expected number of nodes to finish searching a root move

[Uri Blass]

I suggest the following experiment.

Take many positions from games and tell your program to analyze
them for 1,000,000 nodes
suppose that after 1,000,000 nodes you are at iteration number i and search root move X


You have the following information that you write in a table
a)n-number of nodes that you searched move X at iteration i-1

b)m-number of nodes that you already searched move X at iteration i and did not finish to search

c)c-the specific case of searching after 1,000,000 nodes that can be one of the following

1)X is the move that the program wants to play in iteration i and i+1.
2)X is not the move that the program wants to play in iteration i and i+1.
3)X was not the move that the program wanted to play in iteration i but it failed high at iteration i+1 before it got 1,000,000 nodes
4)X was the move that the program wanted to play at iteration i but another move failed high at iteration i and now it is not the first move that the program search at iteration i+1

d)Let A be number of additional nodes that you need to search until you finish searching move X at iteration i
continue the search and find A.


The target is to have an estimate E for A that we do not know after 1,000,000 nodes based on n,m,c

We basically have for every position P n(P) m(P) c(P) and A(P)
when we can estimate the quality of a specific function
E(P)=E(n(P),m(P),c(P)) based on the sum of squares on many positions (E(P)-A(P))^2

[bob]

I suggest the following experiment.

Take many positions from games and tell your program to analyze
them for 1,000,000 nodes
suppose that after 1,000,000 nodes you are at iteration number i and search root move X


You have the following information that you write in a table
a)n-number of nodes that you searched move X at iteration i-1

b)m-number of nodes that you already searched move X at iteration i and did not finish to search

c)c-the specific case of searching after 1,000,000 nodes that can be one of the following

1)X is the move that the program wants to play in iteration i and i+1.
2)X is not the move that the program wants to play in iteration i and i+1.
3)X was not the move that the program wanted to play in iteration i but it failed high at iteration i+1 before it got 1,000,000 nodes
4)X was the move that the program wanted to play at iteration i but another move failed high at iteration i and now it is not the first move that the program search at iteration i+1

d)Let A be number of additional nodes that you need to search until you finish searching move X at iteration i
continue the search and find A.


The target is to have an estimate E for A that we do not know after 1,000,000 nodes based on n,m,c

We basically have for every position P n(P) m(P) c(P) and A(P)
when we can estimate the quality of a specific function
E(P)=E(n(P),m(P),c(P)) based on the sum of squares on many positions (E(P)-A(P))^2

This doesn't happen for me in any game I could find. I set a target time T, and when I reach that time, I require that the current root move be completed before the search stops. I treat this similar to a "fail low" as I want to be sure and search this move to be sure it will not become a new best move. Most of the time I finish almost instantly and stop the search. Occasionally it will use 2-3 x the target time and change its mind at the last minute.