Intelligent data optimization processing system

Abstract

The invention belongs to the technical field of data processing, and discloses an intelligent data optimization processing system. Comprising the steps that original target data are collected, and the original target data comprise structured data and unstructured data; performing feature extraction and coding on the original target data by using a pre-constructed feature network model to obtain a low-dimensional feature vector; aiming at the low-dimensional feature vectors, constructing a multi-objective optimization mathematical model; modeling the multi-objective optimization mathematical model into an intelligent agent of a reinforcement learning environment, and solving the intelligent agent to obtain a preliminary optimal strategy; optimizing the preliminary optimal strategy by adopting an optimization algorithm to obtain an optimal strategy; optimizing the low-dimensional feature vector based on an optimal strategy to obtain an optimized low-dimensional feature vector; and the optimized low-dimensional feature vector is decoded to obtain the optimized original target data, so that the internal structure and features of the original data can be maintained, and the optimization result has good interpretability and generalization ability.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and more specifically, to an intelligent data optimization processing system. Background Art

[0002] The patent application with publication number CN105787597A discloses a data optimization processing system, including: a data preprocessing module, used to select a data subset for optimization processing from the business data to be processed; a communication module, used to perform MPS encoding on the data subset selected by the data preprocessing module to obtain an MPS data packet, parse the MPS data packet according to a preset data structure, obtain structured data that meets the optimization processing requirements, and transmit the obtained structured data to a computing module; a computing module, used to optimize the structured data using an intelligent optimization algorithm, and can quickly and accurately find fact-based decisions from massive business data according to user needs, thereby improving business operation efficiency.

[0003] However, the existing data optimization systems are still inefficient when processing multiple types of data, lack adaptability, and cannot process data efficiently. At the same time, directly optimizing the original data requires huge computational complexity, low optimization efficiency, lacks effective dimensionality reduction and feature extraction mechanisms, and cannot reduce data complexity. Secondly, most optimization methods only target a single goal and cannot balance and weigh multiple goals, resulting in one-sided optimization results and difficulty in adapting to the needs of complex application scenarios. Traditional methods are based on certain rules and models, lack autonomous learning and decision-making capabilities, cannot dynamically adjust optimization strategies according to data distribution, are easily affected by noise and outliers, and have poor robustness and reliability of optimization results. In addition, existing data optimization methods can often only obtain local optimality and it is difficult to ensure global optimality. There is a lack of effective global search and optimization mechanisms and they cannot escape local minima. At the same time, traditional algorithms focus on numerical optimization and ignore the interpretability and generalization of the results, resulting in optimization results that are difficult to understand.

[0004] In view of this, the present invention proposes an intelligent data optimization processing system to solve the above problems. Summary of the invention

[0005] In order to overcome the above defects of the prior art and to achieve the above objectives, the present invention provides the following technical solutions: an intelligent data optimization processing system, comprising: a data acquisition module, for acquiring original target data, the original target data including structured data and unstructured data; The feature extraction module is used to extract and encode the original target data using the pre-built feature network model to obtain a low-dimensional feature vector; Model building module, used to build a multi-objective optimization mathematical model for low-dimensional feature vectors; The preliminary strategy fitting module is used to model the multi-objective optimization mathematical model as an intelligent agent in the reinforcement learning environment and solve the intelligent agent to obtain the preliminary optimal strategy; The optimal strategy output module is used to optimize the preliminary optimal strategy using an optimization algorithm to obtain the optimal strategy; based on the optimal strategy, the low-dimensional feature vector is optimized to obtain the optimized low-dimensional feature vector; The decoding module is used to decode the optimized low-dimensional feature vector to obtain the optimized original target data. The modules are connected to each other by wired and / or wireless means.

[0006] Furthermore, the method of constructing the feature network model includes: The basic framework structure of the feature network model is defined as follows: input layer, dense block, transition layer and fully connected layer; the dense block is composed of M convolutional layers, and the transition layer is located between dense blocks; Initialize the parameters of the feature network model to θ_0, divide the original target data into n non-overlapping subsets D_i; where i is the index of the subset; define a task set T, which is initially an empty set; For each subset D_i, randomly sample some data from D_i as the support set; use the remaining data in D_i as the query set; use the set of support set and query set as the task and add it to the task set T; define the task distribution function PU; sample a series of tasks from the task set T according to the distribution PU as the training task T_j, where j is the index of the sampled task; and each training task contains a support set and a query set; For each training task T_j, use the support set h_j in it to perform a small amount of gradient update on θ_0 to obtain the corresponding adaptability parameters; Use the adaptive parameter θ_j to calculate the loss value on the query set d_j, and take the average value as the update factor LK; minimize the update factor by optimizing the initial parameter θ_0 until the feature network model converges; obtain the optimal parameters of the feature network model; fix the optimal parameters of the feature network model, and complete the construction of the feature network model.

[0007] Furthermore, the formula for updating the small amount of gradients is: ; where α is the preset learning rate, x is the data in the support set h_j, is the low-dimensional feature vector output by the feature network model when input x, GH θ_0 () is the decoder network constructed when the parameter is θ_0; β is the weight parameter, H(θ_0,h_j) is the adaptation function, θ_j is the adaptation parameter corresponding to the training task T_j, is the gradient operator with respect to the parameter θ_0; Fitness function ;in, is the balance parameter; is the total loss value of the feature network model on the query set d_j of the training task T_j; ;in, and are two hyperparameters, and , W_j is the weight of the j-th sampled task, and acc_j is the accuracy of the feature network model on the j-th sampled task.

