CN114386309A - Problem scale unification method for agent optimization in cloud computing environment - Google Patents
Problem scale unification method for agent optimization in cloud computing environment Download PDFInfo
- Publication number
- CN114386309A CN114386309A CN202111424117.9A CN202111424117A CN114386309A CN 114386309 A CN114386309 A CN 114386309A CN 202111424117 A CN202111424117 A CN 202111424117A CN 114386309 A CN114386309 A CN 114386309A
- Authority
- CN
- China
- Prior art keywords
- matrix
- input
- model
- optimization
- target
- 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
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- 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]
- G06F9/5061—Partitioning or combining of resources
- G06F9/5072—Grid computing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/12—Computing arrangements based on biological models using genetic models
- G06N3/126—Evolutionary algorithms, e.g. genetic algorithms or genetic programming
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2111/00—Details relating to CAD techniques
- G06F2111/06—Multi-objective optimisation, e.g. Pareto optimisation using simulated annealing [SA], ant colony algorithms or genetic algorithms [GA]
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Software Systems (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Mathematical Physics (AREA)
- Computer Hardware Design (AREA)
- Geometry (AREA)
- Medical Informatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physiology (AREA)
- Genetics & Genomics (AREA)
- Biomedical Technology (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a problem scale unification method for agent optimization in a cloud computing environment, which is applied to a cloud and comprises the following steps: acquiring an optimization task of a client about a multi-target problem; generating a plurality of input samples, and obtaining an output matrix by using a real function; generating a plurality of weight vectors, and aggregating the multi-target problem into a single-target problem; acquiring a plurality of relevant source models from a cloud database, and constructing a multi-problem agent model; searching a point to be evaluated by using a multi-problem agent model, and evaluating the point to be evaluated by using a real function to update an output matrix to obtain a pareto front surface; judging whether the preset evaluation times are reached; if not, returning to the step of generating a plurality of weight vectors; and if so, returning the finally obtained pareto frontier as an optimization result to the client. The invention uses a multi-problem agent model method to carry out multi-target optimization, introduces a random projection matrix, can solve the problem of inconsistent scale of a source model and a target model, can obtain more accurate results, and reduces the cost.
Description
Technical Field
The invention belongs to the technical field of cloud computing, and particularly relates to a problem scale unification method for agent optimization in a cloud computing environment.
Background
The cloud computing is distributed computing, and can decompose a large amount of data processing programs into a small program through a network cloud, then process and analyze the small program through a system consisting of a plurality of servers, and then send a final result to a client. At present, the cloud service is not just distributed computing, but a result of hybrid evolution and leap of computer technologies such as distributed computing, utility computing, load balancing, parallel computing, network storage, hot backup redundancy, virtualization and the like. In summary, cloud computing is not a completely new network technology, but a completely new network application concept. The core concept of cloud computing is that internet is used as a center, fast and safe cloud computing services and data storage are provided on websites, and every person using the internet can use huge computing resources and data centers on the internet.
The cloud computing platform is very powerful, the cloud database in the cloud computing platform comprises a large number of realized task models, various task models required for realizing the tasks can be found in the cloud database aiming at the tasks sent to the cloud by the user, and the task models are assembled together to meet the user requirements. Currently, for evaluating more expensive Multi-objective problems, optimization using Multi-problem agents (Multi-proxy Surrogates) is a popular research direction.
For example, Alan Tan Wei Min, Yew-Sonon ong, Abhishek Gupta and Chi-Keong Goh propose a Multi-objective Optimization method for efficient Transfer evolution using Multi-Problem agents (Transfer evolution Multi-objective Optimization with Multi-objective optimizations) in the document Multi-Problim optimizations.
However, the solution is implemented on the basis that the target model and the source model are consistent in problem scale, and the case that the problem scale is inconsistent is not considered. However, in many cases, the problem sizes of the source models are different, and it is impossible to completely find a source model that satisfies all the problem sizes of the target models, and therefore, in such a case, the expensive problem of the multi-objective optimization cannot be effectively solved by using the above scheme.
Disclosure of Invention
In order to solve the problems in the prior art, the invention provides a problem scale unification method for agent optimization in a cloud computing environment. The technical problem to be solved by the invention is realized by the following technical scheme:
a problem scale unification method for agent optimization in a cloud computing environment is applied to a cloud and comprises the following steps:
s1, acquiring optimization tasks of the client about the multi-target problem;
s2, generating a plurality of input samples, and obtaining an output matrix by using a real function in the optimization task;
s3, generating a plurality of weight vectors, and aggregating the multi-target problem into a single-target problem by using the plurality of weight vectors and the current output matrix;
s4, acquiring a plurality of relevant source models relevant to the optimization task from a cloud database, and constructing a multi-problem agent model by utilizing the relevant source models;
s5, searching a point to be evaluated by using the multi-problem agent model, evaluating the point to be evaluated by using the real function to update an output matrix, and obtaining a pareto front surface containing a plurality of non-dominant optimal solutions according to the updated output matrix;
s6, judging whether the current evaluation times reach the preset evaluation times;
if not, returning to S3; if yes, executing S7, and returning the pareto frontier finally obtained as an optimization result to the client.
