CN107274029A - A kind of future anticipation method of interaction medium in utilization dynamic scene - Google Patents
A kind of future anticipation method of interaction medium in utilization dynamic scene Download PDFInfo
- Publication number
- CN107274029A CN107274029A CN201710487508.2A CN201710487508A CN107274029A CN 107274029 A CN107274029 A CN 107274029A CN 201710487508 A CN201710487508 A CN 201710487508A CN 107274029 A CN107274029 A CN 107274029A
- Authority
- CN
- China
- Prior art keywords
- mrow
- mover
- medium
- msubsup
- msub
- 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.)
- Withdrawn
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- 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
Abstract
The future anticipation method of interaction medium in a kind of utilization dynamic scene proposed in the present invention, its main contents include：Multisample generation, sample sequence are merged with perfect, scene, its process is, the multisample for assuming prediction is generated first by condition variation self-encoding encoder, the marking and queuing of sample is carried out followed by the long-term favorable decisions method based on reverse optimum control framework and improves processing, introduce Recognition with Recurrent Neural Network and scene fusion is carried out to the medium of image-context, activity, obtained result is fed back in decoder again to be iterated.The present invention can handle the future event foresight activity of a certain medium in scene dynamically or statically and generate, sorts and improve sample there is provided a Recognition with Recurrent Neural Network set, significantly obtain the precise effect of prediction.
Description
Technical field
The present invention relates to action prediction field, more particularly, to interaction medium in a kind of utilization dynamic scene not
Carry out Forecasting Methodology.
Background technology
The response to predict the following scene to this track is carried out according to passing action trail, i.e., future event is predicted, is
Attract the subject under discussion greatly paid close attention in recent years.Under the background of big data increased popularity, man-machine interaction, robot navigation etc. can meet
Many unprecedented scenes, can high degree training smart equipment progress current scene and task with future anticipation method
Real-time distribution with implement, simultaneously for the huge field of task amount, as population surveillance, resource exploration etc. possess hardware facility
Scene will have high application potential, in addition, meeting an urgent need and arranging in multitask command system distributed coordination, a variety of vehicles anticollisions
Apply, also have immeasurable humane and market value in terms of natural calamity escape route forewarning management.
Future anticipation is still a very highly difficult challenge.Because the scene being related in task both includes vision
On dimension, considering in logic is also contains, while also assigning the function of this task Intelligent Recognition, it is necessary to high data volume
Training test process.On the one hand, space-time instability causes Fuzzy Influence to visual reception, while the responsible journey of data volume
Amount of calculation increase can cause the reduction of live effect caused by degree, on the other hand, and the medium interacted too much produces mutual shadow
Ring, it is impossible to peel off carry out quantitative analysis one by one, therefore the prediction of action is judged to cause difficulty.
The present invention proposes a kind of long-term favorable decisions new frame based on reverse optimum control.First by condition variation
Self-encoding encoder assumes the multisample of prediction to generate, followed by the long-term favorable decisions method based on reverse optimum control framework
Carry out the marking and queuing of sample and improve processing, introducing Recognition with Recurrent Neural Network carries out scene to the medium of image-context, activity and melted
Close, obtained result is fed back in decoder again to be iterated.The present invention can handle a certain in scene dynamically or statically
The future event foresight activity of medium generates, sorts and improved sample there is provided Recognition with Recurrent Neural Network set, significantly takes
The precise effect that must be predicted.
The content of the invention
The problem of for solving to predict future-action in complex scene multi interaction medium, it is an object of the invention to
There is provided a kind of future anticipation method of interaction medium in utilization dynamic scene, it is proposed that a kind of based on reverse optimum control
Long-term favorable decisions new frame.
To solve the above problems, the present invention provides a kind of future anticipation side of interaction medium in utilization dynamic scene
Method, its main contents include：
(1) multisample is generated；
(2) sample sequence with it is perfect；
(3) scene is merged.
Wherein, described multisample generation, including universal model construction and condition variation self-encoding encoder.
