A tale to be told

[BB+]

rn1q1r2/p1p1ppkp/1p4p1/8/2BP4/2N2N2/PP3PPP/R1B1K2R b KQ - 0 11
Three minors versus a Queen is never an easy situation (the Queen usually needs targets like weak pawns, or maybe an open board for forks). IvanHoe has -36 centipawns as the material imbalance:

Material [index 360718]: RRBBNNPPPPPPqrrnppppppp -> BBNqp
values: 10:-32:-101:-140 phase 23 -> -36

Rybka 3 appears to have -255 millipawns for this imbalance.

The overall eval() of this position is 19cp for Rybka 3, and is 25 for IvanHoe (from White's perspective in both cases).

However, a 30s search put Rybka 3 as favouring White by ~40cp, while IvanHoe calls it about equal. The reason for this, from what I can determine, seems to be in the difference in king safety evaluation in subsequent positions. For instance, after the IvanHoe PV 11...Nd7 12. O-O Nf6 13. Be3 Qd6 14. Ne5 a6 15. Rad1 b5 16. Bd3 Rac8, there are 3 White pieces attacking a square around the Black King (Be3/Bd3/Ne5). Rybka 3 values these attackers at approximately 95.4 cp (then multiplied by a 17/24 factor for the game phase, the calculation to get 954 being MemLoc0x582960[3]*530/10), while IvanHoe values them at 50 cp (as KingSafetyMult[3]*45/8, with then also a 23/32 game phase factor). This seems to me to be the principal reason behind the disparity where Rybka 3 gives an eval() of this terminal position of 79 cp, while IvanHoe only 15 cp. In general, I could comment that IvanHoe appears to put greater relative weight on majors attacking the opponent King (see Appendix A.2.2 of the R3/IPPOLIT report for brief data).

[BB+]

I am factually wrong on at least 3 points.

• The "-255" in the Rybka material imbalance table is from a neighbouring entry.

• In any event, the values in the Rybka table are imbalances, whereas I read this "-255" as a material value.

• Despite the "eval" output staring me in the face, I forgot to append that IvanHoe multiplies King safety by a further 6/8 in the given terminal position (due to not having a Queen).

The correct R3 table entry appears to be 0x793, or 1939 millipawns. The "base value" of BBN vs QP is -2000 (nine minus eleven), for a net material value of -61 millipawns, compared to -36 cp with IvanHoe.

I still suspect King safety differentials have a good effect on the comparative search output of Rybka 3 and IvanHoe.

The comparative Rybka 2.3.2a material value appears to be -303/3399 pawns, or about -8.9 centipawns. The attack value from BBN seems that it would be (KingAttack[3]*1777)/32 or 3165/3299 of a pawn -- except that Rybka 2.3.2a requires an attacking side to have a Queen for there to be any King danger penalty.

Sorry about any confusion.