Evaluation and decision-making method of computational experiment in parallel simulation system based on deep learning

Abstract

The present application discloses a method for computational experiment evaluation and decision-making of a parallel simulation system based on deep learning, comprising: designing a variety of control schemes for an actual system, and simulating and deducing the various schemes based on a simulation system to obtain the future state of the actual system under each control scheme; using a computational experiment evaluation decision model to analyze the future state of the actual system under each control scheme, and evaluate the preferred control scheme for the actual system; using an execution network to feed back the preferred control scheme to the actual system as a control strategy for the actual system. The method of this embodiment fully considers the uncertainty and dynamic evolution characteristics of complex systems, and can improve the objectivity and accuracy of computational experiment scheme evaluation decisions, thereby improving the accuracy of the guidance of the parallel simulation system on complex actual systems.

Description

Technical Field

[0001] The present application relates to the field of simulation discipline application technology, and in particular to a computational experiment evaluation and decision-making method for a parallel simulation system based on deep learning. Background Art

[0002] Parallel simulation systems are the application of parallel system methods in the field of simulation. They primarily consist of a simulation system and a real system, which operate in parallel. The simulation system conducts simulation computational experiments on various control schemes for the real system, and manages the real system using the optimal computational experiment scheme. The ultimate goal of a parallel simulation system is to precisely control the real system. The key lies in the simulation system's accurate prediction of the real system's future state and the objective and accurate evaluation of the computational experiment schemes. The evaluation and decision-making methods for computational experiment schemes are one of the key technologies in building parallel simulation systems.

[0003] Parallel simulation systems are mainly used for the management and control of complex systems and related auxiliary decision-making. Complex systems have strong uncertainty and dynamic evolution during operation. For parallel simulation systems of complex systems, only when the evaluation and decision-making methods of computational experimental schemes are objective and accurate can effective decisions be made on actual complex systems.

[0004] Existing solutions lack theoretical research on computational experiment evaluation and decision-making methods, and the description of evaluation and decision-making models is unclear, making it difficult to guide the application of parallel simulation systems. For example, the Chinese patent "CN110020750A A Parallel Emergency Evacuation System Based on ACP and Construction Method Thereof" only describes the role of the computational experiment analysis module, making it difficult to provide theoretical guidance for the construction of parallel emergency evacuation systems.

[0005] The Chinese patent "CN112381402A Parallel Intelligent Sand Control Decision Support Method and System" describes a numerical model for calculating the sand control benefit index of a protection and control scheme. This model cannot change with the evolution of the actual system, and the evaluation results of the scheme lack objectivity. Summary of the Invention

[0006] The embodiment of the present application provides a method for evaluating and deciding computational experiments in a parallel simulation system based on deep learning, which fully considers the uncertainty and dynamic evolution characteristics of complex systems, and can improve the objectivity and accuracy of the evaluation decisions of computational experiment plans, thereby improving the accuracy of the guidance of the parallel simulation system on complex actual systems.

[0007] The present application provides a method for evaluating and deciding computational experiments in parallel simulation systems based on deep learning, which is applied to computational experiment evaluation and decision-making in both actual systems and simulation systems, including:

[0008] Design multiple control schemes for the actual system;

[0009] Through the simulation system, simulation deduction is performed based on multiple control schemes to obtain multiple deduction and prediction schemes of the future state of the actual system under each control scheme;

[0010] Using computational experiments to evaluate the decision model, multiple prediction schemes for future states under various control schemes are analyzed to evaluate the optimal prediction control scheme for the actual system;

[0011] Feeding back the preferred prediction and control scheme to the actual system by using an execution network to serve as a control strategy for the actual system;

[0012] The computational experiment evaluation decision model includes an evaluation network and a utility function, wherein the evaluation network is used to analyze multiple deduction and prediction schemes of future states under various control schemes to evaluate the preferred prediction and control scheme for the actual system;

[0013] The utility function is used to perform state calculation on the actual system to adjust the parameters of the evaluation network.

[0014] Optionally, the evaluation network is established based on the CNN network in the following manner:

[0015] The input of the CNN network is set to one dimension, the number of network nodes of the input layer of the CNN network is set to the number of states of the real system, and the convolution kernel of the convolution layer and the sampling window of the pooling layer in the CNN are both set to one dimension;

[0016] Select the required convolution kernel size and pooling layer sampling window size to build the evaluation network.

