One more question. Line 23 states: for i:=node.size()-2 to 0 do. The leaf node should be stored in node[node.size()-1], so why do we start at node.size()-2? Is it not necessary to update the value of the leaf node?
i:=node.size()-1 would be better you're right :-). Experiments made by people in this list (Don if I remember correctly), showed that you even don't have to create the leaf if the parent node has less than a few simulations. 100 simulations seemed to be already safe, and much faster. So the modification in the algorithm, is to stop the descend when the number of simulations of the node is "< threshold" rather than ending on a unseen node. Sylvain
-----Original Message----- From: Don Dailey <[EMAIL PROTECTED]> To: Dmitry Kamenetsky <[EMAIL PROTECTED]>, computer-go <computer-go@computer-go.org> Date: Wed, 21 Feb 2007 12:54:43 -0500 Subject: Re: Re[2]: [computer-go] UCT vs MC > > On Wed, 2007-02-21 at 16:56 +0300, Dmitry Kamenetsky wrote: > > Thank you for your answer. However, I am even more confused now. I > > understand that "-" is for negamax, but I don't understand why it > > became "1-". I am trying to implement your algorithm and I just want > > to know what lines 7, 16 and 26 should be? > > I'm not sure this is what you are looking for, but in negamax, scores > can be negative or positive. The scores are always adjusted so that > you can view positive numbers as "good" and negative as "bad" from the > point of view you are referencing. So to get the score from the > "other" > point of view you simple negate it. > > But in UCT, we don't deal with negative numbers. A score is between > 0 and 1, so 0.001 is almost losing and 0.999 is almost winning for > example. > > To change 0.99 to the other players point of view in this system, where > scores must be between 0 and 1, you must negate it and add 1. So 0.99 > becomes: 1 - 0.99 = 0.01 > > I hope that is what you are asking about and that this explains it. > > - Don > >
_______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
