Linear target combination optimization problem solving model training configuration method and related device

Abstract

This invention discloses a method and related equipment for training and configuring a solution model for a linear objective combinatorial optimization problem. The method includes: performing linear validation processing on user-input natural language data to generate an extended quintuple dataset; performing uncertainty classification processing on the data in the extended quintuple dataset to determine learning paradigm configuration data; determining training component configuration data for the solution model based on the extended quintuple dataset and the learning paradigm configuration data; determining the loss function for the solution model based on the learning paradigm configuration data and the training component configuration data; constructing and training the solution model based on the learning paradigm configuration data, the training component configuration data, and the loss function; and outputting combinatorial optimization decision results based on the trained solution model. Through the above steps, the barriers to manual modeling and training configuration can be reduced, and the accuracy, stability, and efficiency of combinatorial optimization decision output can be improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent detection, and in particular to a method, apparatus, and computer equipment for training and configuring a model for solving a linear objective combinatorial optimization problem. Background Technology

[0002] Combinatorial optimization problems are widely found in industrial scenarios such as logistics scheduling, energy management, financial decision-making, manufacturing scheduling, and route planning. These problems typically require finding the optimal decision result for the objective function within a feasible region that satisfies constraints. With the development of machine learning technology, combinatorial optimization machine learning methods can utilize historical data, contextual features, and uncertain parameters to learn the solution model, enabling the model to output corresponding combinatorial optimization decisions when faced with new problem instances. For combinatorial optimization problems with linear objective functions, training the solution model usually requires first defining the user problem, problem type, objective function structure, optimization model, learning signal, and available data state, and then configuring the training paradigm, model components, and loss function accordingly.

[0003] Existing machine learning training processes for combinatorial optimization problems demand a high level of expertise from users. They typically require manual translation of the problem from natural language into a mathematical model, followed by further steps such as determining whether the objective function satisfies the linear objective condition, identifying uncertain parameters and supervision signals in the data, selecting an appropriate learning paradigm, configuring the encoder and differentiable combinatorial optimization layers, and determining the loss function for training. If this configuration relies heavily on human experience, problems can easily arise, including inconsistencies between the problem description and the optimization model, inaccurate judgment of uncertain data states, mismatch between training strategies and learning signals, incompatibility between model components and the solution interface, and loss functions that fail to effectively reflect decision quality. These issues negatively impact the training accuracy, stability, and decision output efficiency of models for solving linear objective combinatorial optimization problems.

[0004] Therefore, it is urgent to address how to improve the automation, accuracy, and stability of training configuration for solving linear objective combinatorial optimization problems, so as to reduce the difficulty for non-professional users in performing combinatorial optimization machine learning modeling and training configuration. Summary of the Invention

[0005] Therefore, it is necessary to provide a method, apparatus, computer equipment, and readable storage medium for training and configuring a model to solve the above-mentioned technical problems, so as to solve the problems of high professional threshold and low accuracy of training strategies and model component configuration in the existing combinatorial optimization machine learning training and configuration process.

[0006] A method for training and configuring a model for solving a linear objective combinatorial optimization problem, the method comprising: Linear validation processing is performed on the natural language data input by the user to determine the extended quintuple dataset corresponding to the natural language data. The extended quintuple dataset includes user query information, question type label information, learning signal identification information, objective function linear structure information, and optimization model information. Uncertainty classification processing is performed on the data in the extended quintuple dataset to determine the learning paradigm configuration data used for the training strategy configuration of the solution model; Based on the extended quintuple dataset and the learning paradigm configuration data, the training component configuration data for solving the model is determined, including encoder configuration data and differentiable layer configuration data. Based on the learning paradigm configuration data and the training component configuration data, determine the loss function used to solve the model training. Based on the learning paradigm configuration data, the training component configuration data, and the loss function, a solution model is constructed and trained to obtain a trained solution model. The natural language data is input into the trained solution model, which outputs the combinatorial optimization decision result corresponding to the linear objective combinatorial optimization problem.

[0007] Optionally, the step of performing linear validation on the natural language data input by the user to determine the extended quintuple dataset corresponding to the natural language data includes: Semantic features are extracted from the natural language data to obtain semantic feature elements, which include decision variables, objective functions, constraints, feasible regions, and problem types. Based on the objective function, determine the objective function parameters and the corresponding objective function expression; Linearity detection is performed on the nonlinear terms in the objective function expression to obtain the objective function verification result; Linearity testing is performed on the solution interface corresponding to the natural language data to determine that the solution interface uses the objective function parameters as linear objective coefficients and outputs the decision variables that make the objective function optimal, thereby obtaining the solution interface verification results; The feasible region is subjected to finite combination detection to determine the finiteness verification result of the feasible region; When the verification results of the objective function, the solution interface, and the finiteness of the feasible region are all verified, it is determined that the natural language data satisfies the linearity verification condition, and the corresponding extended quintuple dataset is generated.

[0008] Optionally, the step of performing uncertainty classification processing on the data in the extended quintuple dataset to determine the learning paradigm configuration data for solving the model training strategy configuration includes: Obtain the learning signal identification information from the extended quintuple dataset; Based on the learning signal identification information, the data observable state information corresponding to the extended quintuple dataset is determined. The data observable state information includes uncertain parameter information, optimal decision label information, decision cost information, and black box feedback information. Based on the uncertain parameter information, optimal decision label information, decision cost information, and black-box feedback information, the uncertainty classification result is determined. Based on the uncertainty classification results, the learning paradigm configuration data is determined.

[0009] Optionally, determining the training component configuration data for solving the model based on the extended quintuple dataset and the learning paradigm configuration data includes: Based on the problem type label information and learning signal identification information in the extended quintuple dataset, the data structure features corresponding to the extended quintuple dataset are determined; Based on the data structure features and the learning paradigm configuration data, determine the encoder configuration data corresponding to the solution model; Based on the linear structure information of the objective function and the optimization model information in the extended quintuple dataset, the characteristics of the combinatorial optimization solution interface corresponding to the solution model are determined. Based on the combined optimization solution interface features and the learning paradigm configuration data, the differentiable layer configuration data corresponding to the solution model is determined. Based on the encoder configuration data and the differentiable layer configuration data, the training component configuration data of the solution model is determined.

[0010] Optionally, determining the loss function for model training based on the learning paradigm configuration data and the training component configuration data includes: Based on the learning paradigm configuration data, the corresponding training supervision signal type is determined; Based on the training supervision signal type, the supervision type of the loss function is determined. The supervision type of the loss function includes structured supervision loss function, decision cost optimization loss function, and policy optimization loss function. Based on the encoder configuration data, differentiable layer configuration data, and supervision type of the loss function, an adaptation process is performed to obtain the loss function used for model training.

[0011] Optionally, the step of constructing and training a solution model based on the learning paradigm configuration data, the training component configuration data, and the loss function to obtain a trained solution model includes: Based on the learning paradigm configuration data, the training component configuration data, and the loss function, an initial solution model is constructed. The training samples are input into the encoder in the initial solution model to obtain the target function parameter prediction data; The predicted data of the objective function parameters are input into the differentiable layer of the initial solution model to obtain the corresponding combinatorial optimization decision data. Based on the combined optimization decision data and the loss function, the training loss data is determined; The parameters of the initial solution model are updated based on the training loss data, and the training loss, validation regret value, encoder gradient state and differentiable layer output state are monitored during the training process to obtain training monitoring results. When the training monitoring results meet the preset training stop conditions, the trained solution model is determined.

[0012] Optionally, the step of inputting the natural language data into the trained solution model and outputting the combinatorial optimization decision result corresponding to the linear objective combinatorial optimization problem includes: The natural language data is adapted to obtain the input data. The input data is input into the encoder of the trained solution model to obtain the target function parameter prediction data; Based on the predicted data using the objective function parameters, the solution interface is called to obtain the optimal decision vector; Based on the optimal decision vector, the objective function value, constraint satisfaction result, and baseline comparison result are determined. The optimal decision vector, objective function value, constraint satisfaction result, and baseline comparison result are output as the combinatorial optimization decision result corresponding to the linear objective combinatorial optimization problem.

[0013] A training and configuration device for solving a linear objective combinatorial optimization problem model, the device comprising: The first processing module is used to perform linear validation processing on the natural language data input by the user, and determine the extended quintuple dataset corresponding to the natural language data. The extended quintuple dataset includes user query information, question type label information, learning signal identification information, objective function linear structure information, and optimization model information. The second processing module is used to perform uncertainty classification processing on the data in the extended quintuple dataset and determine the learning paradigm configuration data used for the training strategy configuration of the solution model. The first determining module is used to determine the training component configuration data of the solution model based on the extended quintuple dataset and the learning paradigm configuration data. The training component configuration data includes encoder configuration data and differentiable layer configuration data. The second determining module is used to determine the loss function for solving model training based on the learning paradigm configuration data and the training component configuration data. The construction module is used to construct and train the solution model based on the learning paradigm configuration data, the training component configuration data and the loss function, so as to obtain the trained solution model. The output module is used to input the natural language data into the trained solution model and output the combinatorial optimization decision result corresponding to the linear objective combinatorial optimization problem.

[0014] A computer device includes a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor, wherein the processor, when executing the computer-readable instructions, implements the above-described method for training and configuring a solution model for a linear objective combinatorial optimization problem.

[0015] A readable storage medium storing computer-readable instructions, which, when executed by a processor, implement the above-described method for training and configuring a model for solving a linear objective combinatorial optimization problem.

[0016] The aforementioned method for training and configuring a model to solve a linear objective combinatorial optimization problem involves: performing linear validation on the user-input natural language data to generate an extended quintuple dataset; classifying the data in the extended quintuple dataset to determine the learning paradigm configuration data; determining the training component configuration data for the solution model based on the extended quintuple dataset and the learning paradigm configuration data; determining the loss function for the solution model based on the learning paradigm configuration data and the training component configuration data; constructing and training the solution model based on the learning paradigm configuration data, the training component configuration data, and the loss function; and outputting combinatorial optimization decision results based on the trained solution model. This method reduces the barriers to manual modeling and training configuration, and improves the accuracy, stability, and efficiency of the combinatorial optimization decision output. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the training and configuration method for solving a linear objective combinatorial optimization problem in one embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of a training configuration device for solving a linear objective combinatorial optimization problem in one embodiment of the present invention; Figure 3 This is a schematic diagram of a computer device according to an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] In one embodiment, such as Figure 1 As shown, a method for training and configuring a model for solving a linear objective combinatorial optimization problem is provided, including the following steps: 101. Perform linear validation on the natural language data input by the user to determine the extended quintuple dataset corresponding to the natural language data.

