Training method of efficiency prediction model based on cyclic generative adversarial network

Abstract

The invention discloses a training method of an efficiency prediction model based on a cyclic generative adversarial network, belongs to the technical field of efficiency evaluation, and solves the problem of low efficiency prediction efficiency in the prior art. The method comprises the following steps: acquiring a state data sequence and a corresponding coping success rate of two confrontation parties in a confrontation simulation process, and constructing a training sample based on the state data sequence and the corresponding coping success rate; each training sample comprises an initial sequence and a corresponding coping success rate, a global sampling sequence and an efficiency result; a joint learning model is constructed, the joint learning model comprises a cyclic generative adversarial network and an efficiency prediction network, and the cyclic generative adversarial network is used for generating a global sampling sequence based on the initial sequence; the efficiency prediction network is used for performing efficiency prediction based on the initial sequence and the generated global sampling sequence; and performing joint training on the joint learning model based on the training sample set to obtain an efficiency prediction model for predicting efficiency based on the initial sequence. And rapid and accurate efficiency prediction is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of performance evaluation, and in particular to a training method for a performance prediction model based on a cyclic generative adversarial network. Background Art

[0002] In recent years, with the complexity of system functions and structures, simulation has become an important tool for evaluating and analyzing effectiveness. Conducting adversarial effectiveness evaluation through simulation technology and analyzing the expected goal achievement of tasks under specific conditions are of great significance to the design, verification and application of the system. Effectiveness evaluation is an evaluation method to measure the effect of task completion and has been widely used in many fields.

[0003] Effectiveness evaluation emphasizes the ability or success rate of completing tasks in a specific environment. Traditional effectiveness evaluation methods predict based on the data of the entire simulation process, with large sample data and low prediction efficiency. Summary of the invention

[0004] In view of the above analysis, an embodiment of the present invention aims to provide a training method for a performance prediction model based on a recurrent generative adversarial network to solve the problem of low efficiency of existing performance prediction.

[0005] On the one hand, an embodiment of the present invention provides a training method for a performance prediction model based on a recurrent generative adversarial network, comprising the following steps:

[0006] Obtaining state data sequences and corresponding response success rates of the two adversarial parties during the adversarial simulation process, and constructing training samples based on the state data sequences and corresponding response success rates; each of the training samples includes an initial sequence and a corresponding response success rate, a global sampling sequence and an effectiveness result;

[0007] Constructing a joint learning model, the joint learning model includes a recurrent generative adversarial network and a performance prediction network, the recurrent generative adversarial network is used to generate a global sampling sequence based on the initial sequence; the performance prediction network is used to perform performance prediction based on the initial sequence and the generated global sampling sequence;

[0008] The joint learning model is jointly trained based on the training sample set to obtain a performance prediction model for predicting performance based on an initial sequence.

[0009] Based on the further improvement of the above method, the cyclic generative adversarial network includes a first generator, a second generator, a first discriminator and a second discriminator;

[0010] The first generator is used to generate a global sampling sequence based on an input initial sequence to obtain a synthetic global sampling sequence; the first discriminator is used to discriminate the authenticity of the synthetic global sampling sequence;

[0011] The second generator is used to generate an initial sequence based on the synthetic global sampling sequence to obtain a synthetic initial sequence; the second discriminator is used to discriminate the authenticity of the synthetic initial sequence.

[0012] Based on the further improvement of the above method, the following formula is used to calculate the loss of the cyclic generative adversarial network:

[0013]

[0014] in, represents the discriminator loss, represents the generator loss, Indicates consistency loss.

[0015] Based on the further improvement of the above method, the following formula is used to calculate the discriminator loss and generator loss:

[0016]

[0017] Where D1(·) represents the discrimination result of the first discriminator, D2(·) represents the discrimination result of the second discriminator, G1(·) represents the generation result of the first generator, G2(·) represents the generation result of the second generator, and X i1 represents the initial sequence of the i-th sample, X i2 represents the global sampling sequence of the i-th sample, l(·,·) represents the loss function, and n represents the number of samples in the current training batch.

[0018] Based on the further improvement of the above method, the following formula is used to calculate the consistency loss:

[0019]

[0020] Based on the further improvement of the above method, the performance prediction network includes:

[0021] A first feature extraction module, used for extracting a first feature based on an input initial sequence;

[0022] A second feature extraction module, used for extracting a second feature based on an input global sampling sequence;

[0023] A prediction unit, configured to predict performance based on the first feature and the second feature;

[0024] The regression unit is used to regress the response success rate based on the input initial sequence.

