US 12,437,207 B2
Method for selecting task network, system and method for determining actions based on sensing data
Chih-Ming Chen, New Taipei (TW)
Assigned to Wistron Corporation, New Taipei (TW)
Filed by Wistron Corporation, New Taipei (TW)
Filed on Jan. 14, 2021, as Appl. No. 17/148,586.
Claims priority of application No. 109138475 (TW), filed on Nov. 4, 2020.
Prior Publication US 2022/0138587 A1, May 5, 2022
Int. Cl. G06N 3/126 (2023.01); G06F 18/21 (2023.01); G06F 18/2134 (2023.01); G06F 18/2413 (2023.01)
CPC G06N 3/126 (2013.01) [G06F 18/21342 (2023.01); G06F 18/217 (2023.01); G06F 18/2413 (2023.01)] 15 Claims
OG exemplary drawing
 
1. A method for training and selecting a task network based on a coevolution mechanism, comprising:
obtaining a plurality of first individuals that belong to a first population and a plurality of second individuals that belong to a second population, wherein the plurality of first individuals and the plurality of second individuals correspond to each other and are evolved via a coevolution process, wherein the coevolution process comprises one of the following: a cooperative coevolution and a competitive coevolution;
obtaining a plurality of task networks generated by a plurality of pattern-producing networks, wherein each task network of the plurality of task networks includes a first network layer and a second network layer, the first network layer and the second network layer include a plurality of nodes, and the plurality of nodes of the first network layer is connected to the plurality of nodes of the second network layer based on a weight value initialized by a corresponding pattern-producing network;
training the each task network with the plurality of first individuals and the plurality of second individuals based on a multi-target function, wherein the training regarding a first individual and a second individual comprising: feeding an input feature vector corresponding to the first individual into the each task network to generate an output feature vector and adjusting the weight value to enable the output feature vector to be close to a feature vector that corresponds to the second individual, and evaluating an accuracy of the each task network after the training, wherein the multi-target function comprises a plurality of characteristics of the each task network;
determining a fitness score of the corresponding pattern-producing network based on the accuracy of the each task network, wherein the fitness score is positively correlated to the accuracy of the each task network; and
finding a specific pattern-producing network with highest fitness score from among the plurality of pattern-producing networks, and selecting a specific task network that corresponds to the specific pattern-producing network, wherein the specific task network is selected from among the plurality of task networks.