Chapter
Twelve
The evolution of machines
What, then, is a good test of what a machine can do? Among all the possible tests
one can think of, two come up again and again. The first is playing chess. Surely,
people have long thought, if a machine can play chess, then it must be intelligent,
rational, and able to think. Descartes would presumably have been impressed by
such a machine since, like his contemporaries, he prized human rationality far
above things that ‘lower’ animals can do easily, such as walk about and see where
they’re going. So it is perhaps not surprising that in the early days of computing,
it seemed a great challenge to build a computer that could play chess.
After the trick games played by the mechanical Turk, the first serious game took
place in Manchester in 1952, with Turing playing the part of a machine against a
human opponent. He had written a program on numerous sheets of paper and
consulted them at every move, but was easily defeated. In 1958, the first game
with an actual machine was played with an IBM computer, and from then on com-
puter chess improved rapidly. Most chess programs relied on analysing several
moves ahead. This quickly produces a combinatorial explosion (also known as
‘the curse of dimensionality’), because for every possible next move, there are
even more possible next moves after that. Programmers and mathematicians
invented many ways to get around this, but to some extent computer chess got
better just by brute-force computing power. In 1989, the computer Deep Thought
took on the world chess champion Gary Kasparov, who told reporters that he was
defending the human race against the machine. This time the machine lost, but
eventually, in 1997, its successor, Deep Blue, beat Kasparov (for the personal story,
see Kasparov, 2017).
Deep Blue consisted of thirty-two IBM supercomputers connected together and
could evaluate 100 million positions per second, but no human being plays chess
like this. So, was Deep Blue intelligent? Could it think? Searle (1999) said not, argu-
ing that Deep Blue, like the Turk, relied on an illusion, and the real competition
was between Kasparov and the team of engineers and programmers. The team
said that they never thought their machine was truly intelligent. It was an excel-
lent problem-solver in one domain but could not teach itself or learn from its own
games.
In subsequent human-computer battles, another world champion, Vladimir Kram-
nik, was defeated by Deep Fritz, and a whole team of computers beat a strong
human team. More recently, a chess program called Giraffe has been developed
not to rely on brute processing power to search through all possible moves,
but to evaluate the possibilities and narrow down the avenues worth pursuing,
using a deeply layered neural network. Giraffe plays against itself with the aim of
improving its ability to predict how it would evaluate a future position relative to
the outcomes of winning, losing, or drawing. In this way, Giraffe can bootstrap its
way to matching the best chess engines in the world, trained over years often by
human grandmasters.
One of the latest developments is to switch chess for the ancient Chinese game of
Go, which has a total number of possible moves orders of magnitude larger than
the number of atomic particles in the observable universe. In 2016, Google’s pro-
gram called AlphaGo shocked one of the most experienced human players with
a move a human would simply never have thought of doing (Wong and Sonnad,
2016), earning itself an honorary 9-dan black belt. The principles AlphaGo was
trained on were standard ANN with reinforcement learning, using a prior stage to