CN109376862A - A kind of time series generation method based on generation confrontation network - Google Patents
A kind of time series generation method based on generation confrontation network Download PDFInfo
- Publication number
- CN109376862A CN109376862A CN201811268291.7A CN201811268291A CN109376862A CN 109376862 A CN109376862 A CN 109376862A CN 201811268291 A CN201811268291 A CN 201811268291A CN 109376862 A CN109376862 A CN 109376862A
- Authority
- CN
- China
- Prior art keywords
- data
- module
- generator
- time series
- generation
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention proposes a kind of time series generation methods based on generation confrontation network, based on confrontation Network Theory is generated, design a kind of time series generation confrontation network, pass through limited time series data and generate extensive initial training data.This includes based on shot and long term memory models LSTM design generator to generate data, wherein carrying out potential time-dependent relation between learning data using LSTM, to be used to generate with the data with time-dependent relation for meeting data distribution.Based on expansible hypothesis testing design feature discriminator to carry out quality discrimination to generation data, discriminator learns the correspondence reward value of every generated data probability of step among feedback generator by time difference, it is trained based on the generator of LSTM by the Policy-Gradient of intensified learning, wherein, reward value is provided by discriminator return value, and then measures the availability for generating data.
Description
Technical field
The present invention relates to internet areas and deep learning field, and in particular to it is a kind of based on generate confrontation network when
Between sequence generating method.
Background technique
It is based primarily upon the thought for generating confrontation network based on the time series generation method for generating confrontation network, passes through identification
Time series generating process is generated large-scale dataset as a continuous decision process by feedback mechanism.Wherein, discriminator
Extraction time sequence signature simultaneously assesses each feature for the importance of sequence, by instructing to authentic specimen with sample is generated
Practice to measure the quality for generating data.Discriminator learns every generated data probability of step among feedback generator by time difference
Correspondence reward value, be trained based on the generator of LSTM by the Policy-Gradient of intensified learning, wherein reward value is by discriminator
Return value provides.Have closest to technology of the invention:
(1), the sequence data generation method based on SeqGAN: the thought based on confrontation, truthful data is plus generator
Data are generated to train discriminator.Due to the discrete output of generator, discriminator is allowed to be difficult one gradient of passback for more newborn
It grows up to be a useful person, it is therefore desirable to some changes are made, since discriminator is needed to a complete score sequences, so being exactly (to be covered with MCTS
Special Carlow tree search) by the various possibility completions of each sequence, discriminator generates reward, passback to these complete sequences
To generator, learn to update generator by enhancing.It is thus the mode with intensified learning, training one can produce text
The generation network of this sequence.
(2), the prediction generation method based on LSTM: in terms of LSTM itself can be used as the prediction for time series, so
Afterwards since the data and truthful data of its prediction have similitude, it can also be regarded as a time series data and be generated
Model.
(3), based on the sequence generating method of ORGAN: ORGAN be based on goal directed graphic data generation method, can be with
Data are carried out applied to various fields and generate work, based on the thought that confrontation generates, the rule of target are added in discriminator, is made
Generator study generates the data with goal rule, therefore can also be used for time series data generation.
Wherein, the sequence data generation method discrete state this for text based on SeqGAN has preferable effect,
However it is directed to the time series data of consecutive variations, effect is simultaneously bad.Prediction generation method based on LSTM is for time sequence
The generation method of column data is with the extension of time, error will do it accumulative, and the time is longer, and effect is poorer；Sequence based on ORGAN
It is to be directed to regular design to need domain expert personnel participation model construction that column-generation method, which has a defect substantially,
This just proposes very high request to peopleware, needs to have cross discipline knowledge.Spy based on expansible hypothesis testing design
Discriminator is levied, not yet someone uses so far, and this calculating for generating the quality of data is mentioned from truthful data by comparison
The feature taken out carries out.
Summary of the invention
To solve shortcoming and defect in the prior art, the invention proposes a kind of based on the time sequence for generating confrontation network
Column-generation method fights generation time series data according to data characteristics and temporal aspect.
The technical solution of the present invention is as follows:
A kind of time series generation method based on generation confrontation network, generator, discriminator and master-plan, including with
Lower part:
(1), the Maker model G of one θ of training parametrizationθGenerate the time series Y that length is T1:T=(y1,…,
yt,…,yT).In addition, this method also trains the discriminator model D of φ parametrizationφ, for instructing to improve Maker model.
