CN108573309A - A kind of self-adapting regulation method and system of machine learning algorithm - Google Patents
A kind of self-adapting regulation method and system of machine learning algorithm Download PDFInfo
- Publication number
- CN108573309A CN108573309A CN201810008518.8A CN201810008518A CN108573309A CN 108573309 A CN108573309 A CN 108573309A CN 201810008518 A CN201810008518 A CN 201810008518A CN 108573309 A CN108573309 A CN 108573309A
- Authority
- CN
- China
- Prior art keywords
- machine learning
- learning algorithm
- model
- time
- input data
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N5/00—Computing arrangements using knowledge-based models
- G06N5/02—Knowledge representation; Symbolic representation
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The present invention relates to a kind of self-adapting regulation method of machine learning algorithm and systems, including：By analyzing machine learning algorithm, obtaining control, it calculates the controllable parameter of time, and according to the practical calculating time of controllable parameter machine learning algorithm under each concrete numerical value, establishes the quantitative model library that the machine learning algorithm calculates the time；According to the complexity of input data in each time window, coarseness adjustment is carried out to the machine learning algorithm structure, the complexity range of given algorithm model, and quantificational description is carried out to the input data according to the machine learning algorithm, the concrete numerical value of the controllable parameter is determined in the quantitative model library in conjunction with given time restriction, and the concrete numerical value is applied to the machine learning algorithm, to realize the adaptive adjustment of the machine learning algorithm.The invention enables the application on site scenes that machine learning algorithm can adapt to stream data under the conditions of limited computing resource, and rational result of calculation is provided under given time restriction.
Description
Technical field
The present invention relates to machine learning algorithm analyses and algorithm adaptively to adjust field, is related to a kind of machine based on comentropy
The self-adapting regulation method and system of device learning algorithm.
Background technology
Machine learning algorithm is very extensive in the application of data analysis, data mining etc., and achieves very well
Effect.Meanwhile machine learning algorithm is also a kind of very strong algorithm of the specificity for specifically application or specific data, i.e.,
Once completing training using given training data and uniquely determining, model, which would become hard to have, is directed to different data or different application
Generalization ability.Therefore, the adaptive ability for being directed to different data or application to make model have, such as pass through and adjust model
In certain controllable parameters automatically adjust its model structure, be a kind of solution party of highly effective extended model generalization ability
Method.
With the continuous growth of big data streaming scene, the real-time response of online streaming application is often infrangible need
It asks.But the machine learning algorithm in application is in order to pursue better modelling effect offline, often by being continuously increased model
Complexity promote the accuracy of result；And ever-increasing model complexity, inevitably result in machine learning algorithm calculating
The rising of amount；But also model greatly increases the analytic process of data-oriented and the time-consuming of calculating process.Algorithm performs this
One characteristic and the real-time demand of stream data application on site clash.
On the other hand, in order to provide the result of calculation uniquely determined, once for stream data real-time update model without
Method is provided under given time-constrain rational as a result, historical models will be automatically selected, and act on new time window
Interior data.However experiment shows that the accuracy of the model calculation is past when historical models acts on the data at current time
It is past to be greatly reduced；And the naive model that complexity suitably reduces is chosen, it can but be provided more in the case where meeting current time limitation
Rational result of calculation.Although the accuracy of the result of calculation of naive model compared to no time-constrain complex model
Decline, but can ensure that this result of calculation is optimal under currently given time restriction, i.e., the sub-optimal result is much
Than under given time-constrain can not convergent complex model or by historical data train obtained by historical models result more
It is accurate.
In the scene of stream data application on site, traditional machine learning algorithm is the accuracy for pursuing model algorithm,
And the complexity of these model structures is constantly promoted, to introduce volume calculating, time overhead --- this is to online real-time requirement
Bring great impact.And in actual experiment and application, inventor has found answering for appropriate reduction machine learning algorithm model
Miscellaneous degree can reduce the training of model and calculate the time, thereby may be ensured that the calculating time of model disclosure satisfy that application on site
Real-time demand for services.To make machine learning algorithm disclosure satisfy that the real-time demand of online streaming application, can connect
The data processing provided in the time range received and analysis result, when how accurately to weigh the model calculating of machine learning algorithm
Between relationship between model complexity, and how the adjusting model adaptive according to the application scenarios demand of different characteristic
The calculating time, be badly in need of solve the technical issues of.
Invention content
Existing contradiction between time consumption for training for machine learning algorithm and the real-time demand of streaming application, the present invention
Be designed to provide a kind of adaptive adjustment side for the machine learning algorithm disclosure satisfy that the constraint of online streaming application real-time
Method, i.e. machine learning algorithm can be according to its model structures of the adaptive adjustment of time-constrain, in the case where current given time limits
It can obtain the solution of Optimal calculation result.