[0021] In this embodiment of the invention, the above-described method for training and configuring a model for solving a linear objective combinatorial optimization problem can be applied to a training and configuration platform for a linear objective combinatorial optimization problem. This platform has functions such as processing training and configuration data for a linear objective combinatorial optimization problem, sending and receiving training and configuration data for a linear objective combinatorial optimization problem, and storing training and configuration data for a linear objective combinatorial optimization problem. It can be built based on a server or a server cluster, and the server or server cluster can be an electronic device with the capability to process training and configuration data for a linear objective combinatorial optimization problem.

[0022] The aforementioned natural language data can refer to a description of a linear objective combinatorial optimization problem input by a user in natural language form, or it can include dataset entries or data description information corresponding to the problem. For example, a user description of "generating a vehicle scheduling plan based on historical demand and transportation costs" and its associated data can be considered natural language data.

[0023] In this embodiment, the objective function, solution interface, and feasible region parsed from the natural language data can be linearly verified to determine whether the problem satisfies the training configuration conditions for linear objective combinatorial optimization. Generally, it can be verified whether the objective function has a linear structure with respect to the objective function parameters, whether the solution interface can solve the problem using the objective function parameters as linear objective coefficients, and whether the feasible region is a finite set of combinations.

[0024] The aforementioned extended quintuple dataset can refer to structured data used to describe linear objective combinatorial optimization problems and their training configuration conditions, which includes user query information, problem type label information, learning signal identification information, linear structure information of the objective function, and optimization model information.

[0025] Specifically, the aforementioned user query information can refer to the problem description entered by the user and its data entry point. For example, if a user enters "Minimize transportation costs while satisfying vehicle capacity constraints," this problem description and the corresponding dataset entry point can form the user query information.

[0026] The aforementioned problem type label information can refer to the combinatorial optimization problem type or data description context corresponding to the natural language data. For example, vehicle routing problems, assignment problems, knapsack problems, scheduling problems, or energy scheduling problems can all serve as problem type label information.

[0027] The aforementioned learning signal identification information can refer to the identification information of the available supervisory or feedback signals during the training phase. For example, whether historical data contains actual demand, optimal routes, actual transportation costs, or reward values ​​returned by the simulator can be represented by learning signal identification information.

[0028] The aforementioned linear structure information of the objective function can refer to the linear relationship between the objective function and its parameters. For example, in the path selection problem, the objective function can be expressed as a weighted sum of the path edge weights and the path selection decision variables; this linear relationship can serve as the linear structure information of the objective function.

[0029] The aforementioned optimization model information can refer to the model content corresponding to the combinatorial optimization problem, including decision variables, constraints, feasible region, objective function, and solution interface. For example, in the vehicle scheduling problem, the optimization model information may include vehicle route selection variables, vehicle capacity constraints, feasible path set, transportation cost objective function, and solution interface for outputting the optimal path.

[0030] In one possible embodiment, the above-mentioned linear objective combinatorial optimization problem solving model training and configuration platform can perform semantic parsing on the natural language data input by the user, and extract the problem description, data entry, decision variables, objective function, and constraints from the natural language data.

[0031] Furthermore, the aforementioned training and configuration platform for solving linear combinatorial optimization problems can generate an extended quintuple dataset based on the validated natural language data. Specifically, the user-inputted problem description and data entry can be written into the user query information, the identified combinatorial optimization problem category can be written into the problem type label information, the available uncertain parameters, optimal decision labels, decision costs, or black-box feedback in the training data can be written into the learning signal identification information, the linear relationship of the objective function with respect to the objective function parameters can be written into the objective function linear structure information, and the decision variables, constraints, feasible region, objective function, and solution interface can be written into the optimization model information.

[0032] By using the above methods and steps, linear objective adaptation can be determined for combinatorial optimization problems described in natural language before training configuration, and the natural language problem can be transformed into a structured extended quintuple dataset. This provides a unified data foundation for subsequent uncertainty classification, learning paradigm configuration, training component configuration, and loss function configuration, thereby improving the accuracy and automation of training configuration for solving linear objective combinatorial optimization problems.

[0033] In another possible embodiment, during the RAG retrieval enhancement phase, a hierarchical retrieval mechanism can be employed using the following retrieval strategy:

[0034] Here, `retrieve(q)` performs a retrieval operation on query `q`, returning the top-k most similar paradigms in the knowledge base, where `q` is the query vector for the current question (the question description after embedding), and `top-k(·)` sorts the results by similarity and returns the top k results. `cos(embed(q), embed(d)` i The function `embed(·)` queries the cosine similarity between the embedding of document `q` and document `di` in the knowledge base. `embed(·)` is the embedding function `d` that maps the text to a dense vector. i For the i-th problem modeling paradigm document in the knowledge base, specifically, the problem type (coarse-grained) can be retrieved first, followed by the specific constraint pattern (fine-grained) to avoid retrieval noise. Similar historical problem modeling schemes retrieved are injected as few-shot examples into the prompt to improve the accuracy of LLM in modeling complex problems.

[0035] In the linear verification process, the definition of the probabilistic combinatorial optimization layer can be referenced, and the verification can be performed using the following COoracle format:

[0036] The Combinatorial Optimization Oracle (COoracle) returns the optimal decision vector v* given a linear objective coefficient vector θ. i : The target direction vector output by the encoder, i.e., the coefficients of the linear objective function, with dimension d; v is the candidate decision vector. This indicates the operation of finding the maximum value. It is a linear objective function (the inner product of the parameters and the decision variables).

[0037] Specifically, It is a finite set of feasible solutions, and θ is the target direction. Before entering the InferOpt-type differentiable CO layer, it can be guaranteed that the perturbation-type and regularized CO layers used, as well as subsequent loss functions such as FY / SPO+, are within the applicable range of this scheme.

[0038] The PerturbedAdditive method targets the smoothed Jacobian of the oracle output, rather than taking the derivative of the discrete argmax in the ordinary sense. Its estimation formula can be written as:

[0039] It is the Jacobian matrix of the PerturbedAdditive layer with respect to the parameter θ, i.e., the smooth gradient estimate. This is the expected Oracle output after additive perturbation smoothing. e It is a disturbance intensity hyperparameter that controls the smoothness; ε>0. Z Z is an isotropic Gaussian perturbation random variable, Z ~ N(0, I). : The output of COoracle after perturbation, which is a discrete decision vector. E[·]: The expectation of the perturbation random variable Z.

[0040] In engineering implementation, it is approximated using the Monte Carlo method:

[0041] m It is the number of Monte Carlo samples; Z is the independent and identically distributed Gaussian perturbation of the k-th sample. k ~ N(0, I); This is the COoracle output after the k-th perturbation, which is a discrete decision vector; (1 / m Σ is the average of m sampling results, which achieves an unbiased estimate of the expectation. e It is the disturbance intensity hyperparameter, with the same meaning as above.

[0042] The derivation of this formula relies on the linearity of the objective function with respect to θ. If the objective function is nonlinear, the gradient estimate between the perturbed oracle output and the perturbed direction will no longer correspond to the gradient of the linear oracle described above, potentially leading to distorted training directions. For regularization methods, the correct relationship is to first relax the hard oracle to...

[0043] It is the continuous output of the regularized relaxation combinatorial optimization layer (regularized Oracle). conv ( V ): The convex hull (polyhedron) of the feasible solution set V; i It is the score or direction vector output by the encoder; Oh ( m) is a smooth, strictly convex regularization function defined on conv(V).

[0044] m These are continuous relaxed decision variables in conv(V); It is an operation that takes the continuous decision that maximizes the objective value within the convex hull; the above regularization Oracle with respect to θ Jacobian can be written as follows under smooth conditions: ,in Let Ω be a convex conjugate function. 2 Oh * ( i Let be its Hessian matrix (second derivative matrix), which describes the local sensitivity of the relaxation layer output to the parameters. If the objective function contains nonlinear components, the convex relaxation problem will no longer correspond to the linear Oracle approximation of the original problem.

[0045] Therefore, the second derivative describes the local sensitivity of the relaxation layer output to the parameters, rather than the general gradient formula of the original nonlinear objective function. If the objective function contains nonlinear components, the convex relaxation problem... This will no longer correspond to the linear oracle approximation of the original problem.

[0046] 102. Perform uncertainty classification on the data in the extended quintuple dataset to determine the learning paradigm configuration data used for the training strategy configuration of the solution model.

[0047] In this embodiment of the invention, uncertainty classification processing can be used to determine what kind of learning signal the training data in the extended quintuple dataset has, such as whether the uncertain parameters are observable, whether the optimal decision label is available, whether the decision cost is calculable, and whether feedback can only be obtained through a black-box environment.

[0048] For example, in logistics scheduling scenarios, actual demand, transportation costs, or path edge weights can be used as uncertain parameters; historical optimal paths or expert scheduling schemes can be used as optimal decision labels; actual transportation costs or simulation costs can be used as decision costs; and the reward value returned by the simulator can be used as black-box feedback.

[0049] The above-described model training strategy configuration is used to select the corresponding training method based on the uncertainty classification results. If an optimal decision label exists, supervised learning or imitation learning is configured; if there is no optimal decision label but the decision cost is calculable, empirical cost minimization is configured; if uncertain parameters are observable but end-to-end decision loss is unstable, prediction-based optimization is configured; if only black-box feedback is available, reinforcement learning or policy search is configured.

[0050] The aforementioned learning paradigm configuration data represents the result of training policy configuration and is used to indicate the learning paradigm adopted by the subsequent model solution and the required training signals. For example, learning paradigm configuration data may include supervised learning configuration data, empirical cost minimization configuration data, post-prediction optimization configuration data, or reinforcement learning configuration data.

[0051] In one possible embodiment, the above-mentioned linear objective combinatorial optimization problem solving model training configuration platform can obtain the learning signal identification information in the extended quintuple dataset, and perform uncertainty classification based on the uncertain parameters, optimal decision label, decision cost and available state of black box feedback to obtain the uncertainty classification result, and then determine the learning paradigm configuration data for the solving model training strategy configuration based on the uncertainty classification result.

