Adapting Bayesian network parameters on-line in a dynamic environment

• US 7,685,278 B2
• Filed: 12/18/2001
• Issued: 03/23/2010
• Est. Priority Date: 12/18/2001
• Status: Expired due to Fees

• Current Assignee: Micro Focus LLC (Open Text Corporation)
• Original Assignee: Hewlett-Packard Development Company, L.P. (HP Inc.)
• Inventors: Cohen, Ira, Bronstein, Alexandre, Duro, Marsha Prescott
• Primary Examiner(s): Barqadle; Yasin M
• Application Number:
  US10/026,061
• Publication Number:
  US 20030115325A1
• Time in Patent Office: 3,017 Days
• Field of Search: 709217-219, 709223-225, 709203-207, 706/16, 706/25, 706 45- 52, 706/14
• US Class Current: 709/225
• CPC Class Codes: G06N 7/01 Probabilistic graphical models, e.g. probabilistic networks

Abstract

A method for adapting a Bayesian network includes determining a set of parameters for the Bayesian network, for example, initial parameters, and then updating the parameters in response to a set of observation data using an adaptive learning rate. The adaptive learning rate responds to any changes in the underlying modeled environment using minimal observation data.

Claims

1. A method for adapting a Bayesian network, comprising:

generating a set of parameters for the Bayesian network in response to a set of past observation data such that the Bayesian network models an environment having at least hardware elements;

obtaining a set of present observation data from the environment;

determining an estimate of the parameters in response to the present observation data;

adapting a learning rate for the parameters such that the learning rate responds to changes in the environment indicated in the present observation data by increasing the learning rate when an error between the estimate and a mean of the parameters is relatively large and decreasing the learning rate when convergence is reached between the estimate and the mean of the parameters;

updating the parameters in response to the present observation data using the learning rate; and

using the Bayesian network to model the environment and diagnose problems or predict events in the environment.

2. The method of claim 1, wherein adapting the learning rate comprises adapting a different learning rate for each parameter of the Bayesian network.

3. The method of claim 2, wherein the estimate comprises an estimated value for a particular one of the parameters, and the mean comprises a mean value for the particular parameter, and wherein adapting the learning rate is based on the estimated value and mean value for the particular parameter.

4. The method of claim 1, wherein a subset of values in the present observation data is unavailable when updating the parameters.

5. The method of claim 1, wherein the environment is an online environment.

6. The method of claim 5, wherein the online environment is an email system.

7. The method of claim 5, wherein the online environment is an e-commerce system.

8. The method of claim 5, wherein the online environment is a database system.

9. The method of claim 1, wherein updating the parameters comprises determining an initial set of the parameters and then updating the parameters in response to the present observation data using the learning rate.

10. The method of claim 1, wherein the mean of the parameters comprises a mean value for each of the parameters.

11. The method of claim 1, wherein the mean is a running average of values of at least one of the parameters.

12. The method of claim 1, wherein the learning rate is increased in response to the error between the estimate and the mean being greater than a particular value, and the learning rate is decreased in response to the error between the estimate and the mean being less than a threshold.

13. A system, comprising:

an environment having at least hardware elements to generate a set of present observation data;

a Bayesian network to perform automated reasoning for the environment in response to the present observation data;

an adapter to obtain the present observation data from the environment and to determine an estimate of a set of parameters for the Bayesian network in response to the present observation data, the adapter to adapt a learning rate for the parameters to respond to changes in the environment by increasing the learning rate when an error between the estimate and a mean of the parameters is relatively large and decreasing the learning rate when convergence is reached between the estimate and the mean of the parameters, wherein the Bayesian network is configured to model the environment and diagnose problems or predict events in the environment.

14. The system of claim 13, wherein the adapter is configured to use a different learning rate for each parameter of the Bayesian network.

15. The system of claim 14, wherein the estimate comprises an estimated value for a particular one of the parameters, and the mean comprises a mean value for the particular parameter, and wherein adapting the learning rate is based on the estimated value and mean value for the particular parameter.

16. The system of claim 13, wherein the adapter is configured to determine the parameters by determining an initial set of the parameters and then to update the parameters in response to the present observation data using the learning rate.

17. The hardware system of claim 13, wherein a subset of values in the present observation data is unavailable.

18. The hardware system of claim 13, wherein the environment is an email system.

19. The hardware system of claim 13, wherein the environment is an e-commerce system.

20. The hardware system of claim 13, wherein the environment is a database system.

21. The system of claim 13, wherein the mean of the parameters comprises a mean value for each of the parameters.

22. The system of claim 13, wherein the mean is a running average of values of at least one of the parameters.

23. The system of claim 13, wherein the learning rate is increased in response to the error between the estimate and the mean being greater than a particular value, and the learning rate is decreased in response to the error between the estimate and the mean being less than a threshold.