CN111913797A - Intelligent model distribution method, distribution system and application system - Google Patents

Intelligent model distribution method, distribution system and application system Download PDF

Info

Publication number
CN111913797A
CN111913797A CN202010657248.0A CN202010657248A CN111913797A CN 111913797 A CN111913797 A CN 111913797A CN 202010657248 A CN202010657248 A CN 202010657248A CN 111913797 A CN111913797 A CN 111913797A
Authority
CN
China
Prior art keywords
distribution
weight
intelligent
feature
objective
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.)
Granted
Application number
CN202010657248.0A
Other languages
Chinese (zh)
Other versions
CN111913797B (en
Inventor
匡立伟
李文超
尹山
吴军
谢秋红
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Fiberhome Telecommunication Technologies Co Ltd
Original Assignee
Fiberhome Telecommunication Technologies Co Ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Fiberhome Telecommunication Technologies Co Ltd filed Critical Fiberhome Telecommunication Technologies Co Ltd
Priority to CN202010657248.0A priority Critical patent/CN111913797B/en
Publication of CN111913797A publication Critical patent/CN111913797A/en
Application granted granted Critical
Publication of CN111913797B publication Critical patent/CN111913797B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00—Arrangements for program control, e.g. control units
    • G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/46—Multiprogramming arrangements
    • G06F9/50—Allocation of resources, e.g. of the central processing unit [CPU]
    • G06F9/5005—Allocation of resources, e.g. of the central processing unit [CPU] to service a request
    • G06F9/5027—Allocation of resources, e.g. of the central processing unit [CPU] to service a request the resource being a machine, e.g. CPUs, Servers, Terminals
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
    • 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"

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Data Mining & Analysis (AREA)
  • Mathematical Analysis (AREA)
  • Mathematical Optimization (AREA)
  • Computational Mathematics (AREA)
  • Pure & Applied Mathematics (AREA)
  • Software Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Strategic Management (AREA)
  • Operations Research (AREA)
  • Algebra (AREA)
  • Databases & Information Systems (AREA)
  • Human Resources & Organizations (AREA)
  • Economics (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Probability & Statistics with Applications (AREA)
  • Evolutionary Biology (AREA)
  • Development Economics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Game Theory and Decision Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computing Systems (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Marketing (AREA)
  • Quality & Reliability (AREA)
  • Tourism & Hospitality (AREA)
  • General Business, Economics & Management (AREA)
  • Complex Calculations (AREA)

Abstract

An intelligent model distribution method, an intelligent model distribution system and an intelligent model application system relate to the field of optical communication, and the method comprises the steps of determining main characteristics, characteristic types and characteristic attributes of an intelligent model in each distribution scheme, constructing a characteristic table by combining characteristic values obtained by multiple tests, further establishing an evaluation matrix and calculating a normalized foreground matrix of the distribution scheme; calculating objective weight under the condition of maximizing distribution value, calculating an evaluation target according to the subjective weight and the objective weight, gradually updating the objective weight by adopting a gradient method until the evaluation target is smaller than a set value, and calculating the distribution value of a corresponding distribution scheme by taking the objective weight as a balance weight; and comparing the distribution values of different distribution schemes, and selecting the distribution scheme with the maximum distribution value. The invention can obtain the optimal scheduling of the intelligent model and realize the maximization of efficiency.