In one embodiment of the present invention, the generating a plurality of weight vectors comprises:
uniformly generating a plurality of weight vectors in a unit hyperplane; and the number of the weight vectors is equal to the number of numerical values in the current output matrix, and the number of the weights contained in each weight vector is equal to the number of targets in the multi-target problem.
In an embodiment of the present invention, the aggregating the multi-target problem into a single-target problem by using the plurality of weight vectors and the current output matrix includes:
aggregating the numerical values in the current output matrix by using the plurality of weight vectors to obtain a plurality of aggregated numerical values; wherein each aggregate value represents an output value of the corresponding input sample under the single target problem.
In an embodiment of the present invention, the aggregating values in the current output matrix by using the plurality of weight vectors to obtain a plurality of aggregated values includes:
aggregating the numerical values in the current output matrix by using the plurality of weight vectors according to a Chebyshev polymerization method to obtain a plurality of aggregated numerical values;
wherein, for the Chebyshev polymerization method, the following formula is satisfied:
wherein y (x | w) represents an aggregation value, that is, y represents an output value after the multi-target problem is aggregated into a single-target problem, x represents an input sample, and w represents a weight vector; m represents the multi-meshTarget number in target question; w is ajRepresenting the jth weight value in the weight vector; f. ofjA function value representing a jth objective function after x is input to the real function;
the smallest function value in the jth objective function is represented as the reference point of the jth objective.
In an embodiment of the present invention, the building a multi-problem agent model by using the plurality of relevant source models includes:
training an initial Gaussian process model by using a plurality of pairs of sample data obtained by the input samples and the current output matrix;
aiming at each relevant source model, introducing a corresponding random projection matrix, and utilizing the random projection matrix to perform dimensionality reduction on an input matrix formed by the plurality of input samples to obtain a new input matrix corresponding to the relevant source model;
generating an outer sample target feature vector by using the plurality of pairs of sample data;
aiming at each relevant source model, taking the corresponding new input matrix as input to obtain corresponding output values, and forming a source characteristic matrix by the output values of all relevant source models;
calculating a mixing coefficient based on the out-of-sample target feature vector and the source feature matrix;
and constructing a multi-problem agent model by using the mixing coefficient.
In an embodiment of the present invention, the introducing a corresponding random projection matrix for each relevant source model, and performing dimension reduction on an input matrix formed by the plurality of input samples by using the random projection matrix to obtain a new input matrix corresponding to the relevant source model includes:
for each relevant source model, introducing a random projection matrix M with the dimensions of d1 × d2 for the relevant source model by using the input dimension d2 of the relevant source model and the dimension d1 of the input sample; wherein, the elements in the random projection matrix are random numbers between [0,1 ];
using formula XnewCalculating to obtain a new input matrix X corresponding to the relevant source modelnew(ii) a Wherein X represents an input matrix formed by the plurality of input samples.
In an embodiment of the present invention, the generating an out-of-sample target feature vector by using the plurality of pairs of sample data includes:
obtaining a temporary Gaussian process model which is correspondingly constructed after each sample data pair is eliminated by using a de-one method and the initial Gaussian process model;
obtaining a predicted value of each input sample on each obtained temporary Gaussian process model;
obtaining the average value of all predicted values of the same input sample obtained by different temporary Gaussian process models to obtain the predicted average value of the same input sample;
and forming an out-of-sample target feature vector by the prediction mean of all input samples.
In an embodiment of the present invention, the formula for calculating the mixing coefficient includes:
as,j≥0,j＝1,2,3,...,B
aT≥0
wherein n represents the number of input samples; b represents the number of the relevant source models; each as,jAnd aTIs the mixing coefficient;
representing output values of a jth relevant source model in the source feature matrix;
representing a predicted mean of an ith input sample in the out-of-sample target feature vector; y is(i)Represents the output value of the i-th input sample under the single target problem, i.e. the corresponding aggregate value.
In an embodiment of the present invention, the formula for constructing the multi-problem agent model includes:
wherein, y (x)(*)) Representing a multi-problem agent model; x is the number of(*)Representing an unknown input.
In an embodiment of the present invention, the searching for the point to be evaluated by using the multi-problem agent model, and evaluating the point to be evaluated by using the real function to update the output matrix includes:
searching points to be evaluated through selection, intersection and variation evolution by utilizing the multi-problem agent model;
and inputting the point to be evaluated into the real function for evaluation, and adding the obtained output value into the current output matrix to realize the updating of the output matrix.
