Tapered Evaluation Questions

[jayakiran]

Hello!

I had a few of questions about tapered evaluation -

1. Do all the good engines use tapered evaluation? Are there any good ones that don't use tapered evaluation?
2. If you are using tapered evaluation, do you have different piece values for middle-game and end-game? Or do you build the tapering into piece square tables or other feature evaluation that distinguishes between middle-game and end-game? I realize that it shouldn't matter in the end if you split the points between piece value and feature/pst evaluation.... I'm wondering how people prefer to do this.
3. How does your tapered evaluation interact with SEE? I realize that SEE is more of a binary information (good capture or bad capture) - so maybe it's irrelevant. But since I do some move ordering based on piece square tables and SEE, wanted to know how others are doing it.

Thanks in advance!

Jayakiran

[cdani]

Hello!

I had a few of questions about tapered evaluation -

1. Do all the good engines use tapered evaluation? Are there any good ones that don't use tapered evaluation?
2. If you are using tapered evaluation, do you have different piece values for middle-game and end-game? Or do you build the tapering into piece square tables or other feature evaluation that distinguishes between middle-game and end-game? I realize that it shouldn't matter in the end if you split the points between piece value and feature/pst evaluation.... I'm wondering how people prefer to do this.
3. How does your tapered evaluation interact with SEE? I realize that SEE is more of a binary information (good capture or bad capture) - so maybe it's irrelevant. But since I do some move ordering based on piece square tables and SEE, wanted to know how others are doing it.

Thanks in advance!

Jayakiran

Hi.

1-Yes the ones that I know.

2-Is the same but is not the same, because you want to tune pst and piece value separately. Maybe later you will use the piece values for something, like pruning decisions or whatever.

3-It's irrelevant to SEE, as only piece values are involved.

[AlvaroBegue]

Do all the good engines use tapered evaluation?

Yes.

Are there any good ones that don't use tapered evaluation?

No.

Also, tautology club meets on tautology club meeting days.

[kbhearn]

Hello!

I had a few of questions about tapered evaluation -

1. Do all the good engines use tapered evaluation? Are there any good ones that don't use tapered evaluation?
2. If you are using tapered evaluation, do you have different piece values for middle-game and end-game? Or do you build the tapering into piece square tables or other feature evaluation that distinguishes between middle-game and end-game? I realize that it shouldn't matter in the end if you split the points between piece value and feature/pst evaluation.... I'm wondering how people prefer to do this.
3. How does your tapered evaluation interact with SEE? I realize that SEE is more of a binary information (good capture or bad capture) - so maybe it's irrelevant. But since I do some move ordering based on piece square tables and SEE, wanted to know how others are doing it.

Thanks in advance!

Jayakiran

Most engines atm taper every term in their evaluation - they maintain for the bulk of evaluation scores for both phases in one integer such that keeping track of them is effectively free and then do the weighted combination of them at the end of the eval function. You just have to be careful your eval terms cannot cause an overflow/underflow of the range of bits you've assigned them or you'll wrap the score on yourself for that term.

One author working on a fork of gull posted here recently about trying to use 4 phases instead of the now standard 2 (since 16 bits is plenty to maintain a score in centipawns).

SEE is unrelated and just a crude approximation of whether a move is worth considering.