CN107909154A - A kind of air control Model Parameter Optimization method based on web search - Google Patents
A kind of air control Model Parameter Optimization method based on web search Download PDFInfo
- Publication number
- CN107909154A CN107909154A CN201711311720.XA CN201711311720A CN107909154A CN 107909154 A CN107909154 A CN 107909154A CN 201711311720 A CN201711311720 A CN 201711311720A CN 107909154 A CN107909154 A CN 107909154A
- Authority
- CN
- China
- Prior art keywords
- parameter
- mrow
- search
- auc
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- 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
- G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
- G06Q10/063—Operations research, analysis or management
- G06Q10/0635—Risk analysis of enterprise or organisation activities
Abstract
The invention discloses a kind of air control Model Parameter Optimization method based on web search, this method determines air control model parameter to be searched first;Setup parameter stops search rule, and grid search is carried out respectively to parameters, when the valuation to be evaluated of the parameter meets stopping rule, then stops search, otherwise continues search for remaining valuation to be evaluated;Often carry out grid search and produce one group of parameter combination, cross validation is carried out to the parameter combination, obtain assessment index, the parameter in the assessment index corresponding to maximum is optimal model parameters;This method is capable of the time cost of effectively save parameter search, improves resource utilization;The robustness of lift scheme at the same time.
Description
Technical field
The present invention relates to machine learning field, and in particular to a kind of air control Model Parameter Optimization side based on web search
Method.
Background technology
Currently, machine learning model is quite varied in the use of each business scenario.Using sample data set, by adjusting
The hyper parameter combination of model, we can obtain large number of model, then select optimal models by the effect of each model
Disposed as final mask.When model hyper parameter quantity is more, can the huge parameter combination of quantity of formation, for how adjusting
Parameter, the way of industry mainstream is to be based on trellis search method at present.Specific practice is many of each parameter of setting model
A candidate value, traversal combination is carried out by each candidate value, so that more complete parameter combination is obtained, under a parameter combination
Model assessed.
Traditional trellis search technique, needs the value range of the parameter and parameters clearly adjusted first, secondly selects
The candidate variables of fixed each parameter, then each parameter candidate value is traveled through combine.I.e. by parameter to be adjusted in certain sky
Between in the range of be divided into numerous grids, find optimized parameter by traveling through point all in grid.Trellis search method is being sought
Excellent section is sufficiently large and can obtain globally optimal solution in the case that step pitch is sufficiently small;During grid search is carried out, due to
Category of model effect constructed by parameter combination in most of grid is all poor, and the parameter combination only in partial section can make
Category of model is optimal effect.Therefore, the way of all or most of search parameter grid of tradition traversal will greatly increase
Time cost and resources costs.
In terms of relatively parameter combination, using cross-validation method, the effect of the model of each parameter combination structure is assessed, is chosen
Make the optimal parameter combination of modelling effect as final optimized parameter.Common assessment models effectiveness indicator has AUC (under curve
Area value, Area Under Curve), KS values (KSValue), mean square error etc..AUC is defined as ROC curve (Receiver
Operating Characteristic, receiver operating characteristic curve) under area, for evaluating the excellent of two disaggregated model
It is bad.Its value range is between 0.5 and 1, verification the collection AUC, Ke Yiping on model by comparing different parameters combination structure
Estimate the good and bad situation of different models.Parameter is only calculated with cross validation and corresponds to effectiveness indicator of the model on verification collection (such as
AUC value), without taking into full account model in training set and the effect situation of change on verification collection, the optimized parameter obtained from
The robustness of the corresponding model of combination is not good enough.
The content of the invention
It is an object of the invention to:A kind of air control Model Parameter Optimization method based on web search is provided, solves and adopts
Time and resources costs are high, technical problem of robustness difference when carrying out optimal models selection with grid search and cross validation.
The technical solution adopted by the present invention is as follows:
A kind of air control Model Parameter Optimization method based on web search, comprises the following steps:
Step 1:Determine air control model parameter to be searched;
Step 2:Setup parameter stops search rule, and grid search is carried out respectively to parameters, to be evaluated when the parameter
Valuation meets stopping rule, then stops search, and otherwise continues search for remaining valuation to be evaluated;
Step 3:Often carry out a grid search and produce one group of parameter combination, cross validation is carried out to the parameter combination,
Obtain assessment index, the parameter in the assessment index corresponding to maximum is optimal model parameters.