In the scheme provided by the embodiment of the invention, the condition that the scale of the problem of the source model is inconsistent with that of the target model is successfully solved by introducing the random projection matrix, and a more accurate result can be obtained for the multi-target optimization problem. The embodiment of the invention uses a multi-problem agent model method to carry out multi-objective optimization, can effectively relieve the problem of overhigh cost when carrying out multi-objective optimization, changes the iteration termination condition from the original iteration times to the evaluation times, and can further reduce the cost. Moreover, the constructed multi-problem agent model is built by Gaussian process regression, so that the calculation process can be simplified.
The present invention will be described in further detail with reference to the accompanying drawings and examples.
Drawings
Fig. 1 is a schematic flowchart of a problem scale unification method for agent optimization in a cloud computing environment according to an embodiment of the present invention;
FIG. 2 is a schematic diagram of a source problem and a target problem before and after the source problem and the target problem are unified in scale according to an embodiment of the present invention;
fig. 3 is an optimization result obtained by using the source model with uniform problem scale according to the embodiment of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
Multi-objective optimization requires optimal decisions to be made in the case of tradeoffs between two or more conflicting objectives. Multiobjective optimization is currently widely used in many scientific fields, such as engineering, finance, marketing, recommendation, logistics and path planning, etc.
The method aims at the defect that the prior art cannot effectively solve the expensive problem of multi-objective optimization when the problem scales of a target model and a source model are inconsistent. The embodiment of the invention provides a problem scale unification method for agent optimization in a cloud computing environment, which is applied to a cloud. Referring to fig. 1, fig. 1 is a schematic flow chart of a problem scale unification method for agent optimization in a cloud computing environment according to an embodiment of the present invention. The method comprises the following steps:
and S1, acquiring the optimization tasks of the client on the multi-target problem.
The multi-objective problem is an optimization problem containing multiple objectives that the client wants to solve. And the cloud end acquires the optimization tasks containing the multi-objective problem sent by the client end, independently performs calculation processing, and finally returns the optimization result to the client end.
It will be appreciated that the multi-objective problem is represented by a plurality of objective functions to be optimized simultaneously, corresponding to the real functions. The multiple objective functions comprise multiple variables, and optimal output of the multi-objective problem is achieved by seeking optimal solutions of the variables.
And S2, generating a plurality of input samples, and obtaining an output matrix by using a real function in the optimization task.
Embodiments of the present invention generate a plurality of input samples according to the optimization task. Specifically, a plurality of input samples matching the multi-target problem may be generated by sampling or the like.
For example, a latin hypercube sampling method may be used to sample n input samples at a specified interval, each input sample having d1 variables.
The ith input sample may be represented as:
where d1 is the same as the number of variables of the multi-objective problem. n and d1 are natural numbers greater than 0.
The specified interval may be [0,1] interval, etc., and for the specific process of the latin hypercube sampling method, reference is made to the prior art, which is not described herein again.
For convenience of description hereinafter, the n input samples are constructed to obtain an input matrix of n × d1 dimensions, which may be denoted by X.
After obtaining a plurality of input samples, the input samples may be sequentially input into the real function in the optimization task to obtain a corresponding output value, i.e., a real function value. And constructing an output matrix from the output values of all the input samples.
For convenience of description hereinafter, the output matrix may be represented by PevalAnd (4) showing. It will be appreciated that PevalN number of values in (A) can be expressed as Peval＝(P1,P2,...,Pi,...,Pn)。
And S3, generating a plurality of weight vectors, and aggregating the multi-target problem into a single-target problem by using the plurality of weight vectors and the current output matrix.
Since multi-objective problems are computationally expensive, embodiments of the present invention aggregate them into a single-objective problem, which also corresponds to a true function of a single objective. To achieve aggregation, embodiments of the present invention require the generation of multiple weight vectors.
Since the elements in the weight vector are weights between [0,1], embodiments of the present invention can generate the weight vector by randomly generating weights within the value range, and the like.
In an alternative embodiment, the generating a plurality of weight vectors includes:
a plurality of weight vectors are uniformly generated in a unit hyperplane.
And the number of the weight vectors is equal to the number of numerical values in the current output matrix, and the number of the weights contained in each weight vector is equal to the number of targets in the multi-target problem. For example, if there are 3 objective functions, there are 3 weights in each weight vector. And the sum of the weights in each weight vector is 1.
For a specific way of uniformly generating a plurality of weight vectors in a unit hyperplane, please refer to the related art, which will not be described in detail herein.
In an alternative embodiment, the aggregating the multi-target problem into a single-target problem by using the plurality of weight vectors and the current output matrix includes:
and aggregating the numerical values in the current output matrix by using the plurality of weight vectors to obtain a plurality of aggregated numerical values.
