Hmm. Do you know that Leela does something special here? When I look at Leela's analysis output it seems to the search seems not to consider the ladder escape because the policy net assigns a low probability to it (and such a high probability to move in the upper right). Which is the same as in other non-ladder situations when the policy net puts 90+% weight on one move and almost nothing on other moves. If you do get it to search the ladder escape at all, it reads it correctly and likes it.
So It seems to me the issue is primarily the policy net giving too low of a weight to the ladder escape. And the policy net doesn't read, because it's a neural net - I'd expect the most it would be getting is a feature plane that says "is this stone ladder-capturable" performed via a simple recursive search completely independent from the MCTS that shouldn't misread a ladder this simple unless it's outright buggy. On Mon, Apr 17, 2017 at 11:25 AM, Stefan Kaitschick <skaitsch...@gmail.com> wrote: > > On Mon, Apr 17, 2017 at 3:04 PM, David Wu <lightvec...@gmail.com> wrote: > >> To some degree this maybe means Leela is insufficiently explorative in >> cases like this, but still, why does the policy net not put H5 more than >> 1.03%. After all, it's vastly more likely than 1% that that a good player >> will see the ladder works here and try escaping in this position. > > > > I think it's likely that the ladder is actually the first thing that Leela > considered. My guess is that it put a penalty on the move after misreading > the ladder. > > _______________________________________________ > Computer-go mailing list > Computer-go@computer-go.org > http://computer-go.org/mailman/listinfo/computer-go >
_______________________________________________ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
