Database Reference
In-Depth Information
Ultimately, however, neither was successful. The Cybersyn project was brought to an
abrupt end by Pinochet's coup d'etat, while OGAS was successfully blocked by Soviet
bureaucrats who feared the loss of their sinecures. Yet even if these projects had been fully
implemented, their immense complexity would inevitably have led to their ultimate failure.
When later we see how complicated it is to properly control even far simpler systems
(like our recommendation engines), we will appreciate the boldness - but also the
foolhardiness - of these endeavors. For in reality, even for much more straight-forward
tasks like computer chess or machine translation, cybernetics was for the moment unable to
live up to its expectations.
For that reason specific elements of cybernetics began to emerge as separate research
fields. Probably the most widely known of these is Artificial Intelligence (AI), which
initially was hyped in much the same way as cybernetics (although lacking its scientific
merit) but became discredited over time in public opinion. Like most mathematicians, I was
suspicious of AI: I associated it mainly with long-haired gurus who spoke in incompre-
hensible sentences, always ending with the threat that robots would take over the world. In a
word: cranks!
But I changed my mind after reading the classic
Artificial Intelligence: A Modern
Approach
by Stuart Russell and Peter Norvig [RN02]. This topic centers on the concept
of an
agent
communicating with its
environment
. The authors then systematically introduce
different types of agent: planning and non-planning, learning and non-learning, determin-
istic and stochastic, etc. An AI system encompassing a wide array of diverse fields emerges.
What's more, the practical successes of AI can no longer be ignored: computer programs
play better chess than grand masters, call centers work with voice control, and IBM's
Watson computer recently dealt mercilessly with past champions on the American quiz
show
Jeopardy
. There is still a long way to go of course: modern robots still tend to move
like Martians; you have to repeat everything ten times to make voice control work, and
automated Google translation is a source of constant amusement. Yet the advances are
undeniable.
Michael Thess
1.2 Realtime Analytics Systems
The area of realtime data mining (
realtime analytics
,or
online
methods for short) is
currently developing at an exceptionally dynamic pace. Realtime data mining
systems are the counterpart of today's “classical” data mining systems (known as
offline
methods). Whereas the latter learn from historical data and then use it to
deduce necessary actions (i.e., decisions), realtime analytics systems learn and act
continuously and autonomously; see Fig.
1.1
. (Strictly speaking, they should
therefore be called realtime analytics action systems, but we will stick to the
established terms.) In the vanguard of these new analytics systems are
recommen-
dation engines
(REs). They are principally found on the Internet, where all
information is available in real time and an immediate feedback is guaranteed.
Realtime analytics systems mostly use adaptive analytics methods, which means
that they work incrementally: as soon as a new data set has been learned, it can be
deleted. Apart from anything else, the adaptive operating principle is a practical
necessity: if classic analytics methods were used, each learning step would require
an analysis of all historical data. As realtime systems learn in (almost) every
interaction step, the computing time would be unacceptably high.
