There is a mathematical solution to this optimization problem called linear programming, which solves for the best possible allocation of limited resources (in this case, a limited number of pattern recognizers) that would represent all of the cases on which the system has trained.
