Multi-objective optimization method and system based on generative adversarial network

Abstract

The application provides a multi-objective optimization method and system based on a generative adversarial network, comprising: initializing a first generation solution set, initializing network structures and learning rates of a generator and a discriminator; generating a sub-solution set by using genetic operation or the generator; selecting a next generation solution from a union set of a current solution set and the generated sub-solution set; searching for a dominance pair between the next generation solution set and a discarded solution set, training the discriminator by using the dominance pair; and training the generator to predict a solution based on the trained discriminator. The application is based on the learning ability of the generative adversarial network, designs a search method of the dominance pair, trains the discriminator by using the dominance pair to learn a dominance relationship between non-dominated solutions and dominated solutions, and then uses the discriminator to guide the generator to predict better solutions on the basis of existing non-dominated solutions; and better Pareto optimal frontiers of multi-objective optimization problems and corresponding Pareto optimal solution sets are obtained through iteration.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, in particular, to a multi-objective optimization method and system based on generative adversarial network. BACKGROUND

[0002] In actual production and life, many decision-making, optimization and other problems need to consider multiple objectives, and these objectives often have a mutual antagonistic relationship, thus forming a multi-objective optimization problem. By solving the multi-objective optimization problem, the optimal Pareto front can be obtained, thus revealing the compromise relationship between multiple objectives and obtaining the Pareto solution corresponding to different compromises.

[0003] In order to solve the multi-objective optimization problem, a large number of multi-objective optimization methods have been proposed. One common method is to convert multiple objectives into a single objective by weighting, and then solve it by single-objective optimization method, but the setting of weight depends on prior knowledge. Evolutionary algorithms are also used to solve multi-objective problems, such as multi-objective evolutionary algorithm based on decomposition (MOEA / D) and non-dominated sorting genetic algorithm III (NSGA III). In addition, multi-objective evolutionary algorithms based on decision variable clustering, dominance relationship classification and other learning models have also been proposed, but due to the increase in computational complexity and training data requirements, the performance of these algorithms decreases with the increase in decision variables. Therefore, the present application proposes a multi-objective optimization method based on the powerful learning ability of generative adversarial network.

[0004] Patent document CN110533221A (application number: CN201910688044.0) discloses a multi-objective optimization method based on generative adversarial network: randomly sampling to obtain initial samples; selecting Pareto solutions as training set; randomly selecting half of the training set as training samples for preprocessing, constructing generative adversarial network, and obtaining generated samples after iterative training; determining whether further training and optimization are needed according to the evaluation number; comparing the generated samples with the results of other comparative algorithms to evaluate the advantages and disadvantages of the algorithm. However, this invention does not design a search method for dominated pairs. SUMMARY

[0005] In view of the defects in the prior art, the purpose of the present application is to provide a multi-objective optimization method and system based on generative adversarial network.

[0006] According to the multi-objective optimization method based on generative adversarial network provided by the present application, the following steps are included:

[0007] Step S1: the base station collects system information of each device, initializes a first generation solution set according to multiple targets to be considered, initializes network structure and learning rate of the generator and the discriminator;

[0008] Step S2: the base station uses its own computing resources to generate a sub-solution set by using genetic operation or a generator;

[0009] Step S3: the base station selects a next generation solution from a union set of the current solution set and the generated sub-solution set according to the collected system information;

[0010] Step S4: search for a dominant pair between the next generation solution set and the discarded solution set, and train the discriminator by using the dominant pair;

[0011] Step S5: train the generator to predict a solution based on the trained discriminator, and perform resource block and user transmit power allocation.

[0012] Preferably, in the step S1:

[0013] The first generation solution set P1 is initialized according to a feasible region of an optimization variable; the network structure and the number of nodes of the generator G and the discriminator D in the generative adversarial network are set according to the dimension of the optimization variable, and the learning rate l of the input generator and the discriminator G , l D .