[0008] Furthermore, the multi-objective optimization mathematical model is constructed in the following manner: The multi-objective optimization mathematical model is defined as the multi-objective optimization function JQ of the objective function GH and the constraint conditions; Objective Function ; Where FC(z) is the redundancy function, z is the low-dimensional feature vector, FP(z) is the structural balance function, FL(z) is the generalization function; a1, a2 and a3 are the weight parameters of the corresponding items; Redundant functions ; Wherein, GC(z) means mapping z to an m5-dimensional space using a preset encoder network; ε1 is a redundant parameter, ρ(z_q) is a robust loss function, and z_q is the value of the low-dimensional feature vector z in the qth dimension; Define adaptive threshold parameters for z_q ; Where k'' is the control curvature parameter, and X0 is the control displacement parameter; like , then the robust loss function ;like ,but ;in, is the robust weight parameter; The constraints include limiting the feasible domain and equality constraints. The formula of equality constraints is: ; Where SG is the constrained linear transformation matrix and PJ is the offset vector.

[0009] Furthermore, the method for obtaining the structural balance function includes: All low-dimensional feature vectors form a vector set, randomly divide the vector set into m1 data sets, and use the low-dimensional feature vectors in the data set as samples; construct a structural balance function based on the samples in the data set ; Where trace() represents the trace of the matrix, S_w is the internal scatter matrix in the w-th data set; SE is the heterogeneous scatter matrix; β3 is the structure preservation weight; W' is the linear transformation matrix; Internal Scatter Matrix ; where w_r is the low-dimensional feature vector of the r-th sample in the w-th data set, w_μ is the mean vector in the entire w-th data set, (w_r-w_μ) T is the transpose of (w_r-w_μ); Heterogeneous Scatter Matrix ; Where n_w is the total number of samples in the w-th data set, μ' is the mean vector of the vector set, (w_μ-μ') T is the transpose of (w_μ-μ').

[0010] Furthermore, the formula of the generalization function is: Generalization Function ; Where s_q is the statistical vector of the data on the qth dimension, and the statistical vector is the mean, variance, skewness or kurtosis; τ_q is the weight parameter of the qth dimension, τ1 is the interpretability weight parameter, PA(z_q,s_q) is the adaptive robustness function; FHA(z_q) is the capability function, LPT(z,A) is the interpretability measurement function, and A is the auxiliary information matrix; The adaptive robustness function PA(z_q,s_q) is parameterized as a neural network, the input of which is the value of the qth dimension and the corresponding statistic vector s_q, and the output is a scalar value representing the robustness measure on this dimension; Capability Function ; Where N1 is the total number of samples, B is the bandwidth parameter, and z_q' is the q'th element of the low-dimensional feature vector z, that is, z_q' is the value of z in the q'th dimension, and q and q' do not index the same dimension at the same time.

[0011] Furthermore, the modeling method of the intelligent agent includes: Define a reinforcement learning environment and define the optimization goal of the intelligent agent in the reinforcement learning environment; the reinforcement learning environment includes a state space, an action space, environmental dynamics, and a reward function; the state space is a space of low-dimensional feature vectors, and the action space includes a set of actions corresponding to each optimization operator, and the optimization operator is used to update the low-dimensional feature vector z; The environment dynamics includes updating the low-dimensional feature vector through the optimization operator according to the current state and the selected action to obtain a new state; the reward function is defined as the multi-objective optimization function JQ, and the value calculated by the multi-objective optimization function JQ is used as the reward; The optimization goal is to select the best action based on the current state to maximize the accumulated reward, which is the strategy; The input of the agent is defined as the current state, and the output is the selected action, thus completing the modeling of the agent.

[0012] Furthermore, the solution method of the preliminary optimal strategy includes: Divide the agent into an action network and a comment network; define the input of the action network as the current state and the output as the action probability distribution ,σ_π is the trainable parameter of the action network;π() is the action network; Define the input of the comment network as the current state Z and action a, and the output as the state-action value function HJ=Q(Z,a;σ_Q), which represents the expected cumulative reward of executing action a in state Z, σ_Q is the trainable parameter of the comment network; Q() is the comment network; initialize the trainable parameters σ_π and σ_Q of the action network and comment network; Collect a historical training data set, which includes L' groups of randomly perturbed low-dimensional feature vectors and the optimal optimization operator corresponding to the low-dimensional feature vectors, and form a state-action combination with the low-dimensional feature vectors and the optimal optimization operator corresponding to the low-dimensional feature vectors. Randomly sample low-dimensional feature vectors from the historical training data set, select an action a according to the action probability distribution output by the action network, execute the selected action a on the low-dimensional feature vector, update the sampled low-dimensional feature vector, and obtain a new state Z'; Define the policy error function L_Q(σ_Q), calculate the reward R of the new state Z' and update the trainable parameters σ_Q of the review network to minimize the value of the policy error function; Update the trainable parameters σ_π of the action network to maximize the expected accumulated reward; The sampling training is repeated until the predetermined number of training rounds is reached, and all trainable parameters are fixed. After the training process is completed, the action network learns a preliminary optimal strategy.

[0013] The sampling training is repeated until the predetermined number of training rounds is reached, and all trainable parameters are fixed. After the training process is completed, the action network learns a preliminary optimal strategy.

[0014] Furthermore, the formula of the strategy error function L_Q(σ_Q) is: ; where HT is the new state-action value function; a' is the optimal action selected in the new state Z', is the discount factor, and its value range is [0, 1]; τ3 is the target hyperparameter, and fL(Z, a) is the state transition model; New state-action value function ; , where Z_v is the low-dimensional feature vector in the v-th state-action combination sampled from the historical training data set, a_v is the optimization operator in the v-th state-action combination sampled from the historical training data set; ω1 is a positive parameter; α_v is the weight coefficient of the v-th state-action combination sampled from the historical training data set; The formula for updating the trainable parameter σ_π of the action network is: ;in, is the gradient of the multi-objective optimization function JQ with respect to σ_π, For hope, is the weight parameter of the auxiliary head, C' is the conditional information, σ_H is the trainable parameter of the auxiliary head; H(Z,a,C';σ_H) is the auxiliary head; ; Where W_Z is the state-action weight matrix, W_C' is the conditional information-action weight matrix; b1 is the bias term; Z T is the transpose of state Z, C' T is the transposition of the conditional information C'.