[0025] Based on the further improvement of the above method, the following formula is used to calculate the loss of the prediction unit of the prediction network:

[0026]

[0027] Where n represents the number of samples in the current training batch. represents the difference loss between the first feature and the second feature of the i-th sample, represents the prediction loss of the i-th sample, z i1 represents the first feature of the i-th sample, z i2 represents the second feature of the i-th sample, p i represents the performance prediction result of the i-th sample, y i represents the true performance result of i samples, y i =1, y i = 0, α, β, γ and δ represent weight parameters, P represents the distribution of the true effectiveness of the current batch of samples, Q represents the distribution of the predicted effectiveness of the current batch of samples, KL(P|Q) represents the divergence of calculating P and Q, α t represents the sample difficulty weight parameter, and λ represents the adjustment factor.

[0028] Based on the further improvement of the above method, the following formula is used to calculate the loss of the regression unit of the prediction network:

[0029]

[0030] Where n represents the number of samples in the current training batch, m represents the length of the initial sequence, represents the true value of the response success rate of the kth time step of the i-th sample, represents the regression value of the response success rate of the kth time step of the i-th sample, ε and μ are adjustable parameters.

[0031] Based on the further improvement of the above method, the state data sequence is non-uniformly sampled to obtain a global sampling sequence;

[0032] The samples are constructed with the initial sequence and the global sampling sequence as the input data of the samples, the response success rate corresponding to the initial sequence and the effectiveness result as the labels of the samples to obtain the training sample set.

[0033] Based on the further improvement of the above method, the sampling points of non-uniform sampling are obtained in the following way:

[0034] S11, taking the first data point of the state data sequence in the adversarial simulation process as the first sampling point and the last data point as the last sampling point, to obtain a sampling point set;

[0035] S12, each segment obtained by dividing the state data sequence with the sampling points in the current sampling point set is a current segment;

[0036] S13, traverse each time step in the current segment in turn, calculate the maximum likelihood estimate and the sum of the data point of each time step and the two sampling points of the current segment; take the time point with the largest maximum likelihood estimate and the maximum time point as the candidate sampling point of the current segment;

[0037] S14. If there is only one candidate sampling point at present, the candidate sampling point is directly used as the sampling point and added to the sampling point set. Otherwise: for each candidate sampling point, the maximum likelihood estimate sum between adjacent sampling points is calculated if the candidate sampling point is used as the sampling point, and the candidate sampling point with the largest maximum likelihood estimate sum between adjacent sampling points is used as the sampling point and added to the sampling point set.

[0038] S15. If the current number of iterations is reached, stop the iteration to obtain the sampling point combination; otherwise, return to step S12.

[0039] Compared with the prior art, the present invention constructs a training sample set by extracting the initial sequence and the global sampling sequence and the corresponding response success rate from the state data sequence in the adversarial simulation process, and trains a joint learning model based on the constructed training sample set, that is, trains a cyclic generative adversarial network that generates a global sampling sequence based on the initial sequence and a prediction network that performs performance prediction based on the initial sequence and the generated global sampling sequence. Based on the trained joint network, a global sampling sequence can be generated according to the initial sequence, and further performance prediction can be performed based on the initial sequence and the global sampling sequence. Therefore, the entire simulation process is not required, and the final simulation performance result can be quickly and accurately predicted based on the initial sequence in the initial stage, thereby improving the efficiency of performance prediction.

[0040] In the present invention, the above-mentioned technical solutions can also be combined with each other to achieve more preferred combination solutions. Other features and advantages of the present invention will be described in the subsequent description, and some advantages can become obvious from the description, or can be understood by practicing the present invention. The purpose and other advantages of the present invention can be achieved and obtained through the contents particularly pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The accompanying drawings are only used for the purpose of illustrating specific embodiments and are not to be considered as limiting the present invention. In the entire drawings, the same reference symbols represent the same components;

[0042] Figure 1 The present invention is a flowchart of a method for training a performance prediction model based on a cyclic generative adversarial network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0043] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0044] A specific embodiment of the present invention discloses a training method for an effectiveness prediction model based on a recurrent generative adversarial network, such as Figure 1 As shown, the following steps are included:

[0045] S1. Obtain the state data sequence and the corresponding response success rate of both parties during the confrontation simulation process, and construct training samples based on the state data sequence and the corresponding response success rate; each training sample includes an initial sequence and a corresponding response success rate, a global sampling sequence and an effectiveness result;

[0046] S2. Constructing a joint learning model, wherein the joint learning model includes a recurrent generative adversarial network and a performance prediction network, wherein the recurrent generative adversarial network is used to generate a global sampling sequence based on the initial sequence; and the performance prediction network is used to perform performance prediction based on the initial sequence and the generated global sampling sequence;

[0047] S3. Jointly train the joint learning model based on the training sample set to obtain a performance prediction model for predicting performance based on the initial sequence.

[0048] It should be noted that the two opposing parties include the active party and the passive party, and the effectiveness result is whether the active party is successful, which is 1 if successful and 0 if unsuccessful. The initial sequence is the sequence at the initial stage of the simulation.

[0049] Compared with the prior art, the training method of the performance prediction model based on the cyclic generative adversarial network provided in this embodiment obtains the state data sequence in the adversarial simulation process, extracts the initial sequence and the global sampling sequence therein, and constructs a training sample set as well as the corresponding response success rate. The joint learning model is trained based on the constructed training sample set, that is, the cyclic generative adversarial network that generates the global sampling sequence based on the initial sequence and the prediction network that performs performance prediction based on the initial sequence and the generated global sampling sequence are trained. Based on the trained joint network, the global sampling sequence can be generated according to the initial sequence, and the performance prediction can be further performed based on the initial sequence and the global sampling sequence. Therefore, the entire simulation process is not required, and the final simulation performance result can be quickly and accurately predicted based on the initial sequence in the initial stage, thereby improving the efficiency of performance prediction.

[0050] It should be noted that the status data in the confrontation simulation process includes the number of entities, behavioral strategies, speed, position and operation mode, information processing efficiency of each active party, and the number of entities, behavioral strategies, orientation and information processing capabilities of the passive party.

[0051] The response success rate is the probability of the passive party succeeding at each simulation moment, which is calculated using the following formula:

[0052]

[0053] Where N represents the number of active entities at the current simulation time, p i represents the probability that the i-th active party is detected by the passive party, and ρ represents the success rate of the passive party’s response action.

[0054] During implementation, non-uniform sampling is performed on the state data sequence to obtain a global sampling sequence;

[0055] The samples are constructed with the initial sequence and the global sampling sequence as the input data of the samples, the response success rate corresponding to the initial sequence and the effectiveness result as the labels of the samples to obtain the training sample set.

[0056] According to the simulation process data, the change range of the data state characteristics under the full trajectory of the adversary is not significant, and there is no short-term mutation. Therefore, the initial stage data before the specific step length is selected can be extracted from local features to global trajectory features.

[0057] In order to enable the trained model to predict the final simulation results based on the simulation data in the initial stage, the present invention first extracts the state data sequence in the initial stage of the simulation process, namely, the initial sequence X1, and then performs non-uniform sampling on the state data sequence in the adversarial simulation process to obtain the global sampling sequence X2, thereby extracting features that can accurately predict the final simulation results and improving the accuracy of the prediction.

[0058] If the state data sequence in the initial stage includes M time steps, and the state data of each time step is D-dimensional, then X1, X2∈R D*M .

[0059] Specifically, the sampling points of non-uniform sampling are obtained in the following manner:

[0060] S11, taking the first data point of the state data sequence in the adversarial simulation process as the first sampling point and the last data point as the last sampling point, to obtain a sampling point set;

[0061] S12, each segment obtained by dividing the state data sequence with the sampling points in the current sampling point set is a current segment;

[0062] S13, traverse each time step in the current segment in turn, calculate the maximum likelihood estimate and the sum of the data point of each time step and the two sampling points of the current segment; take the time point with the largest maximum likelihood estimate and the maximum time point as the candidate sampling point of the current segment;

[0063] S14. If there is only one candidate sampling point at present, the candidate sampling point is directly used as the sampling point and added to the sampling point set. Otherwise: for each candidate sampling point, the maximum likelihood estimate sum between adjacent sampling points is calculated if the candidate sampling point is used as the sampling point, and the candidate sampling point with the largest maximum likelihood estimate sum between adjacent sampling points is used as the sampling point and added to the sampling point set.