Wherein each aggregate value represents an output value of the corresponding input sample under the single target problem. For convenience of the following description, a plurality of the aggregation values may be represented as (y)(1),y(2),...,y(i),...,y(n))。
As for the aggregation method, a weight sum method, a boundary cross aggregation method, or the like can be employed. The specific polymerization process is not described in detail herein.
For the first iteration, the current output matrix is the output matrix obtained in S2. And aiming at each subsequent iteration, the current output matrix is the output matrix finally updated by the last iteration.
S4, obtaining a plurality of relevant source models relevant to the optimization task from a cloud database, and constructing a multi-problem agent model by utilizing the relevant source models.
Because the real function of the multi-problem optimization calculation is expensive to calculate, and the cloud end may process similar tasks before, a multi-problem agent model can be created based on related tasks or data processed before the cloud end to replace the real function solution of the multi-target problem, so that the purposes of quick calculation and resource saving are achieved.
The cloud database contains a large number of source models, the cloud can search and search the source models related to the multi-target problem in the cloud database by utilizing the multi-target problem, and therefore a plurality of related source models can be obtained. The specific process belongs to the prior art and is not described in detail herein. For convenience of the following description, the number of several source models is denoted by B.
In an alternative embodiment, the building a multi-problem agent model by using the plurality of relevant source models includes the following steps:
and S41, training an initial Gaussian process model by using a plurality of pairs of sample data obtained by the input samples and the current output matrix.
Each input sample and the corresponding output value in the current output matrix constitute a pair of sample data pairs, and thus it will be appreciated that this step has n pairs of sample data. The method for training the Gaussian process model by using n pairs of sample data belongs to the prior art, and is not described in detail herein, for the convenience of distinguishing, the Gaussian process model obtained in the step is named as an initial Gaussian process model, and can be named as hTAnd (4) showing.
And S42, aiming at each relevant source model, introducing a corresponding random projection matrix, and using the random projection matrix to reduce the dimension of the input matrix formed by the plurality of input samples to obtain a new input matrix corresponding to the relevant source model.
Specific procedures for the step can include the following steps:
s421, for each relevant source model, introducing a random projection matrix M with the dimensions d1 × d2 for the relevant source model by using the input dimension d2 of the relevant source model and the dimension d1 of the input sample.
Wherein, the elements in the random projection matrix are random numbers between [0,1 ].
The input dimension d2 of the relevant source model represents the scale of the input problem of the relevant source model and can be directly determined from each relevant source model; the dimension d1 of the input sample represents the target model input problem size, determined by the optimization task.
S422, using formula XnewCalculating to obtain a new input matrix X corresponding to the relevant source modelnew。
Wherein X represents an input matrix formed by the plurality of input samples.
Thus, it will be appreciated that each associated source model will result in a corresponding new input matrix Xnew。
Because the input problem scale d1 of the target model does not match the input problem scale d2 of the related source model, the embodiment of the invention introduces a random projection matrix M for each related source model to perform dimension transformation on the original input matrix X, so that a new input matrix X is obtainednewAnd the input problem scale requirement of the relevant source model is met.
For easy understanding, please refer to fig. 2, where fig. 2 is a schematic diagram of the source problem and the target problem before and after the source problem and the target problem are unified, and is used to show a case after the source problem and the target problem are unified through the random projection matrix when the sizes of the source problem and the target problem are not consistent. The data of the target problem of the left image is three-dimensional, and the data of the left image is changed into the same scale with the source model through a projection matrix, namely, the data is changed into two dimensions.
And S43, generating an outer sample target feature vector by using the plurality of pairs of sample data.
Specifically, the step may include the steps of:
1) and obtaining a temporary Gaussian process model which is correspondingly constructed after each sample data pair is eliminated by using a de-normalization method and the initial Gaussian process model.
As described above, there are n pairs of sample data pairs, and for each sample number pair, a temporary gaussian process model may be constructed by using the remaining n-1 pairs of sample data pairs excluding the sample number pair and the hyper-parameter θ in the initial gaussian process model, and the specific construction process belongs to the prior art and is not described herein.
Thus, for each sample data pair, a corresponding temporary gaussian process model may be constructed. In total, n temporary gaussian process models are obtained.
2) And aiming at each obtained temporary Gaussian process model, obtaining a predicted value of each input sample on the temporary Gaussian process model.
And (4) substituting each input sample into each temporary Gaussian process model respectively to obtain corresponding output as a predicted value. Thus, each provisional gaussian process model yields a total of n predicted values.
3) And solving the mean value of all predicted values of the same input sample obtained by different temporary Gaussian process models to obtain the predicted mean value of the same input sample.
It will be appreciated that n predicted values are obtained for each input sample via n temporary gaussian process models. Then, the n predicted values can be averaged to serve as the predicted average of the input sample under all temporary gaussian process models. For convenience of description hereinafter, the predicted mean value of the ith input sample is used
And (4) showing.