[0014] Preferably, in the step S2:

[0015] The sub-solution set Q j is generated by using genetic operation or the generator G according to the current solution set P j ; when the sub-solution set is generated by using the generator, the mean and the variance of the current solution set P j are calculated, so as to construct a random distribution, and then a random variable is randomly extracted from the constructed distribution as input noise of the generator;

[0016] The genetic operation includes crossover and mutation;

[0017] The genetic operation and the generator are mixed to avoid mode collapse of the generator, and the mixed use includes randomly selecting to use the genetic operation or the generator to generate the next generation solution set with the same probability.

[0018] Preferably, in the step S3:

[0019] Part of the union set P j ∪ Q j of the current solution set P j and the sub-solution set Q j is selected as the next generation solution set P j+1; select based on the dominance relationship between solutions, the solution not dominated is preferentially retained, and the method of calculating the fitness function is used to select P j ∪Q j For each solution in P j ∪Q j , the number of solutions that can be dominated and the density of nearby solutions are calculated, and the sum of the number of solutions that can be dominated and the density of nearby solutions is taken as the fitness of a solution, and the solutions with the maximum fitness are selected from P j+1 .

[0020] Preferably, in the step S4:

[0021] Each solution in the discarded solution set (P j ∪Q j ) / P j+1 can be dominated by some solutions in P j+1 , P j+1 is the non-dominated solution set, and (P j ∪Q j ) / P j+1 is the dominated solution set, and the dominance pairs are searched between the next generation solution set P j+1 and the discarded solution set (P j ∪Q j ) / P j+1 , each dominance pair consists of two solutions, one of which is dominated by the other;

[0022] The discriminator D is trained using the searched dominance pairs, the training of the discriminator adopts a supervised manner, the input of the discriminator contains two parts, and the output generated by the exchange of the two parts is different, the two solutions in the dominance pair are input into the discriminator according to the dominance relationship, and the discriminator learns the dominance relationship between the non-dominated solution and the dominated solution;

[0023] In the step S5:

[0024] The generator G is trained based on the trained discriminator, the generator adopts an unsupervised training manner, and the generator predicts solutions based on existing non-dominated solutions by using the dominance relationship learned by the discriminator.

[0025] According to the multi-objective optimization system based on the generative adversarial network provided by the application, the base station collects system information of each device, initializes a first generation solution set according to a plurality of targets to be considered, and initializes network structures and learning rates of the generator and the discriminator.

[0026] Module M1: The base station collects system information of each device, initializes a first generation solution set according to a plurality of targets to be considered, and initializes network structures and learning rates of the generator and the discriminator.

[0027] Module M2: The base station generates a sub-solution set by using its own computing resources and genetic operation or generator.

[0028] Module M3: The base station selects the next generation solution from the union of the current solution set and the generated sub-solution set based on the collected system information;

[0029] Module M4: Search for dominant pairs between the next generation solution set and the discarded solution set, and use the dominant pairs to train the discriminator;

[0030] Module M5: Based on the training discriminator, the generator is trained to predict the solution and allocate resource blocks and user transmit power.

[0031] Preferably, in the module M1:

[0032] Initialize the first generation solution set P1 according to the feasible domain of the optimization variable; set the network structure and number of nodes of the generator G and discriminator D in the generative adversarial network according to the dimension of the optimization variable, and input the learning rate l of the generator and discriminator G 、l D .

[0033] Preferably, in the module M2:

[0034] According to the current solution set P j , using genetic operations or generator G to generate sub-solution set Q j ; When using the generator to generate a sub-solution set, calculate the current solution set P j The mean and variance of are used to construct a random distribution, and then random variables are arbitrarily sampled from the constructed distribution as the input noise of the generator;

[0035] Among them, genetic operations include crossover and mutation;

[0036] Genetic operations are mixed with generators to avoid mode collapse of the generator. The mixed use method includes randomly selecting genetic operations or generators with the same probability to generate the next generation solution set.

[0037] Preferably, in the module M3:

[0038] From the current solution set P j and sub-solution set Q j The union P of j ∪Q j Select some of them as the next generation solution set P j+1 ; The selection is based on the dominance relationship between solutions, and the undominated solutions are retained first. The fitness function is calculated. For P j ∪Q j For each solution in P, calculate the number of solutions that can be dominated and the density of nearby solutions. The sum of the number of solutions that can be dominated and the density of nearby solutions is taken as the fitness of a solution. j ∪Q j Select the part of the solution with the largest fitness as Pj+1 .

