A practical parameter optimization technique for semiempirical methods

James J. P. Stewart, MrMOPAC@Worldnet.att.net, Stewart Computational Chemistry,
15210 Paddington Circle, Colorado Springs, CO 80921

Although, in principle, parameter optimization is a straightforward problem of
minimizing the value of a function in a multi-parameter space, in practice many
complications occur during the development of semiempirical methods. These range
from faults in the approximations, faulty and incomplete data, and inaccurate
derivatives, to the hazards posed by the possibility of multiple minima in the
parameter hypersurface. The problems encountered during the development of a new
method will be discussed, along with their resolution, and the implications for
increased accuracy illustrated by examples of long standing faults being
corrected.