CN110533221A - Multipurpose Optimal Method based on production confrontation network - Google Patents
Multipurpose Optimal Method based on production confrontation network Download PDFInfo
- Publication number
- CN110533221A CN110533221A CN201910688044.0A CN201910688044A CN110533221A CN 110533221 A CN110533221 A CN 110533221A CN 201910688044 A CN201910688044 A CN 201910688044A CN 110533221 A CN110533221 A CN 110533221A
- Authority
- CN
- China
- Prior art keywords
- sample
- network
- group
- variable
- optimization
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- 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
- 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)
- Theoretical Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biophysics (AREA)
- General Physics & Mathematics (AREA)
- Software Systems (AREA)
- Business, Economics & Management (AREA)
- Mathematical Physics (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Human Resources & Organizations (AREA)
- Economics (AREA)
- Strategic Management (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Development Economics (AREA)
- Game Theory and Decision Science (AREA)
- Genetics & Genomics (AREA)
- Physiology (AREA)
- Entrepreneurship & Innovation (AREA)
- Marketing (AREA)
- Operations Research (AREA)
- Quality & Reliability (AREA)
- Tourism & Hospitality (AREA)
- General Business, Economics & Management (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
A kind of Multipurpose Optimal Method based on production confrontation network disclosed by the invention, solves the problems such as existing optimization algorithm time cost is high, and training difficulty is excessive, and training network is easy collapse.Its implementation: stochastical sampling obtains initial sample；Pareto solution therein is selected as training set；It randomly selects half in training set to be pre-processed as training sample, building production fights network, obtains generating sample after being iterated training；Judge whether to need to be trained optimization again according to evaluation number；The result for generating sample and other comparison algorithms through the invention is compared, evaluation algorithms superiority and inferiority.Present invention reduces time costs, improve network robustness and stability, and effect of optimization is obvious, can be used for the resource allocation of multiple targets, the production scheduling of multiple products and optimization of the multiple performances of various software systems etc..
Description
Technical field
It is specifically a kind of to be based on production the invention belongs to field of computer technology, in particular to multi-target parameter optimizing
Fight the Multipurpose Optimal Method of network, abbreviation MOGAN, can be used for using computer technology to the resource allocations of multiple targets,
The optimization etc. of the production scheduling of multiple products and the multiple performances of various software systems.
Background technique
In industrial production and life, many problems are all made of the multiple targets for conflicting with each other and influencing.People's meeting
Frequently encountering makes multiple targets in given area while optimization problem as optimal as possible, that is, multi-objective optimization question
(multi-objectiveoptimization problem, MOP), existing optimization aim are more than one and need to more
A target is handled simultaneously.Such as in the design of component, it is desirable to which big some stability of power is more by force, high-power at this time and high
Stabilization is exactly a pair of conflicting target, and here it is a MOP problems；For a control system, it has various non-
Functional attributes, such as runing time, handling capacity, processing number per second have handling capacity in a distributed manner for message system Kafka
With two nonfunctional spaces of time delay, also referred to as target, in certain Kafka system, it is desirable to handling capacities a little and to postpone greatly
It is a little bit smaller, the two targets be also it is conflicting, the task of multiple-objection optimization is exactly the configuration by adjusting its control system
The noninferior solution of one group of the two target is found, this is also a MOP problem.There are many more other in industrial production and life
MOP problem, be related to life and industrial every aspect, relevant method for solving to solve politics, finance, it is military,
The programmed decision-making problem of the various aspects such as environment, the manufacturing, social security plays a crucial role, and is modern industry system
The technical issues that need to address in engineering.
In single-object problem, usual optimal solution only one, and fairly simple and common mathematics side can be used
Method finds out its optimal solution.However in multi-objective optimization question, mutually restricted between each target, an Objective may be made
The improvement of energy is often to lose other target capabilities as cost, it is impossible to which there are one to be optimal all target capabilities all
Solution, so for multi-objective optimization question, solution is usually set --- the Pareto disaggregation of a noninferior solution.
In multiple objective programming, due to the phenomenon that there are the conflicts between target with that can not compare, a solution is in some mesh
It is best for putting on, may be poor in other targets.The solution insubjection solution (Non- of Pareto proposition multiple target
Dominated set) concept, is defined as: assuming that it is any two solution S1 and S2 for all targets, S1 is superior to S2, then
We claim S1 to dominate S2, if S1 is not dominated by other solutions, S1 is known as non-domination solution (insubjection solution), also referred to as Pareto
Solution.
Multi-objective Evolutionary Algorithm (MOEA) is an analoglike biological evolution mechanism and the probability optimization of overall importance search that is formed
Method, basic principle are described as follows: the population that multi-objective Evolutionary Algorithm generates at random from one group, by executing to population
The evolutional operations such as selection, intersection and variation are evolved through excessive generation, individual fitness continuous improvement in population, to gradually force
The Pareto optimal solution set of nearly multi-objective optimization question.Than more typical multi-objective Evolutionary Algorithm have NSGA2, PESA2 and
SPEA2.For these three algorithms, advantage is more but its disadvantage is also obvious.
The advantages of NSGA2, is that operational efficiency height, disaggregation have good distributivity, has especially for low-dimensional optimization problem
Preferable performance；Its shortcoming is that disaggregation process has defect in higher-dimension problem, the diversity of disaggregation is undesirable.
The advantages of PESA2, is that the convergence of its solution is fine, is easier close to optimal face, especially in higher-dimension problem feelings
Under condition；But disadvantage is that selection operation can only once choose an individual, time loss is very big, and the multiplicity of class
Property is bad.
The advantages of SPEA2, is that the good disaggregation of degree of distribution can be obtained, especially in the solution of higher-dimension problem,
But its cluster process keeps diversity to take a long time, operational efficiency is not high.
Other than multi-objective Evolutionary Algorithm, there is also multi-objective particle swarm algorithm (MOPSO).Particle group optimizing
(PSO) algorithm is a kind of simulation social action, the evolution technology based on swarm intelligence, with its unique search mechanism, remarkably
Constringency performance, convenient computer realize, be widely used in engineering optimization field, multiple target PSO (MOPSO) calculate
Method has been applied to different optimization fields, but there are computation complexities it is high, versatility is low, convergence is bad the disadvantages of.
Multiple-objection optimization is frequently encountered and to be solved the problems, such as in an engineering, although the above many methods are all attempted pair
It is solved, but there are still variety of problems and deficiency.Generally speaking, some optimum results are not good enough, some training difficulty
It is larger, need sample sizes it is too many, the optimization time also too long, spend cost it is too big.In addition to this each algorithm is also
Other limitations, such as: NSGA2 disaggregation process in higher-dimension problem has defect, and the diversity of disaggregation is undesirable；PESA2 selection
Operation can only once choose an individual, and time loss is very big, and the diversity of class is bad；SPEA2 cluster process is kept
Diversity takes a long time, and operational efficiency is not high；And that there are computation complexities is high, versatility is low, convergence is bad etc. lacks by MOPSO
Point.
Summary of the invention
It is an object of the invention in view of the above shortcomings of the prior art, propose that a kind of optimum results are more preferable, be easy to training,
Multipurpose Optimal Method of the optimizing faster based on production confrontation network.