Specifically, the invention discloses a kind of self-adapting regulation method of machine learning algorithm, including：
Step 1, by analyzing machine learning algorithm, obtaining control, it calculates the controllable parameter of time, and according to
The practical calculating time of controllable parameter machine learning algorithm under each concrete numerical value, when establishing machine learning algorithm calculating
Between quantitative model library；
Step 2, according to the complexity of input data in each time window, coarseness tune is carried out to the machine learning algorithm structure
It is whole, the complexity range of given algorithm model, and quantificational description is carried out to the input data according to the machine learning algorithm, in conjunction with
Given time restriction determines the concrete numerical value of the controllable parameter in the quantitative model library, and the concrete numerical value is applied to this
Machine learning algorithm, to realize the adaptive adjustment of the machine learning algorithm.
The self-adapting regulation method of the machine learning algorithm, the wherein step 1 include：
Can step 101, the implementation procedure according to the machine learning algorithm judge the calculating that quantify the machine learning algorithm
Time；
Step 102, by counting calculating time and result accuracy of the machine learning algorithm under each parameter, be somebody's turn to do
Controllable parameter, and establish the trades space calculated between time and the result precision.
The self-adapting regulation method of the machine learning algorithm, the wherein step 2 include：
Step 201, the comentropy by calculating the input data in time window, the data obtained in current time window are complicated
Degree；
Step 202, the quantificational description for calculating input data in current time window, the controllable ginseng is determined according to the time restriction
Several value set, and by the concrete numerical value of the quantificational description determining controllable parameter in the value set.
The self-adapting regulation method of the machine learning algorithm, the wherein machine learning algorithm are neural network model or decision
Tree-model or Random Forest model or deep learning algorithm model or autoregressive moving-average model.
The self-adapting regulation method of the machine learning algorithm is suitable for those different initializing set parameters to algorithm
The machine learning algorithm that the calculating time has a significant impact.Including neural network model or decision-tree model or Random Forest model or
Deep learning algorithm model or autoregressive moving-average model, but it is not only limited to above-mentioned algorithm.
The self-adapting regulation method of the machine learning algorithm, wherein when the machine learning algorithm is the autoregressive moving average
When model, which includes：Autoregression model exponent number and moving average model exponent number, the quantificational description are the input data
Partial autocorrelation function and auto-correlation function.
The invention also provides a kind of self-adapted adjustment system of machine learning algorithm, including：
Module is established in quantitative model library, for by analyzing machine learning algorithm, obtaining control, it to calculate the time
Controllable parameter, and according to the practical calculating time of controllable parameter machine learning algorithm under each concrete numerical value, establishing should
Machine learning algorithm calculates the quantitative model library of time；
Adaptive adjustment module, for the complexity according to input data in each time window, to the machine learning algorithm knot
Structure carry out coarseness adjustment, the complexity range of given algorithm model, and according to the machine learning algorithm to the input data into
Row quantificational description determines the concrete numerical value of the controllable parameter in conjunction with given time restriction in the quantitative model library, and should
Concrete numerical value is applied to the machine learning algorithm, realizes the adaptive adjustment of the machine learning algorithm.
The self-adapted adjustment system of the machine learning algorithm, wherein the quantitative model library establish module and include：
Judgment module, for the implementation procedure according to the machine learning algorithm, can judgement quantify the machine learning algorithm
The calculating time；
Trades space establishes module, for by counting calculating time and result of the machine learning algorithm under each parameter
Accuracy obtains the controllable parameter, and establishes the trades space calculated between time and the result precision.
The self-adapted adjustment system of the machine learning algorithm, the wherein adaptive adjustment module include：
Complicated dynamic behaviour module obtains the complexity for the comentropy by calculating the input data in time window；
Module is chosen, the quantificational description for calculating input data in current time window, being determined according to the time restriction should
The value set of controllable parameter, and the concrete numerical value of the controllable parameter is chosen by the quantificational description in the value set.
The self-adapted adjustment system of the machine learning algorithm, the wherein machine learning algorithm are neural network model or decision
Tree-model or Random Forest model or deep learning algorithm model or autoregressive moving-average model.
The self-adapted adjustment system of the machine learning algorithm, wherein when the machine learning algorithm is the autoregressive moving average
When model, which includes：Autoregression model exponent number and moving average model exponent number, the quantificational description are the input data
Partial autocorrelation function and auto-correlation function.
