Fingerprints are one of the most used biometric characteristics. They are used in a wide range of applications, as the identification of subjects is within a database. Therefore, the fingerprints need to be compared against those stored in the database, what is often done by law enforcement agencies like the FBI. Their database contains around 230 million fingerprints from inter alia recorded criminals or unresolved traces from crime scenes. Identifying a fingerprint within this database can take a long time. Since the naïve approach of comparing each fingerprint subsequently would result in long identification time, often a pre-selection is used to reduce number of comparisons. Within this work, we will use multiple fingerprints of a subject to create a binning of the subjects towards the classes of those fingerprints, so that only subjects whose fingerprints share same classes need to be compared. The classification of the fingerprints should be done using only global information of the fingerprint pattern, like the orientation of the ridgelines. The first part of this work is therefore a survey on the different approaches for fingerprint classification using those features. After that, the NIST SD9 database was analyzed towards the correlations between the fingerprint classes of the subject. With the derived information we can show, that using multiple instances of fingerprints for the binning of the database can result in up to 94% less comparisons for identification assuming perfect classification. The use of the system for multi-instance classification defined in this work, together with the trained neural networks for fingerprint classification enabled us to reduce the number of required comparisons by up to 80%, while using just 3 fingers of a subject. In addition, it was shown, that the use of classifiers with only moderate classification accuracy allowed a reduction of comparisons.