[0017] Optionally, the network parameters of the evaluation network are adjusted in the following manner:

[0018] For the output layer of the evaluation network, based on the given weight learning rate μ, the weight update function is set to:

[0019]

[0020] Among them, W l indicates, E indicates;

[0021] For the pooling layer of the evaluation network, the following calculation is performed on any feature map in the convolutional layer of the CNN network and the corresponding feature map of the pooling layer:

[0022]

[0023] Among them, β represents the multiplicative bias of the pooling layer, express, express, represents, f'() represents, and up() represents upsampling operation;

[0024] For the convolutional layer of the evaluation network, the convolution kernels in the convolutional layer share weights, and the partial derivative of the error function of the convolution kernel with respect to the convolution kernel weights satisfies:

[0025]

[0026] in, Indicates that the input vector Above, and convolution kernel elements Multiply bit by bit to get the output feature map Input for the value at the upper (u,v) position.

[0027] Optionally, the method further includes using a self-organizing incremental learning neural network (SOINN) to construct the evaluation network using the neural networks of the last two layers of the CNN network;

[0028] In the process of adjusting the parameters of the evaluation network, the output of the convolutional layer-pooling layer of the CNN network is used as the input of the SOINN.

[0029] Optionally, the method further includes adjusting the network parameters of the evaluation network in the following manner:

[0030] Using an error back propagation algorithm to adjust the network parameters of the convolutional layer-pooling layer part of the evaluation network;

[0031] A self-organizing incremental learning algorithm is used to adjust the network parameters of the SOINN.

[0032] Optionally, the SOINN is configured to perform the following steps:

[0033] After initialization, based on the initialized neuron set A, two neurons s1 and s2 similar to the input data sample are determined:

[0034]

[0035]

[0036] Where W c represents, ξ represents the input data sample;

[0037] like or If it holds, a new node r is generated for the neuron set A, let A=A∪{r}, Wr =ξ;

[0038] If there is no connection between the two neurons s1 and s2, then establish a connection between the two neurons s1 and s2, update the edge set C = C∪{(s1,s2)}, and refresh the age parameter of the edge between the two neurons s1 and s2;

[0039] Increment the age parameters of all edges connected to the winner node and accumulate the local quantization error of the winner node;

[0040] If no new node is established, the weights of the two winner nodes are updated;

[0041] Check the current age parameters of all edges. If age(i,j)>agemax, remove the corresponding edge from the edge set C;

[0042] If there is more input, the process repeatedly determines two neurons that are similar to the input data sample.

[0043] An embodiment of the present application also proposes a computer device, including a processor and a memory, wherein a computer program is stored on the memory, and when the computer program is executed by the processor, the steps of the aforementioned deep learning-based parallel simulation system computational experiment evaluation decision-making method are implemented.

[0044] An embodiment of the present application also proposes a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the aforementioned deep learning-based parallel simulation system computational experiment evaluation decision-making method are implemented.

[0045] The embodiments of the present application fully consider the uncertainty and dynamic evolution characteristics of complex systems, and can improve the objectivity and accuracy of computational experimental scheme evaluation decisions, thereby improving the accuracy of the parallel simulation system's guidance of complex actual systems.

[0046] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Various other advantages and benefits will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiment below. The accompanying drawings are for illustration purposes only and are not to be considered as limiting the present application. The same reference symbols are used throughout the drawings to represent the same components. In the drawings:

[0048] Figure 1This is an example of an adaptive dynamic programming governance architecture for a parallel system according to an embodiment of the present application;

[0049] Figure 2 This is an example of a decision-making method flow for evaluating a computational experiment solution according to an embodiment of the present application;

[0050] Figure 3 This is an example of the network structure of the CNN evaluation decision model in the embodiment of the present application;

[0051] Figure 4 Schematic diagram of the CNN evaluation decision model based on SOINN improvement according to an embodiment of the present application;

[0052] Figure 5 This is a schematic diagram of offline training of the CNN evaluation decision model in an embodiment of the present application;