[0052] More specifically, the IL / SL (Imitation Learning / Supervised Learning) paradigm can be obtained by calculating the minimum distance between the predictive decision and the expert decision, which is calculated using the following formula:

[0053] w These are the trainable parameters of the encoder; N It is the number of training samples; The encoder will convert the context features x i Mapped to the input parameters θ of the combinatorial optimization layer i The mapping function; It is θ i = φ w (x i COoracle output decision with parameters ) It is the expert or optimal decision label for the i-th sample; d (·, ·) is a distance metric function for the solution space, such as Hamming distance, cross-entropy, or structured hinge distance.

[0054] FCO stands for COoracle. For expert or optimal decision-making, d represents the distance metric in the solution space, such as Hamming distance, cross-entropy, structured hinge distance, or task-related distance. IL / SL does not require complete observations ξ. i This requires that the quality of the decision labels be high enough.

[0055] The ECM (Empirical Cost Minimization) paradigm can also be calculated by minimizing the empirical decision cost, i.e., when there is no optimal decision label but the decision cost is calculable (g a *=0, g =1), adopting the Empirical Cost Minimization (ECM) paradigm is

[0056] Is the i-th sample in the prediction decision? The actual cost depends on the context x. i and realized uncertainty parameter ξ i； It is the encoder prediction parameter-driven COoracle output decision, i.e. ξ i It is the realized uncertainty parameter of the i-th sample; The encoder will use the context x i The target function parameter vector obtained by mapping; If differentiable layers are approximated using differentiable approximations (θ), then the gradient can be backpropagated using the chain rule.

[0057] in L / The gradient of the loss with respect to the prediction decision, / i Jacobian for the output parameters of differentiable layers. i / w The gradient of the encoder parameters with respect to the weights is represented by three gradients connected by a chain rule to achieve end-to-end training.

[0058] ECM is suitable for white-box or semi-white-box scenarios where parameters or costs can be used to evaluate the quality of decisions. When the system's value is 1 but it cannot stably compute the end-to-end decision loss, or when the user requires a traditional interpretable baseline, the system simultaneously generates a Predict-then-Optimize (PTO) baseline. PTO is not the final end-to-end paradigm of this invention, but as a baseline, it helps to evaluate the regret improvement brought about by CO-ML training. When the reward is 1 and there is no explicit cost function or decision label, the system selects the RL / policy search paradigm. This paradigm obtains a scalar reward by interacting with the environment or simulator and updates the encoder parameters using REINFORCE, Actor-Critic, or other policy gradient methods.

[0059] in, Let be the gradient of the policy expected reward with respect to the encoder parameters w; R is the scalar reward signal obtained from the environment or simulator. The strategy parameterized by the encoder is the probability of outputting decision y given context x; This is the logarithmic policy gradient (score function).

[0060] By using the above methods and steps, the training paradigm can be automatically matched according to the actual observable state of the training data, avoiding inconsistencies between the training strategy and the learning signal, and improving the accuracy and adaptability of the training configuration for solving the model.

[0061] In another possible implementation, uncertainty classification can also be handled by determining whether the scalar reward R can only be obtained through a black-box environment, simulator, or oracle. If there is neither an expert decision label nor an explicit cost function, and the reward can only be obtained through interaction, it can be labeled as "black-box interactive feedback." This situation typically falls into reinforcement learning or policy search paradigms and should not be misjudged as an ECM problem with full parameter supervision.

[0062] Therefore, the first part is the project routing label T∈Type1,Type2,Type3 used in the flowchart; the second part is the more granular learning signal description vector:

[0063] in, To learn the signal description vector, each component is a 0 / 1 binary indicator; Whether the parameter ξ is explicitly observable in the training set (1 = observable, 0 = unobservable); Is the expert or best decision label available (1 = available, 0 = unavailable)? To determine whether the cost function or cost feedback is computable (1 = computable, 0 = not computable); To indicate whether scalar rewards can only be obtained through a black-box environment / emulator (1 = black-box only, 0 = no). Determine whether the problem has a dynamic or multi-stage structure (1 = yes, 0 = no).

[0064] in Indicates whether the parameter ξ is observable. Indicates whether the expert decision-making label is available. Indicates whether the cost function or cost feedback is available. This indicates whether there is only a black-box reward. Indicates whether the problem has a dynamic or multi-stage structure.

[0065] 103. Based on the extended quintuple dataset and learning paradigm configuration data, determine the training component configuration data for solving the model.

[0066] In this embodiment of the invention, the aforementioned training component configuration data can be used to represent the model components and their parameter contents that need to be configured during the model training process. It includes at least encoder configuration data and differentiable layer configuration data. It should be noted that the aforementioned training component configuration data does not involve selecting a single neural network structure. Instead, it encodes the objective function parameters and trains them based on the problem type, learning signal, linear structure of the objective function, and optimization model information in the extended quintuple dataset. For example, in the logistics scheduling problem, if the extended quintuple dataset represents the input data as tabular features such as distance, demand, and time windows, and the learning paradigm configuration data can be used for supervised learning or empirical cost minimization training, then the training component configuration data can include encoder configuration data adapted to tabular features and differentiable layer configuration data adapted to the linear objective COoracle.

[0067] The encoder configuration data described above can be used to configure the encoder structure in the solution model, enabling the encoder to map contextual features to the objective function parameters required by the combinatorial optimization solution interface. Specifically, the encoders are primarily based on the geometric structure of the data, the dependencies between features, the problem size, and the interpretability requirements. For example, for tabular data, problems with clear feature structures, or simple feature dependencies, a GLM encoder or an MLP encoder can be configured. In graph-structured problems, if decision variables and constraints naturally correspond to nodes and edges, a GNN encoder can be configured. For problems involving dynamic scheduling, multi-stage decision-making, or temporal dependencies, an RNN encoder or a Transformer encoder can be configured. For image, sensor data, or physically constrained scenarios, a CNN encoder or a PINN-type encoder can be configured.

[0068] By configuring data through the encoder, the solution model can generate corresponding objective function parameter prediction data based on the data structure of different problems.

[0069] The aforementioned differentiable layer configuration data can be used to configure CO differentiable layers that are compatible with the combinatorial optimization solution interface, enabling the combinatorial optimization solution process, which was originally not directly backpropagated, to participate in model training. Since the output of COoracle usually has piecewise constant characteristics, direct differentiation is difficult. Therefore, it is necessary to encapsulate the combinatorial optimization solution interface into the training pipeline through perturbation-type or regularized differentiable layers.

[0070] For example, when COoracle can accept arbitrary real-valued target coefficients and the problem size is moderate, a PerturbedAdditive layer can be configured; when the objective function parameters need to remain non-negative or the solution interface is sensitive to the parameter sign, a PerturbedMultiplicative layer can be configured; when the training process is sensitive to gradient variance, requires deterministic smooth output, or can accept continuous relaxation computational overhead, a RegularizedGeneric regularization layer can be configured.

[0071] By configuring data in differentiable layers, the objective function parameters output by the encoder can further drive the decision results output by the combinatorial optimization solution interface, and enable the training loss to be back fed back to the encoder.

[0072] In one possible embodiment, the training configuration platform for solving the linear objective combinatorial optimization problem can determine the training component configuration data of the solution model based on the extended quintuple dataset and the learning paradigm configuration data.

[0073] Specifically, the training configuration platform for solving the linear objective combinatorial optimization problem can first determine the data structure and training signal type of the input data based on the problem type label information and learning signal identification information in the extended quintuple dataset, and then determine the training methods that the encoder needs to support based on the learning paradigm configuration data.

[0074] Meanwhile, the aforementioned training and configuration platform for solving linear objective combinatorial optimization problems can determine the parameter input requirements and differentiability requirements of the combinatorial optimization solution interface based on the linear structure information of the objective function and the optimization model information, thereby determining the configuration data for the differentiable layer.

[0075] For example, in vehicle routing optimization scenarios, if the input is tabular or graph-structured data such as distance matrices, demand quantities, and time windows, the aforementioned linear objective combinatorial optimization problem solving model training configuration platform can select a GLM / MLP encoder or a GNN encoder accordingly. If the objective function is a linear combination of path edge weights and path selection variables, and the solution interface can output the optimal path based on the objective function parameters, then the aforementioned linear objective combinatorial optimization problem solving model training configuration platform can be configured with a perturbation-type CO differentiable layer, enabling the optimal path decision to participate in model training.

[0076] By employing the above methods, the encoder and differentiable layers of the solution model can be adapted to the problem data structure, training paradigm, and linear objective combination optimization solution interface simultaneously, thereby improving the accuracy of training component configuration and the stability of subsequent end-to-end training.

[0077] In one possible implementation, a GNN encoder is used as an example. GNNs are suitable for graph structure problems, where decision variables and constraints naturally correspond to nodes and edges of a graph.

[0078] Specifically, each node vi updates its own representation by aggregating the features of its neighbors:

[0079] It is the feature representation of node i at layer l+1 (after update). It is the feature representation of node i at layer l (before update). It is the set of neighboring nodes of node i in the graph. This represents the feature representation of neighbor node j at layer l. `AGGREGATE(·)` is the neighbor feature aggregation function, which can be mean aggregation, maximum aggregation, or attention-weighted aggregation. `UPDATE(·)` is the node feature update function, which can be a non-linear transformation such as MLP or GRU. After L layers of message passing, the node features... Capture the L-hop neighborhood structure information centered on node i, and then transform θ linearly. i = W · + b Output the predicted values ​​for the corresponding parameters.

[0080] Taking an RNN encoder as an example, the RNN uses hidden states h t Capture historical information: ht=RNN(x t ,h t-1 Transformer captures the dependencies between any two positions in a sequence through a self-attention mechanism:

[0081] Q It is a query matrix, obtained by linear transformation of the input sequence. K It is the key matrix, which is obtained by linear transformation of the input sequence. V The value matrix is ​​obtained by linear transformation of the input sequence. It is the dimension of the key vector, used for scaling to prevent the inner product from becoming too large and causing the gradient to vanish. This is the attention weight matrix, which measures the correlation between positions in the sequence.

[0082] The advantage of RNN / Transformer lies in its ability to handle variable-length input sequences and capture long-range temporal dependencies. Typical application scenarios include multi-stage scheduling problems, dynamic demand forecasting, and flight scheduling.

[0083] Taking a CNN encoder as an example, CNNs extract local spatial features through convolution operations:

[0084] It is the feature output of the (l+1)th layer at spatial location (i, j); It is the offset of the l-th layer convolution kernel. Weight parameters at the location; Is the lth layer in position Input features at the location; σ is the bias term of the l-th layer; σ(·) is a non-linear activation function, such as ReLU or sigmoid. k, m are summation subscripts, representing the spatial offset range of the convolution kernel (i.e., summing over all offset positions).

