If you just include games where your engine plays itself and minimise the absolute error instead of the squared error I believe that the tuned engine would outperform the untuned one. Also weighting the positions closer to the end more will probably help too.hgm wrote: ↑Fri Mar 19, 2021 9:31 am Indeed, this is what I see. The fully tuned engine is not just weaker than the untuned one: it is enormously weaker. Basically it loses each and every game against the untuned engine.
It is also obvious why this is: the PST get such high values that it very quickly sacrifices two or three Pawns to put its other pieces on the 'optimal' squares. Problem is that these squares are only optimal for attacking the Palace under conditions where sufficiently many Pawns participate in the attack: in Janggi you need those Pawns to compensate the fact the the opponent's Advisors defend the Palace, and your own Advisors cannot attack it. Without enough Pawns to trade away the Advisors, an attack has zero chance to succeed, no matter how well executed. So the PST bonuses you 'gained' at the expense of a few Pawns are in fact worthless. They are even worse than worthless, because with all hope for a victory gone, your only resort is to salvage the draw by defending your own Palace. And the locations good for attacking the enemy Palace in most cases are very bad for defending your own. So you get the worst of both worlds: you sacrificed Pawns to get your pieces in the worst possible locations positionally.
This is of course all caused by the fact that the evaluation function that was tuned was fundamentally flawed. It assumed the evaluation was purely additive, and did not account for the fact that you need completely different PST when ahead or equal in material (or at least in Pawns) then when behind.
It is always interesting to read your posts.
It would be very interesting to see what results you would get if you orthogonalised your features with for example PCA over your tuning set, then tuned the weights for the new orthogonal basis. The goodness of the position could then be defined as the difference between the Euclidean distances between white and black where all the negative terms for black would be transferred to positive values for white and the same for white before taking the square root.
/Pio