Most Important Evaluation Terms

[Carbec]

Only count moves to safe squares.

I don't get it. Even if a square is controlled by an ennemy pawn, Why not count it ? Because if a ennemy piece land there, Im attacking it.

[JoAnnP38]

Only count moves to safe squares.

I don't get it. Even if a square is controlled by an ennemy pawn, Why not count it ? Because if a ennemy piece land there, Im attacking it.

In general, you just don't count squares defended by enemy pawns, because the exchange already starts out as a negative. That seems to work well as a good compromise, but you are correct. If you had unlimited CPU why not count all moves that aren't losing? I also only consider piece mobility and I don't count pawn moves.

[chrisw]

Only count moves to safe squares.

I don't get it. Even if a square is controlled by an ennemy pawn, Why not count it ? Because if a ennemy piece land there, Im attacking it.

Whether you get it or not is not relevant. Only results count, and many results come from quite counter-intuitive things.
Anyway, what's to stop from testing? Texel tune and test with and without. If without offends you but performs better, well, you can choose.

[Carbec]

Hi,

No offense taken. I am currently trying to improve my evaluation, which is extremely basic at the moment.
I am doing it manually; I saw that many developpers use some sort of automatic tuning. It seems its the thing to do,
but I don't have the slightest idea to begin; but google is my friend))

[chrisw]

Hi,

No offense taken. I am currently trying to improve my evaluation, which is extremely basic at the moment.
I am doing it manually; I saw that many developpers use some sort of automatic tuning. It seems its the thing to do,
but I don't have the slightest idea to begin; but google is my friend))

Using Python.
Make a vector of all your evaluation constants.
Evaluate 10 plus million varied quiet positions (either by running them past a strong engine, or scoring them 0,1 based on eventual game result).
Any well known regression analysis algorithm will generate you a best fit.

[JohnWoe]

I only use these.

1. Mobility
2. PSQT
3. King safety
4. Bishop pairs
5. Tempo
6. Some simple endgame stuff like KPK

I think it's much better to have a fast and simple ( bug free ) evaluation than a massive w/ 1,000,000 terms which take forever to tune.
Tho king safety is the most important, but impossible to model w/ HCE. I find search and evaluation are the same. You can search deeper w/ fast and simple evaluation. Vs searching shallower with big evaluation.

[chrisw]

Only count moves to safe squares.

I don't get it. Even if a square is controlled by an ennemy pawn, Why not count it ? Because if a ennemy piece land there, Im attacking it.

Incidentally, you would pick that up (attacking an enemy piece that could land on the square) when you calculate the mobility of enemy pieces (on which it would have a negative effect).

[AngularMomentum]

Mobility is important as well.

On a slightly different note, do you know how to implement mobility efficiently? My NPS goes from 17mil to 10mil average because of calculating all the sliders in every leaf node.

[Robert Pope]

I store mob[LIGHT] and mob[DARK] from the two most recent ply where I generated moves. It may be off by a piece capture or two, but it's in the right ballpark, and there's no computational cost.

[AngularMomentum]

I store mob[LIGHT] and mob[DARK] from the two most recent ply where I generated moves. It may be off by a piece capture or two, but it's in the right ballpark, and there's no computational cost.

Is this a response to my comment? If so, put a quote so it's more clear.

Wouldn't every leaf node stemming from the same depth 1 node have the exact same mobility value for the enemy (The mobility for the side to move is already calculated because the quiscence search always makes sure that there no captures availivable before evaluating so we can get each piece's attack maps).