[0085] PINN embeds physical equations as regularization terms into the loss function based on neural networks, ensuring that the prediction results meet physical constraints. Typical applications include logistics route planning based on satellite imagery, factory equipment scheduling based on sensor data, and energy pipeline network optimization based on fluid dynamics equations.

[0086] 104. Based on the learning paradigm configuration data and training component configuration data, determine the loss function used to solve the model training.

[0087] In this embodiment of the invention, the aforementioned loss function can be used to measure the difference between the combinatorial optimization decision output by the solution model and the training objective, and serve as the basis for updating the encoder and differentiable layer parameters. It should be noted that in this embodiment, the loss function is not fixed but matched according to the learning paradigm configuration data and training component configuration data: when expert decisions or optimal solution labels exist, FYLoss, S-SVM, or structured supervised loss can be selected; when true objective function parameters exist and the decision cost can be calculated, SPO+, ExpectedRegret, or ECM loss can be selected; when reward feedback can only be obtained through a black-box environment, policy gradient loss or Actor-Critic type loss can be selected.

[0088] In one possible embodiment, the training configuration platform for the linear objective combinatorial optimization problem solving model can determine the type of training supervision signal corresponding to the current training scenario based on the learning paradigm configuration data. If the learning paradigm configuration data indicates that there is an optimal decision label in the current training samples, such as the historical optimal path or expert scheduling scheme in a logistics scheduling scenario, then the training configuration platform for the linear objective combinatorial optimization problem solving model can determine a structured supervision loss function so that the combinatorial optimization decision obtained by passing the objective function parameters output by the encoder through a differentiable layer is as close as possible to the optimal decision label.

[0089] If the learning paradigm configuration data characterizes the existence of real objective function parameters and calculable decision costs in the current training samples, such as real transportation costs, path edge weights, or energy prices, then the training configuration platform for solving the above-mentioned linear objective combinatorial optimization problem can determine the decision cost optimization loss function, so that the training objective directly constrains the actual cost or regret value of the predicted decision.

[0090] If the learning paradigm configuration data representation can only obtain reward feedback from a simulator or black-box environment for the current training samples, then the above-mentioned linear objective combinatorial optimization problem solving model training configuration platform can determine the policy optimization loss function, enabling the solving model to adjust the decision policy according to environmental feedback.

[0091] Furthermore, the training configuration platform for solving the linear objective combinatorial optimization problem can also adapt the loss function based on the training component configuration data. For example, when the training component configuration data includes a PerturbedAdditive differentiable layer, the FYLoss can be adapted to the smooth decision result output by that differentiable layer, so that the loss gradient can be backpropagated to the encoder.

[0092] When the training component configuration data includes differentiable layers for a linear objective COoracle, and the training data can compute decision costs, the SPO+, ExpectedRegret, or ECM loss can be adapted to the objective function parameter prediction data and the combinatorial optimization decision data.

[0093] By following the above steps, the loss function can be kept consistent with the learning paradigm, encoder, and differentiable layer, avoiding a disconnect between the training objective and the quality of combinatorial optimization decisions, and improving the stability of the solution model training and the decision-making effect.

[0094] 105. Construct and train a solution model based on the learning paradigm configuration data, the training component configuration data, and the loss function to obtain a trained solution model.

[0095] In this embodiment of the invention, the above-mentioned solution model is a model constructed based on learning paradigm configuration data, training component configuration data, and loss function, which is used to output the decision result of the linear objective combinatorial optimization problem according to the input data.

[0096] The trained solution model mentioned above refers to a solution model that meets the preset training stopping conditions after iterative updates using training samples. The preset training stopping conditions may include convergence of training loss, cessation of decrease in validation regret value, reaching the upper limit of training epochs, and stable encoder gradient state or output state of differentiable layers meeting preset requirements.

[0097] In one possible embodiment, the above-mentioned linear objective combinatorial optimization problem solving model training configuration platform can determine the training method of the solving model according to the learning paradigm configuration data, instantiate the encoder and differentiable layer according to the training component configuration data, and connect the loss function with the encoder and differentiable layer to construct the initial solving model.

[0098] For example, when the learning paradigm configuration data is supervised learning, the training component configuration data includes a GLM encoder and a Perturbed Additive differentiable layer, and the loss function is FYLoss, the above-mentioned linear objective combinatorial optimization problem solving model training configuration platform can construct an initial solution model consisting of a GLM encoder, a perturbed CO differentiable layer, and FYLoss.

[0099] Specifically, the aforementioned training and configuration platform for solving linear combinatorial optimization problems can input training samples into the encoder of the initial solution model. The encoder then generates predicted objective function parameters based on the contextual features in the training samples. These predicted parameters are then input into a differentiable layer, which calls the combinatorial optimization solution interface or performs differentiable processing on the interface to obtain combinatorial optimization decision data. For example, in a logistics scheduling scenario, the encoder can predict path edge weights based on contextual features such as demand, distance matrix, and time window. The differentiable layer then outputs vehicle route decisions based on these path edge weights.

[0100] Furthermore, the aforementioned training and configuration platform for solving linear objective combinatorial optimization problems can determine training loss data based on combinatorial optimization decision data and loss function, and update the parameters of the initial solution model according to the training loss data.

[0101] If the learning paradigm configuration data representation has an optimal decision label, the model can be updated based on the difference between the predicted decision and the optimal decision label; if the decision cost of the learning paradigm configuration data representation is calculable, the model can be updated based on the actual cost or regret value corresponding to the predicted decision; if the learning paradigm configuration data representation only has black-box feedback, the model can be updated based on the reward value returned by the black-box environment.

[0102] Furthermore, during the training process, the training configuration platform for the above-mentioned linear objective combinatorial optimization problem solving model can monitor the training loss, validation regret value, encoder gradient state, and differentiable layer output state to obtain training monitoring results.

[0103] When the training monitoring results indicate that the training loss continues to decrease, the verification regret value meets the preset requirements, the encoder gradient does not show any abnormalities, and the output state of the differentiable layer is stable, the training configuration platform for solving the above-mentioned linear objective combinatorial optimization problem can save the current solution model; when the training monitoring results indicate that the gradient is abnormal, the verification regret value does not decrease, or the output of the differentiable layer is abnormal, the learning rate, perturbation strength, gradient clipping parameters, or training epochs can be adjusted.

[0104] By using the above methods and steps, a solution model can be automatically constructed and trained based on learning paradigm configuration data, training component configuration data, and loss function. This enables the encoder, differentiable layer, and loss function to form an end-to-end training link, thereby improving the training stability and decision output performance of the solution model for linear objective combinatorial optimization problems.

[0105] More specifically, in one possible embodiment, regarding the selection of the loss function, when there is expert decision or optimal solution label y*, FYLoss, S-SVM, or structured supervised loss are preferred; when there is a true parameter ξ and the decision cost can be calculated, ExpectedRegret, SPO+, or ECM objective are preferred; when there is only reward feedback, policy gradient or Actor-Critic type objective is adopted, and structured surrogate loss is added to stabilize training when feasible.

[0106] Among these, there are target label scenarios. SPO+ (SmartPredict-then-OptimizePlus) Loss is designed specifically for CO-ML, and its core idea is to directly optimize decision quality through proxy loss. SPO+ Loss is defined as:

[0107] in, SpO+ loss is used to predict parameters. The actual parameter θ is the input; The target function parameter vector predicted by the encoder; The true objective function parameter vector (from the dataset); The operation is to maximize the alternative objective function in the feasible solution set V.

[0108] SPO+ is the convex upper bound of the SPO loss, which can be directly applied to... Find the subgradient. This loss function is convex and can serve as an optimizable upper bound for the SPO loss. One of its subgradients can be written as:

[0109] in, y*(θ) and y*(θ) are respectively in the parameters The optimal decision under the true parameters θ is calculated via COOracle.

[0110] Fenchel-Young Loss (FYLoss) is constructed based on the Fenchel-Young inequality and is defined as follows:

[0111] It is the Fenchel-Young loss value, which measures the difference between the predicted decision and the target decision; i It is the score vector output by the encoder. It is the label for the optimal decision. It is the convex conjugate of the regularization function Ω, defined as ; Is the regularization function in target decision The value at; It is the inner product of the score vector and the objective decision.

[0112] in Let Ω be the convex conjugate. FYLoss is nonnegativity, convexity, and differentiability, and its gradient is:

[0113] in, Let FYLoss be the gradient of the encoder output θ; The smoothed output of the probabilistic / regularized CO layer (computed by PerturbedAdditive or RegularizedGeneric); y* is the target optimal decision label; y* is the gradient, which is equal to the difference between the smoothed oracle output and the target decision. It is used to push the prediction decision toward y*.

[0114] FYLoss can be used in combination with PerturbedAdditive, PerturbedMultiplicative, and RegularizedGeneric layers; Fenchel-Young loss combined with additive perturbation, multiplicative perturbation, and semi-square norm regularization can probabilistically regularize the CO layer, interpreting it from the perspective of implicit or explicit regularization, thereby providing a corresponding smooth oracle output for FY loss.

[0115] S-SVM (Structured Support Vector Machine) Loss is based on the structured SVM framework and provides robustness guarantees based on margins.

[0116] This is the S-SVM loss value; i It is the score vector output by the encoder; It is a label for expert decision-making; v It is a candidate decision vector, v ∈ V; It is the task loss (such as Hamming distance) between decision v and expert decision y*. To maximize the score plus penalty term in feasible solutions, a Margin constraint is formed; The S-SVM loss is the inner product of the score vector and the expert decision. The S-SVM loss uses a margin mechanism to ensure a sufficient margin between the predicted decision and the expert decision, making it suitable for scenarios that require stability of the decision boundary.

[0117] In scenarios without target labels, the expected loss directly minimizes the difference between the actual decision cost and the optimal cost:

[0118] It is the ExpectedRegret loss, which directly minimizes the difference in expected cost between the predicted decision and the optimal decision; x It is realized under uncertainty and follows an unknown distribution P; c ( x ) is the stochastic cost vector under uncertainty ξ; Encoder prediction parameters Driven by COoracle output decisions; It is the ex-post optimal decision cost (Oracle solution corresponding cost) when the true parameter ξ is known. Let the distribution of the uncertainty parameter ξ be expected. The gradient of the expected loss is calculated through a differentiable CO layer:

[0119] The chain rule connects the three gradient segments.

