Toga Knight Outpost

[Michel]

52,75% would be a very good result for a relatively small change!

[lucasart]

52,75% would be a very good result for a relatively small change!

Indeed: I would be very surprised if the entire knight/bishop outpost code it even worth 52.75% (with or without the refinement compared to Toga with knight/bishop outpost removed).

[jd1]

Yes, indeed, 52.75% is nearly 20 elo - so like Lucas said I doubt outposts alone are worth that much. However, it is self-play which may exaggerate. Nevertheless, Ben, do you think you could post the code? I would like to test it myself to see whether I can validate your results.

Thanks,
Jerry

[Tennison]

To verify, I retried the test of the DV1.65.

Same conditions & 5000 games.

DV1.65 - Test 2 : 52,7%


I just try this night first version of modified bishop eval.

First try is - it's logical - not good :

DV1.66 : 52.1%

[Tennison]

It's not only Knight Outpost. It's all the Knight Eval in Eval_Piece that is changed.

New Outpost and New Mobility.

[Tennison]

As I said, I'm not a good programmer and code should be optimized maybe.
But the algorythm seems ok.
Maybe needs some tunings again with values.

Original Code :

Code: Select all

            // mobility

            mob = -KnightUnit;

            mob += unit[board->square[from-33]];
            mob += unit[board->square[from-31]];
            mob += unit[board->square[from-18]];
            mob += unit[board->square[from-14]];
            mob += unit[board->square[from+14]];
            mob += unit[board->square[from+18]];
            mob += unit[board->square[from+31]];
            mob += unit[board->square[from+33]];

            op[me] += mob * KnightMobOpening;
            eg[me] += mob * KnightMobEndgame;


			// outpost
			mob = 0;
			if (me == White){
				if (board->square[from-17] == WP)
					mob += KnightOutpostMatrix[me][SquareTo64[from]];
				if (board->square[from-15] == WP)
					mob += KnightOutpostMatrix[me][SquareTo64[from]];
			}
			else{
				if (board->square[from+17] == BP)
					mob += KnightOutpostMatrix[me][SquareTo64[from]];
				if (board->square[from+15] == BP)
					mob += KnightOutpostMatrix[me][SquareTo64[from]];
			}

			op[me] += mob;

DV1.65 Code :

Code: Select all

            mob = -KnightUnit;
            defender=0;
            attacker=0;
            possible_attacker=0;
            KOM_multi=0;
            no_attack=0;

             if (me == White){

               /// Safe Mobility
               if ((board->square[from-18] != BP && board->square[from-16] != BP))    mob+= unit[board->square[from-33]];
               if ((board->square[from-16] != BP && board->square[from-14] != BP))    mob+= unit[board->square[from-31]];
               if ((board->square[from- 3] != BP && board->square[from- 1] != BP))    mob+= unit[board->square[from-18]];
               if ((board->square[from+ 1] != BP && board->square[from+ 3] != BP))    mob+= unit[board->square[from-14]];
               if ((board->square[from+48] != BP && board->square[from+50] != BP))    mob+= unit[board->square[from+33]];
               if ((board->square[from+46] != BP && board->square[from+48] != BP))    mob+= unit[board->square[from+31]];
               if ((board->square[from+33] != BP && board->square[from+35] != BP))    mob+= unit[board->square[from+18]];
			   if ((board->square[from+29] != BP && board->square[from+31] != BP))    mob+= unit[board->square[from+14]];

               /// Outpost
               if (board->square[from-17] == WP) defender+=1;
               if (board->square[from-15] == WP) defender+=1;