[0053] Figure 6 This is an example of a CNN-based computational experiment evaluation decision-making method according to an embodiment of the present application. DETAILED DESCRIPTION

[0054] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0055] The embodiment of the present application provides a method for evaluating and deciding computational experiments in a parallel simulation system based on deep learning, which is applied to a computational experiment evaluation and decision framework including an actual system and a simulation system, such as Figure 1 Shown, including:

[0056] Design multiple control schemes for the actual system;

[0057] Through the simulation system, simulation deduction is performed based on multiple control schemes to obtain multiple deduction and prediction schemes of the future state of the actual system under each control scheme;

[0058] Using computational experiments to evaluate decision-making models, multiple prediction schemes for future states under various control schemes are analyzed, and multiple control schemes for the actual system are evaluated and optimized to determine the optimal predictive control scheme for the actual system.

[0059] The preferred prediction and control scheme is fed back to the actual system by using an execution network to serve as a control strategy for the actual system.

[0060] like Figure 1As shown, the complex real-world system governance framework constructed in the embodiments of this application can not only use the feedback of computational experiment evaluation decision results to control the real-world system, but also use the operating status of the real-world system to perform online optimization and adjustment of the computational experiment evaluation decision model. The framework of this application is based on the concept of adaptive dynamic programming method, has a good practical application foundation, and can better promote the practical application of parallel simulation technology.

[0061] In some embodiments, the computational experiment evaluation decision model includes an evaluation network and a utility function, and the evaluation network is used to analyze multiple deduction and prediction schemes of future states under various control schemes, evaluate and select multiple control schemes for the actual system, and evaluate the preferred predictive control scheme for the actual system.

[0062] The utility function is used to perform state calculation on the actual system to adjust the parameters of the evaluation network.

[0063] The evaluation and decision-making technology principles of computational experiment schemes in the complex real-world system control framework are as follows: Figure 2 As shown, it specifically includes two parts:

[0064] Online evaluation and decision-making of computational experiment schemes: Design diverse computational experiment schemes for various control schemes of actual systems, and rely on simulation systems to simulate and deduce each computational experiment scheme, obtaining deduction and prediction scheme data for the future state of the actual system under each computational experiment. This data is used as input parameters for the computational experiment evaluation and decision-making model, and the advantages and disadvantages of various control schemes of the actual system are compared and evaluated to obtain the optimal deduction and prediction scheme. Considering the nonlinear and complex characteristics of the actual system, the embodiment of this application proposes to establish a computational experiment evaluation and decision-making model based on deep learning.

[0065] Dynamic optimization of evaluation decision model parameters: Taking into account the "indifference" of the evaluation decision model with fixed parameters to the dynamic evolution characteristics of complex actual systems, the embodiment of the present application proposes an online self-organizing learning method for the evaluation decision model parameters (see subsequent implementation methods). After using the optimal control plan after the evaluation decision to regulate the actual system, the actual system generates system state time series data. Based on this part of data, the utility function is used to perform online learning adjustment on the parameters of the evaluation decision model, and the network parameters are dynamically optimized and adjusted.

[0066] Convolutional Neural Network (CNN) is a typical deep learning model that typically uses forward propagation to calculate output values ​​and backpropagation to adjust weights and biases. Traditional CNNs are commonly used for image recognition, with the network input being a two-dimensional image data array. For computational experimental evaluation of decision-making problems, in some embodiments, the evaluation network is established based on the CNN network using the following method:

[0067] The CNN network input is set to one-dimensional, the number of network nodes in the CNN network input layer is set to the number of states of the real system, and the convolution kernels and pooling layer sampling windows in the CNN are also set to one-dimensional. In other words, in this example, a two-dimensional CNN is designed as a one-dimensional CNN to establish an evaluation decision model.

[0068] Select the required convolution kernel size and pooling layer sampling window size to establish the evaluation network. In some specific examples, the established CNN evaluation decision model network (evaluation network) structure is as follows: Figure 3 shown.

