CONTINUOUS CONTROL BY DEEP LEARNING AND REINFORCEMENT Russian patent published in 2019 - IPC G06N3/08 

FIELD: neural networks.

SUBSTANCE: group of inventions relates to neural networks and can be used to learn a neural network executor used to select actions to be performed by an agent interacting with the medium. Method includes obtaining a mini-packet of experimental tuples and updating current values of parameters of the neural network executing, containing for each experimental tuple in a mini-packet: processing the training observation and training action in the experimental tuple using the neural network critic to determine the neural network output for the experimental tuple and determining the predicted output of the neural network for the experimental tuple; updating current values of parameters of a neural network-critic using errors between predicted outputs of a neural network and outputs of a neural network and updating current values of parameters of a neural network executor using a neural network-critic.

EFFECT: high efficiency of learning.

13 cl, 4 dwg