[0015] Furthermore, the method for obtaining the optimal strategy includes: Define the initial frog population, represent the initial optimal strategy as a vector π', and use it as the first frog in the initial frog population. Randomly generate K-1 frogs in the neighborhood of the initial optimal strategy to form the initial frog population P. The size of the initial frog population P is K. For each frog l, according to its strategy π_l, it interacts with the agent in the reinforcement learning environment and obtains the cumulative reward R_l of strategy π_l; Define the frog evolution function fit(l)=fs(R_l), where fs(R_l) is a monotonically increasing function of R_l; For each frog l, calculate the value of the frog evolution function, recorded as the frog evolution degree; sort each frog l from large to small according to the frog evolution degree, the top U% frogs form population X, and the remaining (1-X)% frogs form population Y; For each frog j' in population Y, a frog i' is randomly selected from population X, and a new frog is generated. ; Among them, ∈2 is an adaptive parameter; τl is a noise intensity vector; t(u') is a random vector that obeys the Student's t distribution, u' is a degree of freedom parameter; ρ' is a step size parameter; l_gb is the vector corresponding to the frog with the highest degree of evolution in the entire initial frog population P, l_wa is the average vector of the entire initial frog population, is the element-by-element multiplication symbol; If the frog evolution degree of the new frog j'' is greater than or equal to j', then j' is replaced by j''; then the improved population Y' is obtained; the iteration is repeated until the preset maximum number of iterations is reached; after the iteration is completed, the strategy corresponding to the frog with the highest degree of frog evolution is selected from the frog population as the optimal strategy.

[0016] The technical effects and advantages of an intelligent data optimization processing system of the present invention are as follows: The present invention can process data efficiently and realize multi-objective optimization, has strong data adaptability, and can optimize structured data and unstructured data at the same time; through intelligent modeling and feature extraction, the system maps high-dimensional original data to low-dimensional feature space, which not only reduces data complexity, but also lays the foundation for subsequent optimization; secondly, a multi-objective optimization model is introduced, which can balance and coordinate multiple optimization goals to meet the complex needs in real scenarios; the intelligent optimization algorithm based on reinforcement learning can independently explore and discover the optimal strategy, which significantly improves the optimization efficiency and accuracy; it also integrates bio-inspired algorithms to further optimize the preliminary optimization results to ensure that the global optimal solution is obtained and the high quality of the optimization results is ensured; the inherent structure and characteristics of the original data can be maintained, so that the optimization results have good interpretability and generalization ability; at the same time, it has adaptive ability and can dynamically adjust the optimization strategy according to the data distribution to improve robustness and reliability; finally, the optimized feature vector is decoded into the optimized original data to complete the end-to-end optimization processing flow. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A schematic diagram of an intelligent data optimization processing system of the present invention; Figure 2 It is a schematic diagram of an intelligent data optimization processing method of the present invention. DETAILED DESCRIPTION

[0018] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] Example 1

[0020] See also Figure 1 As shown, the intelligent data optimization processing system described in this embodiment includes: A data acquisition module is used to acquire original target data, which includes structured data and unstructured data; The feature extraction module is used to extract and encode the original target data using the pre-built feature network model to obtain a low-dimensional feature vector; Model building module, used to build a multi-objective optimization mathematical model for low-dimensional feature vectors; The preliminary strategy fitting module is used to model the multi-objective optimization mathematical model as an intelligent agent in the reinforcement learning environment and solve the intelligent agent to obtain the preliminary optimal strategy; The optimal strategy output module is used to optimize the preliminary optimal strategy using an optimization algorithm to obtain the optimal strategy; based on the optimal strategy, the low-dimensional feature vector is optimized to obtain the optimized low-dimensional feature vector; The decoding module is used to decode the optimized low-dimensional feature vector to obtain the optimized original target data. The modules are connected by wire and / or wireless means to realize data transmission between modules.

[0021] The structured data refers to data with a predefined format or inherent structure. The structure of this data is fixed, and each data entity occupies a specific field or column; unstructured data refers to data without a predefined data model, and the data does not follow a specific data structure or pattern.

[0022] Furthermore, the construction methods of the feature network model include: The basic framework structure of the feature network model is defined as follows: input layer, dense block, transition layer and fully connected layer; the shape of the input layer is set according to the size of the original target data; the dense block is composed of M convolutional layers; the transition layer is located between dense blocks, and reduces the dimension of the features through convolution and pooling operations to control the number of model parameters; the fully connected layer maps the features extracted by the dense block to a low-dimensional feature vector space, and the number of neurons in the fully connected layer determines the dimension of the feature vector.

[0023] Initialize the parameters of the feature network model to θ_0, divide the original target data into n non-overlapping subsets D_i; where i is the index of the subset; define a task set T, which is initially an empty set; For each subset D_i, randomly sample some data from D_i as the support set; use the remaining data in D_i as the query set; use the set of support set and query set as the task and add it to the task set T; define the task distribution function PU; sample a series of tasks from the task set T according to the distribution PU as the training task T_j, where j is the index of the sampled task; and each training task contains a support set and a query set; ;in, and are two hyperparameters, and , used to control the weight of each distribution, W_j is the weight of the j-th sampled task, acc_j is the accuracy of the feature network model on the j-th sampled task (based on the value of the loss function, the accuracy of the task on the query set can be calculated, and the loss function can be a cross entropy loss function); the weight of the task is taken as the parameter to be estimated, and the Bayesian method and historical data are used to perform a posteriori inference on the weight to obtain the optimal weight value for each task.