4) And forming an out-of-sample target feature vector by the prediction mean of all input samples.
And (3) forming an out-of-sample target feature vector by the prediction mean of n input samples:
and S44, regarding each relevant source model, taking the corresponding new input matrix as input to obtain corresponding output values, and forming a source characteristic matrix by the output values of all relevant source models.
Inputting the new matrix XnewSubstituting into corresponding correlation source model to obtain corresponding output values, and forming source characteristic matrix y from all output valuess。
S45, calculating a mixing coefficient based on the target feature vector outside the sample and the source feature matrix.
In an optional embodiment, the formula for calculating the mixing coefficient includes:
as,j≥0,j＝1,2,3,...,B
aT≥0
wherein n represents the number of input samples; b represents the number of the relevant source models; each as,jAnd aTIs the mixing coefficient;
representing output values of a jth relevant source model in the source feature matrix;
representing a predicted mean of an ith input sample in the out-of-sample target feature vector; y is(i)Represents the output value of the i-th input sample under the single target problem, i.e. the corresponding aggregate value.
By the above calculation formula, the mixing coefficient a can be calculateds,jAnd aT。
And S46, constructing a multi-problem agent model by using the mixing coefficient.
In an optional embodiment, the formula for constructing the multi-problem agent model includes:
wherein, y (x)(*)) Representing a multi-problem agent model; x is the number of(*)Representing an unknown input.
By the formula, aiming at unknown input, a multi-problem agent model h can be obtained through calculationMPSThe output value of (1).
S5, searching a point to be evaluated by using the multi-problem agent model, evaluating the point to be evaluated by using the real function to update an output matrix, and obtaining a pareto front surface containing a plurality of non-dominant optimal solutions according to the updated output matrix.
The method comprises the steps of utilizing the multi-problem agent model to conduct evolution search in a plurality of input samples through a genetic algorithm, finding a point x to be evaluated, substituting the point x to be evaluated into a real function in the optimization task to conduct evaluation, obtaining an output value, and utilizing the output value to update an output matrix. From the values in the updated output matrix, a pareto frontier can be described, which contains a plurality of non-dominant optimal solutions, each of which represents the best input that can optimize the multi-objective problem.
And S6, judging whether the current evaluation times reach the preset evaluation times.
The iteration stop condition of the embodiment of the present invention is the number of evaluations rather than the number of iterations, because the cost required for one iteration is much greater than the cost required for evaluating several points. The use of the evaluation times as the iteration stop conditions can reduce the cost to the maximum extent.
The current evaluation number is an accumulated value, and the evaluation number may be one point or some points at a time. In a preferred embodiment, one click evaluation is evaluated once.
If not, return is made to S3.
It is understood that if the preset number of evaluations is not reached and the next iteration is needed, the process returns to S3. The updated output matrix of S5 is referred to as the "current output matrix" of S3 when the process returns to S3. Meanwhile, since the output matrix is changed, the weight vector needs to be regenerated to realize matching.
If yes, executing S7, and returning the pareto frontier finally obtained as an optimization result to the client.
It will be appreciated that the iteration stops if a preset number of evaluations is reached. The pareto frontage resulting from each iteration is updated by adding a new non-dominated optimal solution. And finally, obtaining a non-dominant optimal solution contained in the pareto frontier by iteration as a final optimization result. The resulting pareto frontage may thus be returned to the client as an optimization result.
Referring to fig. 3, fig. 3 is a diagram illustrating an optimization result obtained by using a source model with a uniform problem size according to an embodiment of the present invention. The optimization result is a pareto frontier obtained finally by iteration, and the black dots on the upper side represent non-dominant optimal solutions. f. of1,f2,f3Respectively representing a first optimization objective, a second optimization objective and a third optimization objective provided by the client, namely a plurality of objective functions.
In the scheme provided by the embodiment of the invention, the condition that the scale of the problem of the source model is inconsistent with that of the target model is successfully solved by introducing the random projection matrix, and a more accurate result can be obtained for the multi-target optimization problem. The embodiment of the invention uses a multi-problem agent model method to carry out multi-objective optimization, can effectively relieve the problem of overhigh cost when carrying out multi-objective optimization, changes the iteration termination condition from the original iteration times to the evaluation times, and can further reduce the cost. Moreover, the constructed multi-problem agent model is built by Gaussian process regression, so that the calculation process can be simplified.
Some alternative implementations of the embodiments of the invention are described below.