[0064] S15. If the current number of iterations is reached, stop the iteration to obtain the sampling point combination; otherwise, return to step S12.

[0065] For example, the state data sequence in the adversarial simulation process has a total of T time steps S = {x1, x2, ... x T}, the data point x at each time step t The dimension of is D. The state data sequence in the initial stage includes M time steps, so the state data sequence in the adversarial simulation process needs to be divided into M-1 segments.

[0066] The rth segment is represented by [b r-1 ,b r ), where each data point obeys a multi-source Gaussian distribution, and its mean vector and covariance matrix are:

[0067]

[0068] First, in step S11, the first data point of the state data sequence is used as the first sampling point, and the last data point is used as the last sampling point, that is, b0 = 0, b2 = T-1. The sampling points in the current sampling point set {b0, b2} divide the state data sequence into a segment, that is, the state data sequence as a whole is taken as a segment.

[0069] Take the state data sequence as the current segment, and traverse each time step in the current segment in turn, that is, calculate each time step in the current segment except the sampling point, and calculate the maximum likelihood estimate and of the data point of each time step and the two sampling points of the current segment.

[0070] Specifically, the following formula is used to calculate the maximum likelihood estimate of the data point at each time step and the two sampling points of the current segment:

[0071]

[0072] Among them, ψ(b0,t) represents the maximum likelihood estimate of the sampling point b0 and the data point at the t-th time step, ψ(t,b2) represents the maximum likelihood estimate of the data point at the t-th time step and the sampling point b2, b0 and b2 represent the sampling points at the left and right ends of the current segment respectively, S (1) represents the covariance matrix of the data sequence between b0 and t, S (2)represents the covariance matrix of the data sequence between t and b2, logdet(·) represents the logarithm-determinant function, Tr(·) represents the trace of the covariance matrix, η represents the regularization parameter, η≥0.

[0073] Traverse each time step between b0 and b2 and calculate ψ t , take ψ t The largest time point is the candidate sampling point. Since there is only one segment at present, the candidate sampling point set is directly obtained. If there are multiple segments at present, it is necessary to execute step S13 for all the current segments to obtain the corresponding candidate sampling points, form a candidate sampling point set, and then execute step S14. Since there is only one segment at present, corresponding to one candidate sampling point, the obtained candidate sampling point is directly used as a sampling point and added to the sampling point set, and the sampling point set is updated to {b0, b1, b2} (arranged in time order).

[0074] Since the number of iterations has not been reached, the process returns to step S12. At this time, the sampling points in the current sampling point set divide the state data sequence into two segments, namely the current segments [b0, b1) and [b1, b2). Then, step S13 is executed on the segment [b0, b1) to obtain the corresponding candidate sampling point t1, and step S13 is executed on the segment [b1, b2) to obtain the corresponding candidate sampling point t2.

[0075] In step S14, if the candidate sampling point t1 is used as the sampling point, the sampling point set is {b0, t1, b1, b2}, and the maximum likelihood estimate and the sum of adjacent sampling points are calculated according to the above formula, that is, ψ(b0, t1) + ψ(t1, b1) + ψ(b1, b2); if the candidate sampling point t2 is used as the sampling point, the sampling point set is {b0, b1, t2, b2}, and the maximum likelihood estimate and the sum of adjacent sampling points are calculated according to the above formula, that is, ψ(b0, b1) + ψ(b1, t2,) + ψ(t2, b2); the candidate sampling point corresponding to the maximum likelihood estimate and is taken as the sampling point, added to the sampling point set, and the sampling point set is updated to {b0, b1, b2, b3} (arranged in chronological order), and so on, iterate M-2 times to obtain the sampling point.

[0076] By adopting non-uniform sampling, the trend of the state sequence can be better sampled, laying the foundation for accurately extracting prediction features.

[0077] Specifically, the constructed joint learning model includes: a recurrent generative adversarial network and a performance prediction network.

[0078] The cyclic generative adversarial network includes a first generator G1, a second generator G2, a first discriminator D1 and a second discriminator D2;

[0079] The first generator G1 is used to generate a global sampling sequence based on the input initial sequence to obtain a synthetic global sampling sequence; the first discriminator D1 is used to discriminate the authenticity of the synthetic global sampling sequence;

[0080] The second generator G2 is used to generate an initial sequence based on the synthesized global sampling sequence to obtain a synthesized initial sequence; the second discriminator D2 is used to discriminate the authenticity of the synthesized initial sequence.