Further, the rule that stops search is as follows:The step-size in search of setup parameter, stopping wheel number and stopping threshold first
Value;When the valuation to be evaluated of the parameter stops the change of assessment index in wheel number and be respectively less than the outage threshold described, then stop
The only search of the parameter, otherwise continues.
Further, the formula of the acquisition assessment index is as follows:
Wherein, Index is assessment index, AucvalidFor the Auc, Auc of the obtained verification set value of cross validationtrainFor
The Auc values of the obtained training set of cross validation.
In conclusion by adopting the above-described technical solution, the beneficial effects of the invention are as follows:
The present invention selectively attempts candidate parameter combination, rather than all parameter groups of simple traversal in search procedure
Close;By setting the rule that stops search, for each parameter, if one fixed step size rear mold type effect of search is not obviously improved,
It will stop the search of the parameter candidate value, to obtain the candidate value of history optimal effectiveness as the end value of the parameter;So energy
The time cost of enough effectively save parameter searches, improves resource utilization.
The present invention constructs a kind of new assessment index that can weigh the effect situation of change that training set collects with verification, profit
With the assessment index it is contemplated that verification collection and training set Evaluated effect gap problems of too, the robustness of lift scheme.
Embodiment
All features disclosed in this specification, or disclosed all methods or during the step of, except mutually exclusive
Feature and/or step beyond, can combine in any way.
Elaborate below to the present invention.
A kind of air control Model Parameter Optimization method based on web search, comprises the following steps:
Step 1:Determine air control model parameter to be searched;
Step 2:Setup parameter stops search rule, and parameters are carried out with grid search, the rule that stop search respectively
It is then as follows:The step-size in search of setup parameter, stopping wheel number and outage threshold first;When the valuation to be evaluated of the parameter is in the stopping
The change of assessment index in wheel number is respectively less than the outage threshold, then stops the search of the parameter, otherwise continue;
Step 3:Often carry out a grid search and produce one group of parameter combination, cross validation is carried out to the parameter combination,
Obtain assessment index, the parameter in the assessment index corresponding to maximum is optimal model parameters.
The formula of the acquisition assessment index is as follows:
Wherein, Index is assessment index, AucvalidFor the Auc, Auc of the obtained verification set value of cross validationtrainFor
The Auc values of the obtained training set of cross validation.
Specific embodiment
The present embodiment is by taking widely used Xgboost models as an example, since Xgboost models have more ginseng to be searched
Number, the present embodiment only select which part parameter to illustrate.
Step 1:Determine air control model parameter to be searched;
Xgboost models parameter to be searched is as follows:
1) eta, iteration step length;Its value range is [0,1];
2) min_child_weight, child node smallest sample weight and;Its value range is [0, ∞];
3) max_depth, the depth capacity of tree;Its value range is [1, ∞];
4) subsample, the subsample for training pattern account for the ratio of whole sample set;Its value range for (0,
1]。
Step 2:Setup parameter stops search rule, and parameters are carried out with grid search, the rule that stop search respectively
It is then as follows:The step-size in search of setup parameter, stopping wheel number and outage threshold first;When the valuation to be evaluated of the parameter is in the stopping
The change of assessment index in wheel number is respectively less than the outage threshold, then stops the search of the parameter, otherwise continue;
It is α to make step-size in search, and it is m, outage threshold ε to stop wheel number, then sets as follows:
1)eta:αeta=0.02, m=20, ε=0.001, search range are [0,1];
2)min_child_weight:αmin_child_weight=1, m=20, ε=0.001, search range are [0,50];
3)max_depth:αmax_depth=1, m=20, ε=0.001, search range are [1,50];
4)subsample:αsubsample=0.02, m=20, ε=0.001, search range for (0,1].
Step 3:Often carry out a grid search and produce one group of parameter combination, cross validation is carried out to the parameter combination,
Obtain assessment index, the parameter in the assessment index corresponding to maximum is optimal model parameters.