Description

Intelligent model distribution method, distribution system and application system
Technical Field
The present invention relates to the field of optical communications, and in particular, to an allocation method, an allocation system, and an application system for an intelligent model.
Background
In the existing research, the expected utility theory is mainly applied to research the multi-attribute decision problem. But it is expected that the utility theory has inherent drawbacks: it assumes that all decision makers are fully rational and always strive for utility maximization. Thus, the conclusions drawn from it generally do not match the actual decision-making behavior of the decision-maker in uncertain cases, such as the well-known Alykura and Elsberg paradox. And the prospect theory uses a value function v (x) and a decision weight function w (p) to describe the selection behavior of the individual. The value function is related to the selection of the reference point, and the foreground value is determined by the value function and the decision weight function, i.e.
Figure BDA0002577203660000011
The foreground theory mainly analyzes the problem from the perspective of gain and loss, and the basic theoretical viewpoints are as follows: facing "gain", tending to "risk avoidance"; in the face of "losses", there is a tendency to "pursue risks". This assumption is consistent with the decision principle, and therefore the results obtained from it are more consistent with the actual decision behavior in uncertain situations.
For evaluating the characteristic indexes of the intelligent model, due to various different practical conditions, the prior art is lack of the evaluation standard and research of the optical network equipment model, and the model indexes are classified in a fine granularity according to the dimension of resource use by mostly adopting a fuzzy mode identification mode. The method has the following defects: the fuzzy recognition theory is adopted for the mixed multi-attribute situation which is met by a single characteristic attribute value and the optical communication intelligent model scheduling field.
An equipment and control platform distribution model system is constructed, an alternative iterative algorithm based on Lagrange dual decomposition and linear programming is used in the prior art, and the method has the following defects: neglecting that in practice the gain and loss will vary from case to case.
The feature weight and the model weight are information capable of reflecting the relative importance of the feature index and the model, and are descriptions of differences of the feature index and the model in decision positions. In the current research, there are roughly three types of methods for determining attribute weights: (1) the objective weighting method mainly comprises the following steps: entropy method, dispersion maximization method, satisfaction method based on scheme, proximity method based on scheme, and the like. (2) The subjective weighting method mainly comprises the following steps: a point estimation value method, a ring ratio scoring method, a judgment matrix method, an attribute importance ranking method and the like. (3) The combined empowerment method mainly comprises the following steps: a variance maximization weighting method, a combined least square method, an optimal coordination weighting method and the like. The disadvantages of the above method are: the weights are derived by a decision maker (or an analyst) at one time, and dynamic planning is lacked.
At present, no corresponding research and standard exists in the aspect of optical network intelligent model evaluation. The research related to intelligent model scheduling in the field of optical communication is less, the disclosed method related to resource scheduling also has the defects of different degrees, and the method is difficult to obtain ideal effect and low in efficiency when being directly applied to the intelligent model scheduling of optical communication.
Disclosure of Invention
Aiming at the defects in the prior art, the invention aims to provide an intelligent model distribution method, an intelligent model distribution system and an intelligent model application system, so that the optimal scheduling of the intelligent model is obtained, and the efficiency maximization is realized.
In order to achieve the above object, in one aspect, a method for distributing an intelligent model is adopted, including:
determining main characteristics, characteristic types and characteristic attributes of the intelligent model in each distribution scheme, constructing a characteristic table by combining characteristic values obtained by multiple tests, further establishing an evaluation matrix and calculating a normalized foreground matrix of the distribution scheme;
calculating objective weight under the condition of maximizing distribution value, calculating an evaluation target according to the subjective weight and the objective weight, gradually updating the objective weight by adopting a gradient method until the evaluation target is smaller than a set value, and calculating the distribution value of a corresponding distribution scheme by taking the objective weight as a balance weight;
and comparing the distribution values of different distribution schemes, and selecting the distribution scheme with the maximum distribution value.
Preferably, determining the main features comprises:
and calculating a spearman correlation coefficient between every two features, comparing the spearman correlation coefficient with a correlation coefficient threshold value, wherein the correlation coefficient threshold value is less than or equal to the correlation coefficient threshold value, and taking all irrelevant features as main features.
Preferably, the characteristic attribute comprises a constant, a continuous variable and a fuzzy number, the characteristic attribute is quantized into the main characteristics of the continuous variable and the fuzzy number according to a plurality of characteristic values of each main characteristic, and a characteristic table is constructed by the quantized main characteristics;
and establishing an evaluation matrix according to the type and the characteristics of the model, establishing a reference point matrix according to a reference point given by each model, and calculating a normalized foreground matrix under the current distribution scene according to the gain type and the cost type included by the characteristic type.
Preferably, the quantifying the feature with the feature attribute of continuous variable by normal distribution includes:
making a frequency distribution histogram according to a plurality of characteristic values of the characteristics, calculating skewness and kurtosis of the characteristic values, and if the skewness and the kurtosis meet normal distribution, solving a normal distribution function; and if the normal distribution is not satisfied, converting the biased distribution into a normal distribution function.