Further, described universal model construction, for playing interactive object in generation event in certain scene,
No matter static or motion, referred to as medium, if giving n medium, their passing movement track X=<X1,X2,…,Xn
>, calculate their future behaviour track Y=<Y1,Y2,…,Yn>Respective probability then be P<Y|X,I>, especially, I is should
The input at scene lower a moment, the maximum probability obtained according to input is exported as prediction locus.
Further, described condition variation self-encoding encoder, learns a decision function f using depth generation model,
Passing movement track X and input I are mapped on the Y of future behaviour track by the function, specifically：
(1) random latent variable z is introducedi, in input XiUnder conditions of, study to output YiProbability distribution be P (Yi|Xi),
It is respectively identification network Q that this process, which builds corresponding neutral net,φ(zi|Yi,Xi), condition pro-active network Pv(zi|Xi) and it is raw
Into network Pθ(Yi|Xi,zi)；
(2) in the training stage：1. respectively with two Recognition with Recurrent Neural Network to medium i, XiAnd YiEncoded, obtain two
As a result 2. the two results are merged and is delivered to one layer of fully-connected network using nonlinear activation function again；3. again
Generated with two parallel fully-connected networks on ziAverageAnd standard deviationGaussian distribution model is modeled as according to it, and
And be normalized with KL divergences；4. propose that two loss functions are adjusted to self-encoding encoder, respectively rebuild loss function With KL divergence loss functions lKLD=DKL(Qφ(zi|Yi, Xi)||Pv(zi))；
(3) in test phase：1. due to future anticipation trackDo not apply to and be rejected, therefore utilize passing track
With multiple latent variable zi (k)Sample merges；2. it is relative with the training stage, nowWith β (zi (k)) it is input to Recognition with Recurrent Neural Network solution
The hypothesis set of multiple predictions is produced in code device.
Further, described sample sequence with it is perfect, including sample evaluation improves and iterative feedback two parts.
Further, described sample evaluates perfect, is predicted for future event, using the decision-making in enhancing learning framework
Formulating method, a medium can have selected after training one reward for a long time it is maximized action go implement as future thing
Part, now designs a Recognition with Recurrent Neural Network model and goes prediction to being made in each trainingCarry out long term accumulation reward
Measurement, be specially：
(1) medium i is scored, there is K forecast sampleThen the evaluation score s of each sample can
Have：
Wherein,Refer to the forecast sample of other All Medias in addition to i,It is pre- test samples of the medium i in time t
This,It is forecast samples all before knowing timestamp t, ψ is allocated to each time step t reward function, should
Function is realized with the fully connected network network layers for being connected to Recognition with Recurrent Neural Network；
(2) perfect to medium i, in scoring process, Recognition with Recurrent Neural Network model estimates a regression vector simultaneouslyThe regression function η that the estimation procedure is used is：
Wherein, regression function η can accumulate all passing scene environments and the dynamic of all interaction mediums, while
Estimated on whole time dimension optimal
Further, described iterative feedback, optimum regression substituting vector is obtained using formula (2)Go gradually complete
The hypothesis set of kind predictionSpecially：
(1) each iteration,All it is updated toIt is then delivered to reverse optimum control module；
(2) reverse optimum control training in sequence and improve module, they have two kinds loss items, be respectively：Intersect
Entropy loss lCE=H (p, q), wherein q are obtained by activation primitive, i.e.,Return
Lose item
(3) to a training batch in neutral net, its whole multitask loss item is：
Wherein, N is the number of the batch medium.
Further, described scene fusion, including context insertion and interaction feature.
Further, described context insertion, using Recognition with Recurrent Neural Network, receives following on each timestamp t
Input Xt：
Wherein,It isIn the speed of t, γ is one layer of full articulamentum with nonlinear activation function, its handle
Speed is mapped to higher-dimension and characterizes space,It will be located atConvolutional neural networks feature ρ (I) carry out pondization operate,
R is gathered together the hiding vector on other medium spatial domains by a fused layer, self-embedding vectorIt is used as circulation nerve
The initialization hidden state vector of network.