In an alternative embodiment, for S3, the aggregating values in the current output matrix by using the plurality of weight vectors to obtain a plurality of aggregated values includes:
aggregating the numerical values in the current output matrix by using the plurality of weight vectors according to a Chebyshev polymerization method to obtain a plurality of aggregated numerical values; wherein, for the Chebyshev polymerization method, the following formula is satisfied:
wherein y (x | w) represents an aggregation value, that is, y represents an output value after the multi-target problem is aggregated into a single-target problem, x represents an input sample, and w represents a weight vector; m represents the number of targets in the multi-target problem; w is ajRepresenting the jth weight value in the weight vector; f. ofjA function value representing a jth objective function after x is input to the real function;
the smallest function value in the jth objective function is represented as the reference point of the jth objective.
The Chebyshev polymerization method can be applied to not only convex curves but also non-convex curves, and has a wider application range, so that the method can be preferably applied to the embodiment of the invention.
Therefore, by the formula, the aggregation value realizes the conversion of the multi-target problem into the single-target problem.
In an optional implementation manner, the searching for the point to be evaluated by using the multi-problem agent model, and evaluating the point to be evaluated by using the real function to update the output matrix includes the following steps:
and searching points to be evaluated through selection, intersection and variation evolution by using the multi-problem agent model.
The selection, crossing, and variant evolution search are the main steps of the Genetic Algorithm, the specific Genetic Algorithm can adopt Non-dominated targeting Genetic Algorithm (NSGA) and NSGA-II, etc., and the specific evolution search process is not described in detail herein.
Secondly, inputting the point to be evaluated into the real function for evaluation, and adding the obtained output value into the current output matrix to realize the updating of the output matrix.
It will be appreciated that the values in the output matrix are incremented with each evaluation.
For the effect of the method of the embodiment of the present invention, please understand with reference to the comparative experimental data. Table 1 shows experimental results of the present invention compared to existing methods in a cloud computing environment.
TABLE 1 comparison of experimental results of the present invention and existing methods in a cloud computing environment
| Number of evaluations | Not using agent | Existing use agent method | The method of the invention |
| 100 | 7.14e-2 | / | 6.86e-2 |
| 200 | 6.61e-2 | / | 6.33e-2 |
| 300 | 6.57e-2 | / | 6.13e-2 |
Here, "/" indicates that the existing proxy method cannot cope with the problem of inconsistent scale. The evaluation index was IGD.
An Inverted Generational Distance (IGD) evaluation index is an evaluation index for comparing comprehensive performance. IGD mainly employs a method of calculating the minimum sum of distances from each point on the real pareto frontplane to an algorithmically acquired set of individuals. The smaller the IGD value, the better the convergence and distribution performance of the algorithm, i.e. the better the comprehensive performance. The results can be concluded by experimental comparison: under the condition of the same evaluation times, the method provided by the embodiment of the invention can obtain a better optimization result.
In summary, in the multi-objective optimization method of multi-problem agents in the prior art, when the problem scales of the input model and the target model are inconsistent, the existing algorithm has the defect of failure. With the deepening of informatization, a cloud database of the cloud computing platform has a large number of trained models. Therefore, based on the above problems, an embodiment of the present invention provides a problem scale unification method for agent optimization in a cloud computing environment. The source model which meets requirements better is selected on the cloud side, the method of the random projection matrix is introduced, so that the input problem scale meets the input requirement scale of various source models, the limitation that the input problem scale of the source model and the input problem scale of the target model must be consistent in the original method is overcome, the expensive problem of the multi-target optimization algorithm is solved by better applying the multi-problem agent model, and the required cost is greatly reduced under the condition of no accuracy. And moreover, a cloud computing platform is utilized, so that the time and space cost is greatly reduced, the operation is simplified, and the application range expansion is realized.
It is noted that, herein, relational terms such as first and second, and the like may be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Also, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising an … …" does not exclude the presence of other identical elements in a process, method, article, or apparatus that comprises the element.
The above description is only for the preferred embodiment of the present invention, and is not intended to limit the scope of the present invention. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention shall fall within the protection scope of the present invention.
Claims (10)
1. A problem scale unification method for agent optimization in a cloud computing environment is applied to a cloud end, and the method comprises the following steps:
s1, acquiring optimization tasks of the client about the multi-target problem;
s2, generating a plurality of input samples, and obtaining an output matrix by using a real function in the optimization task;
s3, generating a plurality of weight vectors, and aggregating the multi-target problem into a single-target problem by using the plurality of weight vectors and the current output matrix;
s4, acquiring a plurality of relevant source models relevant to the optimization task from a cloud database, and constructing a multi-problem agent model by utilizing the relevant source models;
s5, searching a point to be evaluated by using the multi-problem agent model, evaluating the point to be evaluated by using the real function to update an output matrix, and obtaining a pareto front surface containing a plurality of non-dominant optimal solutions according to the updated output matrix;
s6, judging whether the current evaluation times reach the preset evaluation times;
if not, returning to S3; if yes, executing S7, and returning the pareto frontier finally obtained as an optimization result to the client.