Preferably, the quantifying the feature with the feature attribute as the fuzzy quantity by constructing a language judgment fuzzy set includes:
establishing a level set of language judgment, wherein each level g is represented by an intuitive fuzzy number, and calculating a characteristic judgment average value of each model;
Figure BDA0002577203660000031
wherein p (g) represents the probability that the feature belongs to the class g; α (g), β (g) represent the degree of membership and the degree of non-membership of the level g, respectively.
Preferably, the calculating the normalized foreground matrix in the current distribution scenario includes:
constructing an evaluation matrix (x) from each allocation schemeij)t*nWhich isWherein i represents the number of the intelligent model types, j represents the feature number, t represents the number of the intelligent model types, and n represents the total number of the features; for each feature x under each modelijSetting a reference point OijTo construct a reference point matrix (O) for each intelligent modelij)t*n；
According to each feature xijCharacteristic type of (1), and reference point OijComparing, wherein for the cost type characteristics, the part lower than the reference point is gain, and the part exceeding the reference point is loss; for gain-type features, the part below the reference point is the loss, and the part above the reference point is the gain;
calculating each feature xijRelative to a set reference point OijGain of (1)ijOr loss of lossijWherein, in the step (A),
gainij＝|xij-Oij|0.28，lossij＝-|xij-Oij|0.32，
if the feature xijIf there is gain, the loss is 0; if the feature xijIf there is a loss, the gain is 0; correspondingly putting the gain and the loss into a gain matrix and a loss matrix, and constructing a normalized foreground matrix (V) under the current distribution scene through the gain matrix and the loss matrixij)t*n。
Preferably, the objective weight of the intelligent model comprises an objective model weight wikAnd objective feature weight wijkThe objective weight is solved by the Langerian multiplier method,
Figure BDA0002577203660000041
Figure BDA0002577203660000042
wherein i represents the number of the intelligent model types, j represents the number of the features, t represents the number of the intelligent model types, and n represents the total number of the features; k denotes the number of the allocation scene, andk equals 0 to indicate allocation on the equipment platform, k equals 1 to indicate allocation on the management and control platform, and wikRepresenting the objective model weight, w, of the ith intelligent model in the kth sceneijkRepresenting the objective feature weight m of the jth feature of the ith intelligent model in the kth sceneikRepresents the number, v, of the ith intelligent model in the kth sceneijkAnd the foreground value of the jth feature of the ith intelligent model in the kth scene is represented.
Preferably, the calculation process of adjusting the objective weight by the gradient method to obtain the equalization weight includes:
calculating the information entropy difference delta f of the subjective weight and the objective weight under each distribution scheme; constructing a Jordan matrix under each distribution scheme, and calculating objective weights corresponding to the maximum value and the minimum value; establishing an evaluation target y according to the delta f, the objective weight and the subjective weightm；
When y ismWhen the objective weight is more than or equal to the set value, the objective weight is adjusted and the delta f is recalculated, and then y is recalculatedmUp to ymAnd if the current objective weight is smaller than the set value, outputting the current objective weight as the balance weight.
Preferably, Δ f ═ fMaster and slave-fPassenger(s)，fMaster and slaveInformation of subjective weight, fPassenger(s)Information that is an objective weight;
fmaster and slave＝∑w′iklog(w′ik)+∑w′ijklog(w′ijk)，
fPassenger(s)＝∑wiklog(wik)+∑wijklog(wijk)，
w′ikShowing the subjective model weight, w 'of the ith intelligent model in the kth scene'ijkExpressing the subjective feature weight of the jth feature of the ith intelligent model in the kth scene;
constructing a Jordan matrix under each allocation scheme, and finding the objective weight w corresponding to the maximum value in the Jordan matrixmaxObjective weight w corresponding to minimum valuemin。
Preferably, the evaluation target y is constructed according to Δ f, objective weight and subjective weightmThe method specifically comprises the following steps:
when the value of deltaf is greater than 0,
Figure BDA0002577203660000051
when the value of af <0 is present,
Figure BDA0002577203660000052
wherein, w'max、w′minAre respectively wmaxAnd wminCorresponding to the subjective weight of the same scene under the same intelligent model.
Preferably, when ymWhen the objective weight is larger than or equal to the set value, the objective weight is adjusted according to the learning rate tau, and the method comprises the following steps:
in [ w ]min,wmax]Get the probing point a betweenm、bmAnd a is am<bm
am＝wmin+(1-τ)(wmax-wmin)
bm＝wmin+τ(wmax-wmin)
Wherein, wmax-bm＝am-wmin，
Figure BDA0002577203660000061
Said ym-1Denotes the m-1 th cycle evaluation target, ym-2Denotes the m-2 th cycle evaluation target, ym-3Representing the evaluation target of the m-3 times cycle; the first and second cyclic learning rates tau are 0.3685;
if y (a)m,wmax)<y(wmin,bm) Then maintain wminInvariable, wmax＝bm；
If y (a)m,wmax)>y(wmin,bm) Then maintain wmaxInvariable, wmin＝am。
In another aspect, a distribution system based on an intelligent model of the distribution method is provided, which includes:
the feature selection module is used for determining the main features, feature types and feature attributes of the intelligent models in each distribution scheme and recording feature values obtained by each test;
the characteristic table establishing module is used for establishing a characteristic table according to the content of the characteristic selecting module;
the foreground matrix module is used for establishing an evaluation matrix according to the feature table and calculating a normalized foreground matrix of the distribution scheme;
the calculation module is used for calculating objective weight under the condition of maximizing the distribution value and calculating an evaluation target according to the objective weight and the subjective weight;
the comparison module is used for continuously comparing whether the evaluation target obtained by the calculation module is smaller than a set value or not and sending a comparison result to the calculation module;
