Optosf optimizes the overall smoothing factor (osf) for the Probabilistic Neural Network (PNN) classifier.
The regional weights are optimized using the optrws command. To save optrws runtime, it is suggested that optrws be run using the K-L feature vectors of only a reasonably small set of fingerprints, perhaps a small subset of the full prototype set that will be used in the finished classifier. But then, after the full prototype set of feature vectors is made by transforming previously made orientation arrays using the transform matrix incorporating the optimized regional weights, one can expect that the classifier that uses these feature vectors will be slightly more accurate if it uses an overall smoothing factor slightly larger than 1, to compensate for the fact that the prototype set is larger than it was during optimization of the regional weights. During optimization of the regional weights, no explicit overall smoothing factor is used, since any effect such a factor would have had could equally well be produced by just using different values of the regional weights; so, optrws in effect fixes the overall smoothing factor at 1.
The optosf command is provided to optimize the overall smoothing factor for best results on the full set of prototypes. It optimizes osf by attempting to find a minimum (or at least a local minimum) of an activation error rate that results when a set of finished feature vectors is classified by PNN. The set of prototypes used by the PNN, during this optimization, is a superset of the set on which the activation error rate is computed: both sets start at the beginning of the provided data, but they are of different lengths. Whichever fingerprint the classifier is running on is temporarily left out of the prototypes set, i.e. a leave-one-out method is used in order to simulate a realistic situation.
The optimization method used is a very simple one, consisting of taking steps of an initial size, then halving the stepsize and reversing direction if the error rates ceases to decrease, etc. This method, while obviously not sufficient for the general problem of minimizing a real function of one real variable, appears to be sufficient for this particular problem, since the activation error rate as a function of the osf seems to always have a unimodal form.
A file specifying values of some or all of the parameters. Parameters not specified in this file assume default values. To find out what the parameters are, and as an example of the format of a parameters file, see the file pcasys/parms/optosf.prs in the PARAMETER FILES section below. The user's prsfile must specify values for those parameters that optosf.prs indicates have no defaults; it can also specify default-overriding values for one or more of the parameters that have defaults.
Contains default values of the parameters for optosf (optimize overall smoothing factor command). Parameters with no defaults must be set in the users prsfile.
Default settings in pcasys/parms/optosf.prs
How many (first) features of the feature vectors to use.
Initial value for osf (overall smoothing factor).
Initial step size for osf.
Program stops when step size becomes <= this value.
Size of the table used to avoid redundant computing.
If y, write progress messages to stdout.
A - (hyphen) means let optosf make the description; otherwise, value is the description.
The file containing the prototype feature vectors, each vector stored as one row of the matrix.
The file containing the classes of the prototype feature vectors.
The number of first feature vectors from fvs_file to use as the PNN prototypes when optimizing osf.
The number of first feature vectors from fvs_file to run the PNN on to optimize osf.
The results output file.