The present invention is a kind of Multipurpose Optimal Method based on production confrontation network, which is characterized in that including walking as follows
It is rapid:
(1) initial sample is obtained:
(1a) selects optimization aim: for multi-objective optimization question present in a certain system, selection determination needs to optimize
Multiple targets, it is assumed that have q optimization aim, there is p system variable to produce this q optimization aim in all system variables
Raw to influence, each system variable has the data set and value range of oneself；
The maximum evaluation number of (1b) setting: each group of p system variable is defined as one group of independent variable, is obtained by one group of independent variable
To the value of q optimization aim be defined as dependent variable, primary evaluation, which just refers to, obtains the mistake of corresponding dependent variable by one group of independent variable
Journey；Evaluation number is set in optimization process as e, maximum evaluation number is E, and evaluation number e is initialized as zero；
(1c) determines initial sample size: determining that initializaing variable quantity is m group according to maximum evaluation number E, guarantees optimization
Process can be completed in maximum evaluation number E；
(1d) stochastical sampling obtains initial sample: stochastical sampling is carried out to each system variable for being related to optimization aim,
The initializaing variable of m group p dimension is obtained, evaluates the superiority and inferiority for measuring this m group initializaing variable with the operation method of multiple target place system,
Each group of initializaing variable obtains the value of corresponding q optimization aim, and the initializaing variable of every group of stochastical sampling and the value of optimization aim are total
With the initial sample of composition one group of p+q dimension；Initializaing variable sampling is evaluated and is combined with optimization aim once, and evaluation number e adds 1,
Whole m group initializaing variable is traversed, the final number e that evaluates adds m, obtains the initial sample of m p+q dimension；
(2) production for constructing multiple-objection optimization fights network (MOGAN): it includes generating network G and differentiation network D,
It generates network G and differentiates that network D is all made of three layers of full Connection Neural Network, the two is confronted with each other training, is promoted constantly excellent each other
Change, is built into production confrontation network (MOGAN) of multiple-objection optimization；
(3) select training set from initial sample: all p+ are chosen in the definition solved according to Pareto from initial sample
The Pareto solution of q dimension, then removes the value of q optimization aim in each group of Pareto solution, by the p of these removal optimization target values
The Pareto solution of dimension is used as training set；
(4) production of the multiple-objection optimization of training building fights network (MOGAN):
(4a) determines training sample: from training set randomly choose half as training sample, then to training sample into
The pretreatment of row data normalization, obtains pretreated training sample x；
(4b) inputs training sample into MOGAN: pretreated training sample x is input to production confrontation network
Differentiate in network D；
(4c) generates sample: in the production confrontation network of multi-objective optimization question, using network G is generated, generating p dimension
Generation sample z；
(4d) obtains differentiating result: will generate sample z and training sample x input and differentiates network D, output differentiates result；
(4e) training generates network G and differentiates network D: according to differentiating as a result, fixed generate network G, to differentiate network D into
Row training, is continued to optimize, and is from training sample or to carry out self-generating net until differentiating that network D can be accurately judged to a sample
The sample that network G is generated；Network D is differentiated according to differentiating as a result, fixing, and is trained, is continued to optimize, until sentencing to network G is generated
Other network D cannot judge a sample be from training sample or come self-generating network G generation sample；
It obtains generating sample set after the training of (4f) successive ignition: executing step (4a)-(4e), complete once to generation
The training of formula confrontation network；Judge whether to reach the number of iterations, if not reaching the number of iterations of design, repeats step
(4a)-(4e), continues to train；If reaching the number of iterations of design, m is generated using network G is generated1Group p dimension is initial to be become
Amount, and evaluate and measure this m1The superiority and inferiority of group sample, calculates the value of q dimension optimization aim, and evaluation number e adds m1, then m1Group p dimension
Initializaing variable and q dimension optimization aim merge to obtain m1The generation sample set of group p+q dimension；
(5) judgement evaluation number: judging evaluation number, if evaluation number e reaches maximum evaluation number E at this time,
Then the generation sample set that step (4f) is obtained executes step (9), further verifies effect of optimization as final result collection,
Otherwise, step (6) are executed；
(6) it obtains intersecting result set: by m1The generation sample set and training sample x of group p+q dimension merge, after merging
Sample set in select all p+q dimension Pareto solutions, removal Pareto solves the value of q optimization aim, obtains the initial change of p dimension
The Pareto of amount is solved, and is carried out simulation binary system crossover operation to the Pareto solution of these p dimension initializaing variable, is obtained m2Group p dimension is just
The intersection result set of beginning variable；
(7) variation result set is obtained: to m2The intersection result set of group p dimension initializaing variable carries out multinomial mutation operation, obtains
To m3The variation result set of group p dimension initializaing variable；
(8) evaluation obtains new initial sample: the operation method of system measures this m to evaluate where multiple target3Group p dimension
The superiority and inferiority of initializaing variable, each group of initializaing variable obtain the value of corresponding q optimization aim, every group of p initializaing variable and q optimization
Mesh target value collectively constitutes the initial sample of one group of new p+q dimension, and the every evaluation of initializaing variable is simultaneously combined once with optimization aim, commented
Valence number e adds 1, traverses whole m3The initializaing variable of group p dimension, the final number e that evaluates add m3, obtain m3Group p+q dimension it is new initial
Sample；Carry out step (3)~step (5) again；
(9) effect of optimization is verified:
(9a) has optimization algorithm with other and optimizes to the selected p optimized variable of step (1), obtains corresponding
Results of comparison collection；
(9b) optimizes the present invention by comparing final result collection and the results of comparison collection of other optimization algorithms of the invention
Effect is verified.
Multiple-objection optimization is frequently encountered and to be solved the problems, such as in an engineering, and existing many methods all exist various
Limitation and deficiency, some optimum results are not good enough, and some training difficulty are larger, the sample size of needs is too many, optimization also
Time too long, spend cost it is too big.In addition to this, there are also other limitations for each algorithm, such as: NSGA2 is solved in higher-dimension problem
Collection process has defect, and the diversity of disaggregation is undesirable；PESA2 selection operation can only once choose an individual, time loss
It is very big, and the diversity of class is bad；SPEA2 cluster process keeps diversity to take a long time, and operational efficiency is not high；And
MOPSO there are computation complexities it is high, versatility is low, convergence is bad the disadvantages of.
Compared with prior art, the present invention has the advantage that:
It is optimized using two network dual trainings, as a result well: by the present invention in that being based on two nets with one kind
The mode of network dual training optimizes, and has broken the intrinsic thinking of original multi-objective optimization question solution, has used one
Network carries out simulation and generates characteristic variable, another network judges performance quality, and alternating iteration carries out the side of the two processes
Method optimizes, and as a result well, simultaneously because two networks use three layers of fully-connected network, is easy to trained.
Reduce the demand to training samples number: the present invention randomly chooses half sample characteristics by experiment sample every time
Method, ensure that the diversity and randomness of training sample, it is also ensured that the quality of sample.Simultaneously as avoiding passing through
Many experiments obtain the process of great amount of samples, have saved time cost to the maximum extent.
Sample size is effectively expanded, save time cost: invention introduces the intersection thoughts in genetic algorithm, pass through
Simulation binary system is carried out to the sample of generation to intersect, and is expanded sample size, more effective search has been carried out to sample space, more
The data distribution for having found Pareto solution fastly, has greatly saved time cost.
Avoid falling into local optimum: invention introduces the variation thought in genetic algorithm, by the sample to generation into
The variation of row multinomial, effectively prevents the case where falling into local optimum, is easier to search out global optimum.
Detailed description of the invention
Fig. 1 is implementation flow chart of the invention；
Fig. 2 is the sub-process figure of production confrontation network internal logic in the present invention；
Fig. 3 is to differentiate network D in the present invention and generate the structure chart of network G；
Fig. 4 is the curve graph of the present invention and other optimization algorithms for the optimum results of ZDT1 function；
Fig. 5 is the curve graph of the present invention and other optimization algorithms for the optimum results of ZDT6 function.
Specific embodiment
With reference to the accompanying drawing with example to the detailed description of the invention
Embodiment 1
Multiple-objection optimization is the problem of being frequently encountered in industrial production, and many problems are all by conflicting with each other and influencing
Multiple targets composition, the function for example, Kafka system two conflicting performance objective handling capacities and delay, in circuit design
Rate and stability be also it is conflicting, these are all multi-objective optimization questions.Multi-objective optimization question is related to industrial
Every aspect, relevant method for solving is to various aspects such as solution politics, finance, military affairs, environment, the manufacturing, social securities
Programmed decision-making problem plays a crucial role, and is the technical issues that need to address in modern industry system engineering.Although existing
There are many methods all to solve to multi-objective optimization question trial, but is having variety of problems and deficiency, some optimum results
Not good enough, some training difficulty are larger, the sample size that needs is too many, the optimization time also too long, spend cost too big.This
Invention expands research for this status, proposes a kind of Multipurpose Optimal Method based on production confrontation network.
The present invention be it is a kind of based on production confrontation network Multipurpose Optimal Method include the following steps: referring to Fig. 1
(1) initial sample is obtained:
(1a) selects optimization aim: for multi-objective optimization question present in a certain system, selection determination needs to optimize
Multiple targets, it is assumed that have q optimization aim, there is p system variable to produce this q optimization aim in all system variables
Raw to influence, each system variable has the data set and value range of oneself.
The maximum evaluation number of (1b) setting: each group of p system variable is defined as one group of independent variable, is obtained by one group of independent variable
To the value of q optimization aim be defined as dependent variable, primary evaluation, which just refers to, obtains the mistake of corresponding dependent variable by one group of independent variable
Journey；Evaluation number is set in optimization process as e, maximum evaluation number is E, and evaluation number e is initialized as zero.
(1c) determines initial sample size: determining that initializaing variable quantity is m group according to maximum evaluation number E, guarantees optimization
Process can be completed in maximum evaluation number E.
(1d) stochastical sampling obtains initial sample: stochastical sampling is carried out to each system variable for being related to optimization aim,
The initializaing variable of m group p dimension is obtained, evaluates the superiority and inferiority for measuring this m group initializaing variable with the operation method of multiple target place system,
Each group of initializaing variable obtains the value of corresponding q optimization aim, and the initializaing variable of every group of stochastical sampling and the value of optimization aim are total
With the initial sample of composition one group of p+q dimension；Initializaing variable sampling is evaluated and is combined with optimization aim once, and evaluation number e adds 1,
Whole m group initializaing variable is traversed, the final number e that evaluates adds m, obtains the initial sample of m p+q dimension.