[0039] Preferably, in the module M4:

[0040] The discarded solution set (P j ∪Q j ) / P j+1 Every solution in P j+1 Some solutions in P dominate, j+1 is a non-dominated solution set, and (P j ∪Q j ) / P j+1 is the dominated solution set, in the next generation solution set P j+1 and the discarded solution set (P j ∪Q j ) / P j+1 Search for dominating pairs between them, each dominating pair consists of two solutions, one of which is dominated by the other;

[0041] The discriminator D is trained using the searched dominant pairs. The discriminator is trained in a supervised manner. Its input consists of two parts, and the output produced by swapping the two parts is different. The two solutions in the dominant pair are input into the discriminator according to the dominance relationship, allowing the discriminator to learn the dominance relationship between the non-dominated solution and the dominated solution.

[0042] In the module M5:

[0043] The generator G is trained based on the trained discriminator. The generator adopts an unsupervised training method and uses the dominance relationship learned by the discriminator to enable the generator to predict solutions based on existing non-dominated solutions.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] 1. Based on the learning ability of generative adversarial networks, this paper designs a search method for dominant pairs. The dominant pairs are used to train a discriminator to learn the dominance relationship between non-dominated solutions and dominated solutions. The discriminator is then used to guide the generator to predict better solutions based on existing non-dominated solutions.

[0046] 2. The present invention can obtain a better Pareto optimal frontier and a corresponding Pareto optimal solution set for a multi-objective optimization problem through iteration. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments with reference to the following drawings:

[0048] Figure 1 Search for dominant pairs;

[0049] Figure 2training of the discriminator;

[0050] Figure 3 training of the generator;

[0051] Figure 4 overall flow of the algorithm. DETAILED DESCRIPTION

[0052] The application will be described in detail below with specific embodiments. The following examples will help those skilled in the art to further understand the application, but do not limit the application in any form. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the application. These are within the scope of the application.

[0053] Example 1

[0054] According to the multi-objective optimization method based on a generative adversarial network provided by the application, as shown in Figures 1-4 , comprising:

[0055] Step S1: The base station collects system information of each device, initializes a first generation solution set according to multiple objectives to be considered, and initializes network structure and learning rate of the generator and the discriminator;

[0056] Preferably, in the step S1:

[0057] The first generation solution set P1 is initialized according to the feasible region of the optimization variable; the network structure and the number of nodes of the generator G and the discriminator D in the generative adversarial network are set according to the dimension of the optimization variable, and the learning rate l of the input generator and discriminator G , l D .

[0058] Step S2: The base station uses its own computing resources to generate a sub-solution set by genetic operation or generator;

[0059] Preferably, in the step S2:

[0060] According to the current solution set P j , a sub-solution set Q j is generated by genetic operation or generator G; when the sub-solution set is generated by the generator, the mean and variance of the current solution set P j are calculated, so as to construct a random distribution, and then a random variable is randomly extracted from the constructed distribution as input noise of the generator;

[0061] The genetic operation includes crossover and mutation.

[0062] Genetic operations are mixed with generators to avoid mode collapse of the generator. The mixed use method includes randomly selecting genetic operations or generators with the same probability to generate the next generation solution set.

[0063] Step S3: The base station selects the next generation solution from the union of the current solution set and the generated sub-solution set based on the collected system information;

[0064] Preferably, in step S3:

[0065] From the current solution set P j and sub-solution set Q j The union P of j ∪Q j Select some of them as the next generation solution set P j+1 ; The selection is based on the dominance relationship between solutions, and the undominated solutions are retained first. The fitness function is calculated. For P j ∪Q j For each solution in P, calculate the number of solutions that can be dominated and the density of nearby solutions. The sum of the number of solutions that can be dominated and the density of nearby solutions is taken as the fitness of a solution. j ∪Q j Select the part of the solution with the largest fitness as P j+1 .