[0081] In order to train a network that can more accurately generate a global sampling sequence based on an initial sequence, the present invention adopts a cyclic generative adversarial network, that is, the trained generative adversarial network can not only generate a global sampling sequence based on the initial sequence, but also generate an initial sequence based on the global sampling sequence, thereby avoiding the global sampling sequence generated by the generator for all samples being the same.

[0082] During implementation, both the generator and the discriminator can adopt the long short-term recurrent neural network LSTM structure.

[0083] Specifically, the performance prediction network includes:

[0084] A first feature extraction module, used for extracting a first feature based on an input initial sequence;

[0085] A second feature extraction module, used for extracting a second feature based on an input global sampling sequence;

[0086] A prediction unit, configured to predict performance based on the first feature and the second feature;

[0087] The regression unit is used to regress the response success rate based on the input initial sequence.

[0088] The two feature extraction units of the prediction network extract the features of the initial sequence and the global sampling sequence respectively. The prediction unit is used to predict the final performance result based on the extracted first feature and the second feature. The regression network is used to regress the success rate according to the local features of each time step of the initial sequence. Classification and regression tasks are performed simultaneously, which improves the application flexibility and performance of the prediction network. The regression unit can use LSTM or other recurrent RNN network structures.

[0089] During implementation, the cyclic generative adversarial network can be trained first, and the prediction network can be trained based on the trained cyclic generative adversarial network, that is, the training can be performed in a two-stage training manner. Alternatively, the parameters of the prediction network can be fixed first, and the cyclic generative adversarial network can be trained based on one or more batches of samples, and then the parameters of the cyclic generative adversarial network can be fixed, and the prediction network can be trained based on one or more batches of samples. If the prediction network does not converge, the parameters of the prediction network can be fixed, and the cyclic generative adversarial network can be trained based on one or more batches of samples, and so on, that is, the cyclic generative adversarial network and the prediction network can be trained alternately until the prediction network converges.

[0090] Specifically, the loss of the cyclic generative adversarial network is calculated using the following formula:

[0091]

[0092] in, represents the discriminator loss, represents the generator loss, Denotes the consistency loss. The discriminator loss is the sum of the generation loss of the first discriminator and the discriminative loss of the second discriminator.

[0093] Specifically, the discriminator loss is calculated using the following formula:

[0094]

[0095] Where D1(·) represents the discrimination result of the first discriminator, D2(·) represents the discrimination result of the second discriminator, G1(·) represents the generation result of the first generator, G2(·) represents the generation result of the second generator, and X i1 represents the initial sequence of the i-th sample, X i2 represents the global sampling sequence of the i-th sample, and n represents the number of samples in the current training batch.

[0096] The generator loss is calculated using the following formula:

[0097]

[0098] Among them, l(·,·) represents the loss function, and n represents the number of samples in the current training batch.

[0099] The generator loss is the sum of the generation loss of the first generator and the generation loss of the second generator. The generation loss of the first generator is the loss between the global sampling sequence generated by the first generator based on the initial sequence of the i-th sample and the true global sampling sequence of the i-th sample. The generation loss of the second generator is the loss between the initial sequence generated by the second generator based on the global sampling sequence of the i-th sample and the true initial sequence of the i-th sample. In implementation, l(·,·) can use the mean square error loss.

[0100] In order to improve the accuracy of generated data and avoid generating the same results for different samples, the loss of the cyclic generative adversarial network includes consistency loss.

[0101] Specifically, the following formula is used to calculate the consistency loss:

[0102]

[0103] That is, the initial sequence G2(X) generated by the second generator according to the global sampling sequence of samples i2 ) is input into the first generator to obtain the sequence G1(G2(X i2 )) and the global sampling sequence X i2 Compare the global sampling sequence G1(X i1 ) is input to the second generator to obtain the sequence G2(G1(X i1 )) and the initial sequence X i1 In addition, the first generator is used to generate a global sampling sequence. Even if the input is a global sampling sequence, the corresponding global sampling sequence can be generated. Therefore, the consistency loss includes X i2 With G1(X i2 ) loss. The second generator is used to generate the initial sequence. Even if the input is the initial sequence, the corresponding initial sequence can be generated. Therefore, the consistency loss includes X i1 With G2(X i1 ) loss.

