Re: computer-go: A problem with understanding lookahead

[Don Dailey <drd@xxxxxxxxxxxxxxxxxxxxx>]

All  bets  are off  when    selectivity  is considered.  My    thought
experiment is not valid.

Don


   From: Joan Pons Semelis <joan@xxxxxxxxxxxxxxxxx>

   From: "Mark Boon" <tesuji@xxxxxxxxxxxxxxxxx>

   [snip]
   > Did anyone do any research on this? I mean, trying to quantify theoretically
   > how much the evaluation has to be wrong to get a less reliable tree at a
   > given depth. It might possibly be that it's only the case for really very
   > unreliable functions, which will of course be the case for Go. That would be
   > very useful information.
   > 
   > Mark


   My experience with very simple evaluations but with rather well tuned
   move selectors for the given evaluation (so no full width) show good
   improving with depth, improving the evaluation without modifiying
   the move selector gives worst play till the move selector is tuned
   again. Widening the search also gives better play to a certain point
   where the 'filter' of the selector lets pass moves that the evaluation
   mishandles.

   Joan