Maker model is trained by the Policy-Gradient of intensified learning.Wherein, reward value is provided by discriminator value of feedback, passes through the time
The corresponding reward of intermediate every the generated data probability of step of difference study feedback.Discriminator is by truthful data and generates data progress
Training.
The target of generator is the expectation maximized from original state to the reward value for generating complete time series, is calculated
Formula is as follows:
Wherein, RTIndicate the reward of entire sequence, Q (s=Y1:t-1, a=yt) be formation sequence effect assessment function, table
Show the reward for a that takes action at state s.Gθ(y | s) indicate that generator generates the probability of y at state s.
For complete time series Q (s=YT-1, a=yT)=Dφ(Y1:T), Dφ(Y1:T) it is that discriminator passes through relative entropy
Measuring the feature distribution difference between real sequence and formation sequence is the reward function that the sequence being currently generated provides:
For incomplete sequence, in order to assess the reward value of intermediate state, learn to obtain using time difference unknown
Remaining T-t data,Wherein It is to be based on
Roll-out strategy and current state sample to obtain.Since current state, it is defeated to obtain a batch for operation n times roll-out strategy
Sample out, then effect assessment function is as follows:
Wherein, αhThe learning rate of h step is represented, γ represents attenuation rate, rt+1To reward immediately, D is usedφ(Y1:T) as prize
The advantages of encouraging function is can be iteratively improving generation model by dynamically more newly arriving.Once generating one group more really generates sample
This, updates TSFD parameter phi re -training discriminator by formula (3):
Training objective is to maximize identification authentic specimen Y~Pdata(Y) probability, minimum misrecognition forgery sample Y~
PGθ(Y) probability.After re -training discriminator, the method based on Policy-Gradient is used to maximize the expectation of reward.Objective function
(4) the unbiased esti-mator such as formula of the gradient of J (θ), then passes through formula and (5) updates generator parameter θ:
(2), based on the time series feature extraction of expansible hypothesis testing, using time domain distance, difference distance, nerve net
Network scheduling algorithm extracts various features, establishes μ=1 in sequence in different times ..., nμFeature vector, Xμ.According to equation (7)
To hypothesisIt carries out multiple hypothesis test and obtains each hypothesis testingResult p-valuePass through puppet
Discovery rate rμCorrelation hypothesis is modified, returns and assumesThe feature (the strong feature of importance) being rejected.Discriminator TSFD
By the important feature of extraction time sequence, is measured and generated between sample and authentic specimen feature distribution by relative entropy (9)
Difference.
(3), time series is generated as generator using LSTM, the generator by the Policy-Gradient of intensified learning into
Row training, wherein the reward function of intensified learning is feedback of the discriminator for given time sequence, and LSTM generator use is passed
The renewal function g returned will input x1,…,xTIt is mapped to hidden state h1,…,hT, it may be assumed that
ht=g (ht-1,xt) (10)
Hidden state is mapped in data distribution with a softmax output layer z again.
p(yt|x1,…,xt)=z (ht)=softmax (c+Vht) ⑾
C is bigoted vector, and V is weight matrix.
Beneficial effects of the present invention:
(1) by using LSTM as generator, effectively the time relationship of time series data is learnt；
(2) algorithm uses and carries out similarity-rough set based on feature extraction simultaneously, has more compared to common comparative approach
Good effect, can learn to more accurate initial data feature；
(3) algorithm can efficiently generate big-sample data based on Small Sample Database, provide data for deep learning and support.
Detailed description of the invention
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technical description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this
Some embodiments of invention for those of ordinary skill in the art without creative efforts, can be with
It obtains other drawings based on these drawings.
Fig. 1 is that the present invention is based on the time series generation method structure charts for generating confrontation network.
Fig. 2 is that the present invention is based on the flow charts for the time series generation method for generating confrontation network.
Fig. 3 is that the present invention is based on the overview flow charts for the time series generation method for generating confrontation network.
Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based on
Embodiment in the present invention, it is obtained by those of ordinary skill in the art without making creative efforts every other
Embodiment shall fall within the protection scope of the present invention.
As shown in Figure 1, the system structure of the invention based on the time series generation method for generating confrontation network includes five
A module: maker module, Monte Carlo module, LSTM module (deep learning module) and expansible hypothesis testing feature extraction
Device module.
