heat absorption/loss characteristics
of the building, and so on. Using this
model, which allowed their system
to anticipate needs, and the ability to
pump heat from one part of the building to another, they designed a system
that reduced temperature fluctuations
and was more energy efficient.
Humans do not have the measurement and calculation ability that is
available to a modern computer system; a system that imitates people
won’t do as well as one based on physical models and modern sensors.
Humans solve complex physics problems all the time. For example, running
is complex. Runners maintain balance
intuitively but have no idea how they
do it. A solution to a control problem
should be based on physical laws, and
mathematics, not mimicking people.
Computers can rapidly search complex
spaces completely; people cannot. For
example, a human who wants to drive to
a previously unvisited location is likely
to modify a route to a previously visited
nearby place. Today’s navigation devices can obtain the latest data and calculate a route from scratch and often find
better routes than a human would.
Another approach to creating artificial intelligence is to construct programs that have minimal initial capability but improve their performance
during use. This is called machine
learning. This approach is not new.
Alan Turing speculated about building a program with the capabilities of
a child that would be taught as a child
1 Learning is not magic; it is
the use of data collected during use to
improve future performance. That requires no “intelligence.” Robert Dupchak’s simple penny-matching machine used data about an opponent’s
behavior and appeared to “learn.” Use
of anthropomorphic terms obscures
the actual mechanism.
Building programs that “learn” seems
easier than analyzing the actual problem, but the programs may be untrustworthy. Programs that “learn” often exhibit the weaknesses of “hill-climbing”l
algorithms; they can miss the best
l Hill-climbing algorithms are analogous to hikers
who always walk uphill. They may end up at the
top of a foothill far below the mountain peak.
product. If heuristics are used in critical applications, legal liability will be a
An AI Assembly-Line Assistant
An assembly line could run faster after
tool-handling assistants were hired:
Whenever workers finished using a
tool, they tossed it in a box; when a tool
was needed, the assistant retrieved it
for the workers.
A top research lab was contracted
to replace the human assistants with
robots. This proved unexpectedly difficult. The best computer vision algorithms could not find the desired tool
in the heap. Eventually, the problem
was changed. Instead of tossing the
tool into the box, assemblers handed
it to the robot, which put it in the box.
The robot remembered where the tool
was and could retrieve it easily. The AI
controlled assistant could not imitate
the human but could do more. It is
wiser to modify the problem than to
accept a heuristic solution.
“Artificial Intelligence” in Germanj
When AI was young, a German psychology researcher visited pioneer AI researchers Seymour Papert and Marvin
Minsky (both now deceased) at MIT.
He asked how to say “artificial intelligence” in German because he found
the literal translation (Künstliche Intel-ligenzk) meaningless.
Neither researcher spoke German. However, they invited him to
an AI conference, predicting that he
would know the answer after hearing
the talks. Afterward, he announced
that the translation was “natürliche
Dummheit” (natural stupidity) because AI researchers violated basic
rules of psychology research. He
said that psychology researchers do
not generally ask subjects how they
solve a problem because the answers
might not be accurate; if they do
ask, they do not trust the answers.
In contrast, AI researchers were asking chess players how they decide on
their next move and then writing programs based on the player’s answers.
j I cannot warrant the truth of this story; it was
related to me as true, but I was not present for
the events. I include it because it contains an
k Current terminology in German.
Artificial Neural Networks
Another approach to AI is based on
modeling the brain. Brains are a network of units called neurons. Some
researchers try to produce AI by imitating the structure of a brain. They create models of neurons and use them
to simulate neural networks. Artificial
neural networks can perform simple
tasks but cannot do anything that cannot be done by conventional computers. Generally, conventional programs
are more efficient. Several experiments
have shown that conventional mathematical algorithms outperform neural
networks. There is intuitive appeal to
constructing an artificial brain based
on a model of a biological brain, but no
reason to believe this is a practical way
to solve problems.
The Usefulness of Physics
A researcher presented a paper on using AI for image processing to an audience that included experts in radar signal processing. They observed that the
program used special cases of widely
used signal-processing algorithms
and asked “What is new in your work?”
The speaker, unaware of techniques
used in signal processing, replied, “My
methods are new in AI.” AI researchers
are often so obsessed with imitating
human beings that they ignore practical approaches to a problem.
A study of building temperature-control systems compared an AI approach with one developed by experienced engineers. The AI program
monitored individual rooms and
turned on the cooling/heating as needed. The engineers used a heat-flow
model that included the building’s
orientation, the amount of sunlight
hitting sections of the building, the
Learning is not magic,
it is the use of data
collected during use
to improve future