How do you detect a useless outpost ?

[Henk]

[d]r2q1rk1/3nppbp/1p1p2p1/pPpP4/PnP1PP2/R4N1P/6P1/1BBQ1RK1 b - - 0 1

here is your useless outpost, white wins even if black makes 4 non-capturing moves in a row.

I suppose that some penalization for this knight will be good, but maybe this does not happen much to be able to achieve a measurable elo win. Anyway, I will try something related, and as other times I have done, even if is not a win (should not be a regression) I keep the parameter in Andscacs, as I consider it conceptually good and sure it helps in some positions.

thanks for not insulting me, Daniel!

probably it will fail, it is not a major term and, true, should not happen too often.

the speed decrease will hardly be offset by the value of the term in very short games.

besides, I do not have a clue how you are going to tune this to the other terms: you have to tune to kingside outposts, queenside outposts, general case, whether 3rd and 4th ranks are included, etc., etc. I absolutely have no clue how is it possible for any reasonable search or eval patch to succeed, tuning seems the crux of computer chess.

I think king safety and material are most important terms. Or not ?

[cdani]

I think king safety and material are most important terms. Or not ?

Of course 1st is material. To the top ones, where king safety lies, I would add maybe mobility.

[Henk]

I think king safety and material are most important terms. Or not ?

Of course 1st is material. To the top ones, where king safety lies, I would add maybe mobility.

There was also the question of useless mobility. For instance bishop with excellent mobility but having no effective moves. For instance can't attack a weak peace/important square.

[cdani]

I think king safety and material are most important terms. Or not ?

Of course 1st is material. To the top ones, where king safety lies, I would add maybe mobility.

There was also the question of useless mobility. For instance bishop with excellent mobility but having no effective moves. For instance can't attack a weak peace..

I obtained a little win with some penalization to bishops and knights not able to go to rivals half.

[Henk]

I think king safety and material are most important terms. Or not ?

Of course 1st is material. To the top ones, where king safety lies, I would add maybe mobility.

There was also the question of useless mobility. For instance bishop with excellent mobility but having no effective moves. For instance can't attack a weak peace..

I obtained a little win with some penalization to bishops and knights not able to go to rivals half.

And what are important squares. Squares near king, center and probably many more. Squares blocking passed pawns.

[cdani]

And what are important squares. Squares near king, center and probably many more. Squares blocking passed pawns.

Typical ones:
* of course outposts.
* in general pieces near or attacking near king with more related parameters of king safety.
* bishop pins against queen.
* any attack in front of a passed in different ways.
* penalization for queen in front of rival passer.
...
For reference you can see many open source engines.

[Henk]

Perhaps a defending knight is badly positioned unless on f3, c3, f6, c6 on pawn shield where it also protects the king.

chess will be a tough nut to crack.

tuning few terms leads us nowhere, 200-300 ply reasonable search is impossible, and, in order to at least reasonably approximate evaluation, you need to do full attack evaluation in terms of different pieces for at least some 10-20000 reasonable eval terms.

as engines do not do that, they rely on deeper search to get them closer to the game outcome (deepr search = better eval), but 30 mg plies with a lot of pruning involved and miserable eval certainly leads us nowhere too.

so, rigth now, we are stuck.

maybe Skipper will find something promising.

Skipper? Yes of course. [Aprils fool]

If you need 20000 terms you might as well use neural networks for that is a black box too. No one understands what is happening inside. Tuning is waiting. Can't think of something more boring.

[Henk]

And time is scarce too. So reducing evaluation terms is important.

[Lyudmil Tsvetkov]

And time is scarce too. So reducing evaluation terms is important.

but if you reduce eval terms, you will never get the position rigth.

[d]3rq3/1k1r4/p1p1p3/7n/1PP3p1/4R3/4P1RK/B3Q3 w - - 0 1

for example, above, e6 is backward (general definition for a backward pawn is any pawn, the square in front of which is free, but who can not reasonably advance, because it will be lost), any advance will lose the pawn, g4 is not backward, although a single black piece supports the square in front of the pawn, while 4 enemy pieces (but of bigger value) obstruct it, b4 is also backward, although connected, as b4-b5 loses a pawn ( 2 enemy pawns stop its advance).

how many engines consider such terms?

those are objective, valid terms, and you can not substitute them for any reasonable combination of fewer terms tuned. in great many positions the game/choosing the best move is influenced by such small terms, so basically another reason why false positives migth be frequent with any eval or search patch.

as said, rigth now we are stuck.

[Henk]

And time is scarce too. So reducing evaluation terms is important.

but if you reduce eval terms, you will never get the position rigth.

[d]3rq3/1k1r4/p1p1p3/7n/1PP3p1/4R3/4P1RK/B3Q3 w - - 0 1

for example, above, e6 is backward (general definition for a backward pawn is any pawn, the square in front of which is free, but who can not reasonably advance, because it will be lost), any advance will lose the pawn, g4 is not backward, although a single black piece supports the square in front of the pawn, while 4 enemy pieces (but of bigger value) obstruct it, b4 is also backward, although connected, as b4-b5 loses a pawn ( 2 enemy pawns stop its advance).

how many engines consider such terms?

those are objective, valid terms, and you can not substitute them for any reasonable combination of fewer terms tuned. in great many positions the game/choosing the best move is influenced by such small terms, so basically another reason why false positives migth be frequent with any eval or search patch.

as said, rigth now we are stuck.

Yes they don't have a right evaluation. Only a crude approximation. Nothing to do about it.

Otherwise start the tuner and wait for infinity and hope you are tuning the right terms.

Maybe we/they find a new local optimum with a slightly less crude approximation in the meantime.