[0069] In some embodiments, the network parameters of the evaluation network are adjusted in the following manner:

[0070] For the output layer of the evaluation network, in the embodiment of the present application, the partial derivative of the loss function with respect to the additive bias is taken into account. The partial derivative of the output layer loss function with respect to the bias is:

[0071]

[0072] δ represents the error sensitivity, that is, the error sensitivity of the output layer is equal to the partial derivative of the loss function with respect to the neuron bias. During the reverse conduction process of the neural network, the error sensitivity of the high-level and low-level layers satisfies the following relationship:

[0073]

[0074] in: Represents bitwise multiplication of matrix elements, and the output layer is:

[0075]

[0076] The partial derivative of the loss function with respect to the weight is:

[0077]

[0078] In this embodiment, based on the given weight learning rate μ, the weight update function is set to:

[0079]

[0080] Among them, W l Represents the output layer weight, E represents the output layer error, and the update method of bias b is similar to the weight.

[0081] For the pooling layer of the evaluation network, in the convolutional neural network, the pooling layer is connected after the convolution layer. The error sensitivity δ of the pooling layer l lEach element in is associated with a block of elements (sampling window size) in the output feature map of the previous convolutional layer l-1. To calculate the error sensitivity map of convolutional layer l-1, the downsampled error sensitivity map, i.e., the error sensitivity map of the pooling layer, needs to be upsampled to the same size as the feature map of the convolutional layer. The upsampled error sensitivity map is then element-wise multiplied by the partial inverse of the activation function of layer l.

[0082] Any feature map j in the convolutional layer of the CNN network can be calculated by performing the following calculations on the corresponding pooling layer feature map:

[0083]

[0084] in, represents the multiplicative bias of the pooling layer, represents the error sensitivity of the jth element in the lth pooling layer, represents the input of the jth element in the lth pooling layer, f'() represents the partial derivative of the activation function of the lth layer, and up() represents the upsampling operation, which flattens each element in the input data onto the output feature map according to the sampling window size during the forward propagation. This operation can be quickly implemented using the following Kronecker Product:

[0085]

[0086] For the convolutional layer of the evaluation network, based on the error sensitivity map calculated by the pooling layer, the partial derivative of the error function with respect to the bias in the convolutional layer can be calculated by summing the neuron nodes associated with the sensitivity map in the lth layer:

[0087]

[0088] The convolution kernels in the convolution layer are weight-shared. When calculating the partial derivative of the error function with respect to the convolution kernel weight, it is necessary to find the gradient of all weights associated (shared) with the sensitivity map, and then sum these gradients. The partial derivative of the error function of the convolution kernel with respect to the convolution kernel weight satisfies:

[0089]

[0090] in, Indicates that the input vector Above, and convolution kernel elements Multiply bit by bit to get the output feature map The input of the value at the upper (u, v) position.

[0091] In some embodiments, the evaluation network is constructed by using a self-organizing incremental learning neural network (SOINN) to construct the neural network of the last two layers of the CNN network. Specifically, in this embodiment, the neural network of the last two layers (including the output layer) of the CNN evaluation decision model is constructed by using a self-organizing incremental learning neural network (SOINN), such as Figure 4 As shown in the figure, we implement the online self-organizing learning method of CNN evaluation decision model parameters in computational experiments.

[0092] In the process of adjusting the parameters of the evaluation network, the output of the convolution layer-pooling layer of the CNN network is used as the input of the SOINN. In some embodiments, the network parameters of the evaluation network are adjusted in the following manner: the network parameters of the convolution layer-pooling layer of the evaluation network are adjusted by using the error back propagation algorithm. The network parameters of the SOINN are adjusted by using the self-organizing incremental learning algorithm. In the process of adjusting the parameters of the evaluation decision model, Figure 5 As shown in the figure, the output of the front-end convolutional layer-pooling network layer is used as the input of the back-end SOINN network. The network parameters of the front-end convolutional layer-pooling network are adjusted using the error back propagation algorithm, and the parameters and network topology structure of the back-end SOINN network are adjusted using the self-organizing incremental learning algorithm. This enables the CNN evaluation decision model to autonomously adapt to the dynamic changes of the system, discover new patterns of the actual system operation status and perform self-organizing learning, and improve the evaluation decision model's ability to adapt to the dynamic evolution of the system and environmental uncertainty.

