Is that correct? If I understand the method correctly then I think it is a bit less than that. A random walk within the selective part is (12 - fullDepth) nodes and in the end you add (1 + sum(i=1..fullDepth) (perft(i))) per walk from the root to the overall node count. Would you agree?Michel wrote:I am bit lost. Is it now settled which method is better?

The UCT type methods give a StdDev of about 4e12 for Perft(12) for 100e6 random walks it seems. But a random walk is 12 nodes.

I understand it the same way, based on the pseudo code snippet HGM gave where he increments his "nodes" counter unconditionally at the beginning of each node.Michel wrote:So a cost of

4e12 * sqrt(12 * 100e6)=1.4e17

The data posted by HGM for 5 ply fullwith 500e6 nodes search seems to give a StdDev of 0.012% which comes down to 7.5e12. So a total cost of

7.5e12*sqrt(500e6)=1.7e17

So in this case UCT seems somewhat better. At the cost of a much more complex implementation however.

EDIT:

I am assuming that in HGM's date "nodes" means all nodes visited. Not just leaf nodes.

If we do the count with leaf nodes then UCT is much farther ahead but this is probably unfair as one should really count the number of move generation

calls and not the number of leaf nodes.

As to your comparison of costs: shouldn't we only compare data having the same number of visited nodes?

Sven