[0024] For each training task T_j, a small amount of gradient update is performed on θ_0 using the support set h_j within it to obtain the corresponding adaptability parameters.

[0025] The formula for a small gradient update is: ; where α is the preset learning rate, x is the data in the support set h_j, is the low-dimensional feature vector output by the feature network model when input x, GH θ_0 () is the decoder network constructed when the parameter is θ_0, which maps the low-dimensional feature vector back to the original data space and reconstructs the data corresponding to the original input data x; β is the weight parameter, H(θ_0,h_j) is the fitness function, which measures the adaptability of the model on the support set and encourages the model to quickly adapt to new tasks; θ_j is the fitness parameter corresponding to the training task T_j, is the gradient operator with respect to the parameter θ_0; Fitness function ;in, is a balancing parameter used to control the weight, and its value range is [0, 1]; is the total loss value of the feature network model on the query set d_j of the training task T_j (obtained based on the loss function, the loss function is defined as the cross entropy loss function).

[0026] Use the adaptive parameter θ_j to calculate the loss value on the query set d_j, and take the average value as the update factor LK; minimize the update factor by optimizing the initial parameter θ_0 until the feature network model converges; and obtain the optimal parameters of the feature network model.

[0027] The optimized formula is: ; where ε is the meta-learning rate, is a coordination parameter used to control the weight of the coordination term; C(θ_0) is a learning coordination function used to coordinate parameter updates between different tasks to improve consistency and stability.

[0028] ; Among them, CoordNet() is the coordination network function, the input is the parameter θ_0 and a set θ'_task consisting of a set of adaptive parameters, and the output is a scalar value, which represents the amount of parameter update after coordination; the coordination network function is obtained by parameterizing the coordination network, and the coordination network consists of multiple fully connected layers, convolutional layers, attention layers, etc., and its structure is tuned according to actual conditions; ω1 is the trainable parameter of the coordination network function; the set θ'_task represents the adaptive parameters obtained by gradient updating on the support set of each task based on the parameter θ_0.

[0029] The optimal parameters of the feature network model are fixed, and the original target data is input into the input layer of the feature network model at this time. The feature representation of the original target data is extracted through dense blocks and transition layers, and the extracted feature representation is input into the fully connected layer to output a low-dimensional feature vector.

[0030] Furthermore, the construction methods of the multi-objective optimization mathematical model include: The multi-objective optimization mathematical model is defined as the multi-objective optimization function JQ of the objective function GH and the constraint conditions; Objective Function ; Where FC(z) is the redundancy function, z is the low-dimensional feature vector, FP(z) is the structural balance function, FL(z) is the generalization function; a1, a2 and a3 are the weight parameters of the corresponding items; Redundant functions ; Among them, GC(z) means using a preset encoder network (such as autoencoder, variational autoencoder, etc.) to map z to a lower dimensional space; ε1 is a redundancy parameter used to control the weight of de-redundancy and noise influence; ρ(z_q) is a robust loss function used to reduce the influence of noise and outliers on the feature vector; z_q is the qth element of the low-dimensional feature vector z, that is, z_q is the value of z in the qth dimension.

[0031] Define adaptive threshold parameters for z_q , which is used to control the smoothness of the function; k'' is the control curvature parameter. The larger the value of k'' is, the steeper the function curve changes near X0; the smaller the value of k'' is, the smoother the function curve changes; usually k'' is a positive value; X0 is the control displacement parameter, which determines the horizontal position of the function curve.

[0032] like , then the robust loss function ;like ,but ;in, is a robust weight parameter used to control the weight and further enhance the robustness to outliers and the generalization ability.

[0033] All low-dimensional feature vectors form a vector set, randomly divide the vector set into m1 data sets, and use the low-dimensional feature vectors in the data set as samples; construct a structural balance function based on the samples in the data set ; Where trace() represents the trace of the matrix, i.e., the sum of the diagonal elements of the matrix, S_w is the internal scatter matrix within the w-th data set; SE is the heterogeneous scatter matrix; β3 is the structure preservation weight, which is used to control the weight of preserving the structure; W' is the linear transformation matrix (PCA is performed on the data to obtain the coefficient matrix of the first several principal components, which is used as the initial linear transformation matrix), which is used to map the original target data z' to a low-dimensional space; ||z-W'·z'|| 2 The goal is to keep the structural information of the original data as much as possible during the dimensionality reduction process, that is, to minimize the difference between the low-dimensional feature vectors z and z'.

[0034] Internal Scatter Matrix ; where w_r is the low-dimensional feature vector of the r-th sample in the w-th data set, w_μ is the mean vector in the entire w-th data set, (w_r-w_μ) T is the transpose of (w_r-w_μ); Heterogeneous Scatter Matrix ; Where n_w is the total number of samples in the w-th data set, μ' is the mean vector of the vector set, (w_μ-μ') T is the transpose of (w_μ-μ').

[0035] Generalization Function ; Wherein, s_q is the statistical vector of the data on the qth dimension, which describes the statistical characteristics of the data on this dimension. The statistical vector is the mean, variance, skewness or kurtosis; τ_q is the weight parameter of the qth dimension, which is used to control the weights of the robustness term and the generalization term; τ1 is the interpretability weight parameter, which is used to control the weight of the interpretability measurement term; PA(z_q,s_q) is an adaptive robustness function, which depends not only on the value of the qth dimension, but also on the statistical vector of this dimension, so that the robustness strategy can be adaptively adjusted according to the data distribution; FHA(z_q) is a capability function, which is used to improve the generalization ability of the feature vector on the qth dimension, LPT(z,A) is an interpretability measurement function, A is an auxiliary information matrix, and the interpretability measurement function depends not only on the feature vector z, but also on an auxiliary information matrix, which is used to improve the interpretability of the feature vector.