the calculation module is further used for gradually updating the objective weight by adopting a gradient method when the evaluation target is greater than or equal to the set value until the comparison module compares that the evaluation target is less than the set value, the calculation module calculates the distribution value of the corresponding distribution scheme according to the objective weight at the moment, compares the distribution values of different distribution schemes, and selects the distribution scheme with the maximum distribution value.
On the other hand, an application system of the distribution system based on the intelligent model is further provided, and the application system comprises an intelligent training platform, the distribution system of the intelligent model, a control platform and equipment;
the intelligent training platform is used for training a plurality of intelligent models, the distribution system of the intelligent models obtains a distribution scheme with the maximum distribution value through calculation, and the corresponding various intelligent models are distributed to the control platform and/or the equipment according to the distribution scheme.
The technical scheme has the following beneficial effects:
through intelligent model feature quantization and feature attribute classification, compared with the traditional method that the scheme is measured, the expected benefit is generally used to be maximized. The management and control platform and the equipment are fully combined, performance characteristics such as memories, running time and stability of different models are comprehensively considered, intelligent distribution is achieved, resource waste is reduced, optimal scheduling of the intelligent models is obtained, advantages of the management and control platform and the equipment are brought into play, and model utilization efficiency is maximized.
Drawings
FIG. 1 is a schematic diagram of a foreground matrix constructed in an embodiment of the present invention;
FIG. 2 is a flow chart of obtaining assigned values according to an embodiment of the present invention;
fig. 3 is an application system for intelligent model distribution according to an embodiment of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In addition, the technical features involved in the embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.
The invention provides an embodiment of an intelligent model distribution method, which comprises the following steps:
determining main characteristics, characteristic types and characteristic attributes of the intelligent model in each distribution scheme, constructing a characteristic table by combining characteristic values obtained by multiple tests, further establishing an evaluation matrix and calculating a normalized foreground matrix of the distribution scheme;
calculating objective weight under the condition of maximizing distribution value, calculating an evaluation target according to the subjective weight and the objective weight, gradually updating the objective weight by adopting a gradient method until the evaluation target is smaller than a set value, and calculating the distribution value of a corresponding distribution scheme by taking the objective weight as a balance weight;
and comparing the distribution values of different distribution schemes, and selecting the distribution scheme with the maximum distribution value.
Preferably, the intelligent model can be obtained by training through an intelligent training platform.
The characteristics of the intelligent model can be obtained by collecting the high-frequency characteristics of the intelligent model and selecting the characteristics which meet the actual requirements, and the characteristics can comprise running time, memory, bandwidth, energy consumption, safety and the like.
The characteristic attributes include a constant, a continuous variable and a fuzzy number, and the characteristics of the continuous variable and the fuzzy number need to be quantized. Wherein, the constant refers to a fixed value which is not easily interfered by the outside, such as a memory; the continuous variable refers to a variable which is limited by external factors but is distributed in a fixed interval as a whole according to a certain distribution function, such as running time; the fuzzy number is not easy to be represented by a numerical value, and language judgment fuzzy set mode description needs to be constructed, such as safety and the like.
The characteristic types comprise a gain type and a cost type, and a gain matrix and a loss matrix are established according to the characteristic types to further obtain a foreground matrix. Wherein, the cost type represents that the lower the characteristic value, the better, such as cost, energy consumption, etc.; the gain type indicates that the higher the feature value, the better, such as profit, security, and the like.
Based on the above, an embodiment of a test table is provided, for example, table 1 is a test table of an intelligent model on a device side, and includes main features of memory, runtime, and security, and in table 1, test 2, and test … …, test 5 are all test values obtained by testing various intelligent models under one allocation scheme.
TABLE 1
Figure BDA0002577203660000081
Figure BDA0002577203660000091
Based on the above, an embodiment is also provided in which redundant features are removed and main features are retained by calculating the spearman correlation coefficient between the features. Taking three features in table 1 as an example, the spearman correlation coefficient of each two features is calculated: ρ (memory, runtime) is 0.1, ρ (runtime, security) is 0.2, and ρ (memory, security) is 0.6. In this embodiment, assuming that a correlation coefficient threshold is 0.6, a correlation coefficient not greater than 0.6 indicates that the two features are not correlated, and the memory, the running time, and the security are not correlated with each other, all as main features.
In addition, an embodiment for quantifying the characteristic with the characteristic attribute as the continuous variable is also provided. For example, in table 2, the feature attribute of the intelligent model at the runtime of the device side in the scenario is a continuous variable, which is quantified by a positive-distribution function. Firstly, making a frequency distribution histogram according to a plurality of test values of continuous variable characteristics, calculating skewness and kurtosis of corresponding test values by using SPSS software, and solving a normal distribution function if the skewness and the kurtosis meet normal distribution (namely the skewness is approximately equal to 0 and the kurtosis is approximately equal to 0); and if the skewness and the kurtosis do not meet the normal distribution, converting the skewness distribution into the normal distribution, and then obtaining the normal distribution function.
TABLE 2
Number of tests Run time
Test 1 6.04s