2. The method for problem size unification of agent optimization in a cloud computing environment according to claim 1, wherein said generating a plurality of weight vectors comprises:
uniformly generating a plurality of weight vectors in a unit hyperplane; and the number of the weight vectors is equal to the number of numerical values in the current output matrix, and the number of the weights contained in each weight vector is equal to the number of targets in the multi-target problem.
3. The problem size unification method for agent optimization in a cloud computing environment according to claim 1 or 2, wherein the aggregating the multi-objective problem into a single objective problem by using the plurality of weight vectors and a current output matrix comprises:
aggregating the numerical values in the current output matrix by using the plurality of weight vectors to obtain a plurality of aggregated numerical values; wherein each aggregate value represents an output value of the corresponding input sample under the single target problem.
4. The method of claim 3, wherein the aggregating values in the current output matrix using the plurality of weight vectors to obtain a plurality of aggregated values comprises:
aggregating the numerical values in the current output matrix by using the plurality of weight vectors according to a Chebyshev polymerization method to obtain a plurality of aggregated numerical values;
wherein, for the Chebyshev polymerization method, the following formula is satisfied:
wherein y (x | w) represents an aggregation value, that is, y represents an output value after the multi-target problem is aggregated into a single-target problem, x represents an input sample, and w represents a weight vector;m represents the number of targets in the multi-target problem; w is ajRepresenting the jth weight value in the weight vector; f. ofjA function value representing a jth objective function after x is input to the real function;
the smallest function value in the jth objective function is represented as the reference point of the jth objective.
5. The problem size unification method for agent optimization in a cloud computing environment according to claim 1, wherein said building a multi-problem agent model using said plurality of relevant source models comprises:
training an initial Gaussian process model by using a plurality of pairs of sample data obtained by the input samples and the current output matrix;
aiming at each relevant source model, introducing a corresponding random projection matrix, and utilizing the random projection matrix to perform dimensionality reduction on an input matrix formed by the plurality of input samples to obtain a new input matrix corresponding to the relevant source model;
generating an outer sample target feature vector by using the plurality of pairs of sample data;
aiming at each relevant source model, taking the corresponding new input matrix as input to obtain corresponding output values, and forming a source characteristic matrix by the output values of all relevant source models;
calculating a mixing coefficient based on the out-of-sample target feature vector and the source feature matrix;
and constructing a multi-problem agent model by using the mixing coefficient.
6. The method of claim 5, wherein the introducing a random projection matrix for each relevant source model, and using the random projection matrix to perform dimension reduction on the input matrix formed by the plurality of input samples to obtain a new input matrix corresponding to the relevant source model comprises:
for each relevant source model, introducing a random projection matrix M with the dimensions of d1 × d2 for the relevant source model by using the input dimension d2 of the relevant source model and the dimension d1 of the input sample; wherein, the elements in the random projection matrix are random numbers between [0,1 ];
using formula XnewCalculating to obtain a new input matrix X corresponding to the relevant source modelnew(ii) a Wherein X represents an input matrix formed by the plurality of input samples.
7. The method of claim 6, wherein the generating the off-sample target feature vector using the plurality of pairs of sample data comprises:
obtaining a temporary Gaussian process model which is correspondingly constructed after each sample data pair is eliminated by using a de-one method and the initial Gaussian process model;
obtaining a predicted value of each input sample on each obtained temporary Gaussian process model;
obtaining the average value of all predicted values of the same input sample obtained by different temporary Gaussian process models to obtain the predicted average value of the same input sample;
and forming an out-of-sample target feature vector by the prediction mean of all input samples.
8. The method for unifying problem size of agent optimization in cloud computing environment as claimed in claim 7, wherein the formula for calculating the mixing coefficient comprises:
minimize:
Subject to:
as,j≥0,j＝1,2,3,...,B
aT≥0
wherein n represents the number of input samples; b represents the number of the relevant source models; each as,jAnd aTIs the mixing coefficient;
representing output values of a jth relevant source model in the source feature matrix;
representing a predicted mean of an ith input sample in the out-of-sample target feature vector; y is(i)Represents the output value of the i-th input sample under the single target problem, i.e. the corresponding aggregate value.
9. The method for problem size unification of agent optimization in a cloud computing environment according to claim 8, wherein said multi-problem agent model is constructed according to a formula comprising:
wherein, y (x)(*)) Representing a multi-problem agent model; x is the number of(*)Representing an unknown input.