Further, described interaction feature, using the spatial domain grid operations using pond layer operation, makes different medium not
With the unlikely disappearance of the interaction between sample, it is specially：In moment t medium i, its sample k spatial domain grid cell with
Centered on, then to each grid cell g, the feature of its interior Hide AllAverage pond will all be carried out
Change operation.
Brief description of the drawings
Fig. 1 is the system flow chart of the future anticipation method of interaction medium in a kind of utilization dynamic scene of the invention.
Fig. 2 is the example of the future anticipation method of interaction medium in a kind of utilization dynamic scene of the invention.
Embodiment
It should be noted that in the case where not conflicting, the feature in embodiment and embodiment in the application can phase
Mutually combine, the present invention is described in further detail with specific embodiment below in conjunction with the accompanying drawings.
Fig. 1 is the system flow chart of the future anticipation method of interaction medium in a kind of utilization dynamic scene of the invention.
Mainly include data input；Multisample is generated；Sample sort with it is perfect；Scene is merged.
Wherein, multisample is generated, including universal model construction and condition variation self-encoding encoder.
Universal model is constructed, for playing interactive object in generation event in certain scene, no matter static or fortune
Move, referred to as medium, if giving n medium, their passing movement track X=<X1,X2,…,Xn>, calculate theirs
Future behaviour track Y=<Y1,Y2,…,Yn>Respective probability then be P<Y|X,I>, especially, I is defeated for the scene lower a moment
Enter, the maximum probability obtained according to input is exported as prediction locus.
Described condition variation self-encoding encoder, learns a decision function f using depth generation model, and the function was incited somebody to action
It is mapped to toward movement track X and input I on the Y of future behaviour track, specifically：
(1) random latent variable z is introducedi, in input XiUnder conditions of, study to output YiProbability distribution be P (Yi|Xi),
It is respectively identification network Q that this process, which builds corresponding neutral net,φ(zi|Yi,Xi), condition pro-active network Pv(zi|Xi) and it is raw
Into network Pθ(Yi|Xi,zi)；
(2) in the training stage：1. respectively with two Recognition with Recurrent Neural Network to medium i, XiAnd YiEncoded, obtain two
As a result 2. the two results are merged and is delivered to one layer of fully-connected network using nonlinear activation function again；3. again
Generated with two parallel fully-connected networks on ziAverageAnd standard deviationGaussian distribution model is modeled as according to it, and
And be normalized with KL divergences；4. propose that two loss functions are adjusted to self-encoding encoder, respectively rebuild loss function With KL divergence loss functions lKLD=DKL(Qφ(zi|Yi,Xi)‖Pv(zi))；
(3) in test phase：1. due to future anticipation trackDo not apply to and be rejected, therefore utilize passing track
With multiple latent variable zi (k)Sample merges；2. it is relative with the training stage, nowWith β (zi (k)) it is input to Recognition with Recurrent Neural Network solution
The hypothesis set of multiple predictions is produced in code device.
Sample sort with it is perfect, including sample evaluation improve with iterative feedback two parts.
Sample evaluates perfect, is predicted for future event, using the decision-making method in enhancing learning framework, Jie
Matter can have selected one after training rewards maximized action and goes to implement as following event for a long time, and now design one is followed
Ring neural network model goes the prediction to being made in each trainingThe measurement of long term accumulation reward is carried out, is specially：
(1) medium i is scored, there is K forecast sampleThen the evaluation score s of each sample can
Have：
Wherein,Refer to the forecast sample of other All Medias in addition to i,It is pre- test samples of the medium i in time t
This,It is forecast samples all before knowing timestamp t, ψ is allocated to each time step t reward function, should
Function is realized with the fully connected network network layers for being connected to Recognition with Recurrent Neural Network；
(2) perfect to medium i, in scoring process, Recognition with Recurrent Neural Network model estimates a regression vector simultaneouslyThe regression function η that the estimation procedure is used is：
Wherein, regression function η can accumulate all passing scene environments and the dynamic of all interaction mediums, while
Estimated on whole time dimension optimal
Iterative feedback, optimum regression substituting vector is obtained using formula (2)Go gradually to improve the hypothesis set of predictionSpecially：
(1) each iteration,All it is updated toIt is then delivered to reverse optimum control module；
(2) reverse optimum control training in sequence and improve module, they have two kinds loss items, be respectively：Intersect
Entropy loss lCE=H (p, q), wherein q are obtained by activation primitive, i.e.,Return
Lose item
(3) to a training batch in neutral net, its whole multitask loss item is：
Wherein, N is the number of the batch medium.
