Looking for breakthrough ideas for innovation challenges? Try Patsnap Eureka!

Action model based on deep learning and training method thereof

A technology of deep learning and training methods, applied in the field of deep learning neural network models, which can solve problems such as information loss and inability to guarantee accuracy

Active Publication Date: 2020-02-28
SUN YAT SEN UNIV
View PDF2 Cites 10 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

In practical problems, when people record a sequence, the recorded state may not be completely observable due to interference in the recording process or other factors, but some information is lost and partially observable ( For example, when shooting a video, there is a partial occlusion between the target and the camera), due to the loss of information, the final model training and planning problem solving cannot guarantee the accuracy

Method used

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View more

Image

Smart Image Click on the blue labels to locate them in the text.
Viewing Examples
Smart Image
  • Action model based on deep learning and training method thereof
  • Action model based on deep learning and training method thereof
  • Action model based on deep learning and training method thereof

Examples

Experimental program
Comparison scheme
Effect test

Embodiment 1

[0062] Such as figure 1 Shown is an embodiment of an action model based on deep learning, including a data completion module, a data encoding and decoding module, a state reasoning module and a search planning module;

[0063] The data completion module is used to predict the missing part of the original data P, and supplement it to the original data to generate complete and observable data O;

[0064] The data codec module is used to realize the two-way conversion of the data O in the original form and the data S of the propositional form in the implicit space; The digital codec module includes two submodules: an encoding module and a decoding module; the encoding module is used to convert The data O in the original form is encoded as a proposition S in the hidden space; the decoding module is used to decode the proposition S in the hidden space to obtain the data O in the original form; for the data encoding and decoding module, the present invention uses a variational autoe...

Embodiment 2

[0070] Such as Figure 2-3 Shown is a kind of embodiment of the training method of the action model based on deep learning, for training the action model based on deep learning of embodiment 1, wherein image 3 In order to have collected some observation data of the 8-digit problem, it is recorded as data set X, and the specific steps are as follows:

[0071] Step 1: Input the data set X into the data completion module, the sample form of the data set X is , P is a d-dimensional continuous vector; M is a d-dimensional discrete vector;

[0072] Step 2: The data completion module completes the data set X to obtain the data set X2, and uses the data set to train the codec module;

[0073] Step 3: The encoding sub-module of the data encoding and decoding module encodes the samples of the data set X2 to obtain the data set X3 and use the data set to train the state reasoning module.

[0074] Among them, the training of the data completion module includes the training of the gener...

Embodiment 3

[0113] Such as Figure 4 Shown are the initial observations and target observations of some eight-number problems to be solved, which are recorded as set Q. Input the set Q into the action model that has been trained in Embodiment 2, solve the problems in the set Q, and for each problem, given as Figure 4 The observation picture P of the initial state shown 0 and the position indicator variable M of the missing part 0 , and the observed picture P of the target state g and the position indicator variable M of the missing part g ,Such as Figure 5 As shown, the specific steps in the planning stage are as follows:

[0114] Step 1, the observation picture P of the initial state 0 and the position indicator variable M of the missing part 0 , and the observed picture P of the target state g and the position indicator variable M of the missing part g , respectively input into the generator G of the data completion module to obtain the completed initial state observation pic...

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

PUM

No PUM Login to View More

Abstract

The invention relates to an action model based on deep learning and a training method thereof. The action model comprises a data completion module, a data encoding and decoding module and a state reasoning module. And after the action model is trained, a planning problem can be solved. According to the action model, the missing part of the original data can be complemented, and the problem of pooraccuracy caused by partial missing of the original data when the action model is trained is effectively solved. The action model learns the expression of the proposition form of the state in the implicit space through data training, and learns the reasoning ability in the implicit space, so that the problem can obtain a solution sequence through a search algorithm. The training process is unsupervised learning, the advantage of summarizing rules in a large amount of data of deep learning is efficiently utilized, and the cost of manual modeling is not consumed.

Description

technical field [0001] The present invention relates to the field of deep learning neural network models, and more specifically, to an action model based on deep learning and a training method thereof. Background technique [0002] When applying the techniques of classical planning to real life, there is an inevitable process, namely modeling. Abstract real-life problems into expressions in the form of propositions, and learn the action model (ActionModel), that is, by making full use of their own prior knowledge, and even discovering the laws of the problem domain, human beings can make possible occurrences in the domain The premise (Predicate) and effect (Effect) of the action (Action) are summarized, and a strict model that can perform logical calculations is established. When the problem is relatively simple, the workload of this process is acceptable; but when encountering complex problems, the requirements for modeling work on the modeler are significantly increased, ...

Claims

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

Application Information

Patent Timeline
no application Login to View More
Patent Type & Authority Applications(China)
IPC IPC(8): G06N3/04G06N3/08
CPCG06N3/08G06N3/045
Inventor 蔡佳然卓汉逵
Owner SUN YAT SEN UNIV
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Patsnap Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Patsnap Eureka Blog
Learn More
PatSnap group products

Browse by: Latest US Patents, China's latest patents, Technical Efficacy Thesaurus, Application Domain, Technology Topic, Popular Technical Reports.

© 2024 PatSnap. All rights reserved.Legal|Privacy policy|Modern Slavery Act Transparency Statement|Sitemap|About US| Contact US: help@patsnap.com