[computer-go] Some thoughts about All-as-first

[Łukasz Lew]

You may do weighted average I.e. playout of moves
m1 m2 m3 m4 m5 m6 m7 ...

updates statisic of move m1 with weight 1 (I guess You are using
weighted average)
updates statisic of move m2 with weight 0.9
updates statisic of move m3 with weight 0.8
...

many decay schemes are possible. geometric, linear, etc.
You are proposing stair: 1,1,1,1,1,1,0,0,0,...

Good luck.
Łukasz

On 9/12/06, Don Dailey <drd at mit.edu> wrote:
> It occurred to me to use the idea gradually.   Start with all-as-first
> and gradually, as more simulations are run, taper back to
> "most-as-first", then "a-few-as-first" etc.    Eventually you will
> consider only the first move in the simulation as relevant.
>
> The only value of this idea is when you don't run enough simulations -
> but until that point it's pretty powerful.
>
> For the more sophisticated programs like yours and mine,  I'm not sure
> how to integrate it, but it's probably possible and may be a big win -
> it's one of those things I haven't looked at yet.   It's a little tricky
> when you are building a tree already.
>
> - Don
>
>
>
>
> On Tue, 2006-09-12 at 15:06 +0200, Magnus Persson wrote:
> > Quoting Don Dailey <drd at mit.edu>:
> >
> > > I don't have an easy way to test where the ceiling is - I only suspect
> > > it's between 1500 and 1600.    All as first gets you closer if you are
> > > not doing a LOT of simulations, but probably limits you in the long run,
> > > perhaps at higher levels than we can achieve in practical play.    So
> > > your experiment will help us understand that.
> >
> > Previously I never experimented with the All-as-first-heuristic but I ran some
> > tests with a simple MC-search some days ago and became convinced that it was
> > indeed useful. The trick seems to be to strike a balance between move ordering
> > and move selection. The heuristic is very good for move ordering (if
> > you do not
> > have any prior ideas of which moves might be good) but at some point one must
> > stop using it and stick with First-move-only. It is not clear to me how to do
> > this best since it depends on the search in general.
> >
> > For a basic 10000 simulation 1-ply search one might do the following. Run 2000
> > simulations and update the score for every legal first move with All-as-first.
> > Now one can prune X percent of the worst moves, and perform a First-only eval
> > with the remaining 8000 simulations. I could do those experiments myself, but
> > right now I have no time since I am working on Valkyria so much.
> >
> > In Valkyria I do something different and more complex but the search now
> > actually prunes a lot of bad moves without even searching them once as First.
> > But I have not done any rigorous testing wether it is safe or not. But to my
> > eyes it really looks good. The latest version of Valkyria is much stronger
> > (about 100 Elo) now but that is because I added some tactics to the pseudo
> > simulation code, and need to run some proper experiments to establish if the
> > search is more effective now.
> >
> > -Magnus
> > _______________________________________________
> > computer-go mailing list
> > computer-go at computer-go.org
> > http://www.computer-go.org/mailman/listinfo/computer-go/
>
> _______________________________________________
> computer-go mailing list
> computer-go at computer-go.org
> http://www.computer-go.org/mailman/listinfo/computer-go/
>