[0036] The adaptive robustness function PA(z_q,s_q) is parameterized as a neural network (such as a convolutional neural network, a full neural network, etc.). The input of the neural network is the value of the qth dimension and the corresponding statistic vector s_q, and the output is a scalar value representing the robustness measure on this dimension.

[0037] Capability Function ; Where N1 is the total number of samples, B is the bandwidth parameter that controls the smoothness of the function, z_q' is the q'th element of the low-dimensional feature vector z, that is, z_q' is the value of z in the q'th dimension, and q and q' do not index the same element (dimension) at the same time.

[0038] The interpretability metric function is parameterized as an attention mechanism, which takes z and the auxiliary information matrix A as input, learns the relationship between the two through the attention mechanism, and outputs a scalar value representing the interpretability metric of the feature vector. The auxiliary information matrix A contains the interpretability information related to the original data, and the interpretability information includes semantic labels and structural information.

[0039] The constraints include restricted feasible domain and equality constraints; the restricted feasible domain is defined as a simple set of constraints, such as a hypersphere or a hypercube.

[0040] The formula for the equality constraint is: ; Among them, SG is the constrained linear transformation matrix (it is parameterized as a trainable matrix and trained according to the constraints during the optimization process), which is used to map z to the space where the constraints are located, and PJ is the offset vector; based on the linear transformation, the result is translated and offset so that the final constraint mapping result can meet the constraints.

[0041] It should be noted that semantic labels refer to semantic information or annotations related to the original data, such as topic tags for text data; structural information refers to the inherent structural features of the original data, such as the grammar and sentence structure of the text; it helps to explain the connotation represented by the feature vector.

[0042] Further, the modeling methods of the intelligent agent include: Define a reinforcement learning environment and define the optimization goal of the intelligent agent in the reinforcement learning environment; the reinforcement learning environment includes state space, action space, environment dynamics and reward function; the state space is the space of low-dimensional feature vectors, and the action space includes a set of actions corresponding to each optimization operator. The optimization operators such as gradient descent, Newton's method, conjugate gradient, etc. are used to update the low-dimensional feature vector z.

[0043] The environmental dynamics include updating the low-dimensional feature vector through the optimization operator according to the current state (low-dimensional feature vector z) and the selected action (optimization operator) to obtain a new state; the reward function is defined as a multi-objective optimization function, and the value calculated by the multi-objective optimization function JQ is used as the reward. The higher the reward, the better the low-dimensional feature vector z.

[0044] The optimization goal is to select the best action (optimization operator) based on the current state (low-dimensional feature vector z) to maximize the accumulated reward, which is the strategy; The input of the agent is defined as the current state (low-dimensional feature vector z), and the output is the selected action (optimization operator), thus completing the modeling of the agent.

[0045] Furthermore, the solution to the initial optimal strategy includes: The intelligent agent is divided into an action network and a comment network; the action network selects actions based on the current state, and the comment network is used to evaluate the value function of the combination of the current state and action, guiding the action network to update the strategy.

[0046] Define the input of the action network as the current state (low-dimensional feature vector z), and the output as the action probability distribution , represents the probability of selecting action a in state Z, σ_π is the trainable parameter of the action network; π() is the action network.

[0047] Define the input of the comment network as the current state Z and action a, and the output as the state-action value function HJ=Q(Z,a;σ_Q), which represents the expected cumulative reward of performing action a in state Z. σ_Q is a trainable parameter of the comment network; Q() is the comment network; Initialize the trainable parameters σ_π and σ_Q of the action network and the critic network, usually using random initialization.

[0048] Collect a historical training data set, which includes L' groups of randomly perturbed low-dimensional feature vectors and the optimal optimization operator corresponding to the low-dimensional feature vectors, and form a state-action combination with the low-dimensional feature vectors and the optimal optimization operator corresponding to the low-dimensional feature vectors. Randomly sample low-dimensional feature vectors from the historical training data set, select an action a according to the action probability distribution output by the action network, execute the selected action a (optimization operator) on the low-dimensional feature vector, update the sampled low-dimensional feature vector, and obtain a new state Z'; Define the policy error function L_Q(σ_Q), calculate the reward R of the new state Z' and update the trainable parameters σ_Q of the review network to minimize the value of the policy error function.

[0049] ; where HT is the new state-action value function; a' is the optimal action selected in the new state Z', is the discount factor, with a value range of [0, 1]; τ3 is the target hyperparameter, with a value range of [0, 1], which is used to control the weight; fL(Z,a) is the state transition model; it predicts the probability of the next state (new state Z') based on the current state Z and action a.

[0050] , which represents the state-action value function of the target network to select the optimal action a' in the new state Z'.

[0051] , where Z_v is the low-dimensional feature vector in the v-th state-action combination sampled from the historical training data set, a_v is the optimization operator in the v-th state-action combination sampled from the historical training data set; ω1 is a positive parameter used to control the decay rate of similarity; α_v is the weight coefficient of the v-th state-action combination sampled from the historical training data set.

[0052] Update the trainable parameters σ_π of the action network to maximize the expected accumulated reward. The formula for updating the trainable parameters σ_π of the action network is: ;in, is the gradient of the multi-objective optimization function JQ with respect to σ_π, For hope, is the weight parameter of the auxiliary head, C' is the conditional information, σ_H is the trainable parameter of the auxiliary head; H(Z,a,C';σ_H) is the auxiliary head; the conditional information is metadata related to the original target data, expressed in matrix form.

