[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: computer-go: extracting date to beat 9d from chess and draughtsscene
At 12:20 AM 11/14/99 +0100, you wrote:
In reaction at the GO mailing list to the statement whether
a 9 dan will get beaten in 2010 by a machine
...
To get a bound ...let's have a look in different worlds called draughts
and chess.
Let's distinguish from a program 3 aspects that are needed to
beat an opponent in a game:
- high tactical level
- evaluation of a position
- strategical aspects
I'm not sure what dan level my chessrating of 2254 FIDE compares to, ...
probably amateur 1-2 dan. this is about 4-5 stones from the worst pro.
probably 8 or 9 stones from a 9-dan.
... In the endgame chess is a lot harder than draughts
as the main goal in draughts:
the endgame rules in chess are artificial and not natural. in go, you can
sort of just keep playing stones (like chinese rules)
This convinced me that knowledge was the way to go for
any type of game that can't be searched completely. ...
Now as we all know GO because of branching factor
is a bit harder than this.
its a *lot*harder. in the late fuskei and early midgame. locally you can
easily have 20-30 open spaces, so you are looking at say
25!=15511210043330985984000000 or 1.5*10^25. furthermore, there are major
branch points in most sequences where one side can tenuki and start
something (that may relate) elsewhere.
We need a revolution in both programming languages (to represent
knowledge) and the way in which we evaluate positions.
yes
Apart from that we also need a couple of tens of years before
Go programs get at tactical decent levels.
maybe much sooner, tactical stuff can probably done fairly well. (if you
know the problem is to connect the stones, then you can get real good at
this). the first problem will be to choose the right tactics. later
problems will be more difficult.
Most important thing now might be to estimate how deep the average
combination in GO is, and how many plies the programs lack.
i am an amatuer 1-dan. i can look 5-15 moves ahead (not real good)
depending how much of a one-way street it is, how open/complicated the area
is etc. pros can look a lot further and they know which lines to ignore.
...let's simply assume that
getting the same depth as in draughts&chess to see most combinations
is what we try to achieve.
you won't get that far in an open area of the board (say 30 empty spaces).
you can do real well in the endgame. and nothing at all in the fuseki.
...
If every 2 years the speed of hardware doubles, then that means
that for 6 ply we need about 9^6 = 531441 times faster hardware.
To get 531441 times faster hardware we need to wait:
2 * (log 531441 / log 2) = 2 * 19 = 38 years
That's 2037. I get 64 in that year.
hmmm, thats when the unix clock expires (i'll be 91 or not here :)
maybe we can beat some strong amateurs by 2037.
Ray (will hack java for food) http://home.pacbell.net/rtayek/
hate spam? http://www.blighty.com/products/spade/