Below with reference to Fig. 1 and Fig. 2, the detailed process for the time series generation method for generating confrontation network is carried out specifically
It is bright:
Step (1), in maker module, there are the data being centainly distributed for generating from random noise, for generating
Original data；
Step (2), in LSTM module, generated by using the network of truthful data pre-training with certain data characteristics
Data, and be added to maker module generation noise in formed generate data；
Step (3), in the module of Monte Carlo, the data of generation are subjected to completion according to certain rule, formed one it is complete
Whole time series；
Step (4) will carry out the various features of truthful data in expansible hypothesis testing feature extractor module
It extracts, the model of various truthful data features is preserved in training one；
Step (5), in discriminator module, the quality for generating data will be evaluated according to the feature of really data, it is raw
At value of feedback, value of feedback is fed back to generator by Monte Carlo module and is iterated update.
Of the invention is had based on the time series generation method for generating confrontation network by using LSTM as generator
Effect learns the time relationship of time series data；The algorithm, which is used, simultaneously carries out similarity ratio based on feature extraction
Compared with, compared to common comparative approach have better effect, can learn to more accurate initial data feature；The algorithm can be with
Big-sample data is efficiently generated based on Small Sample Database, sufficiently expands the diversity of data, data is provided for deep learning and supports.
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all in essence of the invention
Within mind and principle, any modification, equivalent replacement, improvement and so on be should all be included in the protection scope of the present invention.
Claims (1)
1. a kind of based on the time series generation method for generating confrontation network, which is characterized in that data preprocessing module, data are raw
At module, link block, data identification module and iteration update module, comprising the following steps:
Step (1), in data preprocessing module, according to data input to data carry out data preprocessing operation include normalization,
The operation such as completion, discretization；
Step (2), in data generation module, pre-training is carried out to generator using a part of truthful data, so that the generator
There is a preferable initiation parameter；
Step (3), in link block, it will generation module export data carry out conversion and completion sequence, formed identify
Data format required for device；
Step (4), in identification module, it will the data of input are compared with truthful data, with identify generate data it is true
Reality, and identification result is fed back into generator, the parameter for adjusting generator updates；
Step (5), in iteration update module, for the update and operation of entire frame system, dispatch modules connection and
The update of parameter and the training of module.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811268291.7A CN109376862A (en) | 2018-10-29 | 2018-10-29 | A kind of time series generation method based on generation confrontation network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811268291.7A CN109376862A (en) | 2018-10-29 | 2018-10-29 | A kind of time series generation method based on generation confrontation network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109376862A true CN109376862A (en) | 2019-02-22 |
Family
ID=65390416
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811268291.7A Pending CN109376862A (en) | 2018-10-29 | 2018-10-29 | A kind of time series generation method based on generation confrontation network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109376862A (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109978550A (en) * | 2019-03-12 | 2019-07-05 | 同济大学 | A kind of credible electronic transaction clearance mechanism based on generation confrontation network |
| CN110175168A (en) * | 2019-05-28 | 2019-08-27 | 山东大学 | A kind of time series data complementing method and system based on generation confrontation network |
| CN110727844A (en) * | 2019-10-21 | 2020-01-24 | 东北林业大学 | Online commented commodity feature viewpoint extraction method based on generation countermeasure network |
| CN111310915A (en) * | 2020-01-21 | 2020-06-19 | 浙江工业大学 | Data anomaly detection and defense method for reinforcement learning |
| CN111370074A (en) * | 2020-02-27 | 2020-07-03 | 北京晶派科技有限公司 | Method and device for generating molecular sequence and computing equipment |
| CN111401556A (en) * | 2020-04-22 | 2020-07-10 | 清华大学深圳国际研究生院 | Selection method of opponent type imitation learning winning incentive function |
| CN111863236A (en) * | 2019-04-24 | 2020-10-30 | 通用电气精准医疗有限责任公司 | Medical machine composite data and corresponding event generation |
| CN116054185A (en) * | 2023-03-30 | 2023-05-02 | 武汉新能源接入装备与技术研究院有限公司 | Control method of reactive power compensator |
-
2018
- 2018-10-29 CN CN201811268291.7A patent/CN109376862A/en active Pending
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109978550A (en) * | 2019-03-12 | 2019-07-05 | 同济大学 | A kind of credible electronic transaction clearance mechanism based on generation confrontation network |
