Granted Patent B2

US 12,361,331 · App. 17/308,000 · Granted Jul 15, 2025

Inventors: Chi Wang (Redmond, WA); John Carol Langford (Scarsdale, NV); Qingyun Wu (Charlottesville, VA); Paul S. Mineiro (Bellevue King, WA); Marco Rossi (Harrisonburg, VA)

Assignee: Microsoft Technology Licensing, LLC

G06N20/20G06F11/3428G06F17/18G06N20/00

Abstract

Claims (59)

1. A method for tuning a hyperparameter for a machine learning model, the method comprising:

receiving a hyperparameter for tuning;

generating a set of challenger configurations based on the hyperparameter;

selecting a subset of challenger configurations from the set of challenger configurations based at least in part on a resource lease associated with each of the challenger configurations having a lowest resource lease;

adding the subset of challenger configurations selected from the set of challenger configurations to a live pool for evaluation, wherein the live pool includes live running configurations of the machine learning model that are being simultaneously evaluated in a live environment;

comparing a loss value derived from a loss function for each of the subset of challenger configurations to a loss value derived from the loss function for a champion configuration simultaneously in the live environment;

selecting a challenger configuration from the subset of the challenger configurations based on the comparison;

replacing the champion configuration with the challenger configuration; and

generating a new set of challenger configurations based on a new champion configuration.

2. The method of claim 1 , wherein a number of challenger configurations scheduled for evaluation is based on a computational budget.

3. The method of claim 2 , further comprising removing at least one challenger configuration from the set of challenger configurations based on a loss value derived from the loss function for the at least one challenger configuration and the loss value derived from the loss function for the champion configuration.

4. The method of claim 2 , wherein the loss value derived from the loss function for the challenger configuration is a probabilistic upper bound for the challenger configuration and the loss value derived from the loss function for the champion configuration is a probabilistic lower bound for the champion configuration.

5. The method of claim 1 , further comprising:

assigning each challenger configuration in the set of challenger configurations a resource lease; and

increasing the resource lease over time.

6. The method of claim 5 , further comprising:

determining that the challenger configuration in the subset of the set of challenger configurations has reached a limit associated with the resource lease;

comparing a loss value derived from the loss function for the challenger configuration in the subset of challenger configurations to a second loss value derived from the loss function for a second challenger configuration in the subset of the set of challenger configurations; and

increasing the resource lease for the challenger configuration in the subset of the set of challenger configurations based on the comparison.

7. The method of claim 6 , further comprising:

based on the comparison, replacing the challenger configuration in the subset of the set of challenger configurations with another challenger configuration from the set of challenger configurations, wherein a resource lease of the another challenger configuration from the set of challenger configurations is the smallest resource lease in the set of challenger configurations.

8. The method of claim 1 , wherein the hyperparameter indicates which namespaces interact together.

9. The method of claim 1 , further comprising:

receiving historical data samples;

generating a prediction based on the historical data samples; and

receiving at least one of full-information feedback or partial-information feedback based on the prediction.

10. A system for tuning a hyperparameter for a machine learning model, the system comprising:

a processor; and

memory including instructions which when executed by the processor, cause the processor to:

receive a hyperparameter for tuning;

receive configuration information associated with generating challenger configurations for the hyperparameter;

generate a set of challenger configurations based on the hyperparameter and the configuration information;

select a subset of challenger configurations from the set of challenger configurations based at least in part on a resource lease associated with each of the challenger configurations having a lowest resource lease;

add the subset of challenger configurations selected from the set of challenger configurations to a live pool for evaluation, wherein the live pool includes live running configurations of the machine learning model that are being simultaneously evaluated in a live environment;

compare a loss value derived from a loss function for each of the subset of the challenger configurations to a loss value derived from the loss function for a champion configuration simultaneously in the live environment;

select a challenger configuration from the subset of the challenger configurations based on the comparison;

replace the champion configuration with the challenger configuration; and

generate a new set of challenger configurations based on a new champion configuration.

11. The system of claim 10 , wherein the instructions, when executed by the processor, cause the processor to assign each challenger configuration in the set of challenger configurations a resource lease and increase the resource lease over time.

12. The system of claim 11 , wherein the instructions, when executed by the processor, cause the processor to determine that the challenger configuration in the subset of the set of challenger configurations has reached a limit associated with the resource lease, compare a loss value derived from the loss function for the challenger configuration in the subset of the set of challenger configurations to a second loss value derived from the loss function for a second challenger configuration in the subset of the set of challenger configurations, and increase the resource lease for the challenger configuration in the subset of the set of challenger configurations based on the comparison.

13. The system of claim 12 , wherein the instructions, when executed by the processor, cause the processor to replace the challenger configuration in the subset of the set of challenger configurations with another challenger configuration from the set of challenger configurations based on the comparison, wherein a resource lease of the another challenger configuration from the set of challenger configurations is the smallest resource lease in the set of challenger configurations.

14. The system of claim 13 , wherein a number of challenger configurations scheduled for evaluation is based on a computational budget.

15. The system of claim 10 , wherein the instructions, when executed by the processor, cause the processor to:

receive historical data samples;

generate a prediction based on the historical data samples; and

receive at least one of full-information feedback or partial-information feedback based on the prediction.

16. A computer-readable storage medium including instructions, when executed by a processor, cause the processor to:

receive a hyperparameter for tuning;

generate a set of challenger configurations based on the hyperparameter;

select a subset of challenger configurations from the set of challenger configurations based at least in part on a resource lease associated with each of the challenger configurations having a lowest resource lease;

add the subset of challenger configurations selected from the set of challenger configurations to a live pool for evaluation, wherein the live pool includes live running configurations of the machine learning model that are being simultaneously evaluated in a live environment;

compare a loss value derived from a loss function for each of the subset of challenger configurations to a loss value derived from the loss function for a champion configuration simultaneously in the live environment;

select a challenger configuration from the subset of the challenger configurations based on the comparison;

replace the champion configuration with the challenger configuration; and

generate a new set of challenger configurations based on a new champion configuration.

17. The computer-readable storage medium of claim 16 , wherein a number of challenger configurations scheduled for evaluation is based on a computational budget.

18. The computer-readable storage medium of claim 17 , wherein the instructions, when executed by the processor, cause the processor to remove at least one challenger configuration from the set of challenger configurations based on a loss value derived from the loss function for the at least one challenger configuration and the loss value derived from the loss function for the champion configuration.

19. The computer-readable storage medium of claim 16 , wherein the instructions, when executed by the processor, cause the processor to assign each challenger configuration in the set of challenger configurations a resource lease.

20. The computer-readable storage medium of claim 19 , wherein the instructions, when executed by the processor, cause the processor to determine that the challenger configuration in the subset of the set of challenger configurations has reached a limit associated with the resource lease, compare a loss value derived from the loss function for the challenger configuration in the subset of the set of challenger configurations to a second loss value derived from the loss function for a second challenger configuration in the subset of the set of challenger configurations, and increase the resource lease for the challenger configuration in the subset of the set of challenger configurations based on the comparison.

Assignments (1)

Continuity (1)

References Cited (38)