The formula of the acquisition assessment index is as follows:
Wherein, Index is assessment index, AucvalidFor the Auc, Auc of the obtained verification set value of cross validationtrainFor
The Auc values of the obtained training set of cross validation;
Setting parameter is combined as:P(eta,min_child_weight,max_depth,subsample)
After carrying out grid search, following parameter combination to be assessed is obtained:
P1(0,0,1,0), P2(0.02,0,1,0), P3(0.04,0,1,0), P4(0.06,0,1,0) ..., PN(1,50,50,
1);
Wherein, N is the number of parameter combination;
Above-mentioned parameter is combined and carries out cross validation, obtains the assessment index of parameters combination, maximum assessment index
Corresponding parameter is optimal model parameters.
Claims (3)
- A kind of 1. air control Model Parameter Optimization method based on web search, it is characterised in that:Comprise the following steps:Step 1:Determine air control model parameter to be searched;Step 2:Setup parameter stops search rule, grid search is carried out respectively to parameters, when the valuation to be evaluated of the parameter Meet stopping rule, then stop search, otherwise continue search for remaining valuation to be evaluated;Step 3:Often carry out a grid search and produce one group of parameter combination, cross validation is carried out to the parameter combination, is obtained Assessment index, the parameter in the assessment index corresponding to maximum is optimal model parameters.
- A kind of 2. air control Model Parameter Optimization method based on web search according to claim 1, it is characterised in that:The rule that stops search is as follows:The step-size in search of setup parameter, stopping wheel number and outage threshold first;When the parameter Valuation to be evaluated is respectively less than the outage threshold in the change for stopping the assessment index in wheel number, then stops searching for the parameter Rope, otherwise continues.
- A kind of 3. air control Model Parameter Optimization method based on web search according to claim 1, it is characterised in that:Institute The formula for stating acquisition assessment index is as follows:<mrow> <mi>I</mi> <mi>n</mi> <mi>d</mi> <mi>e</mi> <mi>x</mi> <mo>=</mo> <mfrac> <mrow> <msub> <mi>Auc</mi> <mrow> <mi>v</mi> <mi>a</mi> <mi>l</mi> <mi>i</mi> <mi>d</mi> </mrow> </msub> </mrow> <mrow> <mn>1</mn> <mo>+</mo> <msup> <mrow> <mo>(</mo> <mfrac> <mrow> <msub> <mi>Auc</mi> <mrow> <mi>v</mi> <mi>a</mi> <mi>l</mi> <mi>i</mi> <mi>d</mi> </mrow> </msub> <mo>-</mo> <msub> <mi>Auc</mi> <mrow> <mi>t</mi> <mi>r</mi> <mi>a</mi> <mi>i</mi> <mi>n</mi> </mrow> </msub> </mrow> <mrow> <msub> <mi>Auc</mi> <mrow> <mi>t</mi> <mi>r</mi> <mi>a</mi> <mi>i</mi> <mi>n</mi> </mrow> </msub> </mrow> </mfrac> <mo>)</mo> </mrow> <mn>4</mn> </msup> </mrow> </mfrac> </mrow>Wherein, Index is assessment index, AucvalidFor the Auc, Auc of the obtained verification set value of cross validationtrainTested to intersect Demonstrate,prove the Auc values of obtained training set.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711311720.XA CN107909154A (en) | 2017-12-11 | 2017-12-11 | A kind of air control Model Parameter Optimization method based on web search |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711311720.XA CN107909154A (en) | 2017-12-11 | 2017-12-11 | A kind of air control Model Parameter Optimization method based on web search |
Publications (1)
Publication Number | Publication Date |
---|---|
CN107909154A true CN107909154A (en) | 2018-04-13 |
Family
ID=61865164
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201711311720.XA Pending CN107909154A (en) | 2017-12-11 | 2017-12-11 | A kind of air control Model Parameter Optimization method based on web search |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107909154A (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108846521A (en) * | 2018-06-22 | 2018-11-20 | 西安电子科技大学 | Shield-tunneling construction unfavorable geology type prediction method based on Xgboost |
CN108921369A (en) * | 2018-05-08 | 2018-11-30 | 阿里巴巴集团控股有限公司 | Conflict rule generation method and device and electronic equipment |
CN109508455A (en) * | 2018-10-18 | 2019-03-22 | 山西大学 | A kind of GloVe hyper parameter tuning method |
