On Tue, Jul 7, 2009 at 3:49 AM, Magnus Persson <magnus.pers...@phmp.se>wrote:
> Quoting Oliver Lewis <ojfle...@gmail.com>:
>
>  Others on this list have reported in the past that the randomness is
>> actually very important.  Playouts that are very heavy, no matter how
>> "clever" they are, actually reduce the performance because they narrow the
>> number of games too much.
>>
>
> I would like to disagree with this statement. I think it is difficult to
> make a good heavy playouts, but it is not impossible.
>
> Failing to make a playout stronger through heaviness does not prove
> anything. It just mean one has failed.
>
> If I could make a heavy playout of 1 Dan strength and then run in MC tree
> search. I am sure it would be stronger than the playout itself.


I agree with everything you said.

In the Mogo papers it is claimed that having stronger playouts does not
NECESSARILY lead to stronger play in general.   I don't believe everything I
read, but I think there may be something to this.    Also, having a random
element to this does not imply weakening the play,  perhaps it's more like
varying the playing style.   If the Mogo team is correct  the formula seems
to be that there is something in the playouts that can compliment the search
in some way not fully understood.

As you gradually go from random to deterministic you cease to have "Monte
Carlo" and you have something that is more like a best first search.    It
may be that in a few years our programs will gradually transition away from
"MC" and more toward "TS" using best first methods.     Maybe that is really
what we have discovered and the MC part is distracting us?


>
>
> The problem I think is to find a good tradeoff between heavyness and speed.
> In my test with Valkyria vs Fuego, Valkyria is superior when the number of
> playouts are the same. But Fuego can play 5 times more playouts per second
> on the hardware that results in Fuego being slightly stronger than Valkyria
> at the moment.


The search is good at some things and poor at other things.   These 2
aspects need to work together in a complimentary way and in my opinion will
mean more domain specific knowledge.     Your observation concerning Fuego
and Valkyria indicate that there is a lot of overlap - you can make up
search with knowledge and knowledge with search.

I believe the huge lack of progress over the years (until recently) has been
the blind failure to recgonize that you cannot cover everything with
knowledge but we must not move too far in the other direction either.
Domain specific GO knowledge is too brittle to do it all,  but should be
provided as hints to a search.   That is exactly how us humans do it.   We
try to back up our positional judgement and knowledge with concrete
analysis.    When the knowledge, understanding and intuition is strong, less
analysis is needed and visa versa.

I have had many discussions with Larry Kaufman,  who works with the Rybka
chess team - currently the strongest chess program in the world.   I think
some of the things we have talked about applies very much to GO.     It is
very interesting to me that even though he is heavily involved in the
knowledge engineering aspect of the program,   he seems to feel that Rybka
is confined by the knowledge part.    He tells me (and this applies to ALL
programs, not just Rybka),  that there are certain positions, ideas or
themes where Rybka is a victim of it's evaluation function.   Adding an
additional order of magnitude more time to the search is not going to change
the basic misconception if the evaluation just doesn't understand the
position.     So you can have 2 equally strong chess programs that are
playing stronger than any human player,  choosing a different move in the
same position and one can be correct and the other wrong - and the one that
is wrong may not find the correct  move even thinking 10X longer.

That is a pretty depressing thought for GO programming because if it's a
problem in Chess, then it can only be worse for GO.   So I suspect that from
your description of the differences between Valkyra and Fuego  there will be
huge differences in which positions Valkyra plays better vs Fuego.

One thing I have found in chess is that each piece of knowledge has
side-effects.   Every new "rule of thumb" that you impose, makes your
program a bit more dogmatic.    I believe the trick is figuring out how not
to make it too dogmatic,  while  still giving it a sensible set of
heuristics to guide it along.


>
>
> It just cannot be that having strong playout leads to worse play in
> general. It is just a matter of not slowing down the code too much and add
> just those heavy elements that do increase playing strength.


Yes,  I guess I just repeated you but in a different way.

- Don




>
>
> -Magnus
>
> --
> Magnus Persson
> Berlin, Germany
>
> _______________________________________________
> computer-go mailing list
> computer-go@computer-go.org
> http://www.computer-go.org/mailman/listinfo/computer-go/
>
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Reply via email to