Method for predicting the areas of information needed to be collected

Patent metadata

Field	Value
Publication number	US-12198035-B2
Application number	US-202318482693-A
Country	US
Kind code	B2
Filing date	Oct 6, 2023
Priority date	Nov 14, 2022
Publication date	Jan 14, 2025
Grant date	Jan 14, 2025

Abstract

Disclosed is a method for predicting areas of environmental information needed to be collected, which is performed by one or more processors of a computing device. The method may include: outputting one or more episodes based on environmental information; measuring uncertainty for each of the one or more episodes; and predicting an area of the environmental information needed to be collected based on the measured uncertainty.

First claim

What is claimed is:

1. A method for predicting an area of environmental information needed to be collected, the method performed by a computing device, the method comprising:
outputting one or more episodes based on environmental information by inputting the environmental information into a first neural network model, wherein the first neural network model corresponds to a neural network model pre-trained based on the environmental information, and wherein the first neural network model includes a reinforcement-learning agent;
measuring an uncertainty for each of the one or more episodes by inputting the one or more episodes into a second neural network model;
calculating an average of measured uncertainties for the one or more episodes;
choosing a largest average of the measured uncertainties among the calculated average of the uncertainties;
selecting an area of the environmental information corresponding to the chosen largest average of the uncertainties; and
predicting an area of the environmental information needed to be collected based on the measured uncertainty and the calculated average of the uncertainties by predicting the selected area of the environmental information as the area of the environmental information needed to be collected; and
training the first neural network model by inputting the predicted area of the environmental information to the reinforcement-learning agent, and wherein the reinforcement-learning agent uses the predicted environmental information for next time-point prediction to predict environmental information at continuous future time points.

2. The method of claim 1 , wherein the environmental information includes external environmental information, internal environmental information, and control information,
the external environmental information includes at least one of a vehicle speed, an external temperature, an air flow of an air conditioning device, an air inflow amount, weather information, or external humidity,
the internal environmental information includes at least one of evaporator information, heater information, cooler information, waste heat recovery information, temperature information, humidity information, air cleanliness information, or air flow information, and
the control information includes at least one of compressor information, valve information, heating amount information, control information for condenser, control information for an evaporator, control information for a radiator, control information for an accumulator, control information for a chiller, control information for an outdoor heat exchanger, control information for an air purifying device, or control information for a waste heat recovery device.

3. The method of claim 1 , wherein the outputting of one or more episodes by inputting the environmental information into the first neural network model includes
adjusting external environmental information included in the environmental information, and
outputting one or more episodes based on the adjusted external environmental information.

4. The method of claim 1 , wherein the first neural network model includes an environment to which the agent belongs, and wherein the environment to which the agent belongs is implemented based on a dynamics model.

5. The method of claim 1 , wherein the second neural network model is pre-trained based on the environmental information, and corresponds to a neural network model in which a Monte Carlo dropout scheme is used.

6. A computer program stored in a non-transitory computer-readable storage medium, wherein when the computer program is executed by one or more processors, the computer program causes the one or more processors to perform operations for predicting an area of environmental information needed to be collected, and the operations comprise:
an operation of outputting one or more episodes based on environmental information by inputting the environmental information into a first neural network model, wherein the first neural network model corresponds to a neural network model pre-trained based on the environmental information, and wherein the first neural network model includes a reinforcement-learning agent;
an operation of measuring uncertainty for each of the one or more episodes by inputting the one or more episodes into a second neural network model;
an operation of calculating an average of measured uncertainties for the one or more episodes;
an operation of choosing a largest average of the measured uncertainties among the calculated average of the uncertainties;
an operation of selecting an area of the environmental information corresponding to the chosen largest average of the uncertainties; and
an operation of predicting the selected area of the environmental information as an area of the environmental information needed to be collected; and
an operation of training the first neural network model by inputting the predicted area of the environmental information to the reinforcement-learning agent, and wherein the reinforcement-learning agent uses the predicted environmental information for next time-point prediction to predict environmental information at continuous future time points.

7. A computing device comprising:
at least one processor; and
a memory,
wherein the at least one processor is configured to
output one or more episodes based on environmental information by inputting the environmental information into a first neural network model, wherein the first neural network model corresponds to a neural network model pre-trained based on the environmental information, and wherein the first neural network model includes a reinforcement-learning agent,
measure uncertainty for each of the one or more episodes by inputting the one or more episodes into a second neural network model,
calculate an average of measured uncertainties for the one or more episodes,
choose a largest average of the measured uncertainties among the calculated average of the uncertainties,
select an area of the environmental information corresponding to the chosen largest average of the uncertainties,
predict an area of the environmental information needed to be collected based on the measured uncertainty and the calculated average of the uncertainties by predicting the selected area of the environmental information as the area of the environmental information needed to be collected, and
train the first neural network model by inputting the predicted area of the environmental information to the reinforcement-learning agent, and wherein the reinforcement-learning agent uses the predicted environmental information for next time-point prediction to predict environmental information at continuous future time points.

Assignees

• Makinarocks Co Ltd
• Hanon Systems

Inventors

• Lee Taeho
• Kim Minseop
• Choi Sanghyeok
• Lee Jeonghoon
• Kim Joongjae
• Ju Ikchan

Classifications

• G06N3/092 Primary

  Reinforcement learning · CPC title

• G06N3/096

  Transfer learning · CPC title

• G06N3/0895

  Weakly supervised learning, e.g. semi-supervised or self-supervised learning · CPC title

• G06N3/088

  Non-supervised learning, e.g. competitive learning · CPC title

• G06N3/09

  Supervised learning · CPC title