The present invention adjusts the controllable parameter in model by the quantization of the calculating time to machine learning algorithm, to change
The structure of model or certain specific algorithms, so as to avoid the great number brought due to introducing the model of overcomplicated
Calculate time overhead.It can adaptively be adjusted according to time restriction it is an advantage of the current invention that providing a kind of machine learning algorithm
The method of integral mould structure so that machine learning algorithm can adapt to the online of stream data under the conditions of limited computing resource
Application scenarios provide rational result of calculation under given time restriction.
Description of the drawings
Fig. 1 is the overall structure figure of self-adapting regulation method of the present invention；
Fig. 2 is the step flow chart of the method for the present invention；
Fig. 3 is the calculating time diagram of different model parameters；
Fig. 4 is the quantitative relationship figure between different model parameters and calculating time；
Fig. 5 is the trades space figure between model calculating time and modelling effect；
Relational graphs of the Fig. 6 between data complexity and model complexity；
Fig. 7 a are the comparison diagram of the actual execution time and complex model and naive model of adaptive model；
Fig. 7 b are the comparison diagram of the practical implementation effect and complex model and naive model of adaptive model.
Specific implementation mode
In order to realize that the adaptive adjustment of machine learning algorithm, the present invention include the following steps：
Step 1：By analyzing machine learning algorithm, the controllable parameter of acquisition its calculating time of control, and according to
The practical calculating time of controllable parameter machine learning algorithm under each concrete numerical value, when establishing machine learning algorithm calculating
Between quantitative model library.The quantitative model library provide different machine learning algorithms the calculating time and these parameter combinations it
Between relationship, to for model adaptation adjust foundation is provided.
Can step 101, the implementation procedure according to the machine learning algorithm judge the calculating that quantify the machine learning algorithm
Time；
Step 102, by counting calculating time and result accuracy of the machine learning algorithm under each parameter, be somebody's turn to do
Controllable parameter, and establish the trades space calculated between time and the result precision.The different possible parameter combinations of input, sentence
Whether these fixed parameters can describe the execution time of algorithm.Calculation amount can be determined in these machine learning algorithm models
The parameter for measuring description, is defined as controllable parameter；
Step 103 establishes these controllable parameters and model and calculates functional relation between the time.
Step 2：According to the complexity of input data in each time window, coarseness tune is carried out to the machine learning algorithm structure
It is whole, the complexity range of given algorithm model, and quantificational description is carried out to the input data according to the machine learning algorithm, in conjunction with
Given time restriction determines the concrete numerical value of the controllable parameter in the quantitative model library, and the concrete numerical value is applied to this
Machine learning algorithm.It specifically includes, the sample based on input calculates its comentropy, with the correlation between data and model complexity
It based on relationship, establishes data and model and controllably saves quantitative relationship between parameter, parameter regulation is carried out as Controlling model
Foundation；
Step 201, the comentropy by calculating the input data in time window calculate the data in the current time window and answer
The complexity of miscellaneous degree, first coarseness cover half type really, the i.e. rational value range of these controllable parameters；
Step 202, the quantificational description for calculating input data in current time window, the controllable ginseng is determined according to the time restriction
Several value set, and by the concrete numerical value of the quantificational description determining controllable parameter in the value set.For different calculations
The description that method reasonably quantifies data, for providing foundation for parameter adjusting method and strength of adjustment.
Wherein in step 101, machine learning algorithm that calculation amount can be quantized refer to those models the calculating time by
Certain parameters influence in model, and these parameters have characteristic below in algorithm practical implementation：Quilt when initialization
Unique assignment, follow-up calculating process, which influences algorithm execution time, to be modified.Meanwhile these parameters are to the shadow of calculating time
Sound can be independent variable using itself by formalized description.Particularly, the determination method of different machines study is different, and is needed
Specific algorithm performs process is made a concrete analysis of, so that it is determined that algorithm can quantized character.
Wherein in step 102, controllable parameter refers to that those can controllable parameter influential on the calculating time of model
Or algorithm etc., include the structural parameters of certain models, such as：The nerve of each layer in the network number of plies of neural network algorithm, network
First interstitial content, the number set in random forests algorithm, the depth etc. each set；Further include certain calculating times different realizations
The selection of algorithm, such as：The selection of different activation primitives in neural network algorithm, the method etc. of parameter Estimation.Controllable ginseng
Number includes, but is not only limited to above-mentioned structural parameters and realizes algorithms selection.
Wherein in step 103, the independent variable of these relationships is made of controllable model parameter, it may be possible to certain given ranges
Continuous natural number, it may be possible to determine practical 0/1 switch function for executing algorithms selection, or determine the end of algorithm trip
Only some threshold value of condition, can take the numerical value of arbitrary zone of reasonableness.In addition, while obtaining formalized description, also need
Provide the value range of these independents variable.These controllable parameters include, but are not only limited to above-mentioned form and value
Range etc..
