Hello, I´m playing with MC-UCT and patterns in MC "random" simulation.
On Sunday 31 December 2006 23:02, David Fotland wrote: > I see. It seems that most of Mogo's strength is due to using pattern > sequences in the Monte Carlo random games. My MC engine got a lot better when I modified the "random" MC games with no so "random" moves. Giving more chances to play ataris and last move´s near positions. Using the previous mentioned, even without using patterns, my engine improved quite a lot. > In your paper, say that each UCT leaf node is evaluated by exactly one > random game (simulation), with a result of 0 or 1. Is this true? I think > you say Mogo does 70,000 random games per move. Does this mean that the > UCT tree for a move has 70,000 nodes? When you say 70,000 games per move, > does that mean total game move made, or game per node evaluation? I´m not sure if this helps but I will tell you what is my issue right now. As far I understand, what they do, is to have a game tree 'a la' minimax. Leaf nodes in the gametree are evaluated with MC. (instead of evaluation function in "traditional" minimax). At each node, UCT is used to decide which child to explore more to gain more confidence or discard it. UCT is fast, is a lot better than just getting the child node in the minimax tree with more win value. The difference, I think, with "classical" minimax, is that the tree is walked a lot of times, trying to gain more and more information where the UCT estimates is better. The thing is, at least what I´m facing now, is how to extend that minimax tree without hogging the engine. The "problem" with UCT is that it requires that every child node should have an initial value. So, if you try to create a big game tree you will hog the engine initialising all the child nodes with initial MC simulations. Thats my current problem, right now my game tree for UCT is only 2 levels deep. (1 ply). If I try to make it deeper, it decreases the play quality I suppose because the engine loose all the time trying to initialise leaf nodes. I hope this helps you (or somebody else in the list). BTW, My engine is playing at KGS!!! First time! Its name is tango9. If interested I can write detailed info about it. It seems its about 15k (doing only 20k MC sims) ... not good, but not bad for 1st try :-) It helped me a lot to fix a couple of not so easy to reproduce bugs. I´m very happy with the feedback. When someone makes a lot of undo´s or when 4 peoples join the game and one say "johndoe will be able to beat it" is priceless :-D :-D Eduardo __________________________________________________ Preguntá. Respondé. Descubrí. Todo lo que querías saber, y lo que ni imaginabas, está en Yahoo! Respuestas (Beta). ¡Probalo ya! http://www.yahoo.com.ar/respuestas _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
