I think that it may be more interesting to test at fixed small number of nodes and not at fixed depth.mjlef wrote:My guess:kiroje wrote:Hi
While fiddling and rewriting(for the billionth time) this question popped into my head :
If you could only choose 5 evaluation terms, what would they be ?
It should be in identical program with regards on search etc.
/Kim Jensen
material
simple "if you have no pawns you need a rook advantage to win-divider" endgame rule (amazing what a simple rule like this catches)
mobility
king safety
pawn structure
But I do not know. From time to time I try turning off some parts of the evaluation to see what effect they have. An interesting experiment would be to take a strong program (say Fruit/Toga) and run fixed depth matches with various evaluation terms turned off, then see how many ELO the terms get you. This would be a worst case test, since it would negate the cost (slowdown) a rule gives. Probably 5 ply searches would do for this test, and it would not even take very long. If you do this, I would love to see the impact of various terms. You could just add ifdefs around various terms and see what happens.
Mark
Note that slow down in getting fixed depth can happen with simpler evaluation not because of speed but because of having more cutoffs.
I am not sure if it happens or maybe the opposite happens but it is better to remove this effect.
Uri