Random Aggressive LMP

[xr_a_y]

How about adding a bit of randomness and aggressiveness to LMP ?

Code: Select all

// classic LMP
if ( isQuiet && quietCount > lmpLimit[depth] ) continue;
// Aggressive Random LMP
const int percentLimit = 66; // to be tuned
const int aggressiveDivisor = 2; // to be tuned
if ( isQuiet && quietCount > lmpLimit[depth]/aggressiveDivisor && random(0,100) > percentLimit) continue;

[amanjpro]

Doesn't this make the engine potentially blind from time to time?

[xr_a_y]

Or maybe this way

Code: Select all

         // aggressive random pruning
         if (depth < lmpMaxDepth && isQuiet && randomInt<int,2909>(0,100) > lmpLimit[depth]/2 + 100 - validQuietMoveCount) {
            continue;
         }

[xr_a_y]

ok I tested quite a lot of tuning on this but nothing seems to make a big difference.
I'm now trying "aggressive random LMR" in the form of

Code: Select all

            // aggressive random reduction
            if (randomInt<int,2909>(0,100) > 100 - 5*validQuietMoveCount) {
               ++reduction;
            }

Here "5" being a constant to be tuned...

[xr_a_y]

ok I tested quite a lot of tuning on this but nothing seems to make a big difference.
I'm now trying "aggressive random LMR" in the form of

Code: Select all

            // aggressive random reduction
            if (randomInt<int,2909>(0,100) > 100 - 5*validQuietMoveCount) {
               ++reduction;
            }

Here "5" being a constant to be tuned...

Looks like I may have some Elo from "Agressive Random LMR"

Code: Select all

Score of minic_3.13_linux_x64_skylake vs minic_dev: 335 - 381 - 788 [0.485]

with this configuration

Code: Select all

if (randomInt(0,100) > 100 + SearchConfig::lmpLimit[improving][depth + pruningDepthCorrection] - 7*validQuietMoveCount) reduction++;

[PK]

Pruning random moves near the leaves improved strength of my old, private and buggy engine called Glass. To be more precise, the intention was to weaken the engine that way, and it was counterproductive towards this goal.

[If you find this information useful, please copy it somewhere outside Talkchess. There is no guarantee that you won't get blocked for half a year, the community accepting this state of affairs]

[xr_a_y]

Let me summarize my experiment here :
- RALMP : no gain but nothing bad either. It seem I cannot find any good parameter set but in general no harm.
- RALMR : at short TC I found a working set of parameter to get 10Elo but this does not seem to hold at longer TC

[BTW, the forum is also inaccessible from France for a long time now, I always have to go through my canadian proxy.]..