> I can't help but feel we are missing something. With UCT we miss those > wonderful beta cutoff's that you get with straightforward alpha beta > pruning - but with alpha beta pruning you are still exploring a lot more > useless nodes. It seems the 2 are just not very compatible. > > Currently UCT does seem like the best choice by far. Yes, in the current state of the art. What would be great is to understand in what extend this would continue to be true if we change the evaluation function. What I believe (but this is only intuition, this have to be proved !), is that UCT is very efficient in Go because our evaluation function is bad. With a good/(very)^+ good evaluation function, alpha-beta would become much better than UCT. I think this is clearly the case if you can reach the end of the game in the tree for example. I don't know if someone tried UCT in chess to see the results. As we often assume that the evaluation function in chess is much better than in Go, it would be a good test case?
> [...] > But UCT may be a rather extreme example of this phenomenon - better to > go deep even without a huge number of simulations. Of course with UCT > none of the evaluations are completely wasted, they all have an impact > on potential root choices. It is an interesting point of view indeed. I did not see UCT on that angle... This would confirm our results where launching several simulations for each node (instead of node) where not improving the results (comparing at equal number of nodes. Comparing at equal number of simulations this was much worse !). Yet, one of the property of UCT which I think is very useful in Go is the "smooth" pruning. The other is that the value is between the mean and the min-max in a smooth manner, while keeping the convergence toward the min-max. Sylvain _______________________________________________ computer-go mailing list [email protected] http://www.computer-go.org/mailman/listinfo/computer-go/