In numerous system variables, the present invention has been determined by p system related with q optimization aim and has become
It measures, then the stochastical sampling in this p system variable, every group is intended to sample, and ties up initializaing variable with the p that these stochastical samplings form
The value of the corresponding q optimization aim of affix behind forms a p+q and ties up initial sample, each p+q ties up initial sample
The information of initializaing variable and corresponding optimization aim is contained, stochastical sampling has generality and result may be very well.
(2) production for constructing multiple-objection optimization fights network (MOGAN): it includes generating network G and differentiation network D,
It generates network G and differentiates that network D is all made of three layers of full Connection Neural Network, the two is confronted with each other training, is promoted constantly excellent each other
Change, is built into production confrontation network (MOGAN) of multiple-objection optimization.
In production confrontation network, training sample, which is input to, to be differentiated in network D, is trained, until differentiating network D energy
Be accurately judged to a sample be from training sample or come self-generating network G generation sample；Random noise is input to life
It in network G, is trained, until differentiating that network D can not judge that a sample is from training sample or to carry out self-generating net
The sample that network G is generated, by the dual training of the two, last whole network output is the generation sample for generating network G and generating.
Unlike other optimization algorithms, present invention employs productions to fight network, based on zero in game theory
And game idea, compared to more traditional model, there are two different networks for it --- and it generates network and differentiates network, rather than
Single network, and training method, using dual training mode, this makes that difficulty is trained to reduce, these features to give birth to
An accepted way of doing sth fights the multiple groups optimal solution for searching out multiple-objection optimization that network can be faster and better.
(3) select training set from initial sample: all p+ are chosen in the definition solved according to Pareto from initial sample
The Pareto solution of q dimension, then removes the value of q optimization aim in each group of Pareto solution, by the p of these removal optimization target values
The Pareto solution of dimension is used as training set.
(4) production of the multiple-objection optimization of training building fights network (MOGAN):
(4a) determines training sample: from training set randomly choose half as training sample, then to training sample into
The pretreatment of row data normalization, obtains pretreated training sample x.The present invention only chooses half as training sample, compares
The sample size that other optimization algorithms need is few.
(4b) inputs training sample into MOGAN: pretreated training sample x is input to production confrontation network
Differentiate in network D.
(4c) generates sample: in the production confrontation network of multiple-objection optimization, using network G is generated, generating the life of p dimension
At sample z.
(4d) obtains differentiating result: will generate sample z and training sample x input and differentiates network D, output differentiates result.
(4e) training generates network G and differentiates network D: according to differentiating as a result, fixed generate network G, to differentiate network D into
Row training, is continued to optimize, and is from training sample or to carry out self-generating net until differentiating that network D can be accurately judged to a sample
The sample that network G is generated；Network D is differentiated according to differentiating as a result, fixing, and is trained, is continued to optimize, until sentencing to network G is generated
Other network D cannot judge a sample be from training sample or come self-generating network G generation sample.
It obtains generating sample set after the training of (4f) successive ignition: executing step (4a)-(4e), complete once to generation
The training of formula confrontation network；Judge whether to reach the number of iterations, if not reaching the number of iterations of design, repeats step
(4a)-(4e), continues to train；If reaching the number of iterations of design, m is generated using network G is generated1Group p dimension is initial to be become
Amount, and evaluate and measure this m1The superiority and inferiority of group sample, calculates the value of q dimension optimization aim, and evaluation number e adds m1, then m1Group p dimension
Initializaing variable and q dimension optimization aim merge to obtain m1The generation sample set of group p+q dimension.
(5) judgement evaluation number: judging evaluation number, if evaluation number e reaches maximum evaluation number E at this time,
Then the generation sample set that step (4f) is obtained executes step (9), further verifies effect of optimization as final result collection,
Otherwise, step (6) are executed.
(6) it obtains intersecting result set: by m1The generation sample set and training sample x of group p+q dimension merge, after merging
Sample set in select all p+q dimension Pareto solutions, removal Pareto solves the value of q optimization aim, obtains the initial change of p dimension
The Pareto of amount is solved, and is carried out simulation binary system crossover operation to the Pareto solution of these p dimension initializaing variable, is obtained m2Group p dimension is just
The intersection result set of beginning variable.The present invention generates new sample, expands sample size, carry out to sample space by intersecting
More effective search, has greatly saved time cost, reduces the cost of cost.
(7) variation result set is obtained: to m2The intersection result set of group p dimension initializaing variable carries out multinomial mutation operation, obtains
To m3The variation result set of group p dimension initializaing variable.
(8) evaluation obtains new initial sample: the operation method of system measures this m to evaluate where multiple target3Group p dimension
The superiority and inferiority of initializaing variable, each group of initializaing variable obtain the value of corresponding q optimization aim, every group of p initializaing variable and q optimization
Mesh target value collectively constitutes the initial sample of one group of new p+q dimension, and the every evaluation of initializaing variable is simultaneously combined once with optimization aim, commented
Valence number e adds 1, traverses whole m3The initializaing variable of group p dimension, the final number e that evaluates add m3, obtain m3Group p+q dimension it is new initial
Sample；Step (3)~step (5) are carried out again, training set is selected again, carries out the training and judgement of a new round.
(9) effect of optimization is verified:
(9a) has optimization algorithm with other and optimizes to the selected p optimized variable of step (1), obtains corresponding
Results of comparison collection；
(9b) optimizes the present invention by comparing final result collection and the results of comparison collection of other optimization algorithms of the invention
Effect is verified.By the experimental results showed that, effect of optimization of the invention is got well than other optimization algorithms.
For the multi-objective optimization question frequently encountered in industrial production, The present invention gives a results more preferably, is easy to
The technical solution of the Multipurpose Optimal Method of training, optimizing faster based on production confrontation network.In this process, first
Pass through stochastical sampling and handled to have obtained training sample, then constructs production and fight network, with obtained training sample
It is trained and is generated as a result, having obtained more samples by carrying out cross and variation to generation result, then again
Secondary to be trained network, until having reached the evaluation number set, the sample generated at this time is exactly final result of the invention.
By the present invention in that being optimized with a kind of mode based on two network dual trainings, original multiple target is broken
The intrinsic thinking of optimization problem solution carries out simulation using a network and generates characteristic variable, another network judgement property
Can be fine or not, and alternating iteration carries out the method for the two processes to optimize, as a result well, simultaneously because two networks make
With three layers of fully-connected network, it is easy to trained.
Embodiment 2
For Multipurpose Optimal Method based on production confrontation network with embodiment 1, step (2) is a kind of based on deep by designing
The production of degree study fights network G AN model, contacts establishing between functional value and variable, and this method is enabled to be based on choosing
The training sample taken, using the potential characteristic for generating model G and learning the solution of Pareto out, and using another differentiate network D into
Row judgement, calculates error, continues to optimize result.Differentiating network D and generating the two networks of network G uses classical three layers complete
Connect network structure, in which:
The production of building multiple-objection optimization of the invention fights network, wherein generating network G, is one and includes input
Three layers of fully-connected network of layer, hidden layer and output layer, the input layer include 5 nodes, and each node is in [- 1,1] range
Random number；The hidden layer has 128 nodes, and has weight relationship between each node and input layer, initialization weight be [-
1,1] random number in range；The output layer contains n node, and n is the quantity of optimization aim, and each node contains activation letter
Number relu, wherein the value of n is the variable number of specific function；
Differentiation network D therein is three layers of fully-connected network comprising input layer, hidden layer and output layer, this is defeated
Entering layer includes n node, and n is the quantity of optimization aim；The hidden layer has 128 nodes, and between each node and input layer
There is weight relationship, initialization weight is also the random number in [- 1,1] range, and each node contains activation primitive sigmoid；
The output layer contains 1 node, indicates the probability of input sample authenticity, and each node contains activation primitive tanh.
The present invention is optimized using the mode based on two network dual trainings, has broken original multi-objective optimization question
The intrinsic thinking of solution carries out simulation using a network and generates characteristic variable, another network judges performance quality, and
Alternating iteration carries out the method for the two processes to optimize, as a result well, simultaneously because two networks use three layers it is complete
Network is connected, is easy to trained.
Embodiment 3
Multipurpose Optimal Method based on production confrontation network is with embodiment 1-2, to generation sample described in step (6)
This collection merges with training sample and carries out simulation binary system crossover operation, obtains m2The intersection result set of group p dimension initializaing variable, tool
Body is that two neighboring sample is successively carried out simulation binary system crossover operation in sequence in generating sample set, first to two samples
This progress probabilistic determination intersect generating two new solutions, inherits aforementioned two if probability is less than the crossing-over rate of setting
A part of feature of a solution, it is possible to very close Pareto solution, the purpose of this step of the invention is for enlarged sample
Quantity preferably searches for entire sample space.Specific crossover process is carried out according to the formula that such as Imitating binary system intersects:
Wherein,It is two new samples after the two neighboring sample cross of jth group, x1j(t)x2j(t) it is
The prechiasmal two neighboring sample of jth group, t indicate the t generation in the genetic algorithm that simulation binary system intersects, and j indicates to intersect behaviour
The jth group of work, γjIt is the intersection of jth group,
ujFor random number, and uj∈ U (0,1), η are profile exponent, η > 0.
