Databases Reference
In-Depth Information
confrontation. It boasted a 30-node massively parallel computer with 480 special-
purpose (CPUs). Deep Blue was able to search to a depth of 12 moves (ply) ahead,
1
evaluating 200,000,000 positions per second, twice as fast as the system Kasparov had
beaten in 1996. It contained an extensive database of opening moves and endgames,
including every possible endgame with five pieces or less. Furthermore, Deep Blue's
chess database was refined by grandmasters Joel Benjamin, Miguel Illescas, John
Fedorovich, and Nick De Firmian. In retrospect it was amazing that Kasparov faired
as well as he did. After the match, Kasparov made a public statement that he was
unfairly disadvantaged by not being able to study a history of the machine's play the
way he would normally have studied an opponent. He felt that despite its awesome
computational power and extensive database of potential moves, Deep Blue was
beatable by the very best human players.
In truth Deep Blue's accomplishments were achieved to a large extent by brute-
force calculations and massive databases of alternatives. At the end of the day it
doesn't matter that Deep Blue doesn't really understand what a king or a knight really
is. It has no understanding of the game of chess; it feels no joy or tension or
excitement. It's simply a computer executing a program that evaluates a dramatic
number of alternatives. This process is not dramatically different from the database
design process, which includes a massive scale of possibilities from which to select a
near-optimal solution.
Computers like Deep Blue have proven successful at chess because chess has quali-
ties that make it suitable for computer play: the game does not require a physical under-
standing; everything can be abstracted and encoded. Second, the game does not require
a complex understanding of language—simply the ability to express where the current
pieces are and where the next piece should be moved. Finally, as a game of strategy, even
before the final checkmate, the strength of a position and the value of a move can be
estimated. This ability to evaluate some positions as better than others lends itself to an
“evaluation function” which forms the basis of most Artificial Intelligence systems.
Compare the qualities of gaming programs to the problem of physical database design.
Databases can be abstracted and modeled; in fact as described earlier during the discus-
sion of query optimizers most query optimizers model physical resource consumption
and have been doing so in the database industry for more than a decade. Secondly,
physical database design does not require understanding of a complex language, but
rather the ability to parse SQL statements and generate DDL. Database servers already
include the former—a basic ability required to process SQL as part of day-to-day pro-
cessing. Finally, there is the all-important notion of an “evaluation function,” which we
will see can be constructed by exploiting variations of existing database components. In
short, just as Deep Blue conquered the domain of chess, high-quality technologies are
emerging to perform automated physical database design. These technologies can be
1
The best human players evaluate roughly 10 moves ahead.