[0120] Its simplified form is:

[0121] in, The gradient of the loss with respect to encoder parameters w is the empirical cost minimization loss. The gradient of the cost function with respect to the prediction decision (task-dependent); For the output of the differentiable CO layer with respect to parameter θ iJacobian (provided by the differentiable layer selected by M7); For encoder parameters θ i The gradient with respect to the weight w (standard automatic differentiation); This is the gradient of the cost function with respect to the prediction decision (task-dependent). Jacobian of the output of the differentiable combinatorial optimization layer with respect to parameter θ (provided by the configured differentiable layer). This represents the gradient of the encoder parameters θ with respect to the weights w (calculated using standard automatic differentiation). In reinforcement learning / policy search scenarios, a reward-based policy gradient method is employed, and the corresponding gradient formula is given in step 102. By ensuring consistency between the loss function and the learning paradigm, encoder, and differentiable layers, the decoupling of training objectives from the quality of combinatorial optimization decisions can be avoided, thereby improving the stability of the solved model training and the decision-making performance.

[0122] RL / policy search scenarios employ reward-based policy gradient methods; if differentiable CO layers or structured surrogate loss exist, Fenchel-Young type loss can be used as an auxiliary regularization term. Scalar Reward Signal The corresponding basic policy gradient is:

[0123] in, Here, J(w) represents the policy gradient (REINFORCE estimator); J(w) is the gradient of the policy expected reward with respect to the encoder parameter w; and R is the scalar reward signal obtained from the environment or simulator. The strategy parameterized by the encoder is the probability of outputting decision y given context x; The gradient is the logarithmic policy gradient. To obtain the expectation of the trajectory distribution (in practice, it is estimated using Monte Carlo samples).

[0124] The gradient is estimated using the REINFORCE algorithm or the Actor-Critic method to achieve end-to-end training in parameter-free label scenarios.

[0125] 106. Input natural language data into the trained solution model and output the combinatorial optimization decision result corresponding to the linear objective combinatorial optimization problem.

[0126] In this embodiment of the invention, the above-mentioned combinatorial optimization decision result can be used to represent the final solution result output by the trained solution model for the linear objective combinatorial optimization problem, including but not limited to the optimal decision vector, objective function value, constraint satisfaction result, and baseline comparison result.

[0127] The above-mentioned optimal decision vector can be used to represent the specific decision scheme recommended by the solution model; the above-mentioned objective function value can be used to represent the optimization objective value corresponding to the decision scheme; the above-mentioned constraint satisfaction result can be used to represent whether the decision scheme satisfies the constraint conditions in the optimization model; the above-mentioned baseline comparison result can be used to represent the performance difference of the decision scheme relative to the preset baseline model or the traditional solution method.

[0128] For example, in a vehicle routing optimization scenario, the combined optimization decision results can include the recommended vehicle route, the corresponding transportation cost, the satisfaction of vehicle capacity constraints and time window constraints, and the regret or cost improvement results relative to the GLM baseline model.

[0129] In one possible embodiment, the aforementioned linear combinatorial optimization problem solving model training and configuration platform can adapt the natural language data before inputting it into the trained solving model, obtaining input data that the trained solving model can recognize. This input data may include problem type, contextual features, constraint information, and objective function parameter-related data. Subsequently, the platform can input this input data into the encoder of the trained solving model, which outputs predicted objective function parameters. Based on these predicted parameters, the platform then calls the solving interface to obtain the optimal decision vector corresponding to the linear combinatorial optimization problem.

[0130] Furthermore, the aforementioned training and configuration platform for solving the linear combinatorial optimization problem can calculate the corresponding objective function value based on the optimal decision vector and optimization model information, and verify whether the optimal decision vector satisfies the constraints, thus obtaining the constraint satisfaction result. Simultaneously, this platform can also compare the objective function value or regret value corresponding to the optimal decision vector with the baseline model output, obtaining the baseline comparison result. Finally, the platform can output the optimal decision vector, objective function value, constraint satisfaction result, and baseline comparison result as the combinatorial optimization decision result.

[0131] By using the above methods and steps, the trained solution model can automatically output executable, verifiable, and comparable combinatorial optimization decision results based on the input data corresponding to the natural language data, thereby improving the usability and interpretability of the solution results for linear objective combinatorial optimization problems.

[0132] In this embodiment of the invention, linear validation is performed on the natural language data input by the user to generate an extended quintuple dataset; uncertainty classification is performed on the data in the extended quintuple dataset to determine the learning paradigm configuration data; based on the extended quintuple dataset and the learning paradigm configuration data, the training component configuration data of the solution model is determined; based on the learning paradigm configuration data and the training component configuration data, the loss function of the solution model is determined; the solution model is constructed and trained according to the learning paradigm configuration data, the training component configuration data, and the loss function; and the combinatorial optimization decision result is output based on the trained solution model. Through the above method steps, the threshold for manual modeling and training configuration can be reduced, and the accuracy, stability, and efficiency of the combinatorial optimization decision output of the solution model training configuration can be improved.

[0133] It is understood that in the specific embodiments of this application, data related to detection images, initial detection images, alarm signals, etc. are involved. When the embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0134] Optionally, in the step of performing linear validation on the user-input natural language data to determine the corresponding extended quintuple dataset, the following steps can also be taken: extracting semantic features from the natural language data to obtain semantic feature elements; determining the objective function parameters and the corresponding objective function expression based on the objective function; performing linearity detection on the nonlinear terms in the objective function expression to obtain the objective function validation result; performing linearity detection on the solution interface corresponding to the natural language data to determine that the solution interface uses the objective function parameters as linear objective coefficients and outputs the decision variables that optimize the objective function, thus obtaining the solution interface validation result; performing finite combination detection on the feasible region to determine the finiteness validation result of the feasible region; and when the objective function validation result, the solution interface validation result, and the feasible region finiteness validation result are all verified, it is determined that the natural language data meets the linear validation condition, and the corresponding extended quintuple dataset is generated.

[0135] In this embodiment of the invention, the aforementioned semantic feature elements may include, but are not limited to, structured problem elements such as decision variables, objective functions, constraints, feasible regions, and problem types. Specifically, the decision variables represent the combinatorial optimization decision content to be selected, the objective function represents the optimization objective, the constraints limit the range of values ​​for the decision variables, the feasible region represents the set of candidate decisions that satisfy the constraints, and the problem type represents the combinatorial optimization type to which the current problem belongs.

[0136] In this embodiment, the aforementioned semantic feature elements can be extracted from the user description by the linear objective combinatorial optimization problem solving model training configuration platform through semantic parsing. This platform can extract information such as decision variables, objective functions, constraints, feasible regions, and problem types from the user description. For example, from the description "selecting a transportation route and reducing transportation costs while satisfying vehicle capacity constraints", the platform can extract vehicle route selection variables, transportation cost objective functions, vehicle capacity constraints, feasible path sets, and vehicle route problem types.

[0137] The aforementioned objective function parameters mainly characterize the parameter data that drives the changes in the objective function. For example, in path optimization scenarios, path edge weights, transportation costs, actual demand, or electricity prices in energy dispatch can all be used as objective function parameters, and the trained solution model can predict these objective function parameters through the encoder.

[0138] The objective function expression described above can be a mathematical expression derived from the optimization objective in natural language, used to represent the relationship between the objective function parameters and decision variables. For example, in the path selection problem, the objective function expression can be represented as a weighted sum of the path edge weights and the path selection decision variables.

[0139] The aforementioned nonlinear terms represent terms in the objective function expression that do not satisfy a linear structure, such as quadratic terms of the objective function parameters, product terms between different objective function parameters, exponential terms, logarithmic terms, or rational terms. If the objective function expression contains such nonlinear terms, it may not meet the usage conditions of the linear objective combinatorial optimization solution interface.

[0140] The aforementioned linearity check can be used to determine whether the objective function expression satisfies a linear structure with respect to the objective function parameters. The training and configuration platform for the linear objective combinatorial optimization problem solving model can detect whether there are quadratic, product, exponential, or logarithmic terms in the objective function expression; if no such nonlinear terms exist, the objective function can be verified as successful.

[0141] The above objective function verification results can be used to indicate whether the objective function expression satisfies the linear objective condition. The objective function verification result can include verification passed or failed; for example, when the objective function can be expressed as the linear inner product of the objective function parameters and the decision variables, the objective function verification result is verified.

[0142] The aforementioned solution interface is used to invoke the combinatorial optimization solution process and output the corresponding optimal decision based on the input objective function parameters. For example, the solution interface for the vehicle routing problem can output the optimal vehicle path based on the path edge weight parameters, while the solution interface for the assignment problem can output the optimal assignment scheme based on the task cost parameters.

[0143] The linearity check of the aforementioned solver interface is used to determine whether the solver interface can solve the problem using the objective function parameters as linear objective coefficients and output the decision variables that optimize the objective function. In other words, this check focuses on whether the solver interface conforms to the calling format of a linear objective COoracle, rather than simply determining whether the solver itself is usable.

[0144] The aforementioned linear objective coefficients are the parameter coefficients in the objective function that undergo linear operations with the decision variables. For example, when the objective function is a weighted sum of path edge weights and path selection variables, the path edge weights can be used as linear objective coefficients.

[0145] The decision variables mentioned above are used to represent the selection outcome that needs to be determined in combinatorial optimization problems. For example, in a vehicle routing optimization scenario, the decision variable can represent whether a vehicle chooses a certain route; in an assignment problem, the decision variable can represent whether a task is assigned to a certain piece of equipment or personnel.

[0146] The above verification results of the solution interface are used to indicate whether the solution interface meets the linear parameterization solution conditions. If the solution interface can accept the objective function parameters as linear objective coefficients and output the decision variables that optimize the objective function, then the verification result of the solution interface is considered successful.

[0147] The aforementioned finite combination detection is used to determine whether the feasible region is a finite set of combinatorial decisions. The training and configuration platform for the linear objective combinatorial optimization problem-solving model can determine whether the feasible region consists of bounded integer variables, binary variables, or finite enumeration variables based on the type of decision variables, constraints, and their value ranges.

[0148] The above feasible region finiteness verification results are used to indicate whether the feasible region satisfies the finite combination set condition. If the vehicle path set, task assignment set, or knapsack selection set are all finite candidate sets, then the feasible region finiteness verification result can be considered as successful.

[0149] Passing the above verifications indicates that the verification results for the objective function, the solution interface, and the finiteness of the feasible region all meet the preset requirements. Only when all the above verifications pass are the natural language data considered to satisfy the training configuration conditions for a linear objective combinatorial optimization problem.