Invention introduces the intersection thoughts in genetic algorithm, pass through the sample after merging to generation sample set with training sample
This collection carries out simulation binary system and intersects, and generates new sample, expands sample size, more effectively search to sample space
Rope makes production confrontation network quickly have found the data distribution of Pareto solution, is conducive to generate better sample, greatly save
About time cost.
Embodiment 4
Multipurpose Optimal Method based on production confrontation network is with embodiment 1-3, to cross knot described in step (7)
Fruit collection carries out multinomial mutation operation, obtains m3The variation result set of group p dimension initializaing variable, specifically concentrates intersection result
Each sample carries out the mutation operation in genetic algorithm, first carries out probabilistic determination to each sample, if probability is less than the change of setting
Different rate then carries out the random variation of small part to sample, and specific change procedure is related with mutation operator, then generates p dimension just
The variation sample of beginning variable realizes the mutation operator form of multinomial mutation operation are as follows:
v'k=vk+δ·(uk-lk), wherein
In formula, vkIndicate father's individual, v'kIndicate son individual, ukIndicate the upper bound of p system variable value range, lk
Indicate the lower bound of p system variable value range, δ1=(vk-lk)/(uk-lk), δ1=(uk-vk)/(uk-lk), k indicates multinomial
In kth generation in the genetic algorithm of formula variation, u is the random number in [0, a 1] section,ηmIt is profile exponent.
Invention introduces the variation thoughts in genetic algorithm, by carrying out multinomial variation to the sample of generation, effectively
The case where falling into local optimum is avoided, is easier to search out global optimum.
The present invention has not only carried out cross and variation operation, also uses newly-generated sample as new initial sample, in this way
It can quickly be scanned in sample space when being trained again, be conducive to find better result faster.
A more detailed example is given below, the present invention is further described.
Embodiment 5
Multipurpose Optimal Method based on production confrontation network is comprised the following steps that with embodiment 1-4 referring to Fig. 1
(1) initial sample is obtained:
(1a) selects optimization aim: for multi-objective optimization question present in a certain system, selection determination needs to optimize
Multiple targets, it is assumed that have q optimization aim, there is p system variable to produce this q optimization aim in all system variables
Raw to influence, each system variable has the data set and value range of oneself.
The maximum evaluation number of (1b) setting: each group of p system variable is defined as one group of independent variable, is obtained by one group of independent variable
To the value of q optimization aim be defined as dependent variable, primary evaluation, which just refers to, obtains the mistake of corresponding dependent variable by one group of independent variable
Journey；Evaluation number is set in optimization process as e, maximum evaluation number is E, and evaluation number e is initialized as zero；
(1c) determines initial sample size: determining that initializaing variable quantity is m group according to maximum evaluation number E, guarantees optimization
Process can be completed in maximum evaluation number E；
(1d) stochastical sampling obtains initial sample: stochastical sampling is carried out to each system variable for being related to optimization aim,
The initializaing variable of m group p dimension is obtained, evaluates the superiority and inferiority for measuring this m group initializaing variable with the operation method of multiple target place system,
Each group of initializaing variable obtains the value of corresponding q optimization aim, and the initializaing variable of every group of stochastical sampling and the value of optimization aim are total
With the initial sample of composition one group of p+q dimension；Initializaing variable sampling is evaluated and is combined with optimization aim once, and evaluation number e adds 1,
Whole m group initializaing variable is traversed, the final number e that evaluates adds m, obtains the initial sample of m p+q dimension；
(2) production for constructing multiple-objection optimization fights network (MOGAN): it includes generating network G and differentiation network D,
It generates network G and differentiates that network D is all made of three layers of full Connection Neural Network, the two is confronted with each other training, is promoted constantly excellent each other
Change, is built into production confrontation network (MOGAN) of multiple-objection optimization；
Production based on deep learning fights network G AN model, contacts establishing between functional value and variable, so that this
Method can learn the potential characteristic of Pareto solution out using model G is generated, and use another based on the training sample chosen
A differentiation network D is judged, is calculated error, is continued to optimize result.
(3) select training set from initial sample: all p+ are chosen in the definition solved according to Pareto from initial sample
The Pareto solution of q dimension, then removes the value of q optimization aim in each group of Pareto solution, by the p of these removal optimization target values
The Pareto solution of dimension is used as training set；
(4) production of the multi-objective optimization question of training building fights network (MOGAN):
(4a) determines training sample: from training set randomly choose half as training sample, then to training sample into
The pretreatment of row data normalization, obtains pretreated training sample x；
Before each iterative process, the half-sample in the training set that previous step selects, and random alignment will be randomly selected,
As the training sample x of this time iterative process, be input in production confrontation network, with guarantee training sample diversity and
Reliability.
Sample data is standardized, the value of each feature is made to meet the normal distribution of (0,1), generates energy
The structural data of model training is enough carried out, so that production confrontation network is more stable, the Euclidean distance between feature is calculated
More rationally.
(4b) inputs training sample into MOGAN: pretreated training sample x is input to production confrontation network
Differentiate in network D；
(4c) generates sample: in the production confrontation network of multi-objective optimization question, using network G is generated, generating p dimension
Generation sample z；
According to the random number that the input layer for generating network G is always in [- 1,1] range, by between hidden layer node
Weight relationship the hidden layer node value of the network is calculated, which is transmitted to output layer, output layer again
Nodal value calculated by relu function, obtain finally with the consistent generation sample z of training sample x form.
(4d) obtains differentiating result: will generate sample z and training sample x input and differentiates network D, output differentiates result；
The hidden layer node value of the network is calculated by the weight relationship with hidden layer, hidden layer node value is passed through
Sigmoid function calculates, and is transmitted to output layer, and output layer nodal value finally passes through the calculating of tanh function, obtains differentiating that network D is closed
In the differentiation probability of two groups of sample authenticities.
(4e) training generates network G and differentiates network D: according to differentiating as a result, fixed generate network G, to differentiate network D into
Row training, is continued to optimize, and is from training sample or to carry out self-generating net until differentiating that network D can be accurately judged to a sample
The sample that network G is generated；Network D is differentiated according to differentiating as a result, fixing, and is trained, is continued to optimize, until sentencing to network G is generated
Other network D cannot judge a sample be from training sample or come self-generating network G generation sample；
The target formula of use is expressed as follows:
Wherein, the diversity factor of V expression generation sample and training sample, G expression generation network, D expression differentiation network, x~
pr(x) distribution about sample characteristics x is indicated, r indicates the number of parameters of sample, z~pn(z) point about sample characteristics z is indicated
Cloth, n indicate the number of parameters of sample, its expectation is asked in E expression；
(4e1) is fixed to generate network G, is optimized by the loss formula of D to differentiation network D；
When to differentiating that network D is optimized, need to maximize the sum of the mean value of two probability, therefore according to the think of of deep learning
Dimension obtains the loss function for differentiating network:
Two probability that step (4d) is acquired substitute into the loss function D_loss for differentiating network, by constantly minimizing this
Loss function, optimization differentiate the weight between network D difference node layer.
(4e2) is fixed to differentiate network D, is optimized by the loss formula of G to network G is generated；
When optimizing to generation network G, need to minimize the mathematical expectation of probability for generating sample, therefore according to the think of of deep learning
Dimension obtains the loss function for generating network:
Two probability that step (4d) is acquired substitute into the loss function G_loss for generating network, by constantly minimizing this
Loss function, optimization generate the weight between network G difference node layer.
It obtains generating sample set after the training of (4f) successive ignition: executing step (4a)-(4e), complete once to generation
The training of formula confrontation network.After primary training, judge whether to reach the number of iterations, if not reaching the number of iterations of design,
Step (4a)-(4e) is repeated, continues to train；If reaching the number of iterations of design, m is generated using network G is generated1Group p
Initializaing variable is tieed up, and evaluates and measures this m1The superiority and inferiority of group sample, calculates the value of q dimension optimization aim, and evaluation number e adds m1, then
m1Group p dimension initializaing variable and q dimension optimization aim merge to obtain m1The generation sample set of group p+q dimension；
(5) judgement evaluation number: judging evaluation number, if evaluation number e reaches maximum evaluation number E at this time,
Then the generation sample set that step (4f) is obtained executes step (9), further verifies effect of optimization as final result collection,
Otherwise, step (6) are executed；
(6) it obtains intersecting result set: by m1The generation sample set and training sample x of group p+q dimension merge, after merging
Sample set in select all p+q dimension Pareto solutions, removal Pareto solves the value of q optimization aim, obtains the initial change of p dimension
The Pareto of amount is solved, and is carried out simulation binary system crossover operation to the Pareto solution of these p dimension initializaing variable, is obtained m2Group p dimension is just
The intersection result set of beginning variable；
(7) variation result set is obtained: to m2The intersection result set of group p dimension initializaing variable carries out multinomial mutation operation, obtains
To m3The variation result set of group p dimension initializaing variable；
(8) evaluation obtains new initial sample: the operation method of system measures this m to evaluate where multiple target3Group p dimension
The superiority and inferiority of initializaing variable, each group of initializaing variable obtain the value of corresponding q optimization aim, every group of p initializaing variable and q optimization
Mesh target value collectively constitutes the initial sample of one group of new p+q dimension, and the every evaluation of initializaing variable is simultaneously combined once with optimization aim, commented
Valence number e adds 1, traverses whole m3The initializaing variable of group p dimension, the final number e that evaluates add m3, obtain m3Group p+q dimension it is new initial
Sample；Step (3)~step (5) are carried out again, training set is selected again, carries out the training and judgement of a new round；
(9) effect of optimization is verified:
(9a) has optimization algorithm with other and optimizes to the selected p optimized variable of step (1), obtains corresponding
Results of comparison collection；
(9b) optimizes the present invention by comparing final result collection and the results of comparison collection of other optimization algorithms of the invention
Effect is verified.