[0066] Step S4: Search for dominant pairs between the next generation solution set and the discarded solution set, and use the dominant pairs to train the discriminator;

[0067] Preferably, in step S4:

[0068] The discarded solution set (P j ∪Q j ) / P j+1 Every solution in P j+1 Some solutions in P dominate, j+1 is a non-dominated solution set, and (P j ∪Q j ) / P j+1 is the dominated solution set, in the next generation solution set P j+1 and the discarded solution set (P j ∪Q j ) / P j+1 Search for dominating pairs between them, each dominating pair consists of two solutions, one of which is dominated by the other;

[0069] The discriminator D is trained using the searched dominant pairs. The training of the discriminator adopts a supervised approach. Its input consists of two parts, and the output produced by swapping the two parts is different. The two solutions in the dominant pair are input into the discriminator according to the dominance relationship, allowing the discriminator to learn the dominance relationship between the non-dominated solution and the dominated solution.

[0070] Step S5: training the generator based on the trained discriminator to predict the solution and allocate the resource blocks and the user transmit power.

[0071] In the step S5:

[0072] The generator G is trained based on the trained discriminator, and the generator adopts an unsupervised training manner to use the dominance relationship learned by the discriminator to make the generator predict the solution based on the existing non-dominated solution.

[0073] Example 2

[0074] Embodiment 2 is a preferred example of Embodiment 1, which more specifically illustrates the present application.

[0075] Those skilled in the art can understand the method for multi-objective optimization based on the generative adversarial network provided by the present application as a specific embodiment of a multi-objective optimization system based on the generative adversarial network, that is, the multi-objective optimization system based on the generative adversarial network can be realized by executing the step flow of the method for multi-objective optimization based on the generative adversarial network.

[0076] According to the multi-objective optimization system based on the generative adversarial network provided by the present application, comprising:

[0077] Module M1: the base station collects system information of each device, initializes a first generation solution set according to multiple targets to be considered, and initializes network structure and learning rate of the generator and the discriminator;

[0078] Preferably, in the module M1:

[0079] The first generation solution set P1 is initialized according to the feasible region of the optimization variable; the network structure and the node number of the generator G and the discriminator D in the generative adversarial network are set according to the dimension of the optimization variable, and the learning rate l of the input generator and the discriminator G , l D .

[0080] Module M2: the base station uses its own computing resources to generate a sub-solution set by genetic operation or generator;

[0081] Preferably, in the module M2:

[0082] According to the current solution set P j , a sub-solution set Q j is generated by genetic operation or generator G; when the sub-solution set is generated by the generator, the mean and variance of the current solution set P j are calculated to construct a random distribution, and then a random variable is randomly extracted from the constructed distribution as the input noise of the generator;

[0083] Among them, genetic operations include crossover and mutation;

[0084] Genetic operations are mixed with generators to avoid mode collapse of the generator. The mixed use method includes randomly selecting genetic operations or generators with the same probability to generate the next generation solution set.

[0085] Module M3: The base station selects the next generation solution from the union of the current solution set and the generated sub-solution set based on the collected system information;

[0086] Preferably, in the module M3:

[0087] From the current solution set P j and sub-solution set Q j The union P of j ∪Q j Select some of them as the next generation solution set P j+1 ; The selection is based on the dominance relationship between solutions, and the undominated solutions are retained first. The fitness function is calculated. For P j ∪Q j For each solution in P, calculate the number of solutions that can be dominated and the density of nearby solutions. The sum of the number of solutions that can be dominated and the density of nearby solutions is taken as the fitness of a solution. j ∪Q j Select the part of the solution with the largest fitness as P j+1 .

[0088] Module M4: Search for dominant pairs between the next generation solution set and the discarded solution set, and use the dominant pairs to train the discriminator;

[0089] Preferably, in the module M4:

[0090] The discarded solution set (P j ∪Q j ) / P j+1 Every solution in P j+1 Some solutions in P dominate, j+1 is a non-dominated solution set, and (P j ∪Q j ) / P j+1 is the dominated solution set, in the next generation solution set P j+1 and the discarded solution set (P j ∪Q j ) / P j+1 Search for dominating pairs between them, each dominating pair consists of two solutions, one of which is dominated by the other;

[0091] The discriminator D is trained by using the searched domination pair, the training of the discriminator is in a supervised manner, the input of the discriminator includes two parts, and the outputs generated by interchanging the two parts are different, the two solutions in the domination pair are input into the discriminator according to the domination relationship, and the discriminator learns the domination relationship between the non-dominated solution and the dominated solution;

[0092] Module M5: training the generator to predict the solution based on the trained discriminator, and allocating the resource block and the user transmission power.

