Crafty Search Extensions

[GothicChessInventor]

I have an old copy of Crafty with the following code:

Code: Select all

#define LimitExtensions(extended,ply)\
            extended = Min(extended,ply);
            if(ply > 2 * iteration_depth) {
              if(ply < 4 * iteration_depth)
        extended = extended*(4*iteration_depth - ply)/(4*iteration_depth);
        else
        extended = 0;
        }

I have been tweaking this a bit on my own (primarily to hunt for cutoffs and also to curtail extensions when hunting for mates) and I introduced some horrible bug of course!



I am wondering, has there been an "update" to this section of code in the latest version of Crafty? If so, I'd enjoy reading about the new code and what it does.

[bob]

This is what we currently are using:

# if !defined(LIMITEXT)
#define LimitExtensions(extended,ply) \
extended=Min(extended,PLY); \
if (ply > 2*shared->iteration_depth && !tree->no_limit) { \
if (ply <= 4*shared->iteration_depth) \
extended=extended*(4*shared->iteration_depth-ply)/ \
(2*shared->iteration_depth); \
else \
extended=0; \
}
# else
#define LimitExtensions(extended,ply) \
extended=Min(extended,PLY); \
if (ply > 2*shared->iteration_depth) { \
if (ply <= 4*shared->iteration_depth) \
extended=extended*(4*shared->iteration_depth-ply)/ \
(2*shared->iteration_depth); \
else \
extended=0; \
}
# endif

Note the conditional compilation option that is an either/or sort of compile-time switch. We have some times of moves that now get extended without the "limit" being imposed...

[GothicChessInventor]

I have to hand it to you. I tried experimenting with all kinds of numerator/denominator combinations for the search extension multipliers, I even simulated it in an Excel worksheet, and you've got the implementation with the best numerics!

I thought for sure ratios of 3.0 and 1.5 would outperform 4 and 2 since it seemed like there were some pointless extensions going on, but statistically, 4/2 is better than 3/1.5

Very nice.

[bob]

I have to hand it to you. I tried experimenting with all kinds of numerator/denominator combinations for the search extension multipliers, I even simulated it in an Excel worksheet, and you've got the implementation with the best numerics!

I thought for sure ratios of 3.0 and 1.5 would outperform 4 and 2 since it seemed like there were some pointless extensions going on, but statistically, 4/2 is better than 3/1.5

Very nice.

we run lots of tests. null move R=2~3 (adaptive null move) is exactly the same deal. I have tried every combination from 1~1, 1~2, 2~2, ..., 5~5. 2~3 is the clear winner. Most everything I have done has been tested like that. I just re-tested the null-move settings as above to be sure that nothing has changed since the last time I tried, such as faster hardware, deeper depth, search reductions. 2~3 is still better by a significant margin.

Ditto for the extension amounts we are using...