By the present invention in that being optimized with a kind of mode based on two network dual trainings, original multiple target is broken
The intrinsic thinking of optimization problem solution carries out simulation using a network and generates characteristic variable, another network judgement property
Can be fine or not, and alternating iteration carries out the method for the two processes to optimize, as a result well, simultaneously because two networks make
With three layers of fully-connected network, it is easy to trained.
Below with reference to specific application example, the present invention and its technical effect are explained again.
Embodiment 6
Multipurpose Optimal Method based on production confrontation network with embodiment 1-5,
Application example: optimizing kafka system using the present invention, obtains two optimization aims of the system --- and it is maximum
The optimum results of handling capacity and minimum delay.
Kafka system be it is a kind of distributed, the message system based on publish/subscribe is write using Scala, it is with can
Horizontal extension and high-throughput and be widely used, by different types of company, more families as a plurality of types of data pipes and
Message system uses.
Step 1, initial sample is obtained by stochastical sampling；
(1a) selects optimization aim: the optimization aim selected in this example refers to for two performances of distributed system Kafka
Mark --- handling capacity and delay, according to official document, determination has 11 system configuration variable x1, x2... ..., x11Shadow is generated to it
It rings, i.e. system variable p=11, as shown in table 1:
The system variable of table 1kafka
| Number | Name variable | Variable description | Value (default) |
| 1 | num.network.threads | Handle the Thread Count of network request | Integer (3) |
| 2 | num.io.threads | Handle the Thread Count of io | Integer (8) |
| 3 | queued.max.requests | Largest request quantity | Integer (500) |
| 4 | num.replica.fetchers | Thread Count for synchronization counterpart | Integer (1) |
| 5 | socket.receive.buffer.bytes | Socket data receiver buffer area byte number | Integer (102400) |
| 6 | socket.send.buffer.bytes | Socket data send buffer area byte number | Integer (102400) |
| 7 | socket.request.max.bytes | Socket requests maximum number of byte | Integer (104857600) |
| 8 | buffer.memory | The memory byte number of buffered message record | Integer (33554432) |
| 9 | batch.size | Batch processing byte number | Integer (16384) |
| 10 | linger.ms | Delay millisecond number when record is sent | Integer (0) |
| 11 | compression.type | Compression algorithm type | Enumerate (none) |
(1b) sets maximum evaluation number: maximum evaluation number E set in this example is 300, once evaluates and refers to change
This 11 system configuration variable x of kafka1, x2... ..., x11, primary system is then run, handling capacity and delay are obtained；Setting is commented
Valence number is e, is initialized as 0；
(1c) determines initial sample size: being 300 according to maximum evaluation number E, determines that the quantity of stochastical sampling is 100；
(1d) stochastical sampling obtains initial sample: to 11 independent variable x1, x2... ..., x30It is carried out between range of variables
Stochastical sampling obtains the random initializaing variable of 100 group of 11 dimension, changes this 11 system configuration variable x of kafka1, x2... ...,
x11, primary system is then run, obtains handling capacity and delay, each run evaluation number e adds 1, by x1, x2... ..., x11With
Handling capacity, delay are merged into the initial sample of one 13 dimension, and the final number e that evaluates adds 100, obtains the initial sample of 100 group of 13 dimension
This.
Obtain and the system variable of kafka and optimization aim of initial sample of the invention be it is closely related, walk herein
In rapid, stochastical samplings are not carried out to more than all 100 a system variables, but according to actual test determined 11 with gulp down
The amount of spitting and the relevant system variable of delay are sampled, and time cost is greatly saved in this.
Step 2, construct production confrontation network (MOGAN) of multi-objective optimization question: it includes generating network G and differentiation
Network D generates network G and differentiates that network D is all made of three layers of full Connection Neural Network, and the two is confronted with each other training, promotes each other
It continues to optimize, is built into production confrontation network (MOGAN) of multi-objective optimization question.
The multi-objective optimization question of system software performance belongs to multi-data processing, therefore optimizes Shi Yaokao to performance
Consider influencing each other between each feature.This example is by designing a kind of production confrontation network G AN based on deep learning
Model is contacted establishing between the performance and feature of Kafka, allows the invention to utilize based on the training sample chosen
The potential characteristic that model G learns the configuration of good performance of Kafka out is generated, and differentiates that network D is judged using another, is counted
Error is calculated, result is continued to optimize.The model does not use the previous performance that carries out and optimizes weight size between necessarily searching different characteristic
The thinking of relationship, and the relationship of performance and feature is sought using the fitness of network, it continues to optimize, and directly obtain matching for optimization
Set parameter.The result shows that the present invention can be obtained by exploring relationship of the different characteristic inside its configuration space so that performance
Preferable feature configuration.
As shown in Fig. 3, production of the invention fights network, comprising: and differentiate network model D and generates network model G,
The two networks use three layers of classical fully-connected network structure, in which:
Generation network model G of the invention is one and includes input layer, hidden layer and output layer as shown in Fig. 3 (b)
Three layers of fully-connected network, the input layer include 5 nodes, each node be [- 1,1] range in random number；The hidden layer
There are 128 nodes, and have weight relationship between each node and input layer, initialization weight is random in [- 1,1] range
Number；The output layer contains n node, and each node contains activation primitive relu, and wherein the value of n is the variable of specific function
Number, in this example, the variable number n of kafka are 11.
Differentiation network model D of the invention is one and includes input layer, hidden layer and output layer as shown in Fig. 3 (a)
Three layers of fully-connected network, the input layer include n node, i.e. 11 nodes；The hidden layer has 128 nodes, and each section
Have a weight relationship between point and input layer, initialization weight is also the random number in [- 1,1] range, and each node contain it is sharp
Function sigmoid living；The output layer contains 1 node, indicates the probability of input sample authenticity, and each node contains activation
Function tanh.
Step 3, training set is selected from initial sample, as training set；
In view of also having the configuration so that better performances around the configuration feature of better performances, it should choose performance most
Good sample training, this example according to actual needs, are selected in sample and are solved about handling capacity and the Pareto of delay, from these
32 (taking whole if less than 32) are chosen in Pareto solution as training set is iterated training.
Step 4, the production of the multi-objective optimization question of training building fights network (MOGAN)；
Referring to Fig. 2, this step is implemented as follows:
(4a) randomly chooses half as training sample from training set, and it is pre- then to carry out data normalization to training sample
Processing, obtains pretreated training sample x.
In each iterative process, the half-sample randomly selected in the training set of step 3 selection carries out random alignment, as
This time training sample x of iterative process, training sample x is 16 in this example.It is input in production confrontation network, to guarantee
The diversity and reliability of training sample.Why half is selected, is because if selection is very little, the potential characteristic of sample is not easy to learn
It practises；If selection is too many, and not can guarantee sample diversity.
The system variable of selection is traversed, first judges whether to be enumerated variable, then directly inputs life if not enumerated variable
An accepted way of doing sth fights network；If it is enumerated variable, then need to use N shapes to enumerated variable progress one-hot coding processing is belonged to
The N number of state of state register pair is encoded, and each state has other independent register-bits, and there was only one when any
Effectively, using a this efficient coding is that classified variable is indicated as binary vector for position；It, can general piece by one-hot coding
The value for lifting variable expands to theorem in Euclid space, some value of enumerated variable just corresponds to some point of theorem in Euclid space, while can be with
Enumerated variable is discretized into the combination of multiple variables, is directly handled with being generated formula confrontation network, so that the Europe between variable
It is more reasonable that formula distance calculates.
Variable-value is standardized, the value of each variable is made to meet the normal distribution of (0,1), generates energy
The structural data of model training is enough carried out, so that production confrontation network is more stable, the Euclidean distance between variable is calculated
More rationally.Why do so, is the distance between variable because in most of machine learning or deep learning algorithm
It calculates or the calculating of similarity is very important, and this example is all in theorem in Euclid space for the calculating of distance or similarity
Similarity calculation is carried out, and production fights network as a kind of deep learning algorithm, the pretreatment for needing to be normalized mentions
Rise algorithm stability and robustness.
