Target-oriented reinforcement learning method and apparatus for performing the same

Patent metadata

Field	Value
Publication number	US-12223695-B2
Application number	US-202017427957-A
Country	US
Kind code	B2
Filing date	Dec 8, 2020
Priority date	Oct 12, 2020
Publication date	Feb 11, 2025
Grant date	Feb 11, 2025

Abstract

A target-oriented reinforcement learning method according to an embodiment includes: collecting data related to the target of reinforcement learning as target data in the process of performing the reinforcement learning; learning the collected target data as auxiliary learning for the reinforcement learning; and incorporating the results of the learning of the target data into the performance of the reinforcement learning.

First claim

The invention claimed is:

1. A reinforcement learning method performed by a target-oriented reinforcement learning model, the reinforcement learning method comprising:
collecting a data related to a target of the reinforcement learning as target data when an event in which an agent achieving the target and receiving a reward occurs, wherein the target data comprises a predetermined number of frame images before an occurrence of the event and a label indicating the target data corresponds to the target;
learning the target data as auxiliary learning for the reinforcement learning; and
incorporating results of the learning of the target data into performance of the reinforcement learning,
wherein the target-oriented reinforcement learning model comprises:
a feature extraction unit implemented by at least one processor and configured to extract features from state data and the target data;
an action module implemented by the at least one processor and configured to output an action and a value according to a policy based on the feature extracted from the state data; and
a classification module implemented by the at least one processor and configured to classify the target data based on the feature extracted from the target data, and

wherein learning the collected target data comprises:
extracting, by the feature extraction unit, a feature from batch data of the target data;
extracting, by the classification module, a predicted value according to the feature extracted from the batch data of the target data;
calculating, by the target-oriented reinforcement learning model, a loss for the auxiliary learning by using the predicted value and the label of the target data; and
learning, by the target-oriented reinforcement learning model, the visual representation of the target data by using the loss for the auxiliary learning.

2. A computer-readable storage medium having stored thereon a program for performing the method set forth in claim 1 .

3. A computer program performed by a computing device and stored in a medium in order to perform the method set forth in claim 1 .

4. A computing device for performing target-oriented reinforcement learning, the computing device comprising:
an input device configured to receive data;
a memory configured to store a program for performing reinforcement learning and target data collected in a process of performing the reinforcement learning; and at least one processor configured to perform the reinforcement learning using the data received through the input device by executing the program;
wherein a target-oriented reinforcement learning model implemented in such a manner that the at least one processor executes the program, collects data related to a target of the reinforcement learning as the target data in a process of performing the reinforcement learning when an event in which an agent achieving the target and receiving a reward occurs, wherein the target data comprises a predetermined number of frame images before an occurrence of the event and a label indicating the target data corresponds to the target, learns the target data as auxiliary learning for the reinforcement learning, and incorporates results of the learning of the target data into performance of the reinforcement learning, and
wherein the target-oriented reinforcement learning model comprises:
a feature extraction unit implemented by the at least one processor and configured to extract features from state data and the target data;
an action module implemented by the at least one processor and configured to output an action and a value according to a policy based on the feature extracted from the state data; and
a classification module implemented by at least one processor and configured to classify the target data based on the feature extracted from the target data, and

wherein when learning the collected target data, the target-oriented reinforcement learning model is operated such that the feature extraction unit extracts a feature from batch data of the target data, the classification module extracts a predicted value according to the feature extracted from the batch data of the target data, the target-oriented reinforcement learning model calculates a loss for the auxiliary learning by using the predicted value and the label of the target data, and the target-oriented reinforcement learning model learns the visual representation of the target data by using the loss for the auxiliary learning.

Assignees

• Seoul Nat Univ R&Db Foundation

Inventors

• Zhang Byoung-Tak
• Kim Kibeom
• Lee Minsu
• Lee Min Whoo
• Kim Yoonsung

Classifications

• G06V2201/07

  Target detection · CPC title

• G06N20/00

  Machine learning · CPC title

• G06V10/776 Primary

  Validation; Performance evaluation · CPC title

• G06V10/40 Primary

  Extraction of image or video features · CPC title