Sometimes, it is desirable to be able to automatically classify facial expressions; for example, to have a function which determines whether or not a facial expression is ``happy'' or ``sane''. This could also be used to automatically derive constraints which define ``valid'' expressions for an application.
One approach which could be used to generate constraints would be to use the Pandemonium paradigm [9]; several randomly-created functions (or demons) are set up, each of which takes as arguments the parameters of a facial expression and returns a real value proportional to the chromosome's ``score'' as a member of a class. Each function is given a weight, with all weights initially being equal. Several iterations are run, with the user classifying randomly-generated facial expressions; the user's choices are compared with those of the various demons; those which predict the user's classifications well have their weights increased, those which fare badly have them decreased. Demons which predict the opposite of the user's classifications are given negative weights. Given enough demons, this will evolve a function which will accurately classify facial expressions, and which can be translated into a mathematical expression defining the region in the expression space which correspond to the required criterion.
Computational learning techniques other than Pandemonium, such as neural networks, could be applied to this problem.