[0150] The aforementioned linearity validation conditions are prerequisites for allowing natural language data to be included in the subsequent model training configuration process. These conditions include at least the objective function satisfying a linear structure, the solution interface meeting the linear objective coefficient input requirements, and the feasible region meeting the requirements of a finite set of combinations.

[0151] In one possible embodiment, the above-mentioned linear objective combinatorial optimization problem solving model training configuration platform can extract semantic features from natural language data to obtain semantic feature elements including decision variables, objective function, constraints, feasible region, and problem type. It can also determine the objective function parameters and the corresponding objective function expression based on the objective function, and perform linearity detection on the nonlinear terms in the objective function expression to obtain the objective function verification result.

[0152] Furthermore, the training and configuration platform for solving the linear objective combinatorial optimization problem can perform linearity detection on the solution interface corresponding to the natural language data, determine whether the solution interface uses the objective function parameters as linear objective coefficients, and output the decision variables that make the objective function optimal, thus obtaining the solution interface verification results. Meanwhile, the aforementioned training and configuration platform for solving linear combinatorial optimization problems can perform finite combination detection on the feasible region to determine the finiteness verification result of the feasible region. When the verification results of the objective function, the solution interface, and the feasible region are all passed, the aforementioned training and configuration platform for solving linear combinatorial optimization problems can determine that the natural language data meets the linearity verification condition and generate the corresponding extended quintuple dataset.

[0153] By using the above methods and steps, linear objective adaptation verification can be performed on the objective function, solution interface, and feasible region corresponding to the natural language data before generating the extended quintuple dataset. This avoids problems that do not meet the linear objective combination optimization conditions from entering the subsequent training configuration process, thereby improving the accuracy and reliability of the solution model training configuration.

[0154] Optionally, in the step of performing uncertainty classification on the data in the extended quintuple dataset and determining the learning paradigm configuration data for solving the model training strategy configuration, the learning signal identification information in the extended quintuple dataset can also be obtained; based on the learning signal identification information, the observable state information of the data corresponding to the extended quintuple dataset can be determined; based on the uncertainty parameter information, optimal decision label information, decision cost information and black box feedback information, the uncertainty classification result can be determined; and based on the uncertainty classification result, the learning paradigm configuration data can be determined.

[0155] In this embodiment of the invention, the above-mentioned observable state information can be used to indicate what learning signals the training data in the extended quintuple dataset can provide during the training phase, including but not limited to uncertain parameter information, optimal decision label information, decision cost information, and black box feedback information, which are used to determine whether the current training data is suitable for learning paradigms such as supervised learning, empirical cost minimization, prediction post-optimization, or reinforcement learning.

[0156] The aforementioned uncertain parameter information can be used to characterize whether the parameters in the objective function that vary with the sample or scenario are observable. For example, in a logistics scheduling scenario, data such as actual demand, transportation costs, and path edge weights that affect the final path decision can be used as uncertain parameter information; in an energy scheduling scenario, real-time electricity prices and load demand can also be used as uncertain parameter information.

[0157] The aforementioned optimal decision label information is used to characterize whether there are optimal or expert decisions in the training data that can serve as supervision targets. For example, historical optimal paths, expert scheduling schemes, and optimal solutions generated by solvers such as Gurobi can all be used as optimal decision label information. If this information is available, the corresponding learning paradigm configuration data for supervised learning or imitation learning can be configured subsequently.

[0158] The aforementioned decision cost information is used to characterize the actual cost, simulation cost, or historical execution cost that can be calculated or observed under a given decision. For example, the actual transportation cost corresponding to a vehicle routing plan and the operating cost corresponding to an energy dispatching plan can both be used as decision cost information. If the decision cost information is calculable, then empirical cost minimization or decision cost optimization learning paradigms can be configured subsequently.

[0159] The aforementioned black-box feedback information is used to characterize reward or evaluation values ​​returned by simulators, environments, or external oracles when complete parameters, optimal decision labels, or explicit cost functions cannot be directly obtained. For example, in a simulated vehicle scheduling environment, inputting a path decision into the simulator and receiving a service quality score or negative cost reward can serve as black-box feedback information. If only black-box feedback information exists, subsequent configuration data for learning paradigms such as reinforcement learning, policy search, or structured reinforcement learning can be configured.

[0160] The uncertainty classification results described above can be used to represent the training scenario category obtained based on observable state information of the data. These uncertainty classification results may include at least one of the following: observable uncertainty parameter type, available optimal decision label type, computable decision cost type, or available black-box feedback type, and serve as the basis for determining the learning paradigm configuration data.

[0161] In this embodiment of the invention, the aforementioned training and configuration platform for solving the linear objective combinatorial optimization problem can acquire learning signal identification information from the extended quintuple dataset and determine the observable state information of the data corresponding to the extended quintuple dataset based on the learning signal identification information. Specifically, the aforementioned training and configuration platform for solving the linear objective combinatorial optimization problem can identify whether there is uncertain parameter information, optimal decision label information, decision cost information, and black-box feedback information in the training data, and determine the uncertainty classification result based on the above information.

[0162] Furthermore, when the uncertainty classification result represents the availability of optimal decision label information, the aforementioned linear objective combinatorial optimization problem solving model training configuration platform can determine the learning paradigm configuration data corresponding to supervised learning or imitation learning; when the uncertainty classification result represents the unavailability of optimal decision label information but the decision cost information is calculable, the learning paradigm configuration data corresponding to minimizing empirical cost can be determined; when the uncertain parameter information is observable but the end-to-end decision loss is unstable, the learning paradigm configuration data corresponding to post-prediction optimization can be determined; when only black-box feedback information exists, the learning paradigm configuration data corresponding to reinforcement learning or policy search can be determined.

[0163] By using the above methods and steps, different uncertain training scenarios can be distinguished based on the available learning signals in the extended quintuple dataset, and the corresponding learning paradigm configuration data can be automatically matched for the solution model. This reduces the mismatch between training strategies and data states, and improves the accuracy and adaptability of the solution model training configuration.

[0164] Optionally, in the step of determining the training component configuration data of the solution model based on the extended quintuple dataset and learning paradigm configuration data, the following steps are further taken: determining the data structure features corresponding to the extended quintuple dataset based on the problem type label information and learning signal identification information in the extended quintuple dataset; determining the encoder configuration data corresponding to the solution model based on the data structure features and learning paradigm configuration data; determining the combinatorial optimization solution interface features corresponding to the solution model based on the objective function linear structure information and optimization model information in the extended quintuple dataset; determining the differentiable layer configuration data corresponding to the solution model based on the combinatorial optimization solution interface features and learning paradigm configuration data; and determining the training component configuration data of the solution model based on the encoder configuration data and differentiable layer configuration data.

[0165] In this embodiment of the invention, the above-mentioned data structure features can be used to represent the organization form and feature relationship of input data. For example, the distance matrix, demand and time window in logistics scheduling can correspond to tabular data or graph structure data, and the electricity price and load change in energy scheduling can correspond to time series data.

[0166] The encoder configuration data mentioned above can include encoder type, input dimension, output dimension, and training parameters corresponding to the learning paradigm. It can be used to configure the encoder in the solution model, enabling the encoder to convert input data into parameters for the objective function. For example, when the data structure is tabular, a GLM encoder or an MLP encoder can be configured; when the data structure is graph-structured, a GNN encoder can be configured; and when the data structure is time-series data or multi-stage decision data, an RNN encoder or a Transformer encoder can be configured.

[0167] The aforementioned linear structure information of the objective function is used to represent the linear relationship between the objective function and its parameters. For example, the weighted sum of path edge weights and path selection variables, or the weighted sum of task cost parameters and assignment variables, can both serve as linear structure information of the objective function.

[0168] The aforementioned features of the combinatorial optimization solution interface are used to represent how the solution interface receives objective function parameters and outputs the optimal decision. For example, the solution interface can receive path edge weight parameters as linear objective coefficients and output the optimal vehicle path. Specifically, the aforementioned features of the combinatorial optimization solution interface may also include whether the solution interface supports perturbation of objective function parameters, whether non-negativity of parameters is required, and whether it can return discrete decision vectors.

[0169] The aforementioned differentiable layer configuration data is used to configure differentiable computational layers that are adapted to the combinatorial optimization solution interface, enabling the discrete decisions output by the solution interface to participate in model training. For example, perturbation-type differentiable layers or regularized differentiable layers can be configured based on whether the solution interface supports parameter perturbation, whether the parameters need to be non-negative, and whether training requires smooth output.

[0170] In this embodiment of the invention, the aforementioned training and configuration platform for solving the linear objective combinatorial optimization problem can determine data structure characteristics based on problem type label information and learning signal identification information in the extended quintuple dataset, and determine encoder configuration data in conjunction with learning paradigm configuration data. Simultaneously, the platform can determine combinatorial optimization solution interface characteristics based on the linear structure information of the objective function and optimization model information, and determine differentiable layer configuration data in conjunction with learning paradigm configuration data. Subsequently, the platform can determine the training component configuration data for the solution model based on the encoder configuration data and differentiable layer configuration data.

[0171] By using the above methods and steps, the encoder and differentiable layer can be adapted to the input data structure and combinatorial optimization solution interface, respectively, thereby improving the accuracy of the configuration of the solution model training components and the training stability.

[0172] Optionally, in the step of determining the loss function for solving model training based on learning paradigm configuration data and training component configuration data, the corresponding training supervision signal type can be determined based on the learning paradigm configuration data; the supervision type of the loss function can be determined based on the training supervision signal type; and an adaptation process can be performed based on encoder configuration data, differentiable layer configuration data, and the supervision type of the loss function to obtain the loss function for solving model training.

[0173] In this embodiment of the invention, the supervision type of the aforementioned loss function includes structured supervision loss function, decision cost optimization loss function, policy optimization loss function, etc., used to represent the training constraint methods corresponding to the loss function. Generally, different supervision types are adapted to different constraint decision conditions. For example, structured supervision type can be used to constrain the predicted decision to be close to the optimal decision label, and FYLoss or S-SVM loss can be used; decision cost optimization type can be used to constrain the actual cost or regret value of the predicted decision, and SPO+, ExpectedRegret, or ECM loss can be used; policy optimization type can be used to adjust the decision policy according to the black-box reward, and policy gradient loss or Actor-Critic type loss can be used.

