[ Feedback & Cybernetics | Levels of Abstraction | Dynamic Modification by Learning | Genetic Algorithms | General Learning Systems & Domains ]
dynamic Modification
means the ability to keep changing as circumstances change. Adaptability, in other
words, is a sign of intelligence because it requires both acquiring information and
deciding on a course of action based on the input. As more and more information is
required to make better decisions, the concept of learning is introduced into the field of
AI. First of all, the story behind dynamic modification begins with feedback systems
and cybernetics.
The notion of feedback has been a long-standing idea in biology and in
mechanics. For example, temperature regulation is a feedback process whereby a
person maintains his 98.7 degrees Fahrenheit internal body temperature by having the brain
constantly check on the current temperature of the body, and either slows down the rate at
which cells metabolizes and release less heat or vice versa. Similar in function, a
thermostat either turns on or turns off the air conditioner based on the temperature of
the room. In both cases, it is the control mechanism(temperature) that determines
how the system reacts which modifies the control mechanism and the new control feeds back
to the system to stop acting.
Mathematician and engineer Norbert of MIT was one of the first Americans to link
how the mind and machines can be similar. Applying the Feedback Theory,
proposed that any system that reacts to a control mechanism can be considered an
information processor. In a sense, the thermostat decides that the room is too warm
and acts accordingly while the brain does the same with the body temperature.
However, the more complex brain also has feedback mechanisms to regulate glucose levels in
the blood, bone mass, and other physiological factors. Creating the field called
cybernetics--the science of control-- even went as far as to speculate that all
intelligent behavior is attributed to highly complex feedback mechanisms. In other
words, intelligence stems from information processing.(Crevier 28)
[ Top ]
Since how one learns is dependent on how one's knowledge base is arranged, one way to
look at learning is by examining how the brain processes information. Basic learning
can be depicted as taking in data from the physical world, translating the data into some
comprehensible form, and fitting the information in an appropriate place within the
knowledge structure. If one assumes the brain's memory to be like a semantic net,
then the information may provide ways to relate the new fact to existing facts in memory.
So a deeper step in learning would be forming new knowledge from new facts.
As mentioned previously in the Inference section, humans have
the ability to summarize facts collectively into generalizations. There could be
many generalizations for the same group of facts--men are generally more aggressive than
women, men usually make more money than women, and women tend to take care of the children
more than men--are all generalized knowledge taken from factual observations about the
sexes. What is learned in generalization is the collective relationships between a
certain amount of facts--another level of abstraction.
In the third level of abstraction, learning not only relates to the forming of
knowledge between facts, but forming knowledge between generaliza
tions. Whether it is by conscious or unconscious thought, the brain can
form more knowledge between a group of facts by perhaps finding similar structures between
two unrelated generalizations. Most people can determine how the play West Side
Story relates to Romeo and Juliet relate to one another because the
similarities in the plot--two lovers(Fact #1) are kept apart(Fact #2) because their
affiliated groups are feuding(Fact #3)--despite the fact that the languages in the plays
are literally significant--the former is written in modern English while the latter is
written in old English. In this case, one has generalized the actions and the
dialogue in each play into their significance, then generalized the meanings in each
section together into a summery the plot or theme, and then compared the plots and/or
themes for similarity. If doing that does not establish a clear relationship between
the two plays, the meanings comprising the plots may show may match up many times which
would lead one to form the overall(top level) generalized conclusion that West Side
Story and Romeo and Juliet are very similar--a new fact plus the similarity
itself is a relation(i.e. knowledge)!
The power of the human brain for generating new knowledge for learning is a compelling
suggestion that the methods designed to replicate intelligence in machines must be as
flexible and robust as a real brain in fact-manipulation. There have been several
ways in which this goal has been approached by various people.
[ Top
Learning is inextricably linked to intelligence because it is both the means and and
end of cognition. The existing structures in the brain allow one to learn in certain
ways. As one gains information through learning, it is the very storage and
arrangement of the knowledge in the brain that redefine how one functions--including how
one learns in the future. Quite simply, knowledge and learning affect one another in
a feedback loop. For example, a person may find that he has better reading
comprehension when he verbalizes the words he reads. First of all, he has learned
that how to read more effectively. With this new knowledge, he can learn more when
he reads from now on. Thus in a broader sense, learning is not just the taking in of
information, but it is the means to change how one thinks and learns later.
[ Top ]
As a subset of computer science, genetic algorithms are programs that can improve
themselves by modifying how they function. The computer doesn't necessary modify the
lines of code in it's program per se, but the program can be robust enough to change how
it works. For example, in the Traveling Salesman Problem, a genetic algorithm can be
used to find out the shortest route a salesman should take to travel through a certain
number of cities. The program randomly tries out different paths and notes how long
it takes(usually based on the number of computer clock cycles) to reach a certain
destination. Taking the path that is the most efficient to the first city to be
traveled, the program starts on the next city and the process starts over again. At
the end of this trial and error process, the computer will merge all the best routes into
a single path that would be the shortest route for the salesman to travel.[cite]
The previous example is a perfect illustration of the ways genetic algorithms deal with
a more general set of problems. The huge combination of ways to travel through all
the cities can be tamed by the evolution of the solution as the genetic algorithm
constantly takes the best mini-solutions that takes it closer to the answer and
"breeds" them to produce a better answer. Thus, genetic algorithms are
another avenue in which computers can learn and improve. However, one of the most
notable examples of genetic algorithms and computer learning is Douglas Lenat's Eurisko.
[ Top ]
Obviously, the script and frame approaches further illustrate how domain-specific a
computer is. A computer could know how to make any French dish, but would not know
who the famous French cook François Mitterrand is. In the script-method, knowledge
could be swapped, but a computer can not have the kind of general knowledge humans
possess. However, in a sense, people are not total general learning systems.
In society, people usually have specialized jobs that require knowledge specific to their
field of work. Firemen know how to put out flames in the safest way, doctors know
what medications to prescribe their patients, and architects are aware of the latest
architectural styles. Even within a specific field, such as a software company like
Microsoft, there are different people with specific knowledge in specific positions like
secretaries, executives, technicians, programmers, etc. One can not expect a
policeman to diagnose a disease in a person nor would one hope to have a doctor direct
traffic.
If people are general learning systems in one sense but are domain-specific in another
sense, then how general does a computer have to be to be considered intelligent as a
person? The answer may lie back in the Turing Test. When one converses with
someone else, it is not expected that the other person can answer every question that is
posed to him. However, the other person should know enough about life to be
considered intelligent as a human. How much "enough" about human life means is still in debate.