Scene is merged, including context insertion and interaction feature.
Context is embedded in, using Recognition with Recurrent Neural Network, receives following input X on each timestamp tt：
Wherein,It isIn the speed of t, γ is one layer of full articulamentum with nonlinear activation function, its handle
Speed is mapped to higher-dimension and characterizes space,It will be located atConvolutional neural networks feature ρ (I) carry out pondization operate,
R is gathered together the hiding vector on other medium spatial domains by a fused layer, self-embedding vectorIt is used as circulation nerve
The initialization hidden state vector of network.
Interaction feature, using the spatial domain grid operations using pond layer operation, makes the phase between the different samples of different medium
Interaction is unlikely to disappear, and is specially：In moment t medium i, its sample k spatial domain grid cell withCentered on, then to every
One grid cell g, the feature of its interior Hide AllAverage pondization operation will all be carried out.
Fig. 2 is the example of the future anticipation method of interaction medium in a kind of utilization dynamic scene of the invention.As schemed
Show, it is observed that by different iterationses, model learning to feedback gradually strengthen, the route of prediction is also increasingly leaned on
Nearly real route, that is, thick line of adding some points.
For those skilled in the art, the present invention is not restricted to the details of above-described embodiment, in the essence without departing substantially from the present invention
In the case of refreshing and scope, the present invention can be realized with other concrete forms.In addition, those skilled in the art can be to this hair
Bright to carry out various changes and modification without departing from the spirit and scope of the present invention, these improvement and modification also should be regarded as the present invention's
Protection domain.Therefore, appended claims are intended to be construed to include preferred embodiment and fall into all changes of the scope of the invention
More and modification.
Claims (10)
1. a kind of future anticipation method of interaction medium in utilization dynamic scene, it is characterised in that mainly including multisample
Generate (one)；Sample sorts and perfect (two)；Scene merges (three).
2. (one) is generated based on the multisample described in claims 1, it is characterised in that become including universal model construction and condition
Divide self-encoding encoder.
3. based on the universal model construction described in claims 2, it is characterised in that for being risen in certain scene in generation event
Interactive object, no matter static or motion, referred to as medium, if giving n medium, their passing movement track
X=＜ X1,X2,…,Xn＞, calculates their future behaviour track Y=＜ Y1,Y2,…,Yn＞ respective probability is then P ＜ Y | X,
I ＞, especially, I are the input at the scene lower a moment, and the maximum probability obtained according to input is exported as prediction locus.
4. based on the condition variation self-encoding encoder described in claims 2, it is characterised in that learnt using depth generation model
Passing movement track X and input I are mapped on the Y of future behaviour track by one decision function f, the function, specifically：
(1) random latent variable z is introducedi, in input XiUnder conditions of, study to output YiProbability distribution be P (Yi|Xi), this mistake
It is respectively identification network Q that journey, which builds corresponding neutral net,φ(zi|Yi,Xi), condition pro-active network Pv(zi|Xi) and generation net
Network Pθ(Yi|Xi,zi)；
(2) in the training stage：1. respectively with two Recognition with Recurrent Neural Network to medium i, XiAnd YiEncoded, obtain two results 2. the two results are merged and is delivered to one layer of fully-connected network using nonlinear activation function again；3. again with simultaneously
Two capable fully-connected networks are generated on ziAverageAnd standard deviationGaussian distribution model is modeled as according to it, and is used
KL divergences are normalized；4. propose that two loss functions are adjusted to self-encoding encoder, respectively rebuild loss function With KL divergence loss functions
(3) in test phase：1. due to future anticipation trackDo not apply to and be rejected, therefore utilize passing trackWith it is many
Individual latent variable zi (k)Sample merges；2. it is relative with the training stage, nowWith β (zi (k)) it is input to Recognition with Recurrent Neural Network decoder
The middle hypothesis set for producing multiple predictions.