[0093] SOINN is a two-layer neural network based on competitive learning. It uses a set of neurons distributed in the feature space to approximate the density distribution of the input data. The connections between these neurons form one or more connected subgraphs, each of which represents a cluster discovered by SOINN and the topological structure of the cluster. SOINN can adaptively adjust the weights of local neurons and the connections between neurons based on the input data and a few simple rules. At the same time, it adopts a discriminant mechanism that adaptively recognizes input patterns that have not been learned before. It can dynamically generate new neurons to represent input patterns, thereby performing online learning on dynamic input data without prior knowledge, discovering new patterns in the data stream without affecting the results of previous learning. In some embodiments, the SOINN is configured to perform the following steps:

[0094] Initialization process: Initialize the neuron set A = {c1, c2}, where the weights W1 and W2 of neurons c1 and c2 are two random initial data samples; initialize the edge set It is an empty set, that is, there is no initial connection between neurons.

[0095] After initialization, based on the initialized neuron set A, a new data sample (or signal, pattern) ξ∈R is input n , determine two neurons s1 and s2 that are similar to the input data sample ξ:

[0096]

[0097]

[0098] Among them, W c represents ξ, and ξ represents the input data sample.

[0099] like or If it holds, a new node r is generated for the neuron set A, let A=A∪{r}, W r =ξ.

[0100] If there is no connection between the two neurons s1 and s2, then establish a connection between the two neurons s1 and s2, update the edge set C = C∪{(s1,s2)}, that is, establish a connection between the two most similar neurons, set age(s1,s2) = 0, and refresh the age parameter of the edge (s1,s2) between the two neurons s1 and s2.

[0101] Increment the age parameters of all edges connected to the winner node and accumulate the local quantization error of the winner node:

[0102] age(s1,i)=age(s1,i)+1

[0103]

[0104] In this example, the winner node is the node whose distance from the input data is ||ξ-W c ||The two nearest nodes.

[0105] If no new node is created, update the weights of the two winner nodes:

[0106]

[0107]

[0108] in:

[0109]

[0110] Check the current age parameters of all edges. If age(i,j)>agemax, remove the corresponding edge from the edge set C, where agemax is a predefined parameter.

[0111] If the total number of data samples currently input is an integer multiple of λ (i.e., one learning cycle has passed), the entire SOINN is checked to perform intra-class node insertion and denoising processes.

[0112] If there is still input, repeat to determine two neurons that are similar to the input data sample, otherwise stop the algorithm and output the neuron set A and connection matrix C.

[0113] The embodiment of the present application uses the CNN network to establish an evaluation decision model, and designs Figure 5 、 Figure 6 The computational experiment evaluation decision structure shown in the figure is used to evaluate the decision model of the computational experiment plan:

[0114] Design computational experiment schemes based on the simulation system and run them in parallel in the simulation system to obtain system state prediction data corresponding to each computational experiment scheme;

[0115] The system prediction data of each computational experiment is used as the input of a CNN evaluation decision model to evaluate and analyze the computational experiment plan and obtain the corresponding system performance index evaluation value;

[0116] Compare and analyze the system performance indicator evaluation values ​​of all computational experiment schemes to form a ranking of the advantages and disadvantages of the computational experiment schemes to support the formulation of actual system management and control strategies.

[0117] This application is aimed at parallel simulation systems and proposes a parallel governance framework based on adaptive dynamic programming. For the evaluation of computational experimental schemes, a method for constructing an evaluation decision model based on a convolutional neural network is proposed. For the online optimization of evaluation decision model parameters, a model parameter adjustment method based on self-organizing incremental learning is proposed, which fully considers the uncertainty and dynamic evolution characteristics of complex systems. The method of this application can improve the objectivity and accuracy of computational experimental scheme evaluation decisions, thereby improving the management accuracy of parallel simulation systems for complex actual systems, and further promoting the engineering application of parallel simulation systems in the management and control of complex systems in social, military and other fields.

[0118] An embodiment of the present application also proposes a computer device, including a processor and a memory, wherein a computer program is stored on the memory, and when the computer program is executed by the processor, the steps of the aforementioned deep learning-based parallel simulation system computational experiment evaluation decision-making method are implemented.

[0119] An embodiment of the present application also proposes a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the aforementioned deep learning-based parallel simulation system computational experiment evaluation decision-making method are implemented.

