Thanks for the replies! On Mon, Jul 24, 2017 at 9:30 AM, Gian-Carlo Pascutto <g...@sjeng.org> wrote:
> On 23-07-17 18:24, David Wu wrote: > > Has anyone tried this sort of idea before? > > I haven't tried it, but (with the computer chess hat on) these kind of > proposals behave pretty badly when you get into situations where your > evaluation is off and there are horizon effects. The top move drops off > and now every alternative that has had less search looks better (because > it hasn't seen the disaster yet). You do not want discounting in this > situation. > > Hmm. Why would discounting make things worse? Do you mean that you want the top move to drop off slower (i.e. for the bot to take longer to achieve the correct valuation of the top move) to give it "time" to search the other moves enough to find that they're also bad? I would have thought that with typical exploration policies, whether the top move drops off a little faster or a little slower, once its winrate drops down close to the other moves, the other moves should get a lot of simulations as well. It's true that a move with a superior winrate than the move with the > maximum amount of simulations is a good candidate to be better. Some > engines will extend the time when this happens. Leela will play it, in > certain conditions. > > I know that there are ways to handle this at the root, via time control or otherwise. The case I described here is when this happens not at the root, but deeper in the tree. At the root, move B still looks much worse than A, it's merely that within B's subtree there's a newly found tactic C that is much better than A. From the root, B still looks worse than A, its winrate has recently started to rise slowly but extremely steadily (with very high statistical confidence, such that one might project it to eventually overtake A).
_______________________________________________ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