5. based on the sample sequence described in claims 1 and perfect (two), it is characterised in that improve and change including sample evaluation
Generation feedback two parts.
6. evaluate perfect based on the sample described in claims 5, it is characterised in that predicted for future event, using enhancing
Decision-making method in learning framework, a medium can have selected after training one reward for a long time it is maximized action go reality
Apply as following event, now design a Recognition with Recurrent Neural Network model and go prediction to being made in each trainingCarry out
The measurement of long term accumulation reward, be specially：
(1) medium i is scored, there is K forecast sampleThen the evaluation score s of each sample can have：
<mrow>
<mi>s</mi>
<mrow>
<mo>(</mo>
<msubsup>
<mover>
<mi>Y</mi>
<mo>^</mo>
</mover>
<mi>i</mi>
<mrow>
<mo>(</mo>
<mi>k</mi>
<mo>)</mo>
</mrow>
</msubsup>
<mo>;</mo>
<mi>I</mi>
<mo>,</mo>
<mi>X</mi>
<mo>,</mo>
<msubsup>
<mover>
<mi>Y</mi>
<mo>^</mo>
</mover>
<mrow>
<mi>j</mi>
<mo>\</mo>
<mi>i</mi>
</mrow>
<mrow>
<mo>(</mo>
<mo>&ForAll;</mo>
<mo>)</mo>
</mrow>
</msubsup>
<mo>)</mo>
</mrow>
<mo>=</mo>
<munder>
<mo>&Sigma;</mo>
<mrow>
<mi>t</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
</munder>
<mi>&psi;</mi>
<mrow>
<mo>(</mo>
<msubsup>
<mover>
<mi>y</mi>
<mo>^</mo>
</mover>
<mrow>
<mi>i</mi>
<mo>,</mo>
<mi>t</mi>
</mrow>
<mrow>
<mo>(</mo>
<mi>k</mi>
<mo>)</mo>
</mrow>
</msubsup>
<mo>;</mo>
<mi>I</mi>
<mo>,</mo>
<mi>X</mi>
<mo>,</mo>
<msubsup>
<mover>
<mi>Y</mi>
<mo>^</mo>
</mover>
<mrow>
<mi>&tau;</mi>
<mo><</mo>
<mi>t</mi>
</mrow>
<mrow>
<mo>(</mo>
<mo>&ForAll;</mo>
<mo>)</mo>
</mrow>
</msubsup>
<mo>)</mo>
</mrow>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>1</mn>
<mo>)</mo>
</mrow>
</mrow>
Wherein,Refer to the forecast sample of other All Medias in addition to i,It is forecast samples of the medium i in time t,It is forecast samples all before knowing timestamp t, ψ is allocated to each time step t reward function, the function
It is to be realized with the fully connected network network layers for being connected to Recognition with Recurrent Neural Network；
(2) perfect to medium i, in scoring process, Recognition with Recurrent Neural Network model estimates a regression vector simultaneouslyShould
The regression function η that estimation procedure is used is：
<mrow>
<mi>&Delta;</mi>
<msubsup>
<mover>
<mi>Y</mi>
<mo>^</mo>
</mover>
<mi>i</mi>
<mrow>
<mo>(</mo>
<mi>k</mi>
<mo>)</mo>
</mrow>
</msubsup>
<mo>=</mo>
<mi>&eta;</mi>
<mrow>
<mo>(</mo>
<msubsup>
<mover>
<mi>Y</mi>
<mo>^</mo>
</mover>
<mi>i</mi>
<mrow>
<mo>(</mo>
<mi>k</mi>
<mo>)</mo>
</mrow>
</msubsup>
<mo>;</mo>
<mi>I</mi>
<mo>,</mo>
<mi>X</mi>
<mo>,</mo>
<msubsup>
<mover>
<mi>Y</mi>
<mo>^</mo>
</mover>
<mrow>
<mi>j</mi>
<mo>\</mo>
<mi>i</mi>
</mrow>
<mrow>
<mo>(</mo>
<mo>&ForAll;</mo>
<mo>)</mo>
</mrow>
</msubsup>
<mo>)</mo>
</mrow>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>2</mn>
<mo>)</mo>
</mrow>
</mrow>
Wherein, regression function η can accumulate all passing scene environments and the dynamic of all interaction mediums, while whole
Estimated on time dimension optimal
7. based on the iterative feedback described in claims 5, it is characterised in that using formula (2) obtain optimum regression replace to
AmountGo gradually to improve the hypothesis set of predictionSpecially：
(1) each iteration,All it is updated toIt is then delivered to reverse optimum control module；
(2) reverse optimum control training in sequence and improve module, they have two kinds loss items, be respectively：Cross entropy is damaged
Lose itemWherein q is obtained by activation primitive, i.e.,Return loss
(3) to a training batch in neutral net, its whole multitask loss item is：
Wherein, N is the number of the batch medium.
