An idea on how to improve evaluations.

[voyagerOne]

The idea is simple.

After determining the benchmark ELO score by testing multiple games against an engine(s).

Run the test again...but this time remove all the bishops. Compare the ELO score with benchmark ELO to see if there is any type of correlations.

One can run the test omitting knights, rooks, queen...etc.

The idea is to find what evaluation piece seems to be weakest and adjust those evaluations.

Obviously one needs to derive an opening book with the omitted piece(s). But it doesn't have to be perfect just decent (balance) enough as the engine will play the same opening for white and black.

What you think?

[Ferdy]

The idea is simple.

After determining the benchmark ELO score by testing multiple games against an engine(s).

Run the test again...but this time remove all the bishops. Compare the ELO score with benchmark ELO to see if there is any type of correlations.

One can run the test omitting knights, rooks, queen...etc.

The idea is to find what evaluation piece seems to be weakest and adjust those evaluations.

Obviously one needs to derive an opening book with the omitted piece(s). But it doesn't have to be perfect just decent (balance) enough as the engine will play the same opening for white and black.

What you think?

I think most programmers are doing similar thing like removing certain eval feature and do some tests, if it does not weaken the engine then remove it. Parameter optimization is also interesting discussed in another thread.

[Sven]

but this time remove all the bishops. Compare the ELO score with benchmark ELO to see if there is any type of correlations.

Do you mean "remove all the bishops" (i.e., play games without any bishops on the board) or "remove all the bishop eval code"?

[voyagerOne]

@ Sven. Playing the game with no bishops on the board for both black and white.

@Ferdy. This is different. The idea is to "simplified" the game. As most testers know...simply changing/tweaking Bishops's parameter...will most likely impact other evaluations...as the different eval components are not independent of each other.

If you play games of SF-Houdini...it will be interesting to see what scores will be if there are no Bishops on the board. Do the same with knights.

[Ferdy]

@ Sven. Playing the game with no bishops on the board for both black and white.

@Ferdy. This is different. The idea is to "simplified" the game. As most testers know...simply changing/tweaking Bishops's parameter...will most likely impact other evaluations...as the different eval components are not independent of each other.

If you play games of SF-Houdini...it will be interesting to see what scores will be if there are no Bishops on the board. Do the same with knights.

I initially thought of removing on one side only. Yes this is a good idea to try and thanks. This should be more effective when testing with different range of opponents. We need to gather starting test positions for every specific themes, no bishops, no knights and others.