[0104] For the prediction network, its training loss is the sum of the losses of the prediction unit and the regression unit.

[0105] Specifically, the following formula is used to calculate the loss of the prediction unit of the prediction network:

[0106]

[0107] Where n represents the number of samples in the current training batch. represents the difference loss between the first feature and the second feature of the i-th sample, represents the prediction loss of the i-th sample, zi1 represents the first feature of the i-th sample, z i2 represents the second feature of the i-th sample, p i represents the performance prediction result of the i-th sample, y i represents the true performance result of i samples, y i =1, y i = 0, α, β, γ and δ represent weight parameters, P represents the distribution of the true effectiveness of the current batch of samples, Q represents the distribution of the predicted effectiveness of the current batch of samples, KL(P|Q) represents the divergence of calculating P and Q, α t represents the sample difficulty weight parameter, and λ represents the adjustment factor.

[0108] Based on the classification prediction loss, in order to make the model learn the similarity between the initial sequence and the global sampling sequence, the similarity constraint To minimize the difference between the initial sequence and the global sampling sequence. The first term in the difference calculation formula represents the distance between the two features; the second term represents the angular difference between the two features, which helps to improve the discriminative ability of the feature space, especially in high-dimensional space; the third term is the regularization term, which is used to control the scale of the feature vector to avoid excessive values ​​that lead to unstable calculations; α, β, γ are adjustable parameters used to control the influence of different terms. By comprehensively considering distance, angle, and normalization, the model not only focuses on minimizing the direct distance between features during learning, but also considers directionality and scale, which helps to improve the generalization ability and stability of the model.

[0109] The loss of the prediction result uses the dynamic balance loss function Not only does it take into account the difficulty of the sample, it also dynamically adjusts the weight of each category to more effectively handle category imbalance. In addition, in order to make the distribution of the predicted results the same as the true results, the loss of the prediction unit also includes the KL divergence loss.

[0110] Specifically, the loss of the regression unit of the prediction network is calculated using the following formula:

[0111]

[0112] Where n represents the number of samples in the current training batch, m represents the length of the initial sequence, represents the true value of the response success rate of the kth time step of the i-th sample, represents the regression value of the response success rate of the kth time step of the i-th sample, ε and μ are adjustable parameters.

[0113] The first term of the regression loss is the conventional MSE mean square error loss, which is used to calculate the basic regression task loss. The second term is an exponential decay term, which is used to impose additional penalties on samples with large differences between the predicted values ​​and the target values, especially samples near the data boundary or outliers, thereby improving training efficiency.

[0114] The back-propagation algorithm is used to calculate the gradient of the loss function relative to the model parameters, iteratively update the model parameters, and minimize the loss function. When the loss reaches the preset accuracy or reaches a predetermined number of iterations, the training is completed and the performance prediction model is obtained.

[0115] For the adversarial simulation task to be predicted, the initial sequence input effectiveness prediction model in the adversarial simulation process is obtained. First, the global adoption sequence is generated through the cyclic generation network, and then the final effectiveness result can be quickly and accurately predicted through the prediction network. According to these prediction results, decision makers can adjust the offensive strategy and improve the success rate.

[0116] Those skilled in the art will appreciate that all or part of the processes of the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, wherein the computer-readable storage medium is a disk, an optical disk, a read-only storage memory, or a random access memory, etc.

[0117] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A training method for an efficacy prediction model based on a recurrent generative adversarial network, characterized in that: The following steps are involved: Obtaining state data sequences and corresponding response success rates of the two adversarial parties during the adversarial simulation process, and constructing training samples based on the state data sequences and corresponding response success rates; each of the training samples includes an initial sequence and a corresponding response success rate, a global sampling sequence and an effectiveness result; Constructing a joint learning model, the joint learning model includes a recurrent generative adversarial network and a performance prediction network, the recurrent generative adversarial network is used to generate a global sampling sequence based on the initial sequence; the performance prediction network is used to perform performance prediction based on the initial sequence and the generated global sampling sequence; The joint learning model is jointly trained based on the training sample set to obtain a performance prediction model for predicting performance based on an initial sequence.

