vickalan wrote:But that part of the result is confusing to me. Guards are worth about average of (knight + bishop), so this would be around 278. But to help their army win the most often they do best when their value is set to 375? So what is their optimal exchange value (to win most games)?
There are two ways a piece can help the player having it win:
1) By inflicting damage to the opponent because it is so powerful that the opponent cannot escape the damage, even though he tries.
2) By fooling the opponent into unforced self-destruction, because he sacrifices valuable pieces to capture the piece-under-test, overrating it.
E.g. Pawns would make the side having them win most often when you set their value to 900. Because the opponent would immediately start to sacrifice his Queens, Rooks and other pieces for the Pawns. Which you cannot possibly prevent. If the opponent has no Pawns, the other side cannot make the same mistake. So when an imbalace of 6 Pawns versus a Rook you would get a 50% score (say) with P=100, you probably would score 100% with a Pawn value of 900, because the side with the Rook would sacrifice 6 more of his pieces for Pawns. But that does not mean there is any reality in a Pawn value of 900.
The game results show that the Guard is worth between N and B (so less than B), even when the opponent thinks it is worth more (and the Guards thus inspire suicidal Bishop tactics). Part of the score is undeserved, brought about by B for G trades that the B side could have easily avoided if he hadnod been deluded to think B is worth more than G. The game result obtained when you set the G value to 278 is less corrupted, because it elimiates the unnecessary Bishop sacs. But the resulting score suggests an even lower value for the Guard.
In summary: in self-play you should believe the game result, not which programmed value does best. If you want to optimize by varying the programmed value, you should do it when playing against an opponent that uses a fixed value, in a symmetric start position. E.g. it would immediately be obvious that a Pawn value of 900 is madness, when you play a standard FIDE game against an opponent that knowsit is more like 100. So you could try games where both sides have Guards, and one thinks they are worth 259, and the other thinks 350, and then measure how they fare. (You would have to make a copy of Fairy-Max for that.)
I notice in Capablanca chess All pieces except pawns have their values set higher than in 8x8 games. Would there be any reason to changing all my piece values to those of Capablanca chess (when testing on a 10 x 8 board)?
For historic reasons the values in normal Chess are lower than usual. This was because Fairy-Max was derived from micro-Max, which was developed for the goal of making the source code as small as possible. For Capablanca Chess I used a more conventional value scale.
On 10x8 it is better to use Capablanca values.