Wherein in step 201, such conclusion can be obtained based on experiment experience, analysis and prediction to complex data etc.
Process generally requires increasingly complex model.This is because the stronger ability to express that complex model often has, it can be to original
The complex data of beginning is preferably portrayed.Opposite, naive model can be supported to portray simple data.It therefore will be to every
The complexity of data is portrayed in a time window, to as the foundation for adjusting model complexity.And comentropy can be with
The complexity of effective quantitative expression different data, therefore choose the module that comentropy is data complexity.
Wherein in step 202, due to machine learning algorithm need to data carry out different dimensions quantificational description, to for
Adjusting of the model structure of machine learning algorithm etc. provides foundation.Such as：In decision-tree model, the entropy for portraying data is needed to make
The foundation adjusted for model parameter；Needed in principal component analysis calculate data between variance as division principal component according to
According to.These portray according to being determined according to the specific algorithm of different machines learning algorithm, are not only limited to above-mentioned refer to
Measure.
To allow features described above of the invention and effect that can illustrate more clearly understandable, special embodiment below, and coordinate
Bright book attached drawing is described in detail below.
Referring to FIG. 1, present invention is generally directed to stream data application on site, a kind of machine learning based on comentropy is proposed
Algorithm self-adapting regulation method, overall structure include mainly two large divisions referring to Fig.1：
First part is substantially carried out carries out off-line analysis to different machine learning algorithms.Main step includes：101)
According to the calculating time of each machine learning algorithm model, judge whether the calculation amount of each machine learning algorithm model can be measured
Change, 102) identification is on the calculating time influential controllable parameter, 103) the quantitative description calculating time and the controllable parameter it
Between relationship.
Second part acts on the real-time adjusting of application on site.Main step includes：201) quantify current time window number
According to complexity, 202) calculate current time window data certain Expressive Features, provide foundation for parameter regulation.
In order to which the purpose, technical scheme and advantage in enabling the present invention to implement are more clear, the present invention will be in the time
Common autoregressive moving-average model in sequence analysis, in conjunction with the attached drawing in the embodiment of the present invention, in the embodiment of the present invention
Technical solution be more clear, be fully described by.Obviously, described embodiment is only used as practical function effect of the present invention
One kind, instead of all the embodiments.
Present invention could apply to different machine learning algorithm models, including random forest, decision-tree model, nerve net
Network, autoregressive moving-average model etc..When the machine learning algorithm is the autoregressive moving-average model, the controllable parameter packet
Include autoregression model exponent number and moving average model exponent number, the quantificational description include the input data partial autocorrelation function and from
Correlation function；When the machine learning algorithm is the neural network model, which includes the activation of neural network algorithm
The neuron node number of each layer in function, the network number of plies and neural network, the quantificational description include the complexity of the input data
Degree, using the entropy of data in the time window as measurement；When the machine learning algorithm is the Random Forest model, the controllable parameter
Including the number set in random forests algorithm and the depth each set, which includes the complexity of the input data, with
The entropy of data is as measurement in the time window；When the machine learning algorithm is the decision-tree model, which includes tree
Depth, which includes the complexity of the input data, using the entropy of data in the time window as measurement.
Autoregressive moving-average model (abbreviation arma modeling) is the important method of search time sequence, including autoregression mould
Type (abbreviation AR models) and moving average model (abbreviation MA models) two parts.Wherein, autoregression model be with time series from
Body does the process of regression variable, i.e., certain moment is random after being described using the linear combination of the stochastic variable at early period at several moment
The linear regression model (LRM) of variable；Moving average model is the value and past white noise variable that linear process is expressed as to current time
Weighted linear combination model.It can be indicated by following formula：
Wherein, XiIndicate value of the time series at the i moment,And θiIndicate parameter to be asked, εiBe meet with 0 for mean value, σ2
For the stochastic variable of the standardized normal distribution of standard deviation.
Such as it is pre- in the experimentation of financial Time Series of Random Macro-price, needing to carry out the data in each time window
Processing --- first-order difference, so that it is guaranteed that data meet stationarity.Arma modeling is reapplied on this basis, sequence is carried out pre-
It surveys and analyzes.I.e.：
Xt-1=Yt-Yt-1
Wherein, YiIndicate original time series in the value at i moment, XiFor differentiated time series.
The flow of method provided by the invention is as shown in Figure 2.Wherein, first part divides to the offline of machine learning algorithm
During analysis, by the analysis to arma modeling, the model can be controlled by, which finding, calculates the controllable parameter of time, and establishes these
Relationship between controllable parameter and calculating time.
