METHOD AND SYSTEM OF DATA MODELLING

• US 20100174514A1
• Filed: 12/30/2009
• Published: 07/08/2010
• Est. Priority Date: 01/07/2009
• Status: Active Grant

• Current Assignee: University of Sydney
• Original Assignee: University of Sydney
• Inventors: Ramos, Fablo Tozeto, Melkumyan, Aman
• Granted Patent: US 8,849,622 B2
• Time in Patent Office: Days
• US Class Current: 703/2
• CPC Class Codes: G06F 17/18 for evaluating statistical data , e.g. average values, frequency distributions, probability functions, regression analysis forecasting specially adapted for a specific administrative, business or logistic context G06Q10/04

Abstract

A system for large scale data modelling is described. The system includes at least one data measurement sensor (230) for generating measured data, a training processor (240) to determine optimized hyperparameter values in relation to a Gaussian process covariance function including a sparse covariance function that is smooth and diminishes to zero outside of a characteristic hyperparameter length. An evaluation processor (260) determines model data from the Gaussian process covariance function with optimised hyperparameter values and measured data. Also described is methods for modelling date, including a method using a Gaussian process including a sparse covariance function that diminishes to zero outside of a characteristic length, wherein the characteristic length is determined from the data to be modelled.

Claims

1. A method for modelling data using a Gaussian process comprising a sparse covariance function that diminishes to zero outside of a characteristic length, wherein the characteristic length is determined from the data to be modelled.

2. The method of claim 1, wherein the sparse covariance function is smooth.

3. The method of claim 1, wherein the sparse covariance function has a continuous fourth derivative.

4. The method of claim 1, wherein the characteristic length is determined through a machine learning process

5. The method of claim 1, wherein the sparse covariance function is multi-dimensional with a separate characteristic length for each dimension.

6. The method of claim 1, wherein the Gaussian process further comprises combining a non-stationary covariance function with the sparse covariance function.

7. The method of claim 1, wherein the non-stationary covariance function is a neural network covariance function and the Gaussian process further comprises combining a non-stationary covariance function with the neural network covariance function.

8. The method of claim 1, wherein the data is measured using one or more measurement sensors.

9. The method as claimed in claim 8, wherein the data represents geographical surface terrain.

10. The method of claim 8, wherein the data represents in-ground geological mineral distribution.

11. A method for modelling large scale data representing a distribution of physical properties, comprising

obtaining a measured dataset,

applying an exactly sparse Gaussian process covariance function to the measured dataset to determine an optimized set of hyperparameter values associated with the Gaussian process covariance function, and

generating selected model data by Gaussian process regression using the optimized hyperparameter set and the measured dataset.

12. A method as claimed in claim 11, wherein the exactly sparse Gaussian process covariance function comprises a sparse covariance function that is smooth and diminishes to zero outside of a characteristic hyperparameter length, and a non-stationary neural network covariance function.

13. A method as claimed in claim 11, wherein the sparse covariance function is multi-dimensional with a separate characteristic hyperparameter length for each dimension.

14. A method as claimed in claim 11, wherein the measured dataset represents geographical surface terrain measurements.

15. A method as claimed in claim 11, wherein the measured dataset represents in-ground geological mineral distribution measurements.

16. A system for large scale data modelling, comprising:

at least one data measurement sensor for generating measured data representing a distribution of physical properties in a selected geographical region;

a training processor adapted to determine for the measured data optimized hyperparameter values in relation to a Gaussian process covariance function comprising a sparse covariance function that is smooth and diminishes to zero outside of a characteristic hyperparameter length; and

an evaluation processor adapted to determine model data from the Gaussian process covariance function with optimised hyperparameter values and measured data.

17. A system as claimed in claim 16, wherein the Gaussian process covariance function also comprises a non-stationary neural network covariance function.

18. A system as claimed in claim 16, wherein the measured data comprises geographical surface terrain measurements.

19. A system as claimed in claim 16, wherein the measured data comprises in-ground ore grade distribution measurements.

20. A computational method for modelling a data space, the method comprising:

using a computational system, applying to the data space a kernel machine, wherein the kernel machine uses a covariance function which is constructed from a smooth basis function equal to zero outside of a defined interval;

one of outputting and storing in memory a model from the kernel machine.

21. The method of claim 20 wherein the kernel machine uses a Gaussian learning process.

22. The method of claim 20, wherein the basis function is a squared cosine wave.

23. The method of claim 20, wherein the basis function is cos²(π

x)H(0.5−

|x|) where H(.) is a step function.

24. The method of claim 20, wherein the multi-dimensional data space has N dimensions and the covariance function is

25. The method of claim 21, wherein the defined interval is a hyperparameter.

26. A computational method for modelling a data space, the method comprising:

using a computational system, applying to the data space a kernel machine, wherein the kernel of the kernel machine is a positive semi-definite function and a symmetric function, equal to zero outside of a defined interval;

at least one of one of outputting and storing in memory a model from the Gaussian regression process.