8. (three) are merged based on the scene described in claims 1, it is characterised in that including context insertion and interaction feature.
9. based on the context insertion described in claims 8, it is characterised in that Recognition with Recurrent Neural Network is used, at each
Between receive following input X on stamp tt：
<mrow>
<msub>
<mi>X</mi>
<mi>t</mi>
</msub>
<mo>=</mo>
<mo>&lsqb;</mo>
<mi>&gamma;</mi>
<mrow>
<mo>(</mo>
<msub>
<mover>
<mi>&upsi;</mi>
<mo>^</mo>
</mover>
<mrow>
<mi>i</mi>
<mo>,</mo>
<mi>t</mi>
</mrow>
</msub>
<mo>)</mo>
</mrow>
<mo>,</mo>
<mi>p</mi>
<mrow>
<mo>(</mo>
<msub>
<mover>
<mi>y</mi>
<mo>^</mo>
</mover>
<mrow>
<mi>i</mi>
<mo>,</mo>
<mi>t</mi>
</mrow>
</msub>
<mo>;</mo>
<mi>&rho;</mi>
<mo>(</mo>
<mi>I</mi>
<mo>)</mo>
<mo>)</mo>
</mrow>
<mo>,</mo>
<mi>r</mi>
<mrow>
<mo>(</mo>
<msub>
<mover>
<mi>y</mi>
<mo>^</mo>
</mover>
<mrow>
<mi>i</mi>
<mo>,</mo>
<mi>t</mi>
</mrow>
</msub>
<mo>;</mo>
<msub>
<mover>
<mi>y</mi>
<mo>^</mo>
</mover>
<mrow>
<mi>j</mi>
<mo>\</mo>
<mi>i</mi>
</mrow>
</msub>
<mo>,</mo>
<msub>
<mi>h</mi>
<msub>
<mover>
<mi>Y</mi>
<mo>^</mo>
</mover>
<mrow>
<mi>j</mi>
<mo>\</mo>
<mi>i</mi>
</mrow>
</msub>
</msub>
<mo>)</mo>
</mrow>
<mo>&rsqb;</mo>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>4</mn>
<mo>)</mo>
</mrow>
</mrow>
Wherein,It isIn the speed of t, γ is one layer of full articulamentum with nonlinear activation function, and it reflects speed
It is mapped to higher-dimension and characterizes space,It will be located atConvolutional neural networks feature ρ (I) carry out pondization and operate, r passes through
One fused layer, the hiding vector on other medium spatial domains is gathered together, self-embedding vectorIt is used as Recognition with Recurrent Neural Network
Initialize hidden state vector.