[0174] The training supervision signal types mentioned above indicate which type of signal can be used to calculate the training error when training the solution model. For example, in a vehicle routing optimization scenario, if the training data contains historical optimal paths or expert path solutions, the training supervision signal type can be the optimal decision label type; if the training data contains real transportation costs, path edge weights, or actual execution costs, the training supervision signal type can be the decision cost type; if only service quality scores or negative cost rewards can be returned through the simulator, the training supervision signal type can be the black box feedback type.

[0175] The above adaptation process is used to match the input-output relationship of the loss function based on the encoder configuration data, differentiable layer configuration data, and the supervision type of the loss function. For example, when the encoder outputs objective function parameters, the differentiable layer outputs smooth decision results, and the supervision type is structured supervision, FYLoss can be adapted to the differentiable layer output and the optimal decision label; when the supervision type is decision cost optimization, the predicted decision, the true objective function parameters, and the actual cost data can be integrated into SPO+, ExpectedRegret, or ECM loss; when the supervision type is policy optimization, the reward returned by the black-box environment can be integrated into the policy optimization loss.

[0176] In one possible embodiment, the training configuration platform for the aforementioned linear objective combinatorial optimization problem solving model can determine the corresponding training supervision signal type based on the learning paradigm configuration data. Specifically, when the learning paradigm configuration data indicates that the current training process adopts supervised learning or imitation learning, the training configuration platform can determine the training supervision signal type as the optimal decision label type; when the learning paradigm configuration data indicates that the current training process adopts empirical cost minimization or prediction-based optimization, the training supervision signal type can be determined as the decision cost type or the true parameter type; when the learning paradigm configuration data indicates that the current training process adopts reinforcement learning or policy search, the training supervision signal type can be determined as the black-box feedback type.

[0177] Furthermore, the aforementioned training and configuration platform for solving the linear objective combinatorial optimization problem can determine the supervision type of the loss function based on the type of training supervision signal, and perform adaptation processing by combining encoder configuration data and differentiable layer configuration data to obtain the loss function used for training the solution model. For example, in a vehicle path optimization scenario where a historical optimal path exists, a structured supervision type can be determined, and the path edge weight parameters output by the encoder, the path decision results output by the differentiable layer, and the historical optimal path can be incorporated into FYLoss; in a scenario where real transportation costs exist, a decision cost optimization type can be determined, and the cost or regret value corresponding to the predicted decision can be incorporated into ECM or ExpectedRegret loss; in a scenario where only simulator reward exists, a policy optimization type can be determined, and simulator feedback can be incorporated into policy gradient loss.

[0178] By using the above methods and steps, the training supervision signal type can be automatically identified based on the learning paradigm configuration data, and the loss function can be matched with the encoder, differentiable layer, and training target. This avoids the problem of inconsistency between the loss function and the actual available learning signal, and improves the stability of model training and decision quality.

[0179] Optionally, in the step of constructing and training a solution model based on the learning paradigm configuration data, training component configuration data, and loss function to obtain a trained solution model, the following steps can be taken: First, an initial solution model is constructed based on the learning paradigm configuration data, training component configuration data, and loss function. Then, training samples are input into the encoder of the initial solution model to obtain predicted objective function parameters. Next, the predicted objective function parameters are input into the differentiable layer of the initial solution model to obtain corresponding combinatorial optimization decision data. Based on the combinatorial optimization decision data and loss function, training loss data is determined. Then, the parameters of the initial solution model are updated based on the training loss data, and the training loss, validation regret value, encoder gradient state, and differentiable layer output state are monitored during training to obtain training monitoring results. Finally, when the training monitoring results meet the preset training stopping condition, the trained solution model is determined.

[0180] In this embodiment of the invention, the initial solution model can be a model to be trained constructed based on learning paradigm configuration data, training component configuration data, and loss function.

[0181] The aforementioned predicted data of the objective function parameters can be used to represent the objective function parameters output by the encoder based on the training samples.

[0182] The training samples described above can be used to represent data units input to the initial solution model during the model training process. These samples may include at least one of the following: contextual features, uncertain parameters, optimal decision labels, decision costs, or black-box feedback. The data content of the training samples can differ depending on the learning paradigm. For example, in a logistics scheduling scenario, training samples may include contextual features such as distance matrices, demand, and time windows. If the historical data also includes actual path edge weights or actual transportation costs, the training samples may further include uncertain parameters or decision costs.

[0183] The aforementioned combinatorial optimization decision data is used to represent the combinatorial optimization decision results obtained by the differentiable layer based on the objective function parameter prediction data.

[0184] The training loss data described above is used to represent the difference between the combinatorial optimization decision data and the training objective. Depending on the learning paradigm configuration data, the training loss data can be determined by the difference between the predicted decision and the optimal decision label, or by the actual cost, regret value, or black-box feedback reward corresponding to the predicted decision.

[0185] The training monitoring results described above can be used to represent the state information during the model training process. These results can include training loss, validation regret value, encoder gradient state, and differentiable layer output state, used to determine whether the training process is stable and whether training needs to be stopped or adjusted.

[0186] In one possible embodiment, the aforementioned training and configuration platform for solving the linear objective combinatorial optimization problem can construct an initial solution model based on learning paradigm configuration data, training component configuration data, and a loss function. Specifically, the platform can instantiate an encoder based on encoder configuration data, instantiate a differentiable layer based on differentiable layer configuration data, and connect the loss function with the outputs of the encoder and the differentiable layer to form a trainable solution model structure.

[0187] Specifically, the training and configuration platform for solving the linear objective combinatorial optimization problem can also input training samples into the encoder in the initial solution model to obtain the objective function parameter prediction data, and input the objective function parameter prediction data into the differentiable layer in the initial solution model to obtain the corresponding combinatorial optimization decision data.

[0188] Furthermore, the aforementioned training and configuration platform for solving linear objective combinatorial optimization problems can determine training loss data based on combinatorial optimization decision data and loss function, and update the parameters of the initial solution model based on the training loss data.

[0189] For example, when the training samples contain historical best paths, the encoder parameters can be updated based on the difference between the predicted path and the historical best path; when the training samples contain actual transportation costs, the encoder parameters can be updated based on the cost or regret value corresponding to the predicted path; when the training samples only contain simulator feedback, the solution model parameters can be updated based on the black-box feedback reward.

[0190] Furthermore, during the training process, the aforementioned training configuration platform for solving the linear objective combinatorial optimization problem can monitor the training loss, validation regret value, encoder gradient state, and differentiable layer output state to obtain training monitoring results. When the training monitoring results indicate that the training loss tends to converge, the validation regret value meets the preset requirements, the encoder gradient state is stable, and the differentiable layer output state is normal, the aforementioned training configuration platform for solving the linear objective combinatorial optimization problem can determine that the preset training stopping condition is met and identify the current solution model as the trained solution model.

[0191] By using the above methods and steps, the encoder, differentiable layer, and loss function can form a complete end-to-end training link, and the loss, regret value, gradient, and output state of differentiable layer can be monitored during the training process, thereby improving the stability of the model training process and the reliability of the training results.

[0192] Optionally, in the step of inputting natural language data into the trained solution model and outputting the combinatorial optimization decision result corresponding to the linear objective combinatorial optimization problem, the natural language data can be adapted to obtain the input data; the input data can be input into the encoder in the trained solution model to obtain the objective function parameter prediction data; based on the objective function parameter prediction data, the solution interface can be called to obtain the optimal decision vector; based on the optimal decision vector, the objective function value, constraint satisfaction result, and baseline comparison result can be determined; and the optimal decision vector, objective function value, constraint satisfaction result, and baseline comparison result can be output as the combinatorial optimization decision result corresponding to the linear objective combinatorial optimization problem.

[0193] In this embodiment of the invention, the above-mentioned input data is used to represent the input data that the trained solution model can recognize and process, and can be obtained by natural language data adaptation.

[0194] The aforementioned predicted objective function parameters can be used to represent the objective function parameters output by the encoder based on the input data. For example, in a path optimization scenario, the predicted objective function parameters can be path edge weight parameters, transportation cost parameters, or demand-related parameters, used to drive subsequent solution interfaces to perform combined optimization solutions.

[0195] The aforementioned optimal decision vector can be used to represent the optimal combination decision of the solution interface based on the predicted data output of the objective function parameters.

[0196] The objective function value described above can be used to represent the optimization objective value corresponding to the optimal decision vector. For example, in the transportation cost minimization problem, the objective function value can be the transportation cost corresponding to the optimal vehicle route; in the revenue maximization problem, the objective function value can be the revenue value corresponding to the optimal decision.

[0197] The above constraint satisfaction results are used to indicate whether the optimal decision vector satisfies the constraints in the optimization model.

[0198] The baseline comparison results described above can be used to represent the difference in performance between the optimal decision vector and the preset baseline method. For example, the decision results output by the trained solution model can be compared with the GLM baseline model, the post-prediction optimization method, or the traditional solution method to obtain results such as cost reduction, regret value improvement, or decision accuracy enhancement.

[0199] In this embodiment of the invention, the above-mentioned linear objective combinatorial optimization problem solving model training and configuration platform can first adapt the natural language data to obtain the input data, input the input data into the encoder in the trained solving model to obtain the objective function parameter prediction data, and call the solving interface based on the objective function parameter prediction data to obtain the optimal decision vector.

[0200] Furthermore, the aforementioned training and configuration platform for solving the linear objective combinatorial optimization problem can also determine the objective function value based on the optimal decision vector, and perform constraint verification on the optimal decision vector in conjunction with the optimization model information to obtain the constraint satisfaction result. Simultaneously, it compares the objective function value or regret value corresponding to the optimal decision vector with the baseline method to obtain the baseline comparison result. Finally, the optimal decision vector, objective function value, constraint satisfaction result, and baseline comparison result are output as the combinatorial optimization decision result corresponding to the linear objective combinatorial optimization problem.

[0201] By using the above methods and steps, natural language data can be converted into an input form that can be processed by a trained solution model, and output combinatorial optimization decision results that include decision schemes, optimization objectives, constraint verification, and baseline comparison, thereby improving the usability and interpretability of the solution results for linear objective combinatorial optimization problems.