[0093] In the module M5:

[0094] The generator G is trained based on the trained discriminator, and the generator is trained in an unsupervised manner, and the generator learns the domination relationship learned by the discriminator to predict the solution based on the existing non-dominated solution.

[0095] Example 3

[0096] Embodiment 3 is a preferred example of Embodiment 1, which more specifically illustrates the present application.

[0097] The flow and algorithm based on the generative adversarial network for solving the multi-objective optimization problem in the present application.

[0098] For a general form of multi-objective optimization problem:

[0099] min x F(x)=[f1(x),f2(x),...,f M (x)]

[0100] s.t.x∈X

[0101] Wherein, x is the optimization variable of the problem, and the value range of x is in the feasible region X. The optimization target is to minimize M objective functions f1(x), f2(x),..., f M (x). The present application is not limited to minimizing the objective function, and the objective function that needs to be maximized can be changed to a minimum objective function by adding a negative sign.

[0102] The present application aims to design a multi-objective optimization method based on a generative adversarial network. Based on the learning ability of GAN (Generative Adversarial Network), this method can fully explore the optimal Pareto solution set of the multi-objective optimization problem and obtain the corresponding optimal Pareto front.

[0103] The technical scheme adopted by the present application mainly includes the following steps:

[0104] Step 1: initializing the first generation solution set P1 according to the feasible region of the optimization variable;

[0105] Step 2: generate the network structure and node number of the generator G and the discriminator D in the generative adversarial network according to the dimension setting of the optimization variable. Input the learning rate l of the generator and the discriminator G , D ;

[0106] Step 3: generate a sub-solution set Q j from the current solution set P j by using genetic operation (crossing and mutation) or the generator G. When the generator is used to generate the sub-solution set, the mean and variance of the current solution set P j are calculated to construct a random distribution (for example, Gaussian distribution), and then a random variable is randomly extracted from the constructed distribution as the input noise of the generator;

[0107] Step 4: select a part of solutions with better performance from the union set P j ∪Q j of the current solution set P j and the sub-solution set Q j as the next generation solution set P j+1 . In this step, the selection of solutions is based on the dominance relationship between solutions, and the solutions that are not dominated are preferentially retained. The present application does not limit the method of selecting solutions in this step, and the method of calculating the fitness function can be used, that is, for each solution in P j ∪Q j , the number of solutions that can be dominated by it and the density of solutions near it are calculated, and the sum of the two is taken as the fitness of a solution, and then a part of solutions with the maximum fitness is selected from P j ∪Q j as P j+1 ;

[0108] Step 5: each solution in the discarded solution set (P j ∪Q j ) / P j+1 can be dominated by some solutions in P j+1 , so P j+1 is a non-dominated solution set, and (P j ∪Q j ) / P j+1 is a dominated solution set. Search for a dominance pair between the next generation solution set P j+1 and the discarded solution set (P j ∪Q j ) / P j+1 , each dominance pair consists of two solutions, one of which is dominated by the other;

[0109] Step 6: Train the discriminator D with the dominance pairs found in step 5. The training of the discriminator is supervised, and its input contains two parts, and the output generated by the exchange of the two parts is different. The two solutions in the dominance pair are input into the discriminator according to the dominance relationship, so that the discriminator learns the dominance relationship between the non-dominated solution and the dominated solution;

[0110] Step 7: Train the generator G based on the discriminator trained in step 6. The generator is trained in an unsupervised manner, and uses the dominance relationship learned by the discriminator to make the generator predict better solutions based on existing non-dominated solutions.

[0111] Steps 3 to 7 are repeated until the algorithm termination condition is met, which can be whether the non-dominated solution set changes or not.