Step 101, according to the related statistical information of code implementation, whether analysis arma modeling algorithm has and can quantify
Property.Experiment finds that different model parameters determines different calculating cycle-indexes, calculates time complexity etc., i.e., different ginsengs
There is significant difference in the calculating time of exponential model.If wherein algorithm does not have quantifiable energy, illustrate that algorithm can not be by regulating and controlling to join
Number carries out the change of complexity.Then the present invention is not suitable for the algorithm that this one kind can not quantify.
Step 102, judged automatically by certain program analysis means, the controllable parameter in recognizer；Or directly exist
In the initializing set stage during algorithm realization, is judged using experience, select different controllable parameters.Those are found in the mould
To calculating the controllable parameter that has a significant impact of time in type.By configuring the combination of different model parameters, hold code is practical
It is counted during row to executing time etc..Judge that the controllable parameter in the model includes autoregression model exponent number, and sliding
Dynamic averaging model exponent number.The variation of controllable parameter and the relationship of algorithm execution time are as shown in Figure 3.
Step 103, structure mathematical model quantifies influence of these parameters to the calculating time, and relationship is as shown in Figure 4.Every
In secondary experiment, also corresponding record cast calculate time and result accuracy as a result, making result precision and calculating the time
Between trades space, as shown in Figure 5.Simultaneously in the assessment for the model, autoregression model exponent number and sliding average mould
Type exponent number can portray the influence for calculating the time by a simple linear function.The trades space provides different models
The relationship between time and accuracy in computation is calculated, Appreciation gist is provided for the simplification of algorithm model.
Second part is mainly used in online stream data processing.For the data analysis in each time window of input
Its complexity carries out the structure of model with the principle of simple data application naive model and complex data application complex model thick
The adjustment of granularity, i.e. data are more complicated, and the parameter setting of machine learning algorithm model is more complicated accordingly.According to the given time
Limitation adjusts model controllable parameter to rational value according to the property to data partial autocorrelation function and auto-correlation function.
Step 201, due to the difference of naive model and the descriptive power of complex model, the complexity of data is different, institute
The model tormulation ability needed also has corresponding differentiation.Find in an experiment, naive model to the descriptive power of simple data with
And support applications demand enough, and complex data then needs more complicated model to carry out better data description, relationship
As shown in Figure 6.Therefore the complexity of metric data has very strong directive significance to the selection of the complexity of model.With
Comentropy is as the module to data complexity, the relationship between foundation data complexity and model complexity, this
Relationship can be used for the adjustment of guidance model parameter.The comentropy of input data in each time window is calculated, when
When the value is more than a certain threshold value (real value is 3.0 in experiment), then it is determined as that this group of input data has high complexity；Instead
It, then be determined as that data have lower complexity.
Wherein HwindowIndicate the entropy of the data in current time window, piIndicate the number for the time series that value is i in sequence
The probability that value occurs in the time window, windowsize indicate the size of time window, that is, the number of the data point occurred, E (p)
Indicate the expectation of p.
Step 202, the present embodiment is chosen arma modeling and is illustrated, due to original arma modeling be by data from
Related and portraying for partial autocorrelation function carrys out selection parameter.Therefore, the auto-correlation function and partial autocorrelation function for choosing data are made
For the foundation adjusted for arma modeling controllable parameter.In this step, need to calculate the input data in current time window
Partial autocorrelation function (PACF) and auto-correlation function (ACF) are limited the reasonable of the parameter for determining model by the time of acquisition result
Value determines specific Parameters of Autoregressive Models according to partial autocorrelation function, determines that rational sliding is flat according to auto-correlation function
Equal model parameter.
Wherein, XiIndicate that value of the time series at the i moment, μ indicate that the mean value of data in current time window, τ indicate and work as
The time interval at preceding moment, E (α) indicate the expectation of sequence α.
This method finally can be so that machine learning algorithm in given time restriction (such as Fig. 7 a) provides one and is meeting
Optimal calculation under current time limitation is as a result, as shown in Figure 7b.This method can control machine learning algorithm when given
Between limit the lower training and calculating for completing model, acquired results are better than using calculating the fixed single simple of time meet demand
Model effectively controls the time calculated while lift scheme shows.
Based on above-mentioned experimental program, the self-adapting regulation method of the machine learning mentioned by present example can be according to working as
The complexity information of the data of real-time change in preceding time window, carries out the adjusting of automatic model parameter, in terms of control algolithm
Evaluation time can adapt to the time restriction that the result of current application obtains, and meet the practical need of the application on site for stream data
It asks.And this method ensure that under current time limitation, the model calculation is optimal, and improves the service quality of application.
