Since an evaluation function could be applied to each state in the lookahead tree we can also apply it to the root note - the current state. Will this value be the same as the dynamic score obtained via the minimax procedure? Generally no. Which score is better - the static score or the dynamic one? Generally, the dynamic score. Why? Because the dynamic score utilises both structural information and behaviour based information (i.e. the possible moves).
One method of improving performance is to alter the weights dynamically in order to adjust the static score for the current node to agree with the dynamic score. There are many ways this can be done: what follows is one of the simplest schemes.
Assume that the evaluation function is determined by two features
(
and
) and that the corresponding weights are
and
. Thus
Now suppose that and the dynamic score is +3. Also suppose
that
and
. We adjust
and
so that
=+3.
We assume (with no real justification) that the increase of 1 in the static score is achieved by increasing the contribution of each feature equally - in this case by 0.5. Hence:
So we have obtained the weights for the new evaluation function
.