Test 2 3.61s
Test 3 4.86s
…… ……
Test 100 5.24s
In addition, an embodiment for quantizing the feature with the feature attribute of fuzzy number is also provided, and in the embodiment, the quantization is performed by establishing a language judgment fuzzy set. First, a level set S ═ of language judgment is established (S)0,S1,…,S4) Not (poor, medium, good), SiExpressed as intuitive fuzzy numbers O ═ g, (. alpha. (g) } (α (g) +. beta. (g) ≦ 1), where g is rank and. alpha. (g), and. beta. (g) represent the degree of membership and non-membership, respectively, of rank g.
Figure BDA0002577203660000101
Where p (g) represents the probability that the feature belongs to the class g.
For example, an intelligent model is characterized by fuzzy number on the device side, firstly given the level set of language judgment as shown in Table 3, and secondly given the formula
Figure BDA0002577203660000102
And calculating a characteristic value, and finally, synthesizing a plurality of characteristics of which the characteristic attributes of the intelligent model at the equipment side are fuzzy numbers to obtain a characteristic evaluation average value statistical table 4.
TABLE 3
Rating of language judgment Fuzzy number
Good taste (0.7，0.3)
Is preferably used (0.6，0.4)
In (0.5，0.5)
Is poor (0.4，0.6)
Difference (D) (0.3，0.7)
TABLE 4
Feature(s) Mean value of feature judgment
Safety feature 0.47
…… ……
Portability 0.21
And counting various intelligent model data, and constructing a feature table of the intelligent model at the equipment side, as shown in table 5.
TABLE 5
Feature 1 … Characteristic n
Intelligent model 1 -70 … 0.48
Intelligent model 2 102 … 0.11
… … … …
Based on the above embodiment, an embodiment of constructing a normalized foreground matrix of each allocation scheme intelligent model according to the feature table is also provided. Constructing an evaluation matrix (x) according to the feature table of each allocation schemeij)t*nWherein i represents the intelligent model type number, j represents the feature number, t represents the intelligent model type number, and n represents the total number of features. For each feature x under each modelijSetting a reference point OijTo construct a reference point matrix (O) for each intelligent modelij)t*n。
As shown in FIG. 1, the evaluation matrix (x)ij)t*nWith reference point matrix (O)ij)t*nComparing one by one, wherein for the cost-type characteristics, the part below the reference point is regarded as 'gain', and the part above the reference point is regarded as 'loss'; the gain type is in contrast, and a portion below the reference point is regarded as "loss", and a portion above the reference point is regarded as "gain".
Calculating each feature xijRelative to a set reference point OijGain of (1)ijOr loss of lossijWherein the gainij＝|xij-Oij|0.28Loss of lossij＝-|xij-Oij|0.32If the feature xijIf there is gain, the loss is 0; if the feature xijIf there is a loss, the gain is 0. Correspondingly putting the gain and the loss into a gain matrix and a loss matrix, counting all characteristic gains and losses under all kinds of intelligent models, and constructing a normalized foreground matrix (V) under the current distribution scene through the gain matrix and the loss matrixij)t*n. Suppose x in FIG. 111Relative to a set reference point O11There is a gainijThen G is11＝gainijAnd L is11Is 0; x is the number of33Relative to a set reference point O33Loss of existenceijThen L is33＝lossij，G33Is 0.
For the foreground matrix of each intelligent model under each scene, aiming at each foreground VijAnd carrying out foreground normalization. For cost-type feature normalization, e.g., feature 1 of intelligent model 1:
Figure BDA0002577203660000121
for gain-type feature normalization, e.g., feature 2 of intelligent model 1:
Figure BDA0002577203660000122
for example, table 6 shows the foreground after normalization of the intelligent model under the scene control platform, and the remaining data portion after removal of the text is the corresponding foreground matrix. Table 7 shows the foreground after the intelligent model normalization in the scene device, and the corresponding foreground matrix is obtained by removing the remaining data part of the text.
TABLE 6
Management and control platform Memory device Run time Safety feature
Intelligent model 1 1 1 0.770
Intelligent model 2 0 0.625 1
Intelligent model 3 0.749 0 0
TABLE 7
Device Memory device Run time Safety feature
Intelligent model 1 0 0.352 1
Intelligent model 2 1 0 0.134
Intelligent model 3 0.749 1 0
Based on the above embodiments, an embodiment for calculating the maximum assigned value of the intelligent model is provided. As shown in fig. 2, for the foreground matrix () under the current distribution scene of the scheme k, firstly, subjective weights are set according to actual use scenes, and then, a double-layer network is constructed to calculate the balance weights.
The objective weights of the intelligent model include an objective feature weight wikAnd objective model weight wijkIn the first step, objective characteristic weight w is solved under the condition of maximizing distribution value according to the number of intelligent models in a distribution scheme and normalized foreground valueikAnd objective model weight wijkThe formula involved is as follows:
Figure BDA0002577203660000131
Figure BDA0002577203660000132
wherein i represents the number of the intelligent model types, j represents the number of the features, t represents the number of the intelligent model types, and n represents the total number of the features; k represents the number of the distribution scene, k is 0 to represent the distribution on the equipment platform, k is 1 to represent the distribution on the management and control platform, and wikRepresenting the objective model weight, w, of the ith intelligent model in the kth sceneijkRepresenting the objective feature weight m of the jth feature of the ith intelligent model in the kth sceneikRepresents the number, v, of the ith intelligent model in the kth sceneijkAnd the foreground value of the jth feature of the ith intelligent model in the kth scene is represented.
The objective weight is solved by a Langerhans multiplier method, and the formula is as follows:
Figure BDA0002577203660000133
Figure BDA0002577203660000134
k＝1,2。
wherein λ1And is λ2Obtaining objective characteristic weight w by using the Langerian multiplier method with parametersikAnd objective model weight wijkAre respectively as
Figure BDA0002577203660000135
Figure BDA0002577203660000136
And secondly, adjusting the objective weight by adopting a gradient method to obtain a balance weight, obtaining the distribution value of each scheme according to the balance weight in each scheme, obtaining and selecting the maximum distribution value maxV from the distribution values, and distributing according to the corresponding distribution scheme. The step of adjusting the objective weight by a gradient method to obtain the balance weight specifically comprises the following steps:
s101, calculating an information entropy difference delta f of the subjective weight and the objective weight under each distribution scheme; constructing a Jordan matrix under each distribution scheme, calculating objective weights corresponding to the maximum value and the minimum value, and constructing an evaluation target y according to the delta f, the objective weights and the subjective weightsm。
Specifically, according to Δ f ═ fMaster and slave-fPassenger(s)Calculating subjective weight information f under each distribution schemeMaster and slaveAnd objective weight information fPassenger(s)The information entropy difference Δ f of (a), wherein,
fmaster and slave＝∑w′iklog(w′ik)+∑w′ijklog(w′ijk)，
fPassenger(s)＝∑wiklog(wik)+∑wijklog(wijk)，
w′ikShowing the subjective model weight, w 'of the ith intelligent model in the kth scene'ijkThe method comprises the steps of representing the subjective feature weight of the jth feature of the ith intelligent model in the kth scene, wherein entropy represents the size of information content contained in the weight.
The evaluation target y is constructed according to the delta f, the objective weight and the subjective weightmThe method specifically comprises the following steps:
first construct Jordan matrix under each allocation scheme
Figure BDA0002577203660000141
Finding the objective weight w corresponding to the maximum value in the Jordan matrixmaxObjective weight w corresponding to minimum valuemin。w′maxIs wmaxCorresponding to subjective weights of the same scene and the same intelligent model (the same characteristic); w'minIs wminThe corresponding onesSubjective weighting under the same intelligent model (same feature) for a scene.
Then, construct wmax，wminMaximizing evaluation objective, wherein the mth cycle evaluation objective is ym。
If Δ f >0, then
Figure BDA0002577203660000142
Figure BDA0002577203660000151
If Δ f <0, then
Figure BDA0002577203660000152
S102, judging ymIf the value is less than the set value, entering S104; if not, the process proceeds to S103. The set value can be specifically set according to experience and requirements, so that the intelligent model achieves actual maximum efficiency and intelligent distribution on the premise of meeting the actual management and control platform and equipment performance.
S103, the objective weight is adjusted according to the learning rate tau, and the process proceeds to S101. Wherein, adjusting the objective weight according to the learning rate τ specifically includes:
in [ w ]min,wmax]Get the probing point a betweenm、bmAnd a is am<bm(ii) a Wherein, am、bmThe value of the value satisfies the following two conditions:
(1)wmax-bm＝am-wmin
(2)bm+1-am+1＝τ(bm-am)
Figure BDA0002577203660000153
y is abovem-1Denotes the m-1 th cycle evaluation target, ym-2Denotes the m-2 th cycle evaluation target, ym-3Representing the evaluation target of the m-3 times cycle; the first and second loop learning rates are fixed values, and in the present embodiment, are 0.3685.
Through calculation:
am＝wmin+(1-τ)(wmax-wmin)
bm＝wmin+τ(wmax-wmin)
if y (a)m,wmax)<y(wmin,bm) Then maintain wminInvariable, wmax＝bm(ii) a If y (a)m,wmax)>y(wmin,bm) Then maintain wmaxInvariable, wmin＝am。
And S104, taking the finally adjusted objective weight as a balance weight, and outputting the balance weight for calculating the distribution value of the corresponding scheme.
For example, table 7 is a foreground table of the assignment scheme, including the features and weights of each feature of the intelligent models, and the foreground values of each intelligent model, and the last row of table 7 gives the calculated assigned values. If the allocation scheme is the scheme with the maximum allocation value, then according to the column of number/weight, the allocation mode can be obtained as follows: deploying one intelligent model 1, two intelligent models 2 and three intelligent models 3 to a management and control platform.
TABLE 7
Figure BDA0002577203660000161
Based on the above embodiment, an intelligent model distribution system is also provided. As shown in fig. 3, the foreground matrix computation module comprises a feature selection module, a feature table establishment module, a foreground matrix module, a computation module and a comparison module.
The feature selection module is used for determining the main features, feature types and feature attributes of the intelligent models in each distribution scheme and recording feature values obtained by each test;
the characteristic table establishing module is used for establishing a characteristic table according to the content of the characteristic selecting module;
the foreground matrix module is used for establishing an evaluation matrix according to the feature table and calculating a normalized foreground matrix of the distribution scheme;
the calculation module is used for calculating objective weight under the condition of maximizing the distribution value and calculating an evaluation target according to the objective weight and the subjective weight;
the comparison module is used for continuously comparing whether the evaluation target obtained by the calculation module is smaller than a set value or not and sending a comparison result to the calculation module;
the calculation module is further used for continuously updating the objective weight by adopting a gradient method when the evaluation target is greater than or equal to the set value until the comparison module compares that the evaluation target is smaller than the set value, calculating the distribution value of the corresponding distribution scheme according to the objective weight at the moment, comparing the distribution values of different distribution schemes, and selecting the distribution scheme with the maximum distribution value.
The intelligent training platform trains a plurality of intelligent models, the distribution system of the intelligent models obtains a distribution scheme with the maximum distribution value through calculation, and the intelligent models are distributed to the control platform and/or the equipment according to the distribution scheme. Specifically, in each allocation scheme, the intelligent models may be all allocated to the management and control platform, or all allocated to the devices; or one part of the management platform can be distributed to the management and control platform, and the other part of the management and control platform can be distributed to the equipment.
The present invention is not limited to the above-described embodiments, and it will be apparent to those skilled in the art that various modifications and improvements can be made without departing from the principle of the present invention, and such modifications and improvements are also considered to be within the scope of the present invention.

