Re: computer-go: A problem with understanding lookahead

[Vincent Diepeveen <diep@xxxxxxxxx>]

At 10:03 PM 1/18/01 +0100, you wrote:
>> The random algorithm plays very badly, but the point  is that is still
>> plays MUCH BETTER than totally random moves.
>
>Just a wondering: the "random evaluation function" must be deterministic
in order for that to be true, isn't it? Deterministic in the meaning that
for a given position it will return the same value every time, otherwise if
the values change every time a position is evaluated (assuming no caching
for the game tree) it doesn't seem to be possible to get something else
than a random play...

When we talk about half a billion of evaluated nodes and 
huge search depths like 12 plies in the random chess
experiment or sometimes bigger search depths
(searching random 
with alfabeta one
can easily search at a simple PIII800 with like 2 million
nodes a second in a chessprogram, in a GO program a magnitude more
of course), then it doesn't matter whether it's
deterministic or non-deterministic.

Let's assume for example a random generator giving scores between
  -1000 and 1000

Now if we get half a billion scores from it instead of
just being busy with weird theoretic assumptions about
3 or 4 random deterministic numbers, then
you start to slowly realize the scale of the experiment.

Whether your random generator is deterministic or nondeterministic
is of course not an issue as long as it's not delivering always
a constant score. a RANDOM score between -1000 and 1000 in the
above example is all we need.

At billions nodes a second talking about worst case behaviour is
also no issue anymore. The size of the search tree
and the number of nodes is so overwhelming that we can easily
generalize and get down to the facts instead of talking about
worst cases and bad luck.

>regards, Vlad