I know that I've missed at least two enhancements: 1. UCB1-Tuned (adding an upper confidence bound on the variance used in calculating upper confidence bounds) 2. First play urgency - Giving an artificial upper confidence bounds to untried moves (I've seen references that 110% win rate is the right setting)
On 10/19/07, Jason House <[EMAIL PROTECTED]> wrote: > > I've only recently implemented my first attempt at UCT and I'm curious > what tricks exist for tweaking performance. > > My rule for promoting a leaf to an interior node is that I must first have > 100 sims of that node, but changing that value to 10 seems to give very > significant performance improvements. (20x-30x sims for "best" move). What > experiences do others have with this? > > Other candidate improvement I've heard of: > 1. Using AMAF/RAVE for initial estimates of winning percentages. This > seems like it'd give a good speed enhancement that would likely offset > estimation errors in the AMAF estimates > 2. Enhancing quality of random games with 3x3 patterns (something I > consider out of scope for what I'm currently working on but likely way too > significant of an enhancement to not mention) > 3. 1ply pruning heuristics (I believe this is what's done by crazy stone. > I think crazy stone does soft pruning). > 4. Heuristics to avoid simulation of all leaves when promoting a leaf node > to an interior node. (I've seen a Mogo paper on this) > 5. Dynamic adjustment of exploration coefficient (I've seen a Mogo papery > on this, but not much discussion on this mailing list) > > > Am I missing any other ones? What experience do people have playing with > these? >
_______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