               if (defender>=1)
                    {
                    if (board->square[from+17] == BP)  attacker+=1;
                    if (board->square[from+15] == BP)  attacker+=1;
                    if ((board->square[from+33] == BP) && (board->square[from+17] != WP)) possible_attacker+=1;
                    if ((board->square[from+31] == BP) && (board->square[from+15] != WP)) possible_attacker+=1;
                    }
            }
            else{
               if ((board->square[from-48] != WP && board->square[from-50] != WP))    mob+= unit[board->square[from-33]];
               if ((board->square[from-46] != WP && board->square[from-48] != WP))    mob+= unit[board->square[from-31]];
               if ((board->square[from-33] != WP && board->square[from-35] != WP))    mob+= unit[board->square[from-18]];
               if ((board->square[from-29] != WP && board->square[from-31] != WP))    mob+= unit[board->square[from-14]];
               if ((board->square[from+18] != WP && board->square[from+16] != WP))    mob+= unit[board->square[from+33]];
               if ((board->square[from+16] != WP && board->square[from+14] != WP))    mob+= unit[board->square[from+31]];
               if ((board->square[from+ 3] != WP && board->square[from+ 1] != WP))    mob+= unit[board->square[from+18]];
               if ((board->square[from- 1] != WP && board->square[from- 3] != WP))    mob+= unit[board->square[from+14]];

               if (board->square[from+17] == BP) defender+=1;
               if (board->square[from+15] == BP) defender+=1;

               if (defender>=1)
                    {
                    if (board->square[from-17] == WP)  attacker+=1;
                    if (board->square[from-15] == WP)  attacker+=1;
                    if ((board->square[from-33] == WP) && (board->square[from-17] != BP)) possible_attacker+=1;
                    if ((board->square[from-31] == WP) && (board->square[from-15] != BP)) possible_attacker+=1;
                    }
            }

            if (defender==1) KOM_multi+= 8;
                else if (defender==2) KOM_multi+= 15; // 1.65 = 15

            if (attacker == 1) no_attack-= 6;
                else  if (attacker == 2) no_attack-= 8;
                else
                      {
                       if (possible_attacker == 1) no_attack-= 3;
                       else  if (possible_attacker == 2) no_attack-= 4;
                      }

            op[me] += (mob * KnightMobOpening) + (KnightOutpostMatrix[me][SquareTo64[from]]*KOM_multi/8)  + no_attack;
            eg[me] += mob * KnightMobEndgame;

/// From Toga 1.3.4 ----------------------------------

            static const int EndgamePawnKnight = 10; 
            static const int EndgamePassedPawnKnight = 10;

            penalty = 0;
			if (((board->pawn_file[opp][FileA] | board->pawn_file[opp][FileB] | board->pawn_file[opp][FileC]) != 0)
				&& ((board->pawn_file[opp][FileF] | board->pawn_file[opp][FileG] | board->pawn_file[opp][FileH]) != 0))
				penalty -= EndgamePawnKnight;

			if (((pawn_info->passed_bits[opp] & 224)) != 0 && ((pawn_info->passed_bits[opp] & 7) != 0))
				penalty -= EndgamePassedPawnKnight;

            eg[me] += penalty;
/// End of Toga 1.3.4 ----------------------------------

The add-on from Toga 1.3.4 is the only change between DV1.62 and DV1.65 (+0,4%)

[Tennison]

I made new tests still being played ...

The first and second tests with DV1.65 gave 52,75% and 52,7% as I said.
The Time Control was 10ms/move. 5000 games each test. AMD Processor.

The current test on AMD is 10" + 0,1"/move, 1000 games.
Partial results give 53,1% after 661 games.

The current test on Intel is 100ms/move, 10.000 games.
Partial results give 53,34% after 2198 games.

As we can see, results seems coherents.

I hope a first released version near Christmas. This will be my gift ...

[jd1]

Hi Ben,

Have tried the code myself - I didn't run a long test but was also able to get significant (~25) improvement at bullet. So congratulations!

Curiously enough your code makes Toga significantly slower, however the increased accuracy of the knight mobility evaluation appears to more than compensate.

Jerry

[lucasart]

Hi Ben,

Have tried the code myself - I didn't run a long test but was also able to get significant (~25) improvement at bullet. So congratulations!

Curiously enough your code makes Toga significantly slower, however the increased accuracy of the knight mobility evaluation appears to more than compensate.

Jerry

In a bitboard based engine,. that code would be so much more concise and elegant (as well as much faster)...

[jd1]

Agreed. Another advantage of bitboards IMO is that evaluation of threats (loose, hanging pieces, etc.) is a lot simpler and more efficient.