Date: October 24, 2023
While working on the more important additions to the engine, I took the opportunity to tweak a few aspects.
My alpha-beta search function is currently called “negamax” and it used a “color” parameter to automatically update (via a -color expression) the player associated with the next search depth.
I have removed it from the list of parameters and now the color is determined within the function by querying the board object. Overall, the search seems to be a bit faster.
The Principal Variation Search starts with a null window search for the PV nodes, the important nodes for which we need an exact evaluation. A null-window search is faster than a full search, but sometimes it fails (the resulting value is between alpha and beta), and in this case it’s necessary to perform a full search.
I did an analysis of the values of alpha, beta, and the evaluation, and it turns out that the re-search can be made faster by increasing alpha by 1, which makes the condition that leads to the re-search a little rarer.
This change has led to more pruning (about 3%) and, apparently, no problem with the result of the search.
To be on the safe side, I implemented this change by using a new constant PVS_LESSRESEARCH = 1 in the config file. The code now looks like this:
if (USE_PVS && move_number > 0) {
// If this is not the first move, do a null window search.
val = -negamax(board, depth-1, -alpha-1, -alpha);
// If the value is within the window, do a full search.
if (val > (alpha+PVS_LESSRESEARCH) && val < beta && depth > PVS_NORESEARCHDEPTH) {
val = -negamax(board, depth-1, -beta, -alpha);
}
}
I just hope it won’t be a problem when I start seriously debugging the search over many different positions.
ViperMate chess engine by Enrico Altavilla