WO2020029851A1 (en) * | 2018-08-08 | 2020-02-13 | 浙江大学 | Workflow-based vibration spectrum analysis model optimization method |
CN111340147A (en) * | 2020-05-22 | 2020-06-26 | 四川新网银行股份有限公司 | Decision behavior generation method and system based on decision tree |
-
2017
- 2017-12-11 CN CN201711311720.XA patent/CN107909154A/en active Pending
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108921369A (en) * | 2018-05-08 | 2018-11-30 | 阿里巴巴集团控股有限公司 | Conflict rule generation method and device and electronic equipment |
CN108846521A (en) * | 2018-06-22 | 2018-11-20 | 西安电子科技大学 | Shield-tunneling construction unfavorable geology type prediction method based on Xgboost |
WO2020029851A1 (en) * | 2018-08-08 | 2020-02-13 | 浙江大学 | Workflow-based vibration spectrum analysis model optimization method |
CN109508455A (en) * | 2018-10-18 | 2019-03-22 | 山西大学 | A kind of GloVe hyper parameter tuning method |
CN109508455B (en) * | 2018-10-18 | 2021-11-19 | 山西大学 | GloVe super-parameter tuning method |
CN111340147A (en) * | 2020-05-22 | 2020-06-26 | 四川新网银行股份有限公司 | Decision behavior generation method and system based on decision tree |
CN111340147B (en) * | 2020-05-22 | 2021-12-07 | 四川新网银行股份有限公司 | Decision behavior generation method and system based on decision tree |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107909154A (en) | A kind of air control Model Parameter Optimization method based on web search | |
CN102693451B (en) | Method for predicting ammonia process flue gas desulphurization efficiency based on multiple parameters | |
CN103440528B (en) | Thermal power unit operation optimization method and device based on power consumption analysis | |
CN107169565A (en) | Yarn quality prediction method based on fireworks algorithm improvement BP neural network | |
CN103426027B (en) | A kind of intelligence of the normal pool level based on genetic algorithm back propagation neural network model method for optimizing | |
CN102684223B (en) | Optimized evaluating method for wind power output under multi-constraint condition for reducing transmission loss | |
CN104732300B (en) | A kind of neutral net wind power short term prediction method theoretical based on Fuzzy divide | |
CN103177288A (en) | Transformer fault diagnosis method based on genetic algorithm optimization neural network | |
CN103020336B (en) | A kind of equivalent LED light source creation method | |
CN106485314A (en) | A kind of optimization method of the flower pollination algorithm based on adaptive Gauss variation | |
CN107705556A (en) | A kind of traffic flow forecasting method combined based on SVMs and BP neural network | |
CN102855185A (en) | Pair-wise test method based on priority | |
CN103455612B (en) | Based on two-stage policy non-overlapped with overlapping network community detection method | |
CN105005878B (en) | A kind of comprehensive estimation method of strong intelligent grid | |
CN106056235A (en) | Power transmission grid efficiency and benefit detection method based on Klee method and matter element extension model | |
CN102184328A (en) | Method for optimizing land use evolution CA model transformation rules | |
CN104008426A (en) | Distributed computing environment performance predicting method based on integrated learning | |
CN106296434A (en) | A kind of Grain Crop Yield Prediction method based on PSO LSSVM algorithm | |
CN103761676A (en) | Method for evaluating economic performance of power plant | |
CN103353895A (en) | Pre-processing method of power distribution network line loss data | |
CN104462797A (en) | Increment integration algorithm used for procedure parameter online testing | |
CN104616072A (en) | Method for improving concentration of glutamic acid fermented product based on interval optimization | |
CN109491709A (en) | A kind of software code degree of controllability integrated evaluating method based on AHP and neural network | |
CN111914488B (en) | Data area hydrologic parameter calibration method based on antagonistic neural network | |
CN104318008B (en) | A kind of condenser Optimization Design based on loose constraint heredity simplex 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 | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180413 |