[0112] Preferably, in step 3, genetic operations are mixed with the generator to avoid pattern collapse of the generator. The mixed use includes but is not limited to randomly selecting genetic operations or generators to generate the next generation of solution sets with the same probability. Preferably, in step 4, the selection of solutions depends on the dominance relationship between solutions. For the general form of multi-objective optimization problem in example 3, the definition of the dominance relationship is: for two unequal solutions x, y ∈ X, if f m (x)≤f m (y) for all m = 1, 2,..., M, then x dominates y.

[0113] Example 4

[0114] Example 4 is a preferred example of example 1 to more specifically illustrate the present application.

[0115] For example, a multi-objective optimization problem with two objectives:

[0116] min x f1(x)

[0117] min x f2(x)

[0118] s.t.x∈R n

[0119] Step 1: Initialize the first generation of solution set P1 according to the feasible region of the optimization variable;

[0120] Step 2: For the case where x is an n-dimensional vector, the generator G and the discriminator D in the generative adversarial network can use a fully connected neural network. In particular, the number of input and output nodes of G is n, and the number of input nodes of D is 2n, and the number of output nodes is 1. The learning rate of the input generator and discriminator is l G ,l D ;

[0121] Step 3: According to the current solution set P j , randomly use genetic operations (crossover and mutation) or generator G to generate sub-solution set Q with the same probability j When using the generator to generate a sub-solution set, calculate the current solution set P j The mean and variance of are used to construct a Gaussian distribution, and then random variables are arbitrarily extracted from the constructed distribution as the input noise of the generator;

[0122] Step 4: Use the method of calculating the fitness function to select the solution set P from the current solution set P. j and sub-solution set Q j The union P of j ∪Q j Select half of the solutions with better performance as the next generation solution set P j+1 ;

[0123] Step 5: Solve the next generation of P j+1 and the discarded solution set (P j ∪Q j ) / P j+1 Search for a dominant pair between Figure 1 As shown. For each point in the non-dominated solution set The points contained in the intersection of its dominated neighborhood and the dominated solution set To improve training efficiency, we can select a portion of the dominated solutions closest to the reference line to form a dominant pair.

[0124] Step 6: Use the dominant pair found in step 5 to train the discriminator D. The discriminator adopts a supervised training method. Figure 2 As shown, assuming Complete domination Then it will dominate When input to the discriminator, the output The label of is 1. Conversely, the input is When The label is 0;

[0125] Step 7: After completing the training of the discriminator, fix its parameters and use the learned dominance relationship to train the generator G. Figure 3 As shown in Figure 1, the generator adopts an unsupervised training method. One input of the discriminator is the current non-dominated solution, and the other input is the solution generated by the generator. The label is set to 1, so that the generator can predict a better solution based on the existing non-dominated solution.

[0126] The above steps 3 to 7 are repeated until the change of the non-dominated solution set is less than the set threshold. The overall step flow is as follows Figure 4 shown.

[0127] Example 5:

[0128] Embodiment 5 is a preferred embodiment of Embodiment 1, to more specifically illustrate the present application.

[0129] The present application can be widely applied to multi-objective optimization problems in the fields of wireless communication, control theory, etc. For example, the allocation of resource blocks and transmission power in D2D communication will affect the energy efficiency (EE) and spectral efficiency (SE) of the communication system, and there is a trade-off relationship between EE and SE, so it can be modeled as a multi-objective optimization problem as follows.

[0130]

[0131]

[0132] s.t.ρ k,m ∈{0,1},

[0133]

[0134]

[0135] where ρ k,m =1 indicates that the mth resource block is allocated to the kth D2D user pair, and ρ k,m =0 is the opposite; p k represents the transmission power of the transmitter in the kth D2D user pair. We limit each D2D user pair to only occupy one resource block, while the maximum transmission power is constrained.