Based on above-described embodiment, the feature for the foundation time series which can be adaptive, adaptive adjusting is certainly
Regression model exponent number and moving average model exponent number, the calculating time to adjust model are strictly controlled at less than time restriction
Under, and provide the optimal solution under current limit.
It is below system embodiment corresponding with above method embodiment, this implementation system can be mutual with the above embodiment
Cooperation is implemented.The above-mentioned relevant technical details mentioned in mode of applying are still effective in this implementation system, in order to reduce repetition, this
In repeat no more.Correspondingly, the relevant technical details mentioned in this implementation system are also applicable in the above embodiment.
The invention also provides a kind of self-adapted adjustment system of machine learning algorithm, including：
Module is established in quantitative model library, for by analyzing machine learning algorithm, obtaining control, it to calculate the time
Controllable parameter, and according to the practical calculating time of controllable parameter machine learning algorithm under each concrete numerical value, establishing should
Machine learning algorithm calculates the quantitative model library of time；
Adaptive adjustment module, for the complexity according to input data in each time window, to the machine learning algorithm knot
Structure carry out coarseness adjustment, the complexity range of given algorithm model, and according to the machine learning algorithm to the input data into
Row quantificational description determines the concrete numerical value of the controllable parameter in conjunction with given time restriction in the quantitative model library, and should
Concrete numerical value is applied to the machine learning algorithm, realizes the adaptive adjustment of the machine learning algorithm.
The self-adapted adjustment system of the machine learning algorithm, wherein the quantitative model library establish module and include：
Judgment module, for the implementation procedure according to the machine learning algorithm, can judgement quantify the machine learning algorithm
The calculating time；
Trades space establishes module, for by counting calculating time and result of the machine learning algorithm under each parameter
Accuracy obtains the controllable parameter, and establishes the trades space calculated between time and the result precision.
The self-adapted adjustment system of the machine learning algorithm, the wherein adaptive adjustment module include：
Complicated dynamic behaviour module obtains the complexity for the comentropy by calculating the input data in time window；
Module is chosen, the quantificational description for calculating input data in current time window, being determined according to the time restriction should
The value set of controllable parameter, and the concrete numerical value of the controllable parameter is chosen by the quantificational description in the value set.
The self-adapted adjustment system of the machine learning algorithm, the wherein machine learning algorithm are neural network model or decision
Tree-model or Random Forest model or deep learning algorithm model or autoregressive moving-average model.
The self-adapted adjustment system of the machine learning algorithm, wherein when the machine learning algorithm is the autoregressive moving average
When model, which includes autoregression model exponent number and moving average model exponent number, which includes the input number
According to partial autocorrelation function and auto-correlation function；When the machine learning algorithm is the neural network model, the controllable parameter packet
Include the neuron node number of each layer in activation primitive, the network number of plies and the neural network of neural network algorithm, the quantificational description
Complexity including the input data；When the machine learning algorithm is the Random Forest model, which includes random
The number set in forest algorithm and the depth each set, the quantificational description include the complexity of the input data；When the engineering
When habit algorithm is the decision-tree model, which includes the depth of tree, which includes the complexity of the input data
Degree.
Although the present invention is disclosed with above-described embodiment, specific examples are only used to explain the present invention, is not used to limit
The present invention, any those skilled in the art of the present technique, in change that some without departing from the spirit and scope of the invention, can be made and complete
It is kind, therefore the scope of the present invention is subject to claims.
Claims (10)
1. a kind of self-adapting regulation method of machine learning algorithm, which is characterized in that including：
Step 1, by analyzing machine learning algorithm, obtaining control, it calculates the controllable parameter of time, and can according to this
The practical calculating time for controlling parameter machine learning algorithm under each concrete numerical value establishes the machine learning algorithm and calculates the time
Quantitative model library；
Step 2, according to the complexity of input data in each time window, coarseness tune is carried out to the structure of the machine learning algorithm
It is whole, the complexity range of given algorithm model, and quantificational description is carried out to the input data according to the machine learning algorithm, in conjunction with
Given time restriction determines the concrete numerical value of the controllable parameter in the quantitative model library, and the concrete numerical value is applied to this
Machine learning algorithm, to realize the adaptive adjustment of the machine learning algorithm.
2. the self-adapting regulation method of machine learning algorithm as described in claim 1, which is characterized in that the step 1 includes：
Step 101, the implementation procedure according to the machine learning algorithm, when can judgement quantify the calculating of the machine learning algorithm
Between；
Step 102, by counting calculating time and result accuracy of the machine learning algorithm under each parameter, it is controllable to obtain this
Parameter, and establish the trades space calculated between time and the result precision.
