The key point, good or bad, is that I allocate all the legal children nodes as a group. I manage my own memory pool and just hand out nodes sequentially. I don't need to maintain pointers to each child, only a single pointer to the first child. So instead of needing 4 extra bytes per node for pointers to the data, I need something that amortizes to something near zero.
Your method is to allocate 1 node when it's been visited once or twice - very natural I agree. My method is to allocate all the children at once, and wait until the parent has been visited some number of times (currently 100). If there are 50 legal moves, that gives on average about 1 node allocation every 2 visits, which is what you said you do. Of course my allocation is not exact like yours. You expand exactly once every 2 visits, but mine is just an average, some nodes may get expanded after only 1 visit and others may get visited several times before getting expanded. I don't really know if any of this matters much. - Don Petr Baudis wrote: > On Mon, Mar 10, 2008 at 05:36:14PM -0700, Christoph Birk wrote: > >> On Mon, 10 Mar 2008, Christoph Birk wrote: >> >>> On Tue, 11 Mar 2008, Petr Baudis wrote: >>> >>>> On Mon, Mar 10, 2008 at 06:57:07PM -0400, Don Dailey wrote: >>>> >>>>> I think you may still have a bug. You should get well over 1700 with >>>>> 110,000 playouts, even if they are light playouts. >>>>> >>> I will run myCtest with 110k-playout, c=0.25 and node creation >>> after the 2nd visit ... let's see what its ELO rating will be >>> in a couple of days. >>> >> Sorry, I just realized I cannot do 110k playouts because my >> implementation is too slow. >> I suggest you run a 'pachi-0.25-light-50k' that just uses >> 50000 playouts. That way you can compare it to 'myCtest-V-0003'. >> > > Sure, it now won its first game as pachi1-p0.25-li50k. It will be > interesting how much difference number of playouts makes for me too. > > >> BTW: I count the new-node threshold like Don from the parent >> node, so 50 not far from your '2'. >> > > Hmm, I really wonder where this comes from, the idea seems quite > unnatural to me. > > _______________________________________________ computer-go mailing list [email protected] http://www.computer-go.org/mailman/listinfo/computer-go/
