Robotics Reference
In-Depth Information
In 1986 an annual computer Go tournament was started called the
Ing Computer Go Congress, sponsored by the Ing Foundation.
30
Just
as the computer Chess tournaments of the 1970s and 1980s did much
to help advance the strength of the best programs, so the Ing tourna-
ments started along the same path in computer Go. For some years now
there have been a number of regularly held computer Go tournaments,
including a World Championship.
The strength of the best Go programs has improved slowly since
1984. Mick Reiss, who has long been one of the world's leading Go
programmers, estimates that the best programs in 2004 were playing at
7-Kyu and that in recent years the rate of improvement had been around
half a Kyu per year on average.
31
If progress continues to be made at
the rate of half a Kyu per year, the leading programs will reach 7-Dan
amateur level and 1-2 Dan professional around the year 2030 and are
likely to be ready to challenge the Garry Kasparov of the Go world, with
a reasonable chance of success, around the year 2035.
32
Given the magnitude of the computation problem, and the impossi-
bility of compiling a complete database of Go positions, it is hardly sur-
prising that Go programming has not yet progressed to the point where
strong human players have anything to fear. But the magnitude of these
particular problems is not the only important factor in the relative lack
of progress in computer Go. Other factors include the limited size of the
few openings databases that have been created and the great difficulty
experienced in developing an accurate evaluation function.
Go is the best example from the field of games that demonstrates the
difficulty of incorporating human expertise in an evaluation function.
33
30
Ing Chang-Ki was a Taiwanese industrialist and lifelong Go enthusiast whose foundation sup-
ported computer Go up to the year 2000.
31
(Mick Reiss, private communication.) These estimates appear to be fairly consistent with the
improvement from the Nemesis program's 20-Kyu in 1984 to the current best of 7-Kyu in 2004, a
rise of 13 Kyu over 20 years, an average of 0.65 Kyu per year.
32
It is said that the difference between two adjacent professional Dan grades is as third as much as
that between two adjacent amateur Dan grades, in which case the progress from 1-Dan professional
to 9-Dan is likely to take a further five years or thereabouts.
33
In Chess it is relatively straightforward for a strong player to explain to a weaker one (or to
a computer programmer) the salient features in a position and why one side or the other has an
advantage, but the same appears to be untrue in Go. The science of knowledge engineering—
extracting knowledge from human experts in a form in which it can be programmed—appears to
have made virtually no progress in Go programming. An expert Go player can look at a position and
know, intuitively, whether the formation on the board favours Black or White, and why. But as for
extracting the knowledge that allows him to do this—thus far nothing has been achieved. Perhaps
the most difficult concept to explain in Go is “good shape”. It is relatively easy to explain to a human