10. The method for problem size unification of agent optimization in a cloud computing environment according to claim 1, wherein said searching for a point to be evaluated using said multi-problem agent model, and evaluating said point to be evaluated using said truth function to update an output matrix comprises:
searching points to be evaluated through selection, intersection and variation evolution by utilizing the multi-problem agent model;
and inputting the point to be evaluated into the real function for evaluation, and adding the obtained output value into the current output matrix to realize the updating of the output matrix.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111424117.9A CN114386309B (en) | 2021-11-26 | 2021-11-26 | Agent optimization problem scale unification method in cloud computing environment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111424117.9A CN114386309B (en) | 2021-11-26 | 2021-11-26 | Agent optimization problem scale unification method in cloud computing environment |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114386309A true CN114386309A (en) | 2022-04-22 |
| CN114386309B CN114386309B (en) | 2023-08-08 |
Family
ID=81196271
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111424117.9A Active CN114386309B (en) | 2021-11-26 | 2021-11-26 | Agent optimization problem scale unification method in cloud computing environment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114386309B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114327859A (en) * | 2021-11-18 | 2022-04-12 | 西安电子科技大学 | Source model cluster selection method for cloud computing environment large-scale problem agent optimization |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20160034530A1 (en) * | 2014-07-29 | 2016-02-04 | Metanautix, Inc. | Systems and methods for a query optimization engine |
| US20170011143A1 (en) * | 2015-07-08 | 2017-01-12 | The Boeing Company | Search and poll method for solving multi-fidelity optimization problems |
| CN112749495A (en) * | 2021-01-27 | 2021-05-04 | 中国人民解放军国防科技大学 | Multipoint-point-adding-based proxy model optimization method and device and computer equipment |
-
2021
- 2021-11-26 CN CN202111424117.9A patent/CN114386309B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20160034530A1 (en) * | 2014-07-29 | 2016-02-04 | Metanautix, Inc. | Systems and methods for a query optimization engine |
| US20170011143A1 (en) * | 2015-07-08 | 2017-01-12 | The Boeing Company | Search and poll method for solving multi-fidelity optimization problems |
| CN112749495A (en) * | 2021-01-27 | 2021-05-04 | 中国人民解放军国防科技大学 | Multipoint-point-adding-based proxy model optimization method and device and computer equipment |
Non-Patent Citations (1)
| Title |
|---|
| 姜东;唐秋华;李梓响;吴玲;: "多目标模拟退火算法求解混装线平衡与排序", 机械设计与制造, no. 09 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114327859A (en) * | 2021-11-18 | 2022-04-12 | 西安电子科技大学 | Source model cluster selection method for cloud computing environment large-scale problem agent optimization |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114386309B (en) | 2023-08-08 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Ding et al. | A bi-objective load balancing model in a distributed simulation system using NSGA-II and MOPSO approaches | |
| Li et al. | Topology-aware neural model for highly accurate QoS prediction | |
| Azzouz et al. | Steady state IBEA assisted by MLP neural networks for expensive multi-objective optimization problems | |
| Li et al. | An alternative adaptive differential evolutionary algorithm assisted by expected improvement criterion and cut-HDMR expansion and its application in time-based sheet forming design | |
| Zhang et al. | Adaptive infinite impulse response system identification using an enhanced golden jackal optimization | |
| Zhang et al. | IFOA4WSC: a quick and effective algorithm for QoS-aware servicecomposition | |
| Song et al. | An adaptive failure boundary approximation method for reliability analysis and its applications | |
| CN114386309B (en) | Agent optimization problem scale unification method in cloud computing environment | |
| Adjakossa et al. | Kalman recursions aggregated online | |
| Chen et al. | Neural skyline filter for accelerating skyline search algorithms | |
| Zhang et al. | Fuzzy Adaptive NSGA-III for Large-Scale Optimization Problems | |
| CN111898628A (en) | Novel T-S fuzzy model identification method | |
| Hristov et al. | Deriving explicit control policies for Markov decision processes using symbolic regression | |
| Van Kien et al. | Cascade training multilayer fuzzy model for nonlinear uncertain system identification optimized by differential evolution algorithm | |
| WO2019144046A1 (en) | Distributed high performance computing using distributed average consensus | |
| Xie | Estimating civil aircraft’s research and manufacture cost by using grey system model and neural network algorithm | |
| US11487917B2 (en) | Automatic design of mechanical assemblies using estimation of distribution algorithm | |
| Zhaoming et al. | Improving the efficiency of multi-objective grasshopper optimization algorithm to enhance ontology alignment | |
| Cao et al. | A subassembly identification method for truss structures manufacturing based on community detection | |
| Fan | Load balance-aware dynamic cloud-edge-end collaborative offloading strategy | |
| Jun et al. | An improved dynamic structure-based neural networks determination approaches to simulation optimization problems | |
| Yang et al. | Non-matching meshes data transfer using Kriging model and greedy algorithm | |
| Morales et al. | Gaussian Processes for CVaR Approximation in Robust Aerodynamic Shape Design | |
| De Moraes et al. | An evolutive hybrid approach to cloud computing provider selection | |
| Jiang et al. | Developing Nonlinear Customer Preferences Models for Product Design Using Opining Mining and Multiobjective PSO-Based ANFIS Approach |
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 |