US 12,321,827 B2
Method and device for transfer learning between modified tasks
Danny Stoll, Rheinfelden (DE); Diane Wagner, Freiburg (DE); Frank Hutter, Freiburg im Breisgau (DE); Joerg Franke, Freiburg (DE); and Simon Selg, Freiburg (DE)
Assigned to ROBERT BOSCH GMBH, Stuttgart (DE)
Filed by Robert Bosch GmbH, Stuttgart (DE)
Filed on Aug. 5, 2021, as Appl. No. 17/394,793.
Claims priority of application No. 102020210352.6 (DE), filed on Aug. 14, 2020.
Prior Publication US 2022/0051138 A1, Feb. 17, 2022
Int. Cl. G06N 20/00 (2019.01); G06F 18/214 (2023.01); G06N 3/08 (2023.01)
CPC G06N 20/00 (2019.01) [G06F 18/214 (2023.01); G06N 3/08 (2013.01)] 10 Claims
OG exemplary drawing
 
1. A computer-implemented method for transfer learning of hyperparameters of a machine learning algorithm, comprising the following steps:
i) providing a current search space, a previous search space, a loss function to be optimized, and a ranking of evaluated hyperparameter configurations of a previous optimization step utilizing the previous search space with respect to the loss function, the current and previous search spaces each being defined based on predetermined value ranges of the hyperparameters;
ii) creating a reduced search space, value ranges of the hyperparameters of the reduced search space corresponding to the value ranges of the hyperparameters of the current search space, limited as a function of values of a predeterminable number of the best hyperparameter configurations from the ranking;
iii) drawing candidate configurations repeatedly at random from the reduced search space and from the current search space, and utilizing the machine learning algorithm, parameterized in each case with the candidate configurations, to optimize the loss function;
iv) creating a Tree Parzen Estimator (TPE) as a function of the candidate configuration and the results of the machine learning algorithm applied to the loss function to be optimized;
v) multiple repeated drawing of further candidate configurations from the current search space using the TPE and applying the machine learning algorithm, parameterized in each instance with the candidate configurations, to the loss function;
vi) creating a new ranking of the further candidate configurations and selecting one configuration from the new ranking as a hyperparameter configuration for the machine learning algorithm.