[0053] ; Where W_Z is the state-action weight matrix, which is used to capture the interaction between state Z and action a, W_C' is the conditional information-action weight matrix, which is used to capture the interaction between conditional information C' and action a; b1 is the bias term, which is a scalar; Z T is the transpose of state Z, C' T is the transposition of the conditional information C'.

[0054] It should be noted that σ_H={W_Z,W_C',b1} are all trainable parameters of the auxiliary head.

[0055] Repeat the sampling training until the predetermined number of training rounds is reached, and all trainable parameters are fixed. After the training process is completed, the action network learns a preliminary optimal strategy. The preliminary optimal strategy selects the optimal action (optimization operator) according to the current state, that is, obtains the optimal feature vector representation.

[0056] Furthermore, the method for obtaining the optimal strategy includes: Define the initial frog population, represent the initial optimal strategy as a vector π', and use it as the first frog in the initial frog population (initial optimal strategy). Randomly generate K-1 frogs (other strategies) in the neighborhood of the initial optimal strategy (an area obtained by setting a neighborhood radius) to form the initial frog population P. The size of the initial frog population P is K. The random generation method is to generate a Gaussian noise vector and add the vector π' to the Gaussian noise vector.

[0057] For each frog l, according to its strategy π_l, it interacts with the agent in the reinforcement learning environment and obtains the cumulative reward R_l of strategy π_l; Define the frog evolution function fit(l)=fs(R_l), where fs(R_l) is a monotonically increasing function of R_l, for example, fs(R_l)=R_l or other monotonically increasing function forms.

[0058] For each frog l, calculate the value of the frog evolution function, recorded as the frog evolution degree; sort each frog l from large to small according to the frog evolution degree, the first U% of frogs constitute population X, and the remaining (1-X)% of frogs constitute population Y; control the ratio of populations X and Y, usually U<50.

[0059] For each frog j' in population Y, a frog i' is randomly selected from population X, and a new frog is generated. ; Among them, ∈2 is an adaptive parameter; τl is a noise intensity vector (trainable); t(u') is a random vector that obeys the Student's t distribution, u' is a degree of freedom parameter that controls the tail characteristics of the distribution; ρ' is a step size parameter that controls the step size of the displacement; l_gb is the vector corresponding to the frog with the highest degree of evolution in the entire initial frog population P, l_wa is the average vector of the entire initial frog population, is the element-wise multiplication symbol.

[0060] If the frog evolution degree of the new frog j'' is greater than or equal to j', then j' is replaced by j''; then the improved population Y' is obtained; the iteration is repeated until the preset maximum number of iterations is reached; after the iteration is completed, the strategy corresponding to the frog with the highest degree of frog evolution is selected from the frog population as the optimal strategy.

[0061] The method of optimizing the low-dimensional feature vector is to follow the guidance of the optimal strategy and execute the corresponding optimization operator to update the low-dimensional feature vector until the preset maximum number of optimization steps T_max is reached.

[0062] The process of decoding the optimized low-dimensional feature vector is as follows: define a decoder model Dec, map the optimized low-dimensional feature vector back to the space of the original target data, and reconstruct the optimized original target data corresponding to the original target data. The decoder model Dec is a neural network composed of a fully connected layer and a convolutional layer. The parameters σ_Dec of the decoder model are obtained through pre-training.

[0063] This embodiment can process data efficiently and realize multi-objective optimization, has strong data adaptability, and can optimize structured data and unstructured data at the same time; through intelligent modeling and feature extraction, the system maps high-dimensional raw data to low-dimensional feature space, which not only reduces data complexity, but also lays the foundation for subsequent optimization; secondly, a multi-objective optimization model is introduced, which can balance and coordinate multiple optimization goals to meet the complex needs in real scenarios; the intelligent optimization algorithm based on reinforcement learning can independently explore and discover the optimal strategy, which significantly improves the optimization efficiency and accuracy; it also integrates biological inspiration algorithms to further optimize the preliminary optimization results to ensure that the global optimal solution is obtained and the high quality of the optimization results is ensured; it can maintain the inherent structure and characteristics of the original data, so that the optimization results have good interpretability and generalization capabilities; at the same time, it has adaptive capabilities and can dynamically adjust the optimization strategy according to the data distribution to improve robustness and reliability; finally, the optimized feature vector is decoded into the optimized original data to complete the end-to-end optimization processing flow.

[0064] Example 2

[0065] See also Figure 2 As shown, the part not described in detail in this embodiment is described in Example 1, which provides an intelligent data optimization processing method, including: S1. Collect original target data, which includes structured data and unstructured data; S2, using the pre-built feature network model to extract and encode the original target data to obtain a low-dimensional feature vector; S3. Construct a multi-objective optimization mathematical model for low-dimensional feature vectors; S4, model the multi-objective optimization mathematical model as an intelligent agent in a reinforcement learning environment, and solve the intelligent agent to obtain a preliminary optimal strategy; S5. Optimize the preliminary optimal strategy using an optimization algorithm to obtain an optimal strategy; optimize the low-dimensional feature vector based on the optimal strategy to obtain an optimized low-dimensional feature vector; S6. Decode the optimized low-dimensional feature vector to obtain the optimized original target data.

[0066] Example 3

[0067] This embodiment discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the operation mode of the intelligent data optimization processing method provided above is implemented.

