A different way of summing evaluation features

[syzygy]

In practice, this may or may not work better than current evaluation functions, but in theory it's a viable technique at least.

But... you are trying to approximate an existing evaluation function as well as possible with a function having fewer terms. The OP is suggesting that evaluation features should be summed differently, which would result in a different evaluation function.

[Rein Halbersma]

In practice, this may or may not work better than current evaluation functions, but in theory it's a viable technique at least.

But... you are trying to approximate an existing evaluation function as well as possible with a function having fewer terms. The OP is suggesting that evaluation features should be summed differently, which would result in a different evaluation function.

If you don't do the N->K truncation, then it's just a change of basis in which you compute your evaluation. In any case, what I outlined above seems to me the only sensible use of using PCA in evaluation. The OP's suggestion seems not viable.

[Pio]

Although the OP's proposal is rather vague, there still could be something in it. The point is that in a root-mean-square combination the most extreme feature tends to dominate. This might actually correspond to what you need in Chess. E.g. when you have the choice between weakening your opponent's King safety or his Pawn structure by an amount that in itself would give about an equal advantage (in terms of win probability), it is probably better to go for further weakening his King safety when that was already compromised and his Pawn structure was average, and vice versa. So some non-linearity in combiningthe eval terms, like squaring them before addition, is not completely crazy.

The idea as given seems non-sensical though, as negative values for a certain term would give a total evaluation that is indistinguishable from when it had an equal positive value. Adding all terms corresponding to a certain aspect of the position, and applying a non-linear correction to that, does make sense, though. And to determine King safety from the number of individual attacks on the King neighborhood, this is actually somewhat standard.

Focus on the most imbalanced aspect, because that is most likely to decide the game. Aspects could be white and black King safety, Pawn structure, material advantage. When the opponent King safety is low, you want to lower it further at the expense of material. While when it was good, but you were ahead in material, you would probably be happy to gain that same amount of material for allowing him to increase his King safety by the same amount.

Hi!

I always like your posts and read them in whatever forum they are posted. I realized exactly the same problems that you did and my proposal was/is that when a value gets negative you transfer that term to the other sides plus account so you will not get into the trouble you described or that you get into trouble trying to taking the square root of a negative number.

There are probably better ways to do this than what I proposed and I am fully aware of that.

I just thought that it would be natural to sum up the orthogonalized (in a statistical sense) values in a weighted Euclidean way, since that is what you normally do in classical physics. Probably most people are more in economics and think in terms of additive money.

I also like your comments, seeing problems and possibilities with different ideas and not like others just saying that it is completely stupid and that I do not come up with any original ideas like some other people say.

I think my ideas are very original but I do not say that they will work

Good luck with your engines!