10. based on the interaction feature described in claims 8, it is characterised in that use the spatial domain grid using pond layer operation
Operation, makes the interaction between the different samples of different medium is unlikely to disappear, is specially：In moment t medium i, its sample k's
Spatial domain grid cell withCentered on, then to each grid cell g, the feature of its interior Hide All
Average pondization operation will all be carried out.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710487508.2A CN107274029A (en) | 2017-06-23 | 2017-06-23 | A kind of future anticipation method of interaction medium in utilization dynamic scene |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710487508.2A CN107274029A (en) | 2017-06-23 | 2017-06-23 | A kind of future anticipation method of interaction medium in utilization dynamic scene |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107274029A true CN107274029A (en) | 2017-10-20 |
Family
ID=60069234
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710487508.2A Withdrawn CN107274029A (en) | 2017-06-23 | 2017-06-23 | A kind of future anticipation method of interaction medium in utilization dynamic scene |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107274029A (en) |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108777140A (en) * | 2018-04-27 | 2018-11-09 | 南京邮电大学 | Phonetics transfer method based on VAE under a kind of training of non-parallel corpus |
| CN109472412A (en) * | 2018-11-09 | 2019-03-15 | 百度在线网络技术（北京）有限公司 | A kind of prediction technique and device of event |
| CN109710507A (en) * | 2017-10-26 | 2019-05-03 | 北京京东尚科信息技术有限公司 | A kind of method and apparatus of automatic test |
| CN110414989A (en) * | 2019-07-29 | 2019-11-05 | 中国工商银行股份有限公司 | Method for detecting abnormality and device, electronic equipment and computer readable storage medium |
| CN110929644A (en) * | 2019-11-22 | 2020-03-27 | 南京甄视智能科技有限公司 | Heuristic algorithm-based multi-model fusion face recognition method and device, computer system and readable medium |
| CN111563548A (en) * | 2020-04-30 | 2020-08-21 | 鹏城实验室 | Data preprocessing method and system based on reinforcement learning and related equipment |
| CN111599431A (en) * | 2020-03-31 | 2020-08-28 | 太原金域临床检验有限公司 | Report sheet-based data coding model generation method, system and equipment |
| CN111626472A (en) * | 2020-04-17 | 2020-09-04 | 国网甘肃省电力公司信息通信公司 | Scene trend judgment index computing system and method based on deep hybrid cloud model |
| CN111954860A (en) * | 2018-01-21 | 2020-11-17 | 斯塔特斯公司 | System and method for predicting fine-grained antagonistic multi-player movements |
| CN112418421A (en) * | 2020-11-06 | 2021-02-26 | 常州大学 | Roadside end pedestrian trajectory prediction algorithm based on graph attention self-coding model |
| CN113015983A (en) * | 2018-10-24 | 2021-06-22 | Hrl实验室有限责任公司 | Autonomous system including continuous learning world model and related methods |
-
2017
- 2017-06-23 CN CN201710487508.2A patent/CN107274029A/en not_active Withdrawn
Non-Patent Citations (1)
| Title |
|---|
| NAMHOON LEE等: ""DESIRE: Distant Future Prediction in Dynamic Scenes with Interacting Agents"", 《网页在线公开:HTTP://EXPORT.ARXIV.ORG/ABS/1704.04394》 * |
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109710507A (en) * | 2017-10-26 | 2019-05-03 | 北京京东尚科信息技术有限公司 | A kind of method and apparatus of automatic test |
| CN109710507B (en) * | 2017-10-26 | 2022-03-04 | 北京京东尚科信息技术有限公司 | Automatic testing method and device |
| CN111954860A (en) * | 2018-01-21 | 2020-11-17 | 斯塔特斯公司 | System and method for predicting fine-grained antagonistic multi-player movements |
| CN108777140A (en) * | 2018-04-27 | 2018-11-09 | 南京邮电大学 | Phonetics transfer method based on VAE under a kind of training of non-parallel corpus |
| CN113015983A (en) * | 2018-10-24 | 2021-06-22 | Hrl实验室有限责任公司 | Autonomous system including continuous learning world model and related methods |
| CN109472412A (en) * | 2018-11-09 | 2019-03-15 | 百度在线网络技术（北京）有限公司 | A kind of prediction technique and device of event |
| CN110414989A (en) * | 2019-07-29 | 2019-11-05 | 中国工商银行股份有限公司 | Method for detecting abnormality and device, electronic equipment and computer readable storage medium |
