Page 70 -

243_MasterPieces_02a.QXD 4/18/03 6:55 PM Page 42

42 Masterpiece 2 • The Learning Brick Sorter

Think of learning how to ski using a book with text and schematics. Would it be
easy? Surely humorous, but not easy. Having a teacher that shows you what to do makes
the learning process much easier: as any of us know from direct experience, an example is
worth a thousand words.This applies to many other activities, especially those abilities
that involve learning how to use our arms, legs or body for a specific task. However, the
mechanism of learning by example is so powerful that even when we approach concep-
tual matters we tend to consolidate the theory with analogies taken from the real world.
Unfortunately, examples are by themselves not sufficient. If this were the case, we could
learn how to ski by simply watching other skiers, Which, as anyone who has skied is
aware, is a sport not as easy at it looks. We need a way to understand how close we are to
the behavior we want to mimic to distinguish what’s right and what’s wrong in our
actions. In other words, we need feedback. Sometimes the feedback comes from a person
who helps us to learn a new skill, other times the feedback is the environment, as a skier
discovers by falling a few times.
This learning mechanism has another big advantage: We not only learn, we also learn
how to learn. In other words, when we acquire a new ability, we also acquire a model that
we can apply—with a few changes—to different situations, a property called adaptability.
Adaptability is what allows an experienced skier to perform relatively well at his first time
water skiing: He just has to learn the differences between the two activities instead of
having to learn from scratch.Adaptability is what saves parents and teachers from
explaining every single detail of the universe.
As feedback and adaptability work for human beings, they seem a desirable and prom-
ising approach to use with robots too.A robot learns to solve a task without having to be
programmed, and this makes it suitable for a wide range of applications.Additionally, it
allows people with no programming experience to instruct it through simple demonstra-
tions and feedback. Is this a dream? No, it’s reality. Many industrial robots can already be
“programmed” this way, where during a learning session, a human trainer manually acti-
vates the robot, forcing it to perform the movements it will need to perform during the
production cycle; the robot will then be able to replicate the same movements an unlim-
ited number of times. Some artificial vision systems include a more sophisticated learning
mechanism: the operator shows the vision system a variety of objects, each one at a
variety of angles, and every time supplies the software with the name — or code — of
the shown object. When the training procedure is finished, the system is able to recognize
the different objects and tell their names by just “looking” at them. Finally, an increasing
number of software applications—including computer games—incorporate Artificial
Intelligence (AI) techniques to learn from their users’ behavior and become more respon-
sive or more competitive.Actually, learning and adaptability have been part of AI research
since the beginning, because we cannot even conceive a form of intelligence that does
not incorporate these concepts.
On the subject of computer programs, we are used to thinking of computer games as
sequences of simple instructions, and this is what they ultimately are: add the number X