[0136] The present application can be used to solve the Pareto solution set of the above problem, and the specific steps are as follows:

[0137] Step 1: The base station collects system information such as channel state, maximum transmission power, available resource blocks, etc. of each device, so as to obtain the necessary information for calculating SE and EE;

[0138] Step 2: Use a fully connected neural network as the generator G and discriminator D in the generative adversarial network. Initialize the solution set, i.e. N groups of solutions, each group of solutions consisting of the decision variables ρ k,m ,p k of all users;

[0139] Step 3: The base station uses its own computing resources to generate a child solution set Q j from the current solution set P j according to the current solution set P j , with the same probability randomly using genetic operations (crossing and mutation) or generator G to generate a child solution set Q j ;

[0140] Step 4: The base station calculates the current solution set P based on the collected system information j and sub-solution set Q j Each solution of SE and EE corresponds to, and then according to the value of SE and EE from P j With Q j From the union of the two sets, half of the solutions with better performance are selected as the next generation solution set P j+1 Since the above problem is to maximize SE and EE simultaneously, the retained solutions often have larger EE under the same SE than the discarded solutions;

[0141] Step 5: Solve the next generation of P j+1 and the discarded solution set (P j ∪Q j ) / P j+1 The algorithm searches for dominant pairs between the two networks and uses the found dominant pairs to train the discriminator. After the discriminator is trained, its parameters are fixed and the learned dominance relationship is used to train the generator. The training of the generator and discriminator relies on the CPU computing resources of the base station.

[0142] Step 6: Repeat steps 2 to 5 above to obtain the Pareto curve of the trade-off relationship between SE and EE. Each point on the curve corresponds to a set of solutions, namely, the decision variables ρ of all users. k,m ,p k According to the service requirements of the communication system, a suitable compromise point is selected from the Pareto curve, and the base station can allocate resource blocks and user transmit power according to the solution corresponding to this point.

[0143] Those skilled in the art will appreciate that, in addition to implementing the system, device, and various modules provided by the present invention in purely computer-readable program code, it is entirely possible to implement the same program in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, embedded microcontrollers, and the like by logically programming the method steps. Therefore, the system, device, and various modules provided by the present invention can be considered a hardware component, and the modules included therein for implementing various programs can also be considered structures within the hardware component; the modules for implementing various functions can also be considered both software programs for implementing the method and structures within the hardware component.

[0144] The above describes specific embodiments of the present invention. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art may make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. The embodiments of this application and the features in the embodiments may be combined with each other in any manner unless there is a conflict.

Claims

1. A multi-objective optimization method based on a generative adversarial network, characterized in that, Comprising: Step S1: the base station collects system information of each device, initializes a first generation solution set according to multiple objectives to be considered, initializes network structure and learning rate of the generator and discriminator; Step S2: the base station uses its own computing resources to generate a sub-solution set by using genetic operation or the generator; Step S3: the base station selects a next generation solution from the union set of the current solution set and the generated sub-solution set according to the collected system information; Step S4: search for a dominant pair between the next generation solution set and the discarded solution set, and train the discriminator by using the dominant pair; Step S5: train the generator to predict solutions based on the trained discriminator, and allocate resource blocks and user transmit power.

2. The generative adversarial network-based multi-objective optimization method according to claim 1, characterized in that, In the step S1: Initialize the first generation solution set P1 according to the feasible region of the optimization variable; set the network structure and node number of the generator G and the discriminator D in the generative adversarial network according to the dimension of the optimization variable, and input the learning rate l of the generator and the discriminator G , D . 3.The multi-objective optimization method based on a generative adversarial network according to claim 1, wherein, In the step S2: According to the current solution set P j , a genetic operation or a generator G is used to generate a sub-solution set Q j ; when the generator is used to generate the sub-solution set, the mean and variance of the current solution set P j are calculated, so as to construct a random distribution, and then a random variable is randomly extracted from the constructed distribution as input noise of the generator; The genetic operation includes crossover and mutation; The genetic operation and the generator are used in a mixed manner to avoid pattern collapse of the generator, and the mixed use includes randomly selecting the genetic operation or the generator to generate the next generation solution set with the same probability. 4.The multi-objective optimization method based on a generative adversarial network according to claim 1, wherein, In the step S3: From the current solution set P j and sub-solution set Q j The union P of j ∪Q j Select some of them as the next generation solution set P j+1 ; The selection is based on the dominance relationship between solutions, and the undominated solutions are retained first. The fitness function is calculated. For P j ∪Q j For each solution in P, calculate the number of solutions that can be dominated and the density of nearby solutions. The sum of the number of solutions that can be dominated and the density of nearby solutions is taken as the fitness of a solution. j ∪Q j Select the part of the solution with the largest fitness as P j+1 .