[0068] Since the electronic device introduced in this embodiment is an electronic device used to implement an intelligent data optimization processing method in the embodiment of the present application, based on the intelligent data optimization processing method introduced in the embodiment of the present application, the technical personnel of the field can understand the specific implementation of the electronic device of the present embodiment and its various variations, so how the electronic device implements the method in the embodiment of the present application is not described in detail here. As long as the technical personnel of the field implement the electronic device used in the intelligent data optimization processing method in the embodiment of the present application, it belongs to the scope of protection of this application.

[0069] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters and thresholds in the formula are set by technicians in this field according to actual conditions.

[0070] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technical users in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. An intelligent data optimization processing system, characterized in that: include: A data acquisition module is used to acquire original target data, which includes structured data and unstructured data; The feature extraction module is used to extract and encode the original target data using the pre-built feature network model to obtain a low-dimensional feature vector; Model building module, used to build a multi-objective optimization mathematical model for low-dimensional feature vectors; The preliminary strategy fitting module is used to model the multi-objective optimization mathematical model as an intelligent agent in the reinforcement learning environment and solve the intelligent agent to obtain the preliminary optimal strategy; The optimal strategy output module is used to optimize the preliminary optimal strategy using an optimization algorithm to obtain the optimal strategy; based on the optimal strategy, the low-dimensional feature vector is optimized to obtain the optimized low-dimensional feature vector; The decoding module is used to decode the optimized low-dimensional feature vector to obtain the optimized original target data. The modules are connected to each other by wired and / or wireless means.

2. The intelligent data optimization processing system according to claim 1 is characterized in that: The construction method of the feature network model includes: The basic framework structure of the feature network model is defined as follows: input layer, dense block, transition layer and fully connected layer; the dense block is composed of M convolutional layers, and the transition layer is located between dense blocks; Initialize the parameters of the feature network model to θ_0, divide the original target data into n non-overlapping subsets D_i; where i is the index of the subset; define a task set T, which is initially an empty set; For each subset D_i, randomly sample some data from D_i as the support set; use the remaining data in D_i as the query set; use the set of support set and query set as the task and add it to the task set T; define the task distribution function PU; sample a series of tasks from the task set T according to the distribution PU as the training task T_j, where j is the index of the sampled task; and each training task contains a support set and a query set; For each training task T_j, use the support set h_j in it to perform a small amount of gradient update on θ_0 to obtain the corresponding adaptability parameters; Use the adaptive parameter θ_j to calculate the loss value on the query set d_j, and take the average value as the update factor LK; minimize the update factor by optimizing the initial parameter θ_0 until the feature network model converges; obtain the optimal parameters of the feature network model; fix the optimal parameters of the feature network model, and complete the construction of the feature network model.

3. The intelligent data optimization processing system according to claim 2 is characterized in that: The formula for the small amount of gradient update is: ; where α is the preset learning rate, x is the data in the support set h_j, is the low-dimensional feature vector output by the feature network model when input x, GH θ_0 () is the decoder network constructed when the parameter is θ_0; β is the weight parameter, H(θ_0,h_j) is the adaptation function, θ_j is the adaptation parameter corresponding to the training task T_j, is the gradient operator with respect to the parameter θ_0; Fitness function ;in, is the balance parameter; is the total loss value of the feature network model on the query set d_j of the training task T_j; ;in, and are two hyperparameters, and , W_j is the weight of the j-th sampled task, and acc_j is the accuracy of the feature network model on the j-th sampled task.

4. The intelligent data optimization processing system according to claim 3 is characterized in that: The multi-objective optimization mathematical model is constructed in the following manner: The multi-objective optimization mathematical model is defined as the multi-objective optimization function JQ of the objective function GH and the constraint conditions; Objective Function ; Where FC(z) is the redundancy function, z is the low-dimensional feature vector, FP(z) is the structural balance function, FL(z) is the generalization function; a1, a2 and a3 are the weight parameters of the corresponding items; Redundant functions ; Wherein, GC(z) means mapping z to an m5-dimensional space using a preset encoder network; ε1 is a redundant parameter, ρ(z_q) is a robust loss function, and z_q is the value of the low-dimensional feature vector z in the qth dimension; Define adaptive threshold parameters for z_q ; Where k'' is the control curvature parameter, and X0 is the control displacement parameter; like , then the robust loss function ;like ,but ;in, is the robust weight parameter; The constraints include limiting the feasible domain and equality constraints. The formula of equality constraints is: ; Where SG is the constrained linear transformation matrix and PJ is the offset vector.

5. The intelligent data optimization processing system according to claim 4 is characterized in that: The method for obtaining the structural balance function includes: All low-dimensional feature vectors form a vector set, randomly divide the vector set into m1 data sets, and use the low-dimensional feature vectors in the data set as samples; construct a structural balance function based on the samples in the data set ; Where trace() represents the trace of the matrix, S_w is the internal scatter matrix in the w-th data set; SE is the heterogeneous scatter matrix; β3 is the structure preservation weight; W' is the linear transformation matrix; Internal Scatter Matrix ; where w_r is the low-dimensional feature vector of the r-th sample in the w-th data set, w_μ is the mean vector in the entire w-th data set, (w_r-w_μ) T is the transpose of (w_r-w_μ); Heterogeneous Scatter Matrix ; Where n_w is the total number of samples in the w-th data set, μ' is the mean vector of the vector set, (w_μ-μ') T is the transpose of (w_μ-μ').

