Multi-kernel neural network concurrent learning, monitoring, and forecasting system

• US 6,647,377 B2
• Filed: 04/09/2001
• Issued: 11/11/2003
• Est. Priority Date: 11/19/1997
• Status: Expired due to Fees

• Current Assignee: Metricly, Inc. (Virtana Corp.)
• Original Assignee: Netuitive, Inc. (Virtana Corp.)
• Inventors: Jannarone, Robert J.
• Primary Examiner(s): Davis, George B.
• Application Number:
  US09/829,018
• Publication Number:
  US 20020016781A1
• Time in Patent Office: 946 Days
• Field of Search: 706/16, 706/21
• US Class Current: 706/16
• CPC Class Codes: G06N 3/045 Combinations of networks

Abstract

A multi-kernel neural network computing architecture configured to learn correlations among feature values 34, 38 as the network monitors and imputes measured input values 30 and also predicts future output values 46. This computing architecture includes a multi-kernel neural network array 14 with the capability to learn and predict in real time. The CIP 10 also includes a manager 16 and an input-output transducer 12 that may be used for input-output refinement. These components allow the computing capacity of the multi-kernel array 14 to be reassigned in response to measured performance or other factors. The output feature values 46 computed by the multi-kernel array 14 and processed by an output processor 44 of the transducer 12 are supplied to a response unit 18 that may be configured to perform a variety of monitoring, forecasting, and control operations in response to the computed output values.

Claims

1. A method for determining whether refinement operations are indicated using learning correlations, comprising the steps of:

(a) receiving an iteration of measured input values for a current time trial;

(b) providing a vector of input feature values based on the measured input values to a multi-kernel processor, each kernel of the processor operative for;

receiving one or more of the input feature values,

retrieving connection specifications, connection weights, and learning weights, and

computing output feature values based on the received input feature values, the connection weights, and the connection specifications;

(c) responding to a vector of computed output values based on the output feature values computed by each kernel; and

(d) determining whether the refinement operations are indicated using the learning correlations and based on the output feature values.

2. The method of claim 1, wherein providing a vector of input feature values includes computing the input feature values based on the measured input values and input feature specifications.

3. The method of claim 2, wherein computing the input feature values comprises a mathematical operation selected from the group including:

computing the input feature values based on algebraic combinations of the measured input values;

computing input feature values based on coefficients corresponding to a polynomial approximating a function defined by the measured input values;

computing input feature values based on coefficients corresponding to a differential equation corresponding to the function defined by measured input values; and

computing input feature values based on coefficients corresponding to a frequency-domain function corresponding to a function defined by the measured input values.

4. The method of claim 2, wherein responding to a vector of computed output values includes computing the output values based on output feature values and output feature specifications.

5. The method of claim 4, if refinement operations are indicated, further comprising performing one or more refinement operations selected from the group including.

deleting ineffective input or output feature values;

combining redundant input or output feature values;

specifying new input or output feature values;

recomputing the input feature specifications based on the measured input values and the computed output values for a plurality of time trials;

recomputing the learning weights based on the measured input values and the computed output values for a plurality of time trials;

recomputing the connection specifications based on the measured input values and the computed output values for a plurality of time trials;

recomputing the output feature specifications based on the measured input values and the computed output values for a plurality of time trials; and

reassigning functionality among the kernels.

6. The method of claim 4, wherein computing the output feature values includes:

imputing out feature values for a current time trial based on input feature values for one or more historical time trials,

computing monitored output feature values based on the input feature values for the current time travel;

computing deviance values based on the imputed output feature values and the monitored output feature values; and

basing the computed output values on the monitored output feature values.

7. The method of claim 6, wherein responding to the vector of computed output values includes:

comparing each deviance value to an associated threshold values; and

if one of the deviance values exceeds its associated threshold value, performing one or more deviance operation selected from the group including,

indicating an alarm condition, and

basing the computed the output values on the imputed output feature values rather than the monitored the output feature values for the output feature value associated with the deviance value that exceeds its associated threshold value.

8. The method of claim 1, wherein computing output feature values includes predicting output feature values for future time trials.

9. The method of claim 1, wherein computing output feature values includes:

imputing output feature values for the current time trial based on the input feature values for one or more historical time trials; and

predicting output feature values for future time trials.

10. The method of claim 1, wherein responding to the vector of computed output values includes performing one or more control operations selected from the group including:

displaying a representation of the computed output values on a display device, and

actuating a controlled parameter to compensate for a condition indicated by the computed output values.

11. The method of claim 1, wherein the multi-kernel processor comprises:

an array of spatially-dedicated kernels corresponding to a spatially-contiguous field from which input values are measured and for which output values are predicted; and

each kernel configured to compute one of the computed output values based on a set of adjacent measured input values.

12. The method of claim 11, wherein:

each kernel of the multi-kernel processor corresponds to a mutually-exclusive time-specific price forecast based on the measured input values; and

each kernel configured to predict its corresponding mutually-exclusive time-specific price forecast based on the assured input values.

13. The method of claim 1, wherein:

each kernel of the multi-kernel processor corresponds to a pixel in a visual image;

each measured input value corresponds to a measured intensity of one of the pixels in the visual image; and

each computed output value corresponds to a computed intensity of one of the pixels of the visual image.

14. The method of claim 1, wherein the multi-kernel processor comprises:

an array of temporally-dedicated kernels corresponding to a time-based index from which input values are measured and for which output values are predicted; and

each kernel configured to predict a mutually-exclusive one of the time-based index values based on the measured input values.

15. The method of claim 14, wherein the measured input values comprise:

the commodity price index; and

price indices for currencies and other commodities.

16. The method of claim 1, wherein the multi-kernel processor comprises:

a plurality of kernel groups each comprising a plurality of temporally-dedicated kernels corresponding to a time-based index from which input values are measured and for which output values are predicted;

each kernel group comprising a plurality of individual kernels, each configured to predict a component of a mutually-exclusive time-based index value based on the measured input values;

the plurality of groups of temporarily-dedicated kernels defining an array of spatially-dedicated kernel groups; and

each kernel group configured to compute a component of the time-based index.

17. The method of claim 1, wherein:

the time-based index comprises an electricity demand index;

each kernel group corresponds to a plurality of electricity delivery points; and

the measured input values comprise electricity demand and weather data.

18. The method of claim 1, wherein:

each kernel of the processor is operative for computing updated connection weights based on the received input values, the connection weights, the connection specifications, and learning weights;

the connection weights comprise the elements of an inverse covariance matrix;

and computing updated connection weights comprise one or more steps from the group consisting of,

updating the inverse covariance matrix, and

inverting the updated covariance matrix.

19. An apparatus for determining whether refinement operations are indicated using learning correlations, comprising:

(a) means for receiving an iteration of measured input values for a current trial;

(b) means for providing a vector of input feature values based on the measured input values to a multi-kernel processor, each kernel of the processor operative for;

receiving one or more of the input feature values,

receiving correction specifications, connection weights, and learning weights, and

computing output feature values based on the received input feature values, the connection weights, and the connection specifications;

(c) means for responding to a vector of computed output values based on the output feature values computed by each kernel; and

(d) means for determining whether the refinement operations are indicated using the learning correlations and based on the output feature values.