Claims (13)

1. A method for distributing intelligent models, comprising:
determining main characteristics, characteristic types and characteristic attributes of the intelligent model in each distribution scheme, constructing a characteristic table by combining characteristic values obtained by multiple tests, further establishing an evaluation matrix and calculating a normalized foreground matrix of the distribution scheme;
calculating objective weight under the condition of maximizing distribution value, calculating an evaluation target according to the subjective weight and the objective weight, gradually updating the objective weight by adopting a gradient method until the evaluation target is smaller than a set value, and calculating the distribution value of a corresponding distribution scheme by taking the objective weight as a balance weight;
and comparing the distribution values of different distribution schemes, and selecting the distribution scheme with the maximum distribution value.
2. The method of assigning intelligent models according to claim 1, wherein determining the primary features comprises:
and calculating a spearman correlation coefficient between every two features, comparing the spearman correlation coefficient with a correlation coefficient threshold value, wherein the correlation coefficient threshold value is less than or equal to the correlation coefficient threshold value, and taking all irrelevant features as main features.
3. The method of assigning intelligent models according to claim 1, wherein:
the characteristic attributes comprise constants, continuous variables and fuzzy numbers, the main characteristics of which the characteristic attributes are the continuous variables and the fuzzy numbers are quantized according to a plurality of characteristic values of each main characteristic, and a characteristic table is constructed by the quantized main characteristics;
and establishing an evaluation matrix according to the type and the characteristics of the model, establishing a reference point matrix according to a reference point given by each model, and calculating a normalized foreground matrix under the current distribution scene according to the gain type and the cost type included by the characteristic type.
4. The method of assigning an intelligent model according to claim 3, wherein quantifying the feature whose feature attribute is a continuous variable by a normal distribution comprises:
making a frequency distribution histogram according to a plurality of characteristic values of the characteristics, calculating skewness and kurtosis of the characteristic values, and if the skewness and the kurtosis meet normal distribution, solving a normal distribution function; and if the normal distribution is not satisfied, converting the biased distribution into a normal distribution function.
5. The method for assigning an intelligent model according to claim 3, wherein the quantifying of the feature whose feature attribute is a fuzzy quantity by constructing a language judgment fuzzy set comprises:
establishing a level set of language judgment, wherein each level g is represented by an intuitive fuzzy number, and calculating a characteristic judgment average value of each model;
Figure FDA0002577203650000021
wherein p (g) represents the probability that the feature belongs to the class g; α (g), β (g) represent the degree of membership and the degree of non-membership of the level g, respectively.
6. The intelligent model distribution method of claim 3, wherein computing the normalized foreground matrix in the current distribution scenario comprises:
constructing an evaluation matrix (x) from each allocation schemeij)t*nWherein i represents the intelligent model type number, j represents the feature number, t represents the intelligent model type number, and n represents the total number of features; for each feature x under each modelijSetting a reference point OijTo construct a reference point matrix (O) for each intelligent modelij)t*n；
According to each feature xijCharacteristic type of (1), and reference point OijComparing, wherein for the cost type characteristics, the part lower than the reference point is gain, and the part exceeding the reference point is loss; for gain-type features, the part below the reference point is the loss, and the part above the reference point is the gain;
calculating each feature xijRelative to a set reference point OijGain of (1)ijOr loss of lossijWherein, in the step (A),
gainij＝|xij-Oij|0.28，lossij＝-|xij-Oij|0.32，
if the feature xijIf there is gain, the loss is 0; if the feature xijIf there is a loss, the gain is 0; correspondingly putting the gain and the loss into a gain matrix and a loss matrix, and constructing a normalized foreground matrix (V) under the current distribution scene through the gain matrix and the loss matrixij)t*n。
7. The method of assigning an intelligent model according to claim 1, wherein the objective weights of the intelligent model comprise an objective model weight wikAnd objective feature weight wijkThe objective weight is solved by the Langerian multiplier method,
Figure FDA0002577203650000031
Figure FDA0002577203650000032
wherein i represents the number of the intelligent model types, j represents the number of the features, t represents the number of the intelligent model types, and n represents the total number of the features; k represents the number of the distribution scene, k is 0 to represent the distribution on the equipment platform, k is 1 to represent the distribution on the management and control platform, and wikRepresenting the objective model weight, w, of the ith intelligent model in the kth sceneijkRepresenting the objective feature weight m of the jth feature of the ith intelligent model in the kth sceneikRepresents the number, v, of the ith intelligent model in the kth sceneijkAnd the foreground value of the jth feature of the ith intelligent model in the kth scene is represented.
8. The method of claim 7, wherein the step of calculating the objective weights adjusted by the gradient method to obtain the balance weights comprises:
calculating the information entropy difference delta f of the subjective weight and the objective weight under each distribution scheme; at each distributionConstructing a Jordan matrix under the scheme, and calculating objective weights corresponding to the maximum value and the minimum value; establishing an evaluation target y according to the delta f, the objective weight and the subjective weightm；
When y ismWhen the objective weight is more than or equal to the set value, the objective weight is adjusted and the delta f is recalculated, and then y is recalculatedmUp to ymAnd if the current objective weight is smaller than the set value, outputting the current objective weight as the balance weight.
9. The method of assigning intelligent models according to claim 8, wherein: f isMaster and slave-fPassenger(s)，fMaster and slaveInformation of subjective weight, fPassenger(s)Information that is an objective weight;
fmaster and slave＝∑w′iklog(w′ik)+∑w′ijklog(w′ijk)，
fPassenger(s)＝∑wiklog(wik)+∑wijklog(wijk)，
w′ikShowing the subjective model weight, w 'of the ith intelligent model in the kth scene'ijkExpressing the subjective feature weight of the jth feature of the ith intelligent model in the kth scene;
constructing a Jordan matrix under each allocation scheme, and finding the objective weight w corresponding to the maximum value in the Jordan matrixmaxObjective weight w corresponding to minimum valuemin。
10. The method of assigning an intelligent model according to claim 9, wherein said constructing an evaluation objective y based on Δ f, objective weight and subjective weightmThe method specifically comprises the following steps:
when the value of deltaf is greater than 0,
Figure FDA0002577203650000041
when the value of deltaf is less than 0,
Figure FDA0002577203650000042
wherein, w'max、w′minAre respectively wmaxAnd wminCorresponding to the subjective weight of the same scene under the same intelligent model.
11. The method of assigning intelligent models according to claim 9, wherein when y ismWhen the objective weight is larger than or equal to the set value, the objective weight is adjusted according to the learning rate tau, and the method comprises the following steps:
in [ w ]min，wmax]Get the probing point a betweenm、bmAnd a is am＜bm
am＝wmin+(1-τ)(wmax-wmin)
bm＝wmin+τ(wmax-wmin)
Wherein, wmax-bm＝am-wmin，
Figure FDA0002577203650000043
Said ym-1Denotes the m-1 th cycle evaluation target, ym-2Denotes the m-2 th cycle evaluation target, ym-3Representing the evaluation target of the m-3 times cycle; the first and second cyclic learning rates tau are 0.3685;
if y (a)m，wmax)＜y(wmin，bm) Then maintain wminInvariable, wmax＝bm；
If y (a)m，wmax)＞y(wmin，bm) Then maintain wmaxInvariable, wmin＝am。
12. A distribution system based on an intelligent model of the distribution method according to any one of claims 1 to 11, comprising:
the feature selection module is used for determining the main features, feature types and feature attributes of the intelligent models in each distribution scheme and recording feature values obtained by each test;
the characteristic table establishing module is used for establishing a characteristic table according to the content of the characteristic selecting module;
the foreground matrix module is used for establishing an evaluation matrix according to the feature table and calculating a normalized foreground matrix of the distribution scheme;
the calculation module is used for calculating objective weight under the condition of maximizing the distribution value and calculating an evaluation target according to the objective weight and the subjective weight;
the comparison module is used for continuously comparing whether the evaluation target obtained by the calculation module is smaller than a set value or not and sending a comparison result to the calculation module;
the calculation module is further used for gradually updating the objective weight by adopting a gradient method when the evaluation target is greater than or equal to the set value until the comparison module compares that the evaluation target is less than the set value, the calculation module calculates the distribution value of the corresponding distribution scheme according to the objective weight at the moment, compares the distribution values of different distribution schemes, and selects the distribution scheme with the maximum distribution value.
13. An application system of the distribution system based on the intelligent model of claim 12, which is characterized by comprising an intelligent training platform, the distribution system of the intelligent model, a management and control platform and equipment;
the intelligent training platform is used for training a plurality of intelligent models, the distribution system of the intelligent models obtains a distribution scheme with the maximum distribution value through calculation, and the corresponding various intelligent models are distributed to the control platform and/or the equipment according to the distribution scheme.
CN202010657248.0A 2020-07-09 2020-07-09 Intelligent model distribution method, distribution system and application system Active CN111913797B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010657248.0A CN111913797B (en) 2020-07-09 2020-07-09 Intelligent model distribution method, distribution system and application system