Pretreated training sample x is input to production confrontation network by (4b)；
(4c), which is used, generates network G, generation and the consistent generation sample z of training sample x dimension,
According to the random number that the input layer for generating network G is always in [- 1,1] range, by between hidden layer node
Weight relationship the hidden layer node value of the network is calculated, which is transmitted to output layer, output layer again
Nodal value calculated by relu function, obtain finally it is consistent with training sample x form 11 dimension generation sample z.
(4d) will generate the training sample x that sample z and step (4b) are selected and input differentiation network D respectively, by with it is hidden
The hidden layer node value of the network is calculated in weight relationship containing layer, and hidden layer node value is calculated by sigmoid function, passes
It is delivered to output layer, output layer nodal value finally passes through the calculating of relu function, obtains differentiating network D about two groups of sample authenticities
Differentiate probability.
(4e) optimizes production confrontation network according to target formula:
The target formula is expressed as follows:
Wherein, V indicates to generate the diversity factor of sample z and training sample x, and G indicates to generate network, and D indicates to differentiate network, x
~pr(x) data distribution about training sample x is indicated, r indicates the variable quantity of training sample, z~pn(z) it indicates about life
At the data distribution of sample z, n indicates to generate the variable quantity of sample,The equal of training sample x data distribution is sought in expression
Value,The mean value of sample z data distribution is sought survival into expression.
(4e1) optimizes differentiation network D:
It can be seen that by above-mentioned target formula, when to differentiating that network D is optimized, need to maximize the mean value of two probability
The sum of, therefore according to the thinking of deep learning, obtain the loss function for differentiating network D:
Two probability that step (4c) is acquired substitute into the loss function D_loss for differentiating network, by constantly changing difference
Weight between node layer, i.e. w in Fig. 3 (a), constantly minimizes the loss function, is optimized with this and differentiates network D.
(4e2) is optimized to network G is generated:
It can be seen that by above-mentioned target formula, when optimizing to generation network G, need to minimize the probability for generating sample
Mean value, therefore according to the thinking of deep learning, obtain the loss function for generating network G:
Two probability that step (4d) is acquired substitute into the loss function G_loss for generating network, by constantly changing difference
Weight between node layer, i.e. w in Fig. 3 (b), constantly minimizes the loss function, is optimized with this and generates network D.
By the two processes, the ability for generating network generation authentic specimen can be promoted, and differentiate that network judgement sample is true
The ability of property can also be promoted.
(4f) repeats (4a) and arrives (4e), until reaching the number of iterations for meeting and setting, according to actual needs, saves last
The generation sample z' that network finally generates is generated several times, and as final optimization pass as a result, in this example, z' expression generates 100
The initializaing variable of 11 dimension of group, then changes this 11 system configuration variable x of kafka1, x2... ..., x11, primary system is run, is obtained
To handling capacity and delay, merge to obtain the generation sample set of 100 group of 13 dimension with initializaing variable.
The number of iterations of the present invention rule of thumb value, this example value 300000 times.
Step 5, judge to evaluate number；
In above-mentioned steps, it is once to comment that 11 dimension system variables of each pair of kafka system, which are once changed and run,
Valence, using the generation sample set of step 4 as final result collection, carries out step 9, further if evaluation number reaches limitation at this time
Optimum results are verified, otherwise, carry out step 6, training optimization production fights network again.
Step 6, binary system crossover operation is simulated；
Generation sample set and training sample x that 100 group 13 in step 4 is tieed up are merged, from the sample set after merging
13 all dimension Pareto solutions are selected, removal Pareto solves the value of 2 optimization aims, obtains the Pareto of 11 dimension initializaing variables
Solution carries out simulation binary system crossover operation to the Pareto solution of these 11 dimension initializaing variables, obtains m211 dimension initializaing variable of group
Intersect result set.m2For the quantity for intersecting result set, in operating process because the reason of crossing-over rate, some samples intersect,
Some is not intersected, m2Quantity be not fixed.
Intersection is the step key operation in genetic algorithm, simulates producing offspring in nature, refers to two phases
The chromosome mutually matched is exchanged with each other its portion gene in some way, to form two new individuals.Refer in the present invention
Be two generate solution between data cross, generate new solution, a part of feature of both of the aforesaid solution inherited, so having
It may closely Pareto solve, the purpose of this step is preferably to search for entire sample space for enlarged sample quantity.
In this example, the system variable tieed up to two group 11 carries out data cross, generates two groups of 11 new dimension system variables,
A part of feature for inheriting aforementioned two groups of system variables, it is possible to the very close Pareto about handling capacity and delay
Solution.
Step 7, multinomial makes a variation；
To m in step 62The intersection result set of 11 dimension initializaing variable of group carries out multinomial mutation operation, obtains m311 dimension of group is just
The variation result set of beginning variable.The reason of in operating process because of aberration rate, some sample variations, some does not make a variation, so m3
Quantity be not fixed.
Variation is another step key operation of genetic algorithm, simulates the gene mutation in nature, i.e. child replicates father
Possible (probability of very little) generates certain copy errors when female gene, and variation generates new chromosome, shows new character.In
Middle finger of the present invention solves the random variation of progress a part to generating, and generates new solution, the purpose of this step is in order to avoid falling into
Locally optimal solution is easier to search out globally optimal solution.
In this example, the data for exactly tieing up system variable to a certain group 11 are changed at random.
Step 8, evaluation obtains new initial sample；
For m in step 73The variation result set of 11 dimension initializaing variable of group, obtains this m by operating system3Group initializaing variable
Handling capacity and delay, each run evaluation number e all add 1, by x1, x2... ..., x11One is merged into handling capacity, delay
The initial sample of 13 dimensions, the final number e that evaluates add m3, obtain m3The initial sample of 13 dimension of group, carries out step 2~step 5 again,
Training optimization production fights network.
Step 9, effect of optimization is verified；
According to production fight network generate final result collection, on kafka again operation to be tested, obtain as
The HV value of the following table 2.HV is hypervolume index, and what it was measured is the non-dominant disaggregation and reference obtained by multi-objective optimization algorithm
The volume in the dimension region in the object space that point surrounds, the value of HV is bigger, illustrates that optimum results are better.
Table 2: the HV value of the present invention and other optimization algorithms
| Experimental data Different Optimization method | The present invention | NSGAII | IBEA | MOEA/D | MOPSO | MOEA/D-EGO | K-RVEA | PAREGO | CSEA |
| First time data | 0.9742 | 0.9547 | 0.9559 | 0.9578 | 0.9566 | 0.9657 | 0.9584 | 0.9665 | 0.9676 |
| Second of data | 0.9705 | 0.9658 | 0.9659 | 0.9653 | 0.9642 | 0.9682 | 0.9587 | 0.9677 | 0.9665 |
As seen from Table 2, the MOGAN algorithm in the present invention is bigger compared to other algorithms HV value, this illustrates that the present invention is better than
Other existing optimization algorithms also demonstrate the Multipurpose Optimal Method for fighting network the present invention is based on production and are directed to kafka
The validity and reasonability of performance optimization problem.
Embodiment 7
With embodiment 1-6, this example is one and is applied to solve Multipurpose Optimal Method based on production confrontation network
The application example of the multi-objective optimization question of ZDT1.
Application example: ZDT1 minimum function value f1 and f2 are found using the present invention
ZDT is a kind of test function, wherein contain multi-objective optimization question, share ZDT1, ZDT2 ..., ZDT6 six
Function describes entire optimization process and as a result, the formula of ZDT1 is as follows in this example by taking ZDT1 as an example:
f1=x1
0≤xi≤ 1, i=1 ..., n
There are the system variable between 30 0-1, the system variable x in this example in this example1, x2... ..., x30It indicates, In
In this example, x1, x2... ..., x30Namely independent variable has an impact optimization aim, i.e. p=30, and optimization aim is ZDT1's
Two functional values f1 and f2, that is, dependent variable, i.e. q=2.Optimization aim of the invention be f1 and f2 are minimized, but
It is that cannot be directly changed f1 and f2, f1 and f2 can only be changed by changing independent variable x, realize the minimum to f1 and f2.This
Equally set maximum evaluation number E in example is 300, determines the sample that initial sample is tieed up in this example for 100 group 32.
By the identical step of embodiment 6, after carrying out constantly training optimization to production confrontation network, available fixation
The generation sample (generating sample only includes 30 independents variable, does not include f1 and f2) of 30 dimensions of number, by the formula of ZDT1 into
The f1 and f2 for generating sample is calculated in row, draws the function curve of optimization sample f1 and f2, compares of the invention with other
The functional value size of algorithm compares, as a result if Fig. 4, Fig. 4 are the present invention and other optimization algorithms for ZDT1 function
The curve graph of optimum results, the abscissa of Fig. 4 are the values of f1, and ordinate is the value of f2, share MOGAN and k-rvea of the present invention,
Moead, nsga, mopso and ibea etc. totally 6 optimization algorithms as a result, wherein curve of the invention is no any symbol
Solid line.
