|
BGonline.org Forums
Self-correcting rollout algorithm?
Posted By: Timothy Chow In Response To: Self-correcting rollout algorithm? (Maik Stiebler)
Date: Thursday, 10 September 2009, at 11:58 p.m.
Maybe the UCT method programmers would have some valuable experience to share.
Undoubtedly, and if I were to devote some serious effort to writing a backgammon program (not likely, I'm afraid) then I would certainly study that literature.
As someone pointed out in that discussion of computer go, however, backgammon is in the envious position of having good 0-ply evaluations available. This opens up possibilities that are not available in computer go, at least at the current state of the art.
The ideas I sketched in my last message might not be quite right, but I am confident that they're in the right direction, and that a significant improvement in bot strength will result if someone takes the time to implement an adaptive, self-correcting algorithm along the lines I sketched. After all, we know that bots are sometimes making errors during the course of a rollout, and we know that these could be reduced if the n-ply analysis during the rollout were replaced judiciously by mini-rollouts. We also know that some positions are easier than others and that the current standard rollout strategy is not optimally allocating computer time to positions according to how much time they "deserve." So there is a clear path here to an improved algorithm, and the only question is how much improvement can be expected. My personal prediction is that it will be significant---though only time will tell.
|
BGonline.org Forums is maintained by Stick with WebBBS 5.12.