5. The multi-objective optimization method based on the generative adversarial network according to claim 1, characterized in that: In the step S4: P j ∪Q j ) / P j+1 P j+1 P j+1 P j ∪Q j ) / P j+1 P j+1 P j ∪Q j ) / P j+1 P The discriminator D is trained by using the searched dominant pair, the training of the discriminator adopts a supervised manner, the input of the discriminator contains two parts, and the outputs generated by the interchange of the two parts are different, the two solutions in the dominant pair are input into the discriminator according to the dominance relationship, and the discriminator learns the dominance relationship between the non-dominated solution and the dominated solution; In the step S5: The generator G is trained based on the trained discriminator, the generator adopts an unsupervised training manner, and the generator learns the dominance relationship learned by the discriminator, so that the generator predicts solutions based on the existing non-dominated solutions.

6. A multi-objective optimization system based on a generative adversarial network, characterized by, Comprising: Module M1: the base station collects system information of each device, initializes a first generation solution set according to multiple objectives to be considered, initializes network structure and learning rate of the generator and discriminator; Module M2: the base station uses its own computing resources to generate a sub-solution set by using genetic operation or the generator; Module M3: the base station selects a next generation solution from the union set of the current solution set and the generated sub-solution set according to the collected system information; Module M4: search for a dominant pair between the next generation solution set and the discarded solution set, and train the discriminator by using the dominant pair; Module M5: train the generator to predict solutions based on the trained discriminator, and allocate resource blocks and user transmit power.

7. The generative adversarial network-based multi-objective optimization system of claim 6, wherein, In the module M1: Initialize the first generation solution set P1 according to the feasible region of the optimization variable; set the network structure and node number of the generator G and the discriminator D in the generative adversarial network according to the dimension of the optimization variable, and input the learning rate l of the generator and the discriminator G , D .

8. The generative adversarial network-based multi-objective optimization system of claim 6, wherein, In the module M2: According to the current solution set P j , a genetic operation or a generator G is used to generate a sub-solution set Q j ; when the generator is used to generate the sub-solution set, the mean and variance of the current solution set P j are calculated, so as to construct a random distribution, and then a random variable is randomly extracted from the constructed distribution as input noise of the generator; The genetic operation includes crossover and mutation; The genetic operation and the generator are used in a mixed manner to avoid pattern collapse of the generator, and the mixed use includes randomly selecting the genetic operation or the generator to generate the next generation solution set with the same probability.

9. The generative adversarial network-based multi-objective optimization system of claim 6, wherein, In the module M3: From the current solution set P j and sub-solution set Q j The union P of j ∪Q j Select some of them as the next generation solution set P j+1 ; The selection is based on the dominance relationship between solutions, and the undominated solutions are retained first. The fitness function is calculated. For P j ∪Q j For each solution in P, calculate the number of solutions that can be dominated and the density of nearby solutions. The sum of the number of solutions that can be dominated and the density of nearby solutions is taken as the fitness of a solution. j ∪Q j Select the part of the solution with the largest fitness as P j+1 .

10. The multi-objective optimization system based on the generative adversarial network according to claim 6, characterized in that: In the module M4: P j ∪Q j ) / P j+1 P j+1 ∪Q j+1 P j ∪Q j P j+1 ∪Q j+1 P j ∪Q j P j+1 ​ The discriminator D is trained by using the searched dominant pair, the training of the discriminator adopts a supervised manner, the input of the discriminator contains two parts, and the outputs generated by the interchange of the two parts are different, the two solutions in the dominant pair are input into the discriminator according to the dominance relationship, and the discriminator learns the dominance relationship between the non-dominated solution and the dominated solution; In the module M5: The generator G is trained based on the trained discriminator, and the generator adopts an unsupervised training manner, and uses the dominance relationship learned by the discriminator to enable the generator to predict solutions on the basis of existing non-dominated solutions.

Images