Since the target in this example is to keep f1 and f2 as small as possible, and from fig. 4, it can be seen that under ZDT1 function, this hair
Other comparison algorithms of 2.3-4.3 range internal ratio want low to curve representated by MOGAN in bright within the scope of 0-0.5, on f2 on f1
Very much, this illustrates that MOGAN method of the present invention is better than other existing optimization algorithms, also demonstrates that the present invention is based on production confrontation
The validity and reasonability of the Multipurpose Optimal Method of network.
Embodiment 8
With embodiment 1-7, this example is one and is applied to solve Multipurpose Optimal Method based on production confrontation network
The application example of the multi-objective optimization question of ZDT6.
Application example: ZDT6 minimum function value f1 and f2 are found using the present invention
ZDT is the common test function of multi-objective optimization question, share ZDT1, ZDT2 ..., six functions of ZDT6, In
In this example by taking ZD6 as an example, entire optimization process is described and as a result, the formula of ZDT6 is as follows:
0≤xi≤ 1, i=1 ..., n
There are the system variable between 30 0-1, the system variable x in this example in this example1, x2... ..., x30It indicates, In
In this example, x1, x2... ..., x30Namely independent variable has an impact optimization aim, i.e. p=30, and optimization aim is ZDT6's
Two functional values f1 and f2, that is, dependent variable, i.e. q=2.Optimization aim of the invention be f1 and f2 are minimized, but
It is that cannot be directly changed f1 and f2, f1 and f2 can only be changed by changing independent variable x, realize the minimum to f1 and f2.This
Equally set maximum evaluation number E in example is 300, determines the sample that initial sample is tieed up in this example for 100 group 32.
By the identical step of embodiment 6, after carrying out constantly training optimization to production confrontation network, available fixation
The generation sample (generating sample only includes 30 independents variable, does not include f1 and f2) of 30 dimensions of number, by the formula of ZDT6 into
The f1 and f2 for generating sample is calculated in row, draws the function curve of optimization sample f1 and f2, compares of the invention with other
The functional value size of algorithm compares, as a result if Fig. 5, Fig. 5 are the present invention and other optimization algorithms for ZDT1 function
The curve graph of optimum results, the abscissa of Fig. 5 are the values of f1, and ordinate is the value of f2, share MOGAN and k-rvea of the present invention,
Moead, nsga, mopso and ibea etc. totally 6 optimization algorithms as a result, wherein curve of the invention is no any symbol
Solid line.
Since the target in this example is to keep f1 and f2 as small as possible, and from fig. 5, it can be seen that under ZDT6 function, this hair
Other comparison algorithms of curve ratio representated by bright MOGAN want low, this illustrates that MOGAN method of the present invention is better than other existing optimizations
Algorithm, also demonstrate the present invention is based on production confrontation network Multipurpose Optimal Method validity and reasonability.
Embodiment 9
Multipurpose Optimal Method based on production confrontation network with embodiment 1-8,
Application example: the maximum detection rates of image detecting method and maximum detection accuracy are optimized using the present invention
Specific implementation step is same as Example 6, and in image detection, improving detection rates and improving detection accuracy is two
A conflicting target, i.e. optimization aim q=2, when improving detection rates, detection accuracy will be reduced, and improve detection essence
When spending, detection rates will be reduced.After optimization of the invention, it is last it is available about recognition rate and accuracy of identification this
Multiple optimal solutions of two targets therefrom select a detection rates and all very high combination of detection accuracy, can be used for specific
Practical application in.
In brief, a kind of Multipurpose Optimal Method (MOGAN) based on production confrontation network disclosed by the invention, it is main
It is high to solve prior art time cost, training difficulty is excessive, and training network is easy the problems such as collapsing.Its implementation: (1) it obtains
Take initial sample；(2) production for constructing multiple-objection optimization fights network；(3) training set is selected from initial sample；(4) it instructs
The production for practicing the multiple-objection optimization of building fights network；(5) judgement evaluation number；(6) it obtains intersecting result set；(7) it obtains
Make a variation result set；(8) evaluation obtains new initial sample；(9) effect of optimization is verified.Present invention reduces time costs, improve
Network robustness and stability, effect of optimization is obvious, can be used for the resource allocation of multiple targets, the production scheduling of multiple products
With the optimization etc. of the multiple performances of various software systems.
Claims (4)
1. a kind of Multipurpose Optimal Method based on production confrontation network, which comprises the steps of:
(1) initial sample is obtained:
(1a) selects optimization aim: for multi-objective optimization question present in a certain system, selecting determining needs to optimize more
A target, it is assumed that have q optimization aim, there is p system variable to generate shadow to this q optimization aim in all system variables
It rings, each system variable has the data set and value range of oneself；
The maximum evaluation number of (1b) setting: each group of p system variable is defined as one group of independent variable, is obtained by one group of independent variable
The value of q optimization aim is defined as dependent variable, and primary evaluation, which just refers to the process of, obtains corresponding dependent variable by one group of independent variable；It is excellent
Evaluation number is set during changing as e, maximum evaluation number is E, and evaluation number e is initialized as zero；
(1c) determines initial sample size: determining that initializaing variable quantity is m group according to maximum evaluation number E, guarantees optimization process
It can be completed in maximum evaluation number E；
(1d) stochastical sampling obtains initial sample: carrying out stochastical sampling to each system variable for being related to optimization aim, obtains m
The initializaing variable of group p dimension evaluates the superiority and inferiority for measuring this m group initializaing variable with the operation method of multiple target place system, each
Group initializaing variable obtains the value of corresponding q optimization aim, common group of value of the initializaing variable of every group of stochastical sampling and optimization aim
The initial sample tieed up at one group of p+q；Initializaing variable sampling is evaluated and is combined with optimization aim once, and evaluation number e adds 1, traversal
Whole m group initializaing variables, the final number e that evaluates add m, obtain the initial sample of m p+q dimension；
(2) construct multiple-objection optimization production fight network: it include generate network G and differentiate network D, generate network G with
Differentiate that network D is all made of three layers of full Connection Neural Network, the two is confronted with each other training, and promotion is continued to optimize each other, is built into more
The production of objective optimization fights network；
(3) select training set from initial sample: all p+q dimensions are chosen in the definition solved according to Pareto from initial sample
Pareto solution, then remove the value of q optimization aim in each group of Pareto solution, by these removal optimization target values p dimension
Pareto solution be used as training set；
(4) production of the multiple-objection optimization of training building fights network:
(4a) determines training sample: randomly choosing half from training set as training sample, then counts to training sample
It is pre-processed according to standardization, obtains pretreated training sample x；
(4b) inputs training sample into MOGAN: pretreated training sample x being input to the differentiation of production confrontation network
In network D；
(4c) generates sample: in the production confrontation network of multiple-objection optimization, using network G is generated, generating the generation sample of p dimension
This z；
(4d) obtains differentiating result: will generate sample z and training sample x input and differentiates network D, output differentiates result；
(4e) training generates network G and differentiates network D: according to differentiation as a result, fixed generate network G, instructing to differentiation network D
Practice, continue to optimize, is from training sample or to carry out self-generating network G until differentiating that network D can be accurately judged to a sample
The sample of generation；Network D is differentiated according to differentiating as a result, fixing, and is trained, is continued to optimize again to network G is generated, until differentiating
Network D cannot judge a sample be from training sample or come self-generating network G generation sample；
It obtains generating sample set after the training of (4f) successive ignition: executing step (4a)-(4e), complete once to production pair
The training of anti-network；Judge whether to reach the number of iterations, if not reaching the number of iterations of design, repeat step (4a)-
(4e) continues to train；If reaching the number of iterations of design, m is generated using network G is generated1Group p ties up initializaing variable, and
This m is measured in evaluation1The superiority and inferiority of group sample, calculates the value of q dimension optimization aim, and evaluation number e adds m1, then m1Group p dimension is initial to be become
Amount and q dimension optimization aim merge to obtain m1The generation sample set of group p+q dimension；
(5) judgement evaluation number: judging evaluation number, if evaluation number e reaches maximum evaluation number E at this time, will walk
Suddenly then the generation sample set that (4f) is obtained executes step (9), further verifies effect of optimization as final result collection, otherwise,
It executes step (6)；
(6) it obtains intersecting result set: by m1The generation sample set and training sample x of group p+q dimension merge, from the sample after merging
This concentration selects all p+q dimension Pareto solutions, and removal Pareto solves the value of q optimization aim, obtains p dimension initializaing variable
Pareto solution carries out simulation binary system crossover operation to the Pareto solution of these p dimension initializaing variable, obtains m2Group p dimension is initial to be become
The intersection result set of amount；
(7) variation result set is obtained: to m2The intersection result set of group p dimension initializaing variable carries out multinomial mutation operation, obtains m3Group
The variation result set of p dimension initializaing variable；
(8) evaluation obtains new initial sample: the operation method of system measures this m to evaluate where multiple target3Group p dimension is initial
The superiority and inferiority of variable, each group of initializaing variable obtain the value of corresponding q optimization aim, every group of p initializaing variable and q optimization aim
Value collectively constitute the initial sample of one group of new p+q dimension, the every evaluation of initializaing variable simultaneously combine once with optimization aim, is evaluated secondary
Number e adds 1, traverses whole m3The initializaing variable of group p dimension, the final number e that evaluates add m3, obtain m3The new initial sample of group p+q dimension
This；Carry out step (3)~step (5) again；
(9) effect of optimization is verified:
(9a) has optimization algorithm with other and optimizes to the selected p optimized variable of step (1), obtains corresponding control
Result set；
(9b) passes through the results of comparison collection for comparing final result collection and other optimization algorithms of the invention, to effect of optimization of the present invention
It is verified.