[0120] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or apparatus comprising the element.

[0121] The serial numbers of the above embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0122] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus the necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a number of instructions for enabling a terminal (which can be a mobile phone, computer, server or network device, etc.) to execute the methods described in each embodiment of the present application.

[0123] The embodiments of the present application are described above in conjunction with the accompanying drawings, but the present application is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of this application, ordinary technicians in this field can also make many forms without departing from the purpose of this application and the scope of protection of the claims, all of which are protected by this application.

Claims

1. A parallel simulation system computational experiment evaluation decision-making method based on deep learning, characterized in that: Computational experiments applied to evaluate and make decisions on both actual and simulated systems include: Design multiple control schemes for the actual system; Through the simulation system, simulation deduction is performed based on multiple control schemes to obtain multiple deduction and prediction schemes of the future state of the actual system under each control scheme; Using computational experiments to evaluate the decision model, multiple prediction schemes for future states under various control schemes are analyzed to evaluate the optimal prediction control scheme for the actual system; Feeding back the preferred prediction and control scheme to the actual system by using an execution network to serve as a control strategy for the actual system; The computational experiment evaluation decision model includes an evaluation network and a utility function, wherein the evaluation network is used to analyze multiple deduction and prediction schemes of future states under various control schemes to evaluate the preferred prediction and control scheme for the actual system; The utility function is used to perform state calculation on the actual system to adjust the parameters of the evaluation network; The method further includes establishing the evaluation network based on the CNN network in the following manner: The input of the CNN network is set to one dimension, the number of network nodes of the input layer of the CNN network is set to the number of states of the real system, and the convolution kernel of the convolution layer and the sampling window of the pooling layer in the CNN are both set to one dimension; Select the required convolution kernel size and pooling layer sampling window size to establish the evaluation network; The network parameters of the evaluation network are adjusted in the following manner: For the output layer of the evaluation network, based on the given weight learning rate , set the weight update function to: in, represents the output layer weight, represents the output layer error; For the pooling layer of the evaluation network, the following calculation is performed on any feature map in the convolutional layer of the CNN network and the corresponding feature map of the pooling layer: in, represents the multiplicative bias of the pooling layer, represents the error sensitivity of the jth element in the lth pooling layer, represents the input of the jth element in the lth pooling layer, represents the partial derivative of the activation function of the lth layer, and up() represents the upsampling operation; For the convolutional layer of the evaluation network, the convolution kernels in the convolutional layer share weights, and the partial derivative of the error function of the convolution kernel with respect to the convolution kernel weights satisfies: in, Indicates that the input vector Above, and convolution kernel elements Multiply bit by bit to get the output feature map superior Input of the value of the position; The method further includes using a self-organizing incremental learning neural network (SOINN) to construct the evaluation network using the neural networks of the last two layers of the CNN network; In the process of adjusting the parameters of the evaluation network, the output of the convolutional layer-pooling layer of the CNN network is used as the input of the SOINN; The method further includes adjusting the network parameters of the evaluation network in the following manner: Using an error back propagation algorithm to adjust the network parameters of the convolutional layer-pooling layer part of the evaluation network; Adopting a self-organizing incremental learning algorithm to adjust the network parameters of the SOINN; The SOINN is configured to perform the following steps: After initialization, based on the initialized neuron set A, two neurons similar to the input data sample are determined. : in represents the neuron weight, represents the input data sample; like or If true, a new node r is generated for the neuron set A, and ; If two neurons There is no connection between the two neurons Establish connections between them and update the edge collection , and refresh the two neurons The age parameter of the edge between them; Increment the age parameters of all edges connected to the winner node and accumulate the local quantization error of the winner node; If no new node is established, the weights of the two winner nodes are updated; Check the current age parameters of all edges. If , then remove the corresponding edge from the edge set C; If there is more input, the process repeatedly determines two neurons that are similar to the input data sample.

2. A computer device, characterized in that: It includes a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the steps of the parallel simulation system computational experiment evaluation decision method based on deep learning as claimed in claim 1 are implemented.

3. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the deep learning-based parallel simulation system computational experiment evaluation and decision-making method as claimed in claim 1.

Images