Meta-automated machine learning with improved multi-armed bandit algorithm for selecting and tuning a machine learning algorithm

Patent metadata

Field	Value
Publication number	US-12056587-B2
Application number	US-202318186307-A
Country	US
Kind code	B2
Filing date	Mar 20, 2023
Priority date	Jan 17, 2020
Publication date	Aug 6, 2024
Grant date	Aug 6, 2024

Abstract

A method for automated machine learning includes controlling execution of a plurality of instantiations of different automated machine learning frameworks on a machine learning task each as a separate arm in consideration of available computational resources and time budget. During the execution by the separate arms, a plurality of machine learning models are trained and performance scores of the plurality of trained machine learning models are computed such that one or more of the plurality of trained machine learning models are selectable for the machine learning task based on the performance scores. This invention can be used for predicting patient discharge, predictive control in buildings for energy optimization, and so on.

First claim

What is claimed is:

1. A method for automated machine learning, the method comprising:
controlling execution of a plurality of instantiations of different automated machine learning frameworks on a machine learning task each as a separate arm in consideration of available computational resources and time budget, whereby, during the execution by the separate arms, a plurality of machine learning models are trained and performance scores of the plurality of trained machine learning models are computed such that one or more of the plurality of trained machine learning models are selectable for the machine learning task based on the performance scores.

2. The method according to claim 1 , wherein the machine learning models are trained using patient health data, physiological measurements, general patient information and lengths of stays of patients as input, and wherein the machine learning task is to predict a length of stay in a hospital for a new or already admitted patient.

3. The method according to claim 1 , wherein the machine learning models are trained using medical data, and wherein the machine learning task is to classify a sickness of a patient or an emergency-submittance risk of the patient.

4. The method according to claim 3 , further comprising determining a treatment for the patient based on the classification.

5. The method according to claim 1 , wherein the machine learning models are trained using medical and/or patient health data, and wherein the machine learning task is an automated healthcare task.

6. The method according to claim 1 , wherein the machine learning models are trained using data from sensors measuring traffic, and wherein the machine learning task is to predict traffic demographics.

7. The method according to claim 6 , further comprising implementing a dynamic traffic control decision based on the solution.

8. The method according to claim 1 , wherein the machine learning models are trained using data from sensors in a building and weather data, and wherein the machine learning task is to predict energy usage of the building.

9. The method according to claim 1 , wherein the machine learning models are trained using data from sensors measuring traffic and/or weather, and wherein the machine learning task is to predict air quality.

10. The method according to claim 1 , wherein the machine learning models are trained using fundamental data of securities and time series of trading prices, and wherein the machine learning task is to classify securities or predict future security prices.

11. The method according to claim 1 , wherein, during the execution of the separate arms, the performance scores are extrapolated for a remainder of the time budget based on achieved performances of respective ones of the arms during a time interval of the execution which is a portion of the time budget, and wherein the computational resources are assigned to the arms during the remainder of the time budget based on the extrapolated performance scores such that at least one of the arms having a higher extrapolated performance score than another one of the arms receives more of the computational resources.

12. The method according to claim 11 , wherein the performance scores are extrapolated by fitting a learning curve function to past rewards of the respective ones of the arms and extrapolating the past rewards until an end of the remainder of the time budget.

13. The method according to claim 11 , further comprising freezing the execution of at least one of the arms based on the extrapolated performance scores.

14. The method according to claim 1 , wherein, during the execution of the separate arms, at least some of the arms are executed by time multiplexing using a selection mechanism to allocate the computational resources to the arms during the time budget.

15. The method according to claim 1 , further comprising selecting an ensemble of at least two of the plurality of trained machine learning models having the highest performance scores to perform the machine learning task.

16. The method according to claim 1 , wherein each of the arms are executed as a microservice component of a cloud computer system architecture in a docker container which has a container image for a respective one of the automated machine learning frameworks.

17. The method according to claim 1 , further comprising receiving the machine learning task and a dataset for the machine learning task from a user, and selecting or implementing at least one of the trained machine learning models to provide a solution to the machine learning task.

18. The method according to claim 1 , wherein, during the execution of the separate arms, a learning curve is constructed for each of the arms during a time interval of the execution within the time budget, the performance scores of each of the arms are extrapolated until a remainder of the time budget, and execution of at least some of the arms is frozen or disabled based on the extrapolated performance scores, and wherein the learning curves are constructed based on maximum performance scores achieved by respective ones of the arms during the time interval.

19. A system comprising one or more processors which, alone or in combination, are configured to provide for execution of a method comprising:
controlling execution of a plurality of instantiations of different automated machine learning frameworks on a machine learning task each as a separate arm in consideration of available computational resources and time budget, whereby, during the execution by the separate arms, a plurality of machine learning models are trained and performance scores of the plurality of trained machine learning models are computed such that one or more of the plurality of trained machine learning models are selectable for the machine learning task based on the performance scores.

20. A tangible, non-transitory computer-readable medium having instructions thereon which, upon being executed by one or more processors, alone or in combination, provide for execution of a method comprising:
controlling execution of a plurality of instantiations of different automated machine learning frameworks on a machine learning task each as a separate arm in consideration of available computational resources and time budget, whereby, during the execution by the separate arms, a plurality of machine learning models are trained and performance scores of the plurality of trained machine learning models are computed such that one or more of the plurality of trained machine learning models are selectable for the machine learning task based on the performance scores.

Assignees

• Nec Corp

Inventors

• Schmidt Mischa
• Gastinger Julia

Classifications

• G06N3/0985

  Hyperparameter optimisation; Meta-learning; Learning-to-learn · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06F18/217

  Validation; Performance evaluation; Active pattern learning techniques · CPC title

• G06F18/285

  Selection of pattern recognition techniques, e.g. of classifiers in a multi-classifier system · CPC title

• G06F9/3836

  Instruction issuing, e.g. dynamic instruction scheduling or out of order instruction execution · CPC title