Online learning method in a decision system
View Patent ↗A learning model is initiated during start-up learning to activate operation of a decision system. During operation of the decision system, data is qualified for use in online learning. Online learning allows a system to adapt or learn application dependent parameters to optimize or maintain its performance during normal operation. Methods for qualifying data for use in online learning include thresholding of features, restriction of score space for qualified objects, and using a different source of information than is used in the decision process. Clustering methods are used to improve the quality of the learning model. Using the cumulative distribution function to compare two distributions and produce a measure of similarity derives a metric for learning maturity.
1. A computerized automatic online learning method for a decision system application having a learning model comprises the following steps:
(a) Means for collecting a plurality of unlabeled data during operation of the decision system where it is being applied to real use;
(b) Means for applying an online learning qualifier to the collected plurality of unlabeled data to assure that defective objects are not included and automatically obtain at least one qualified non-defect object;
(c) Means for updating the non-defect learning model of the decision system using the at least one qualified non-defect object.
2. The method of claim 1 wherein the decision system update uses the at least one qualified object to update the atom of a kernel-based model consisting of atoms, kernel distribution, and the weights of atoms.
3. The method of claim 2 wherein the kernel-based model update includes a clustering method grouping the atoms to represent them using a smaller number of atoms corresponding to cluster centers.
4. The method of claim 3 wherein the clustering method performs a warping function to decrease distances in the less interesting range and magnify distances in the range of interest in the kernel-based model.
5. A computerized automatic maturity estimation method for a decision system application for continuously updated view of the system's self-evaluation of performance having a learning model distribution comprises the following steps:
(a) Means for obtaining a learning model probability density distribution of at least one feature score during system learning using the learning data;
(b) Means for collecting a plurality of online learning set objects during operation of the system after learning to form an operational model probability density distribution of the at least one feature score;
(c) Means for comparing the learning model probability density distribution of the at least one feature score to the operational model probability density distribution of the at least one feature score to generate a measure of dissimilarity;
(d) Means for outputting a learning maturity estimate from the measure of dissimilarity.
6. The method of claim 5 wherein the measure of dissimilarity is the difference between the average CDF value of the operational model distribution and the 0.5 CDF value of the learning model distribution wherein feature scores are mapped using the learned CDF function to generate CDF values.
7. A computerized automatic learning method for a decision system application comprises the following steps:
(a) means for inputting a plurality of expert labeled data;
(b) means for performing startup learning using the expert labeled data to create a learning model;
(c) means for performing online learning to update the learning model;
(d) means for collecting a plurality of objects during operation of the decision system;
(e) means for applying an online learning qualifier to the collected plurality of objects to generate at least one qualified object; and
(f) means for updating the learning model using the at least one qualified object.
8. The method of claim 7 wherein the learning model is a kernel-based model consisting of atoms, kernel distribution, and the weights of atoms.
9. The method of claim 7 wherein startup learning implements a functional model as an empirical model using empirical feature distribution function to calculate the atom locations and weights.
10. The method of claim 7 wherein the computerized automatic online learning qualifier uses different information in the decision-making process.
11 .The method of claim 7 wherein the learning model update uses the at least one qualified object to update the atom of a kernel-based model consisting of atoms, kernel distribution, and the weights of atoms.
12. The method of claim 11 wherein the kernel-based model update includes a clustering method grouping the atoms to represent them using a smaller number of atoms corresponding to cluster centers.
13. The method of claim 12 wherein the clustering method performs a warping function to decrease distances in the less interesting range and magnify distances in the range of interest in the kernel-based model.