| US11984201B2 (en) | 2019-04-24 | 2024-05-14 | GE Precision Healthcare LLC | Medical machine synthetic data and corresponding event generation |
| CN111863236A (en) * | 2019-04-24 | 2020-10-30 | 通用电气精准医疗有限责任公司 | Medical machine composite data and corresponding event generation |
| CN110175168B (en) * | 2019-05-28 | 2021-06-01 | 山东大学 | Time sequence data filling method and system based on generation of countermeasure network |
| CN110175168A (en) * | 2019-05-28 | 2019-08-27 | 山东大学 | A kind of time series data complementing method and system based on generation confrontation network |
| CN110727844A (en) * | 2019-10-21 | 2020-01-24 | 东北林业大学 | Online commented commodity feature viewpoint extraction method based on generation countermeasure network |
| CN110727844B (en) * | 2019-10-21 | 2022-07-01 | 东北林业大学 | Online commented commodity feature viewpoint extraction method based on generation countermeasure network |
| CN111310915A (en) * | 2020-01-21 | 2020-06-19 | 浙江工业大学 | Data anomaly detection and defense method for reinforcement learning |
| CN111310915B (en) * | 2020-01-21 | 2023-09-01 | 浙江工业大学 | Data anomaly detection defense method oriented to reinforcement learning |
| CN111370074B (en) * | 2020-02-27 | 2023-07-07 | 北京晶泰科技有限公司 | Method and device for generating molecular sequence and computing equipment |
| CN111370074A (en) * | 2020-02-27 | 2020-07-03 | 北京晶派科技有限公司 | Method and device for generating molecular sequence and computing equipment |
| CN111401556A (en) * | 2020-04-22 | 2020-07-10 | 清华大学深圳国际研究生院 | Selection method of opponent type imitation learning winning incentive function |
| CN116054185A (en) * | 2023-03-30 | 2023-05-02 | 武汉新能源接入装备与技术研究院有限公司 | Control method of reactive power compensator |
| CN116054185B (en) * | 2023-03-30 | 2023-06-02 | 武汉新能源接入装备与技术研究院有限公司 | Control method of reactive power compensator |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109376862A (en) | A kind of time series generation method based on generation confrontation network | |
| CN107132516B (en) | A kind of Radar range profile's target identification method based on depth confidence network | |
| CN109120652A (en) | It is predicted based on difference WGAN network safety situation | |
| CN106469560A (en) | A kind of speech-emotion recognition method being adapted to based on unsupervised domain | |
| CN101334893A (en) | Fused image quality integrated evaluating method based on fuzzy neural network | |
| CN108229718A (en) | A kind of information forecasting method and device | |
| CN108924836A (en) | A kind of edge side physical layer channel authentication method based on deep neural network | |
| CN110097095A (en) | A kind of zero sample classification method generating confrontation network based on multiple view | |
| Kinsler | Beyond levels and growth: Estimating teacher value-added and its persistence | |
| CN108957418A (en) | A kind of radar target identification method based on Recognition with Recurrent Neural Network model | |
| CN107392164A (en) | A kind of Expression analysis method based on the estimation of Facial action unit intensity | |
| CN114611670A (en) | Knowledge distillation method based on teacher-student cooperation | |
| CN108073978A (en) | A kind of constructive method of the ultra-deep learning model of artificial intelligence | |
| Reichstein et al. | Modelling landsurface time-series with recurrent neural nets | |
| Burrows et al. | Characterizing the relative importance assigned to physical variables by climate scientists when assessing atmospheric climate model fidelity | |
| CN108073979A (en) | A kind of ultra-deep study of importing artificial intelligence knows method for distinguishing for image | |
| Pei et al. | Self-Attention Gated Cognitive Diagnosis for Faster Adaptive Educational Assessments | |
| Epstein et al. | The perils of tweaking: how to use macrodata to set parameters in complex simulation models | |
| Casallas-Lagos et al. | Characterizing the temporal evolution of the high-frequency gravitational wave emission for a core collapse supernova with laser interferometric data: A neural network approach | |
| CN108073985A (en) | A kind of importing ultra-deep study method for voice recognition of artificial intelligence | |
| Green et al. | Federated learning with highly imbalanced audio data | |
| Mazinani et al. | Prediction of success or fail of students on different educational majors at the end of the high school with artificial neural networks methods | |
| Hu et al. | Misclassification and the hidden silent rivalry | |
| Farshi | The Principle of Luck Conservation: Unlucking the Secrets of Luck Using Conjugate Variables | |
| Ma et al. | Analysis And Prediction Of Body Test Results Based On Improved Backpropagation Neural Network Algorithm |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| CB03 | Change of inventor or designer information | ||
| CB03 | Change of inventor or designer information |
Inventor after: Zhang Yafei Inventor after: Zhang Weishan Inventor after: Lin Weixian Inventor after: Liu Cuan Inventor after: Geng Zukun Inventor before: Zhang Weishan Inventor before: Zhang Yafei Inventor before: Lin Weixian Inventor before: Liu Cuan Inventor before: Geng Zukun |
|
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20190222 |