2. The training method of the performance prediction model based on the cyclic generative adversarial network according to claim 1, characterized in that: The cyclic generative adversarial network includes a first generator, a second generator, a first discriminator and a second discriminator; The first generator is used to generate a global sampling sequence based on an input initial sequence to obtain a synthetic global sampling sequence; The first discriminator is used to discriminate the authenticity of the synthesized global sampling sequence; The second generator is used to generate an initial sequence based on the synthetic global sampling sequence to obtain a synthetic initial sequence; the second discriminator is used to discriminate the authenticity of the synthetic initial sequence.

3. The training method of the performance prediction model based on the cyclic generative adversarial network according to claim 2 is characterized in that: The loss of the cyclic generative adversarial network is calculated using the following formula: in, represents the discriminator loss, represents the generator loss, Indicates consistency loss.

4. The training method of the performance prediction model based on the cyclic generative adversarial network according to claim 3 is characterized in that: The discriminator loss and generator loss are calculated using the following formula: Where D1(·) represents the discrimination result of the first discriminator, D2(·) represents the discrimination result of the second discriminator, G1(·) represents the generation result of the first generator, G2(·) represents the generation result of the second generator, and X i1 represents the initial sequence of the i-th sample, X i2 represents the global sampling sequence of the i-th sample, l(·,·) represents the loss function, and n represents the number of samples in the current training batch.

5. The training method of the performance prediction model based on the cyclic generative adversarial network according to claim 3 is characterized in that: The consistency loss is calculated using the following formula:

6. The training method of the performance prediction model based on the cyclic generative adversarial network according to claim 1, characterized in that: The performance prediction network includes: A first feature extraction module, used for extracting a first feature based on an input initial sequence; A second feature extraction module, used for extracting a second feature based on an input global sampling sequence; A prediction unit, configured to predict performance based on the first feature and the second feature; The regression unit is used to regress the response success rate based on the input initial sequence.

7. The training method of the performance prediction model based on the cyclic generative adversarial network according to claim 6, characterized in that: The loss of the prediction unit of the prediction network is calculated using the following formula: Where n represents the number of samples in the current training batch. represents the difference loss between the first feature and the second feature of the i-th sample, represents the prediction loss of the i-th sample, represents the first feature of the i-th sample, represents the second feature of the i-th sample, p i represents the performance prediction result of the i-th sample, y i represents the true performance result of i samples, y i =1, y i = 0, α, β, γ and δ represent weight parameters, P represents the distribution of the true effectiveness of the current batch of samples, Q represents the distribution of the predicted effectiveness of the current batch of samples, KL(P|Q) represents the divergence of calculating P and Q, α t represents the sample difficulty weight parameter, and λ represents the adjustment factor.

8. The training method of the performance prediction model based on the cyclic generative adversarial network according to claim 6, characterized in that: The loss of the regression unit of the prediction network is calculated using the following formula: Where n represents the number of samples in the current training batch, m represents the length of the initial sequence, represents the true value of the response success rate of the kth time step of the i-th sample, represents the regression value of the response success rate of the kth time step of the i-th sample, ε and μ are adjustable parameters.

9. The training method of the performance prediction model based on the recurrent generative adversarial network according to claim 1, characterized in that: Performing non-uniform sampling on the state data sequence to obtain a global sampling sequence; The samples are constructed with the initial sequence and the global sampling sequence as the input data of the samples, the response success rate corresponding to the initial sequence and the effectiveness result as the labels of the samples to obtain the training sample set.

10. The training method of the performance prediction model based on the cyclic generative adversarial network according to claim 9, characterized in that: The sampling points of non-uniform sampling are obtained in the following way: S11, taking the first data point of the state data sequence in the adversarial simulation process as the first sampling point and the last data point as the last sampling point, to obtain a sampling point set; S12, each segment obtained by dividing the state data sequence with the sampling points in the current sampling point set is a current segment; S13, traverse each time step in the current segment in turn, calculate the maximum likelihood estimate and the sum of the data point of each time step and the two sampling points of the current segment; take the time point with the largest maximum likelihood estimate and the maximum time point as the candidate sampling point of the current segment; S14. If there is only one candidate sampling point at present, the candidate sampling point is directly used as the sampling point and added to the sampling point set. Otherwise: for each candidate sampling point, the maximum likelihood estimate sum between adjacent sampling points is calculated if the candidate sampling point is used as the sampling point, and the candidate sampling point with the largest maximum likelihood estimate sum between adjacent sampling points is used as the sampling point and added to the sampling point set. S15. If the current number of iterations is reached, stop the iteration to obtain the sampling point combination; otherwise, return to step S12.

Images