3. the self-adapting regulation method of machine learning algorithm as claimed in claim 1 or 2, which is characterized in that the step 2 is wrapped
It includes：
Step 201, the comentropy by calculating the input data in time window, obtain the complexity of data in current time window；
Step 202, the quantificational description for calculating input data in current time window, the controllable parameter is determined according to the time restriction
Value set, and by the concrete numerical value of the quantificational description determining controllable parameter in the value set.
4. the self-adapting regulation method of machine learning algorithm as claimed in claim 3, which is characterized in that the machine learning algorithm
Including：Neural network model or decision-tree model or Random Forest model or autoregressive moving-average model.
5. the self-adapting regulation method of machine learning algorithm as claimed in claim 4, which is characterized in that
When the machine learning algorithm is the autoregressive moving-average model, which includes autoregression model exponent number and cunning
Dynamic averaging model exponent number, which includes the partial autocorrelation function and auto-correlation function of the input data；
When the machine learning algorithm be the neural network model when, the controllable parameter include neural network algorithm activation primitive,
The neuron node number of each layer in the network number of plies and neural network, the quantificational description include the complexity of the input data；
When the machine learning algorithm is the Random Forest model, the controllable parameter include the number set in random forests algorithm and
The depth each set, the quantificational description include the complexity of the input data；
When the machine learning algorithm is the decision-tree model, which includes the depth of tree, which includes should
The complexity of input data.
6. a kind of self-adapted adjustment system of machine learning algorithm, which is characterized in that including：
Module is established in quantitative model library, for by analyzing machine learning algorithm, obtain control its calculate the time can
Parameter is controlled, and according to the practical calculating time of controllable parameter machine learning algorithm under each concrete numerical value, establishes the machine
Learning algorithm calculates the quantitative model library of time；
Adaptive adjustment module, for the complexity according to input data in each time window, to the structure of the machine learning algorithm
Coarseness adjustment, the complexity range of given algorithm model are carried out, and the input data is carried out according to the machine learning algorithm
Quantificational description determines the concrete numerical value of the controllable parameter in conjunction with given time restriction in the quantitative model library, and by the tool
Body numerical applications realize the adaptive adjustment of the machine learning algorithm to the machine learning algorithm.
7. the self-adapted adjustment system of machine learning algorithm as claimed in claim 6, which is characterized in that the quantitative model library is built
Formwork erection block includes：
Judgment module judges the calculating time energy of the machine learning algorithm for the implementation procedure according to the machine learning algorithm
No quantization；
Trades space establishes module, for accurate by counting calculating time and result of the machine learning algorithm under each parameter
Property, the controllable parameter is obtained, and establish the power of the model that current controllable parameter determines calculated between time and result precision
Weighing apparatus space.
8. the self-adapted adjustment system of machine learning algorithm as claimed in claims 6 or 7, which is characterized in that the adaptive tune
Mould preparation block includes：
Complicated dynamic behaviour module obtains number in current time window for the comentropy by calculating the input data in time window
According to complexity；
Module is chosen, the quantificational description for calculating input data in current time window determines that this is controllable according to the time restriction
The value set of parameter, and the concrete numerical value of the controllable parameter is chosen by the quantificational description in the value set.
9. the self-adapted adjustment system of machine learning algorithm as claimed in claim 8, which is characterized in that the machine learning algorithm
Including：Neural network model or decision-tree model or Random Forest model or deep learning algorithm model or autoregression sliding
Averaging model.