6. The intelligent data optimization processing system according to claim 5, characterized in that: The formula of the generalization function is: Generalization Function ; Where s_q is the statistical vector of the data on the qth dimension, and the statistical vector is the mean, variance, skewness or kurtosis; τ_q is the weight parameter of the qth dimension, τ1 is the interpretability weight parameter, PA(z_q,s_q) is the adaptive robustness function; FHA(z_q) is the capability function, LPT(z,A) is the interpretability measurement function, and A is the auxiliary information matrix; The adaptive robustness function PA(z_q,s_q) is parameterized as a neural network, the input of which is the value of the qth dimension and the corresponding statistic vector s_q, and the output is a scalar value representing the robustness measure on this dimension; Capability Function ; Where N1 is the total number of samples, B is the bandwidth parameter, and z_q' is the q'th element of the low-dimensional feature vector z, that is, z_q' is the value of z in the q'th dimension, and q and q' do not index the same dimension at the same time.

7. The intelligent data optimization processing system according to claim 6, characterized in that: The modeling method of the intelligent agent includes: Define a reinforcement learning environment and define the optimization goal of the intelligent agent in the reinforcement learning environment; the reinforcement learning environment includes a state space, an action space, environmental dynamics, and a reward function; the state space is a space of low-dimensional feature vectors, and the action space includes a set of actions corresponding to each optimization operator, and the optimization operator is used to update the low-dimensional feature vector z; The environment dynamics includes updating the low-dimensional feature vector through the optimization operator according to the current state and the selected action to obtain a new state; the reward function is defined as the multi-objective optimization function JQ, and the value calculated by the multi-objective optimization function JQ is used as the reward; The optimization goal is to select the best action based on the current state to maximize the accumulated reward, which is the strategy; The input of the agent is defined as the current state, and the output is the selected action, thus completing the modeling of the agent.

8. The intelligent data optimization processing system according to claim 7, characterized in that: The solution method of the preliminary optimal strategy includes: Divide the agent into an action network and a comment network; define the input of the action network as the current state and the output as the action probability distribution ,σ_π is the trainable parameter of the action network;π() is the action network; Define the input of the comment network as the current state Z and action a, and the output as the state-action value function HJ=Q(Z,a;σ_Q), which represents the expected cumulative reward of executing action a in state Z, σ_Q is the trainable parameter of the comment network; Q() is the comment network; initialize the trainable parameters σ_π and σ_Q of the action network and comment network; Collect a historical training data set, which includes L' groups of randomly perturbed low-dimensional feature vectors and the optimal optimization operator corresponding to the low-dimensional feature vectors, and form a state-action combination with the low-dimensional feature vectors and the optimal optimization operator corresponding to the low-dimensional feature vectors. Randomly sample low-dimensional feature vectors from the historical training data set, select an action a according to the action probability distribution output by the action network, execute the selected action a on the low-dimensional feature vector, update the sampled low-dimensional feature vector, and obtain a new state Z'; Define the policy error function L_Q(σ_Q), calculate the reward R of the new state Z' and update the trainable parameters σ_Q of the review network to minimize the value of the policy error function; Update the trainable parameters σ_π of the action network to maximize the expected accumulated reward; The sampling training is repeated until the predetermined number of training rounds is reached, and all trainable parameters are fixed. After the training process is completed, the action network learns a preliminary optimal strategy.

9. The intelligent data optimization processing system according to claim 8, characterized in that: The formula of the strategy error function L_Q(σ_Q) is: ; where HT is the new state-action value function; a' is the optimal action selected in the new state Z', is the discount factor, and its value range is [0, 1]; τ3 is the target hyperparameter, and fL(Z, a) is the state transition model; New state-action value function ; , where Z_v is the low-dimensional feature vector in the v-th state-action combination sampled from the historical training data set, a_v is the optimization operator in the v-th state-action combination sampled from the historical training data set; ω1 is a positive parameter; α_v is the weight coefficient of the v-th state-action combination sampled from the historical training data set; The formula for updating the trainable parameter σ_π of the action network is: ;in, is the gradient of the multi-objective optimization function JQ with respect to σ_π, For hope, is the weight parameter of the auxiliary head, C' is the conditional information, σ_H is the trainable parameter of the auxiliary head; H(Z,a,C';σ_H) is the auxiliary head; ; Where W_Z is the state-action weight matrix, W_C' is the conditional information-action weight matrix; b1 is the bias term; Z T is the transpose of state Z, C' T is the transposition of the conditional information C'.

10. The intelligent data optimization processing system according to claim 9, characterized in that: The method for obtaining the optimal strategy includes: Define the initial frog population, represent the initial optimal strategy as a vector π', and use it as the first frog in the initial frog population. Randomly generate K-1 frogs in the neighborhood of the initial optimal strategy to form the initial frog population P. The size of the initial frog population P is K. For each frog l, according to its strategy π_l, it interacts with the agent in the reinforcement learning environment and obtains the cumulative reward R_l of strategy π_l; Define the frog evolution function fit(l)=fs(R_l), where fs(R_l) is a monotonically increasing function of R_l; For each frog l, calculate the value of the frog evolution function, recorded as the frog evolution degree; sort each frog l from large to small according to the frog evolution degree, the top U% frogs form population X, and the remaining (1-X)% frogs form population Y; For each frog j' in population Y, a frog i' is randomly selected from population X, and a new frog is generated. ; Among them, ∈2 is an adaptive parameter; τl is a noise intensity vector; t(u') is a random vector that obeys the Student's t distribution, u' is a degree of freedom parameter; ρ' is a step size parameter; l_gb is the vector corresponding to the frog with the highest degree of evolution in the entire initial frog population P, l_wa is the average vector of the entire initial frog population, is the element-by-element multiplication symbol; If the frog evolution degree of the new frog j'' is greater than or equal to j', then j' is replaced by j''; then the improved population Y' is obtained; the iteration is repeated until the preset maximum number of iterations is reached; after the iteration is completed, the strategy corresponding to the frog with the highest degree of frog evolution is selected from the frog population as the optimal strategy.

Images