On the predictive power of database classifiers formed by a small network of interacting chemical oscillators

The predictive ability of database classifiers constructed with a network of interacting chemical oscillators is studied. Databases considered here are composed of records, where each record contains a number of parameters (predictors) characterizing the case and the output variable describing the case type. In a series of recent papers we have discussed the top-down design of database classifiers, that return output information as the number of excitations observed at a selected droplet of the network. The design is based on an evolutionary algorithm that optimizes the network to achieve maximum mutual information between the network evolution and the record types. Here we discuss results illustrating that such classifiers do have a predicting ability. They are able to give a correct classification for database records that are not included into the process of classifier training. As example databases we consider points inside a multidimensional unit cube and test if they belong to a ball located at the mid of cube.