10. the self-adapted adjustment system of machine learning algorithm as claimed in claim 9, which is characterized in that
When the machine learning algorithm is the autoregressive moving-average model, which includes autoregression model exponent number and cunning
Dynamic averaging model exponent number, which includes the partial autocorrelation function and auto-correlation function of the input data；
When the machine learning algorithm be the neural network model when, the controllable parameter include neural network algorithm activation primitive,
The neuron node number of each layer in the network number of plies and neural network, the quantificational description include the complexity of the input data；
When the machine learning algorithm is the Random Forest model, the controllable parameter include the number set in random forests algorithm and
The depth each set, the quantificational description include the complexity of the input data；
When the machine learning algorithm is the decision-tree model, which includes the depth of tree, which includes should
The complexity of input data.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810008518.8A CN108573309A (en) | 2018-01-04 | 2018-01-04 | A kind of self-adapting regulation method and system of machine learning algorithm |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810008518.8A CN108573309A (en) | 2018-01-04 | 2018-01-04 | A kind of self-adapting regulation method and system of machine learning algorithm |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108573309A true CN108573309A (en) | 2018-09-25 |
Family
ID=63576436
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810008518.8A Pending CN108573309A (en) | 2018-01-04 | 2018-01-04 | A kind of self-adapting regulation method and system of machine learning algorithm |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108573309A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109871868A (en) * | 2019-01-11 | 2019-06-11 | 中国船舶重工集团公司第七二四研究所 | A kind of space multiple density point Data Association under the conditions of Limited computational resources |
| CN110210558A (en) * | 2019-05-31 | 2019-09-06 | 北京市商汤科技开发有限公司 | Assess the method and device of neural network performance |
| CN111326254A (en) * | 2018-12-17 | 2020-06-23 | 第四范式（北京）技术有限公司 | Method, apparatus, device and medium for index intervention |
-
2018
- 2018-01-04 CN CN201810008518.8A patent/CN108573309A/en active Pending
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111326254A (en) * | 2018-12-17 | 2020-06-23 | 第四范式（北京）技术有限公司 | Method, apparatus, device and medium for index intervention |
| CN111326254B (en) * | 2018-12-17 | 2024-06-07 | 第四范式（北京）技术有限公司 | Method, device, equipment and medium for index intervention |
| CN109871868A (en) * | 2019-01-11 | 2019-06-11 | 中国船舶重工集团公司第七二四研究所 | A kind of space multiple density point Data Association under the conditions of Limited computational resources |
| CN110210558A (en) * | 2019-05-31 | 2019-09-06 | 北京市商汤科技开发有限公司 | Assess the method and device of neural network performance |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109002686B (en) | Multi-grade chemical process soft measurement modeling method capable of automatically generating samples | |
| Attarzadeh et al. | Proposing a new software cost estimation model based on artificial neural networks | |
| CN108304679A (en) | A kind of adaptive reliability analysis method | |
| CN108573309A (en) | A kind of self-adapting regulation method and system of machine learning algorithm | |
| CN110378052A (en) | It is looked to the future the equipment method for predicting residual useful life of operating condition based on Recognition with Recurrent Neural Network | |
| DE102016011532A1 (en) | Machine learning device and machine learning method for optimizing the frequency of tool correction of a machine tool and machine tool with the machine learning device | |
| CN108764568B (en) | Data prediction model tuning method and device based on LSTM network | |
| CN108764520A (en) | A kind of water quality parameter prediction technique based on multilayer circulation neural network and D-S evidence theory | |
| CN108763398A (en) | Database configuration parameters processing method, device, computer equipment and storage medium | |
| Truong et al. | Design of an advanced time delay measurement and a smart adaptive unequal interval grey predictor for real-time nonlinear control systems | |
| CN116582596B (en) | Data processing method and device and PLC data transmission system based on Internet of things | |
| CN110308658A (en) | A kind of pid parameter setting method, device, system and readable storage medium storing program for executing | |
| CN111626844A (en) | Enterprise credit assessment method and device based on big data analysis | |
| KR20210027590A (en) | multilayer artificial neural network performance evaluation apparatus including multiple sub models and method therefor | |
| CN110162403A (en) | A kind of hardware resource distribution method and system based on artificial neural network | |
| CN117473833A (en) | Method and device for predicting rolling force of rod and wire based on finite element and neural network | |
| CN108537322A (en) | Neural network interlayer activation value quantization method and device | |
| CN109451522A (en) | A kind of method for predicting and device towards Bluetooth gateway | |
| CN116303626B (en) | Well cementation pump pressure prediction method based on feature optimization and online learning | |
| CN111539508A (en) | Generator excitation system parameter identification algorithm based on improved wolf algorithm | |
| CN114897274A (en) | Method and system for improving time sequence prediction effect | |
| Kramer et al. | An adaptive penalty function with meta-modeling for constrained problems | |
| CN114330119A (en) | Deep learning-based pumped storage unit adjusting system identification method | |
| CN110110853B (en) | Deep neural network compression method and device and computer readable medium | |
| Jakimowicz et al. | The mechanism of transformation of global business cycles into dynamics of regional real estate markets |
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 | ||
| EE01 | Entry into force of recordation of patent licensing contract | ||
| EE01 | Entry into force of recordation of patent licensing contract |
Application publication date: 20180925 Assignee: Zhongke Yuanshu (Beijing) Technology Co., Ltd. Assignor: Institute of Computing Technology, Chinese Academy of Sciences Contract record no.: X2019990000091 Denomination of invention: Adaptive adjusting method and system of machine learning algorithm License type: Common License Record date: 20190919 |
|
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20180925 |