2. the Multipurpose Optimal Method according to claim 1 based on production confrontation network, which is characterized in that step
(2) production of the building multiple-objection optimization described in fights network, wherein generating network G, is one and includes input layer, hidden layer
With three layers of fully-connected network of output layer, which includes 5 nodes, and each node is the random number in [- 1,1] range；
The hidden layer has 128 nodes, and has weight relationship between each node and input layer, and initialization weight is in [- 1,1] range
Random number；The output layer contains p node, and p is the quantity of system variable, and each node contains activation primitive relu；
Differentiation network D therein is three layers of fully-connected network comprising input layer, hidden layer and output layer, the input layer
Comprising p node, p is the quantity of system variable；The hidden layer has 128 nodes, and has the right between each node and input layer
Series of fortified passes system, initialization weight is also the random number in [- 1,1] range, and each node contains activation primitive sigmoid；This is defeated
Layer contains 1 node out, indicates the probability for differentiating network D input sample authenticity, and each node contains activation primitive tanh.
3. the Multipurpose Optimal Method according to claim 1 based on production confrontation network, which is characterized in that step
(6) simulation binary system crossover operation is carried out to generation sample set described in, obtains m2The intersection result of group p dimension initializaing variable
Two neighboring sample is successively carried out simulation binary system crossover operation in sequence specifically in generating sample set by collection, first right
Two samples carry out probabilistic determination, if probability is less than the crossing-over rate of setting, intersect generating two new solutions, specific to intersect
Process is carried out according to the formula that such as Imitating binary system intersects:
Wherein,It is two new samples after the two neighboring sample cross of jth group, x1j(t) x2jIt (t) is to intersect
The preceding two neighboring sample of jth group, t indicate the t generation in the genetic algorithm that simulation binary system intersects, and j indicates crossover operation
Jth group, γjIt is the crossing-over rate of jth group,
ujFor random number, and uj∈ U (0,1), η are profile exponent, η > 0.
4. the Multipurpose Optimal Method according to claim 1 based on production confrontation network, which is characterized in that step
(7) multinomial mutation operation is carried out to generation sample set described in, obtains m3The variation result set of group p dimension initializaing variable, tool
Body is the mutation operation carried out in genetic algorithm to each generation sample generated in sample set, first carries out probability to each sample
Judgement carries out the random variation of small part, specific change procedure and variation to sample if probability is less than the aberration rate of setting
Operator is related, then generates the variation sample of p dimension initializaing variable, realizes the mutation operator form of multinomial mutation operation are as follows:
v'k=vk+δ·(uk-lk), wherein
In formula, vkIndicate father's individual, v'kIndicate son individual, ukIndicate the upper bound of p system variable value range, lkIndicate p
The lower bound of a system variable value range, δ1=(vk-lk)/(uk-lk), δ1=(uk-vk)/(uk-lk), the variation of k representative polynomial
Genetic algorithm in kth generation, u is the random number in [0, a 1] section,ηmIt is profile exponent.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910688044.0A CN110533221A (en) | 2019-07-29 | 2019-07-29 | Multipurpose Optimal Method based on production confrontation network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910688044.0A CN110533221A (en) | 2019-07-29 | 2019-07-29 | Multipurpose Optimal Method based on production confrontation network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110533221A true CN110533221A (en) | 2019-12-03 |
Family
ID=68661960
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910688044.0A Pending CN110533221A (en) | 2019-07-29 | 2019-07-29 | Multipurpose Optimal Method based on production confrontation network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110533221A (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111481935A (en) * | 2020-04-08 | 2020-08-04 | 网易（杭州）网络有限公司 | Configuration method, device, equipment and medium for AI models of games with different styles |
| CN112488486A (en) * | 2020-11-25 | 2021-03-12 | 吉林大学 | Multi-criterion decision method based on zero sum game |
| CN112508093A (en) * | 2020-12-03 | 2021-03-16 | 北京百度网讯科技有限公司 | Self-training method and device, electronic equipment and readable storage medium |
| WO2021128805A1 (en) * | 2019-12-24 | 2021-07-01 | 浙江大学 | Wireless network resource allocation method employing generative adversarial reinforcement learning |
| CN113434459A (en) * | 2021-06-30 | 2021-09-24 | 电子科技大学 | Network-on-chip task mapping method based on generation of countermeasure network |
-
2019
- 2019-07-29 CN CN201910688044.0A patent/CN110533221A/en active Pending
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2021128805A1 (en) * | 2019-12-24 | 2021-07-01 | 浙江大学 | Wireless network resource allocation method employing generative adversarial reinforcement learning |
| CN111481935A (en) * | 2020-04-08 | 2020-08-04 | 网易（杭州）网络有限公司 | Configuration method, device, equipment and medium for AI models of games with different styles |
| CN111481935B (en) * | 2020-04-08 | 2023-04-18 | 网易（杭州）网络有限公司 | Configuration method, device, equipment and medium for AI models of games with different styles |
| CN112488486A (en) * | 2020-11-25 | 2021-03-12 | 吉林大学 | Multi-criterion decision method based on zero sum game |
| CN112488486B (en) * | 2020-11-25 | 2022-04-15 | 吉林大学 | Multi-criterion decision method based on zero sum game |
| CN112508093A (en) * | 2020-12-03 | 2021-03-16 | 北京百度网讯科技有限公司 | Self-training method and device, electronic equipment and readable storage medium |
| CN112508093B (en) * | 2020-12-03 | 2022-01-28 | 北京百度网讯科技有限公司 | Self-training method and device, electronic equipment and readable storage medium |
| CN113434459A (en) * | 2021-06-30 | 2021-09-24 | 电子科技大学 | Network-on-chip task mapping method based on generation of countermeasure network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110533221A (en) | Multipurpose Optimal Method based on production confrontation network | |
| Cantú-Paz et al. | An empirical comparison of combinations of evolutionary algorithms and neural networks for classification problems | |
| Derrac et al. | IFS-CoCo: Instance and feature selection based on cooperative coevolution with nearest neighbor rule | |
| Tian et al. | A fast clustering based evolutionary algorithm for super-large-scale sparse multi-objective optimization | |
| CN115481727A (en) | Intention recognition neural network generation and optimization method based on evolutionary computation | |
| Hu | Flow-based tolerance rough sets for pattern classification | |
| Pangilinan et al. | Pareto-optimality of oblique decision trees from evolutionary algorithms | |
| De Matteis et al. | A new approach to fuzzy random forest generation | |
| Yu et al. | Unsupervised euclidean distance attack on network embedding | |
| Farooq | Genetic algorithm technique in hybrid intelligent systems for pattern recognition | |
| Parvin et al. | A scalable method for improving the performance of classifiers in multiclass applications by pairwise classifiers and GA | |
| Siwik et al. | Elitist evolutionary multi-agent system in solving noisy multi-objective optimization problems | |
| CN112183598A (en) | Feature selection method based on genetic algorithm | |
| Hong et al. | Improving performance insensitivity of large-scale multiobjective optimization via Monte Carlo tree search | |
| CN113704570B (en) | Large-scale complex network community detection method based on self-supervision learning type evolution | |
| CN114819181A (en) | Multi-target federal learning evolution method based on improved NSGA-III | |
| Hu et al. | Apenas: An asynchronous parallel evolution based multi-objective neural architecture search | |
| Davarynejad et al. | Accelerating convergence towards the optimal pareto front | |
| Moyano et al. | Tree-shaped ensemble of multi-label classifiers using grammar-guided genetic programming | |
| Van Truong et al. | A Coevolutionary approach for classification problems: Preliminary results | |
| Khaparde et al. | A new distributed differential evolution algorithm | |
| Joldos et al. | A parallel evolutionary approach to community detection in complex networks | |
| Antonelli et al. | Multi-objective evolutionary design of fuzzy rule-based systems | |
| Harrison | Network similarity measures and automatic construction of graph models using genetic programming | |
| Berlanga et al. | A novel genetic cooperative-competitive fuzzy rule based learning method using genetic programming for high dimensional problems |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20191203 |