However, such approach introduces many other concerns, such as: how
much information and in what form needs to be transferred between the modules?
For example, thefeature vector sent by eyes could describe every single
pixel to the brain, but it could also provide as little information as,
say, a number representing the average intensity of the received light.
Nilsson states that the split between perception and action is arbitrary,
with some solutions more optimal than others, because then the feature vector
can be reused for many different action functions. However, he does not
go into any specifics which would illustrate how one would actually decide
on the design of the feature vector and the action function in some general
case of an S-R agent. It might be interesting to look into a couple of examples
of S-R agents and step through the exercise of designing some goal oriented
behaviors.
Ben Sprecher
This week's readings and lectures, by begining to present the building blocks
of AI, gave me some idea of where our AI projects will be heading this semester.
It's nice to see some of the algothims we will be implemeting broken down
into steps that make sense and that I know I can program. I'm very anxious
to start making our robot do more than spin its wheels.
I'm also intrigued by the Neural Network stuff, because it seems to me to
be the most direct route to replicating the way our minds work. I don't
claim that something must work like our minds in order to be intelligent,
but replicating the structure of a system we know posseses intelligence
will make it easier to experiment with how to create intelligence. Also,
it can help expand our understanding of ourselves. If, for example, you
could create a hardware or software neural network that exactly mimiced
the structure of the human brain, then exposed it to years of sensory input
and social interaction from humans around it, you could determine whether
an intelligence developed. If it did, then it would have drastic implications
for the current concepts of what makes us human. If intelligence is structural,
can there be such a thing as a soul? Or a God? I can imagine arguments for
and against both points. But, if no intelligence develops, then the question
remains; what beyond the mere neurons and synapses of our brain makes us
intelligent? Is it physical? Is it embodied in the exact configuration and
properties of the flesh? Or is it something beyond that - a divine spark,
or a soul, or something which can never be created by an individual that
is controlled by a spark of its own? Some of the potential experiments in
this area drive to the heart of the most basic and sweeping questions facing
humanity: Who are we? What makes us the way we are? Why are we here?
Emily Sweeney-Samuelson
This week clarified the world of Artificial Intelligence quite a bit for
me. I was pleased to get down to detailed methods - I feel more grounded,
now that I have a little specific knowledge of the categories and programming
procedures of robots. It is exciting to learn; it makes me want to build
them! (I haven't been able to attend the lab yet.)
I want to learn more about machine learning. Neural networks, especially,
are one of the reasons I took this class. The concept of emulating the human
brain on any level, however 'incorrectly', fascinates me. It seems to be
the ultimate challenge, and a central subject in the field of AI. I am not
sure about subsumption architecture; it seems like that type of behavior
would be clean and rewarding to program, but mostly early on. The appeal
of it might wear off when the desired complexity of behavior can no longer
be reached using it. But I'm sure I won't reach that point for a long time;
I am just glad to have some procedures of AI available to me, to decrease
the mystery of the field.
Tim Waring
After having read chapter 2, i understand the logic algebra. I also understand
the basis for neural networks, and what a threshold logic unit is. i understand
production systems and the idea of subsumption architecture.
I don't understand what a feature vector is. is
it just a way of describing sensory inputs in a susinct and sensible way?
also in subsumption architecture the highest (most complex or imformation
demanding behavior gets cued first, then if that's not fullfilled then the
second one can go, etc... all the way down. it seems like that is the way
that animals behave a lot of the time. humans however have wierd and hard
to describe thought porcesses going through them, ... it seems that subsumption
is a good mode of programing instinctual behavior, then true intelligence
must be a neural net or other high level, complicated, unpredictable system.
How do programmers have the "neurons" strengthen or weaken their
responses based on learning?
Sarah Waziruddin
One of the things that interested me last week was the emphasis on the environment
and its reaction with the machines/robots/agents. I didn't think the environment,
or world, would be such a factor in determining intelligent behavior. The
idea of a complex world and simple machines with apparently complex behavior
is greatly interesting to me. This notion makes me question the complexity
in human beings. Are we actually simple machines and only appear so complex
because of the complex world around us? Has our behavior become more complex
as the world around us has become more complex?
Another concept this week that interested me was the idea of a training
phase. This was discussed when we covered Networks and TLUs. The idea that
reinforcement has any affect on a machine is new and strange to me. I would
be very interested in actually programming a machine that is affected by
reinforcement.
The breakdown of tasks to make a fully functioning robot were very helpful
to me. This breakdown allowed me to see how a robot would be programmed
to perform certain tasks. I was also surprised by how many different approaches
there are to implementing action functions. These different approaches all
have different advantages and disadvantages which would be helpful to know
when achieving certain tasks.
Leslie Zavisca
I found the reading and lecture on networks and, in particular, TLU's, to
be the most interesting. This seems to me to be the foundation of imitating
the human brain in its most broken down form. Even though I know we won't
be doing it for this class, I am particularly interested in the determination
of the weights that different sensory inputs carry and whether or not, for
humans, those weights are more or less stable or if certain ones vary depending
on conditions (although this would defeat our purposes). When trying to
understand the applications of TLU's I came up with the very simple example
of how a human being might determine what speed to drive on a highway. The
different sensory inputs would be things like numer of passengers, time
of day, temperature, wind speed, and then I think of more and more abstract
and comples things like mood. How will we ever be able to quantify a variable
such as 'mood' or 'emotions'. Even psychologists are worlds away from agreeing
on issues concerning emotions etc. How will AI ever be able to conquer it?
I am noticing that the more I learn about the advances in AI, the more I
am convinced that we are so far away from accomplishing our goals. Although
it is all very exciting and intriguing and challenging, it is also somewhat
frustrating.
Back to CS372
Course Materials