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010657248.0A CN111913797B (en) 2020-07-09 2020-07-09 Intelligent model distribution method, distribution system and application system

Publications (2)

Publication Number Publication Date
CN111913797A true CN111913797A (en) 2020-11-10
CN111913797B CN111913797B (en) 2023-01-31

Family

ID=73226213

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010657248.0A Active CN111913797B (en) 2020-07-09 2020-07-09 Intelligent model distribution method, distribution system and application system

Country Status (1)

Country Link
CN (1) CN111913797B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117933832A (en) * 2024-03-25 2024-04-26 中国人民解放军63921部队 Index weight evaluation method for spacecraft ground equivalence test

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107784394A (en) * 2017-10-30 2018-03-09 长安大学 Consider that the highway route plan of prospect theory does not know more attribute method for optimizing
US20180307935A1 (en) * 2015-03-24 2018-10-25 Hrl Laboratories, Llc System for detecting salient objects in images
US20190251401A1 (en) * 2018-02-15 2019-08-15 Adobe Inc. Image composites using a generative adversarial neural network
CN110222973A (en) * 2019-05-31 2019-09-10 国网安徽省电力有限公司经济技术研究院 A kind of integrated energy system evaluation method and system based on optimal weights combination
CN110991856A (en) * 2019-11-28 2020-04-10 武汉大学 Electric vehicle charging demand analysis method considering user limitation

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180307935A1 (en) * 2015-03-24 2018-10-25 Hrl Laboratories, Llc System for detecting salient objects in images
CN107784394A (en) * 2017-10-30 2018-03-09 长安大学 Consider that the highway route plan of prospect theory does not know more attribute method for optimizing
US20190251401A1 (en) * 2018-02-15 2019-08-15 Adobe Inc. Image composites using a generative adversarial neural network
CN110222973A (en) * 2019-05-31 2019-09-10 国网安徽省电力有限公司经济技术研究院 A kind of integrated energy system evaluation method and system based on optimal weights combination
CN110991856A (en) * 2019-11-28 2020-04-10 武汉大学 Electric vehicle charging demand analysis method considering user limitation

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
GUIWU WEI: "Grey relational analysis model for dynamic hybrid multiple attribute decision making", 《KNOWLEDGE-BASED SYSTEMS》 *
阎曼婷: "基于前景理论的多属性决策方法研究", 《电脑知识与技术》 *

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117933832A (en) * 2024-03-25 2024-04-26 中国人民解放军63921部队 Index weight evaluation method for spacecraft ground equivalence test

Also Published As

Publication number Publication date
CN111913797B (en) 2023-01-31

Similar Documents

Publication Publication Date Title
CN108280479B (en) Power grid user classification method based on load characteristic index weighted clustering algorithm
CN113242568A (en) Task unloading and resource allocation method in uncertain network environment
CN113610303A (en) Load prediction method and system
CN113938488A (en) Load balancing method based on dynamic and static weighted polling
CN109151953B (en) Network access selection calculation method based on user and network bilateral benefits
CN112633316A (en) Load prediction method and device based on boundary estimation theory
CN112819054A (en) Slice template configuration method and device
CN111913797B (en) Intelligent model distribution method, distribution system and application system
CN115525230A (en) Storage resource allocation method and device, storage medium and electronic equipment
CN111047163A (en) Energy storage strategy data processing system, method, device and medium
CN117220281A (en) Electric automobile access power grid adjustment capability quantitative evaluation method and system
CN112329923A (en) Model compression method and device, electronic equipment and readable storage medium
Zhang et al. User power interaction behavior clustering analysis that is based on the self-organizing-center K-means algorithm
Li Optimization of task offloading problem based on simulated annealing algorithm in MEC
CN115860388A (en) Multi-load regulation and control method, device, terminal and storage medium
CN114037218A (en) Power supply and demand pattern evaluation method and device based on comprehensive entropy weight method-Delphi method
CN110311743B (en) Method for estimating main user duty ratio through variation inference
Lu et al. Evaluation of black-start schemes based on prospect theory and improved TOPSIS method
CN110162747A (en) A kind of pretreatment and recommended method based on feature
CN109670227A (en) A kind of methods of evaluation of the simulation mathematical model parameter pair based on big data
Cao et al. Spectrum trading in heterogeneous cognitive radio networks
CN117938716A (en) Network monitoring data processing system, method and storage medium based on cloud computing
CN118300074A (en) Accurate carbon data driven lower load side energy consumption cluster analysis method based on correlation coefficient measurement
CN118261464A (en) Evaluation method and evaluation device for consumption behaviors of natural gas clients
Zhang et al. Notice of Violation of IEEE Publication Principles: User Power Interaction Behavior Clustering Analysis That is Based on the Self-Organizing-Center K-Means 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
GR01 Patent grant
GR01 Patent grant