On Fri, 2006-11-24 at 10:47 +0000, Jacques Basaldúa wrote: > Chrilly wrote: (thread was "Positions illustrative of computer stupidity") > > > With an infinite fast chip chess programs would be "infinite" > > strong. Most current Go programs would only play infinite fast. > > Its an interesting question if Monte-Carlo programs would also > > play infinite strong. > > I think it is so important, it deserves its own thread. > > I think the answer is *NO*. Monte-Carlo programs do an n-ply deep > search of nodes evaluated by a simulation. It has been well > established that the evaluation by simulation of any node > converges asymptotically to a value. (Excuse me for mentioning > what you all know.)
The current type of Monte Carlo programs are really tree searchers. They will converge on the correct value unless there are bugs. The will play perfect GO given enough time. They don't go N-ply deep either, they are variable depth and with infinite power even the tree search part goes to the end of the game. - Don > Now the question is: Do we use the infinite power to increase n? > If we do, its no longer a Monte Carlo program, its an "oracle" > because it only evaluates finished games. That's obvious. > > But _if we keep n constant_ and increase the number of simulations, > assuming current programs get enough CPU to approach the asymptotic > limit, there will be almost no difference to their current > achievements. > > If the value obtained by simulation would represent the value > of the position, the n-ply search would be unnecessary. It only > represents _the value of random play_. > > Jacques. > > _______________________________________________ > computer-go mailing list > [email protected] > http://www.computer-go.org/mailman/listinfo/computer-go/ _______________________________________________ computer-go mailing list [email protected] http://www.computer-go.org/mailman/listinfo/computer-go/
