US 12,242,980 B2
Machine learning with multiple constraints
Parikshit Ram, Atlanta, GA (US); Dakuo Wang, Cambridge, MA (US); Deepak Vijaykeerthy, Bangalore (IN); Vaibhav Saxena, New Delhi (IN); Sijia Liu, Somerville, MA (US); Arunima Chaudhary, Boston, MA (US); Gregory Bramble, Larchmont, NY (US); Horst Cornelius Samulowitz, White Plains, NY (US); and Alexander Gray, Yonkers, NY (US)
Assigned to International Business Machines Corporation, Armonk, NY (US)
Filed by INTERNATIONAL BUSINESS MACHINES CORPORATION, Armonk, NY (US)
Filed on Sep. 9, 2020, as Appl. No. 17/015,243.
Prior Publication US 2022/0076144 A1, Mar. 10, 2022
Int. Cl. G06N 5/04 (2023.01); G06F 9/38 (2018.01); G06F 18/243 (2023.01)
CPC G06N 5/04 (2013.01) [G06F 9/38 (2013.01); G06F 18/24323 (2023.01)] 20 Claims
OG exemplary drawing
 
1. A computer-implemented method comprising:
detecting a user uploading data and one or more constraints;
collecting the data and the one or more constraints; and
determining that one or more model pipelines satisfies all of the one or more constraints based on applying one or more algorithms to the collected data, the collected constraints, and to multiple model pipelines, wherein the one or more algorithms use an alternating direction method of multipliers (ADMM) based joint optimization approach in determining that one or more of the model pipelines satisfies all of the one or more constraints, and the ADMM based joint optimization approach decomposes the determining into sub-problems solved iteratively, the sub-problems comprising:
one or more hyperparameter problems solving hyperparameter optimization for the one or more model pipelines, and
a pipeline selection problem solved by a combinatorial multi-armed bandit algorithm,
wherein the ADMM based joint optimization approach uses explicit Boolean variables to encode selection of the multiple model pipelines.