| CN110929644A (en) * | 2019-11-22 | 2020-03-27 | 南京甄视智能科技有限公司 | Heuristic algorithm-based multi-model fusion face recognition method and device, computer system and readable medium |
| CN111599431A (en) * | 2020-03-31 | 2020-08-28 | 太原金域临床检验有限公司 | Report sheet-based data coding model generation method, system and equipment |
| CN111626472A (en) * | 2020-04-17 | 2020-09-04 | 国网甘肃省电力公司信息通信公司 | Scene trend judgment index computing system and method based on deep hybrid cloud model |
| CN111626472B (en) * | 2020-04-17 | 2023-10-27 | 国网甘肃省电力公司信息通信公司 | Scene trend judgment index computing system and method based on depth hybrid cloud model |
| CN111563548A (en) * | 2020-04-30 | 2020-08-21 | 鹏城实验室 | Data preprocessing method and system based on reinforcement learning and related equipment |
| CN111563548B (en) * | 2020-04-30 | 2024-02-02 | 鹏城实验室 | Data preprocessing method, system and related equipment based on reinforcement learning |
| CN112418421A (en) * | 2020-11-06 | 2021-02-26 | 常州大学 | Roadside end pedestrian trajectory prediction algorithm based on graph attention self-coding model |
| CN112418421B (en) * | 2020-11-06 | 2024-01-23 | 常州大学 | Road side end pedestrian track prediction algorithm based on graph attention self-coding model |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107274029A (en) | A kind of future anticipation method of interaction medium in utilization dynamic scene | |
| Mousavi et al. | Traffic light control using deep policy‐gradient and value‐function‐based reinforcement learning | |
| CN109635917B (en) | Multi-agent cooperation decision and training method | |
| Bhattacharyya et al. | Multi-agent imitation learning for driving simulation | |
| Christiano et al. | Supervising strong learners by amplifying weak experts | |
| Castillo et al. | Hybrid intelligent systems for time series prediction using neural networks, fuzzy logic, and fractal theory | |
| Papageorgiou et al. | A review of fuzzy cognitive maps research during the last decade | |
| Papageorgiou | Review study on fuzzy cognitive maps and their applications during the last decade | |
| EP3602402B1 (en) | Learning visual concepts using neural networks | |
| Foster et al. | Structure in the space of value functions | |
| Papageorgiou | Review study on fuzzy cognitive maps and their applications during the last decade | |
| CN110147456A (en) | A kind of image classification method, device, readable storage medium storing program for executing and terminal device | |
| CN110991027A (en) | Robot simulation learning method based on virtual scene training | |
| Ghaderi et al. | Behavioral simulation and optimization of generation companies in electricity markets by fuzzy cognitive map | |
| Ulhaq et al. | Efficient diffusion models for vision: A survey | |
| Zheng et al. | Synthetic dynamic PMU data generation: A generative adversarial network approach | |
| Groumpos | Modelling business and management systems using fuzzy cognitive maps: A critical overview | |
| Rabe et al. | Combining a discrete-event simulation model of a logistics network with deep reinforcement learning | |
| Hu et al. | Adaptive exploration strategy with multi-attribute decision-making for reinforcement learning | |
| Andersen et al. | The dreaming variational autoencoder for reinforcement learning environments | |
| Daniel et al. | A perceptually-validated metric for crowd trajectory quality evaluation | |
| CN114116995A (en) | Session recommendation method, system and medium based on enhanced graph neural network | |
| CN117271768A (en) | False news detection method and device based on large language model analysis and guidance | |
| Sayama et al. | Beyond social fragmentation: Coexistence of cultural diversity and structural connectivity is possible with social constituent diversity | |
| Shafik et al. | A reawakening of machine learning application in unmanned aerial vehicle: future research motivation |
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 | ||
| WW01 | Invention patent application withdrawn after publication | ||
| WW01 | Invention patent application withdrawn after publication |
Application publication date: 20171020 |