Page 230 - Operations Strategy

PRoCEss TECHnology sHould REflECT volumE And vARiETy 205

example go figure 2

It was a significant event in the development of artificial intelligence (AI). Between 9 and 15
March 2016 a five-game match was played in the South Korean capital Seoul between arguably
the best professional ‘Go’ player called Lee Sedol and AlphaGo, a computer Go program devel-
oped by Google DeepMind. AlphaGo won the contest by 4 games to 1. Some commentators
saw the event as a continuation of the ‘man versus machine’ chess battles that started when
chess master Garry Kasparov lost to a computer named Deep Blue in a six-game match played
in 1997. In fact, games like chess really are a handy way to gauge a computer’s evolution towards
genuine artificial intelligence. Which is where Go comes in. Although seemingly simple, it is
a far more complex game than chess. Played all over East Asia, it is particularly popular with AI
researchers, in particular, for whom the idea of truly mastering ‘Go’ has become something of
an obsession. Why? Because compared with Go, teaching computers to master chess is easy. The
size of a Go board means that the number of games that can be played on it is colossal: probably
around 10 170 , which is almost a hundred of orders of magnitude greater than the number of
80
atoms in the observable universe (estimated to be around 10 ). As one of DeepMind’s creators,
Dr Demis Hassabis points out; simply using raw computing power cannot master Go. Much
more than chess, Go involves recognising patterns that result from groups of stones surround-
ing empty spaces. Players can refer to seemingly vague notions such as ‘light’ and ‘heavy’ pat-
terns of stones. ‘Professional Go players talk a lot about general principles, or even intuition,’
says, Dr Hassabis, ‘whereas if you talk to professional chess players they can often do a much
better job of explaining exactly why they made a specific move.’
However, ideas such as ‘intuition’ are much harder to describe algorithmically than the
formal rules of any game. Which is why, before AlphaGo was developed; the best GO programs
were little better than a skilled amateur. The breakthrough of AlphaGo was to combine some
of the same ideas as the older programs with new approaches that focused on how the com-
puter could develop its own ‘instinct’ about the best moves to play. It uses a technique that
its makers have called ‘deep learning’ that allows the computer to develop an understanding
of the instinctive rules of the game that experienced players can understand but cannot fully
explain. It develops this leaning by playing games against itself (or a slightly different version
of itself) and analysing the vast amounts of data to sort out these ‘intuitive’ rules. However,
as well as masses of data ‘deep learning’ also requires plenty of processing power. Yet it is the
‘deep learning’ that was being seen as the exciting development that would lead to further
applications. Such an approach could help computers to do complex tasks like accurate face
recognition or translate subtleties of meaning from one language to another. But, although
the techniques used by AlphaGo is an important step in the progress to, what in Dr Hassabis’s
view, is the ‘same sort of broad, fluid intelligence as a human being’, they still lack some of
the abilities that humans take for granted. Arguably the most important of these is the ability
to apply lessons learned in one situation in another, what AI researchers call ‘reasoning by
analogy’ or ‘transfer learning’.

about the balance between people and technology. The choice is often between empha-
sising the power, speed and general physical abilities of automation against the flexible,
intuitive and analytical abilities of human beings. However, an increasing number of
purely information transformation processes are entirely automated (including most
processing technology in the financial services sector, for instance). We need a different
metric to differentiate between different information processing technologies that are
100 per cent ‘automated’, or very close to it.

M06 Operations Strategy 62492.indd 205 02/03/2017 13:05