BGonline.org Forums

Before Stick et All ask...

Posted By: playBunny
Date: Sunday, 14 June 2009, at 1:47 a.m.

In Response To: Before Stick et All ask... (Bob Koca)

It's an idea that may eventually pay dividends but before it bears fruit there are a couple of problems.

The first would diminish with time and it's that the chances of two people wanting a rollout for the same position, match score and dice roll is slimmer than slim (unless the position comes up on a public board or is in a popular publicly available match record). As an indication of this, I have a database of moves from Monte Carlo matches (I was interested in something similar to your idea). There are about 32000 moves in it so far but only about 20 positions have occurred more than once. Almost all, unsurprisingly, are opening moves.

In fact it's exactly the difficulty of creating a databases of moves that make neural nets so powerful a technique.

The second would become more of a problem over time and that's the size of the database. A rollout wouldn't be just of a position and a dice roll but the top N possible moves, maybe only 2 or 3 but maybe as many as 10. The storage requirement for GnuBg rollout data isn't insignificant and so the database would become humongous before it started to be pay out.

Messages In This Thread

• Before Stick et All ask...
  Michael Petch -- Tuesday, 9 June 2009, at 5:03 a.m.
  • Before Stick et All ask...
    Bob Koca -- Tuesday, 9 June 2009, at 5:37 a.m.
    • Before Stick et All ask...
      playBunny -- Sunday, 14 June 2009, at 1:47 a.m.
      • Before Stick et All ask...
        Bob Koca -- Sunday, 14 June 2009, at 9:46 p.m.