[0202] In one embodiment, a training and configuration device for solving a linear objective combinatorial optimization problem is provided, which corresponds one-to-one with the training and configuration method for solving a linear objective combinatorial optimization problem in the above embodiments. For example... Figure 2 As shown, the training configuration device for solving the linear objective combinatorial optimization problem includes: The first processing module 201 is used to perform linear validation processing on the natural language data input by the user, and determine the extended quintuple dataset corresponding to the natural language data. The extended quintuple dataset includes user query information, question type label information, learning signal identification information, objective function linear structure information, and optimization model information. The second processing module 202 is used to perform uncertainty classification processing on the data in the extended quintuple dataset and determine the learning paradigm configuration data used for the training strategy configuration of the solution model. The first determining module 203 is used to determine the training component configuration data of the solution model based on the extended quintuple dataset and the learning paradigm configuration data. The training component configuration data includes encoder configuration data and differentiable layer configuration data. The second determining module 204 is used to determine the loss function for solving model training based on the learning paradigm configuration data and the training component configuration data. Module 205 is used to construct and train a solution model based on the learning paradigm configuration data, the training component configuration data and the loss function, so as to obtain a trained solution model. The output module 206 is used to input the natural language data into the trained solution model and output the combinatorial optimization decision result corresponding to the linear objective combinatorial optimization problem.

[0203] Optionally, the acquisition module 201 is further configured to: Semantic features are extracted from the natural language data to obtain semantic feature elements, which include decision variables, objective functions, constraints, feasible regions, and problem types. Based on the objective function, determine the objective function parameters and the corresponding objective function expression; Linearity detection is performed on the nonlinear terms in the objective function expression to obtain the objective function verification result; Linearity testing is performed on the solution interface corresponding to the natural language data to determine that the solution interface uses the objective function parameters as linear objective coefficients and outputs the decision variables that make the objective function optimal, thereby obtaining the solution interface verification results; The feasible region is subjected to finite combination detection to determine the finiteness verification result of the feasible region; When the verification results of the objective function, the solution interface, and the finiteness of the feasible region are all verified, it is determined that the natural language data satisfies the linearity verification condition, and the corresponding extended quintuple dataset is generated.

[0204] Optionally, the second processing module 202 is further configured to: Obtain the learning signal identification information from the extended quintuple dataset; Based on the learning signal identification information, the data observable state information corresponding to the extended quintuple dataset is determined. The data observable state information includes uncertain parameter information, optimal decision label information, decision cost information, and black box feedback information. Based on the uncertain parameter information, optimal decision label information, decision cost information, and black-box feedback information, the uncertainty classification result is determined. Based on the uncertainty classification results, the learning paradigm configuration data is determined.

[0205] Optionally, the first determining module 203 is further configured to: Based on the problem type label information and learning signal identification information in the extended quintuple dataset, the data structure features corresponding to the extended quintuple dataset are determined; Based on the data structure features and the learning paradigm configuration data, determine the encoder configuration data corresponding to the solution model; Based on the linear structure information of the objective function and the optimization model information in the extended quintuple dataset, the characteristics of the combinatorial optimization solution interface corresponding to the solution model are determined. Based on the combined optimization solution interface features and the learning paradigm configuration data, the differentiable layer configuration data corresponding to the solution model is determined. Based on the encoder configuration data and the differentiable layer configuration data, the training component configuration data of the solution model is determined.

[0206] Optionally, the second determining module 204 is further configured to: Based on the learning paradigm configuration data, the corresponding training supervision signal type is determined; Based on the training supervision signal type, the supervision type of the loss function is determined. The supervision type of the loss function includes structured supervision loss function, decision cost optimization loss function, and policy optimization loss function. Based on the encoder configuration data, differentiable layer configuration data, and supervision type of the loss function, an adaptation process is performed to obtain the loss function used for model training.

[0207] Optionally, the building module 205 is further configured to: Based on the learning paradigm configuration data, the training component configuration data, and the loss function, an initial solution model is constructed. The training samples are input into the encoder in the initial solution model to obtain the target function parameter prediction data; The predicted data of the objective function parameters are input into the differentiable layer of the initial solution model to obtain the corresponding combinatorial optimization decision data. Based on the combined optimization decision data and the loss function, the training loss data is determined; The parameters of the initial solution model are updated based on the training loss data, and the training loss, validation regret value, encoder gradient state and differentiable layer output state are monitored during the training process to obtain training monitoring results. When the training monitoring results meet the preset training stop conditions, the trained solution model is determined.

[0208] Optionally, the output module 206 is further configured to: The natural language data is adapted to obtain the input data. The input data is input into the encoder of the trained solution model to obtain the target function parameter prediction data; Based on the predicted data using the objective function parameters, the solution interface is called to obtain the optimal decision vector; Based on the optimal decision vector, the objective function value, constraint satisfaction result, and baseline comparison result are determined. The optimal decision vector, objective function value, constraint satisfaction result, and baseline comparison result are output as the combinatorial optimization decision result corresponding to the linear objective combinatorial optimization problem.

[0209] Specific limitations regarding the training configuration device for solving linear combinatorial optimization problems can be found in the limitations on the training configuration method for solving linear combinatorial optimization problems described above, and will not be repeated here. Each module in the aforementioned training configuration device for solving linear combinatorial optimization problems can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0210] In one embodiment, a computer device is provided, which may be a terminal device, and its internal structure diagram may be as follows: Figure 3 As shown, the computer device includes a processor, memory, and network interface connected via a system bus. The processor provides computational and control capabilities. The memory includes a readable storage medium storing computer-readable instructions. The network interface communicates with external terminals via a network connection. When executed by the processor, the computer-readable instructions implement a training and configuration method for a linear objective combinatorial optimization problem solution model. The readable storage medium provided in this embodiment includes both non-volatile and volatile readable storage media.

[0211] In this application embodiment, a computer device is provided, including a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor. When the processor executes the computer-readable instructions, it implements the steps of the above-described linear objective combinatorial optimization problem solution model training and configuration method.

[0212] In one embodiment of the application, a readable storage medium is provided, which stores computer-readable instructions. When the computer-readable instructions are executed by a processor, they implement the steps of the above-described method for training and configuring a model for solving a linear objective combinatorial optimization problem.

[0213] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by instructing related hardware with computer-readable instructions. These computer-readable instructions can be stored in a non-volatile readable storage medium or a volatile readable storage medium. When executed, these computer-readable instructions can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0214] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0215] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for training a model for solving a linear combinatorial optimization problem, characterized in that, The method includes: Linear validation processing is performed on the natural language data input by the user to determine the extended quintuple dataset corresponding to the natural language data. The extended quintuple dataset includes user query information, question type label information, learning signal identification information, objective function linear structure information, and optimization model information. Uncertainty classification processing is performed on the data in the extended quintuple dataset to determine the learning paradigm configuration data used for the training strategy configuration of the solution model; Based on the extended quintuple dataset and the learning paradigm configuration data, the training component configuration data for solving the model is determined, including encoder configuration data and differentiable layer configuration data. Based on the learning paradigm configuration data and the training component configuration data, determine the loss function used to solve the model training. Based on the learning paradigm configuration data, the training component configuration data, and the loss function, a solution model is constructed and trained to obtain a trained solution model. The natural language data is input into the trained solution model, which outputs the combinatorial optimization decision result corresponding to the linear objective combinatorial optimization problem. 2.The method of claim 1, wherein, The linear validation process performed on the user-input natural language data to determine the extended quintuple dataset corresponding to the natural language data includes: Semantic features are extracted from the natural language data to obtain semantic feature elements, which include decision variables, objective functions, constraints, feasible regions, and problem types. Based on the objective function, determine the objective function parameters and the corresponding objective function expression; Linearity detection is performed on the nonlinear terms in the objective function expression to obtain the objective function verification result; Linearity testing is performed on the solution interface corresponding to the natural language data to determine that the solution interface uses the objective function parameters as linear objective coefficients and outputs the decision variables that make the objective function optimal, thereby obtaining the solution interface verification results; The feasible region is subjected to finite combination detection to determine the finiteness verification result of the feasible region; When the verification results of the objective function, the solution interface, and the finiteness of the feasible region are all verified, it is determined that the natural language data satisfies the linearity verification condition, and the corresponding extended quintuple dataset is generated.

3. The linear target combination optimization problem solution model training configuration method of claim 1, wherein, The uncertainty classification processing of the data in the extended quintuple dataset to determine the learning paradigm configuration data used for solving the model training strategy configuration includes: Obtain the learning signal identification information from the extended quintuple dataset; Based on the learning signal identification information, the data observable state information corresponding to the extended quintuple dataset is determined. The data observable state information includes uncertain parameter information, optimal decision label information, decision cost information, and black box feedback information. Based on the uncertain parameter information, optimal decision label information, decision cost information, and black-box feedback information, the uncertainty classification result is determined. Based on the uncertainty classification results, the learning paradigm configuration data is determined.

4. The method for training and configuring a model for solving a linear objective combinatorial optimization problem as described in claim 1, characterized in that, The step of determining the training component configuration data for solving the model based on the extended quintuple dataset and the learning paradigm configuration data includes: Based on the problem type label information and learning signal identification information in the extended quintuple dataset, the data structure features corresponding to the extended quintuple dataset are determined; Based on the data structure features and the learning paradigm configuration data, determine the encoder configuration data corresponding to the solution model; Based on the linear structure information of the objective function and the optimization model information in the extended quintuple dataset, the characteristics of the combinatorial optimization solution interface corresponding to the solution model are determined. Based on the combined optimization solution interface features and the learning paradigm configuration data, the differentiable layer configuration data corresponding to the solution model is determined. Based on the encoder configuration data and the differentiable layer configuration data, the training component configuration data of the solution model is determined.

5. The linear target combination optimization problem solving model training configuration method of claim 4, wherein, The step of determining the loss function for model training based on the learning paradigm configuration data and the training component configuration data includes: Based on the learning paradigm configuration data, the corresponding training supervision signal type is determined; Based on the training supervision signal type, the supervision type of the loss function is determined. The supervision type of the loss function includes structured supervision loss function, decision cost optimization loss function, and policy optimization loss function. Based on the encoder configuration data, differentiable layer configuration data, and supervision type of the loss function, an adaptation process is performed to obtain the loss function used for model training.

6. The linear target combination optimization problem solution model training configuration method of claim 1, wherein, The step of constructing and training a solution model based on the learning paradigm configuration data, the training component configuration data, and the loss function to obtain a trained solution model includes: Based on the learning paradigm configuration data, the training component configuration data, and the loss function, an initial solution model is constructed. The training samples are input into the encoder in the initial solution model to obtain the target function parameter prediction data; The predicted data of the objective function parameters are input into the differentiable layer of the initial solution model to obtain the corresponding combinatorial optimization decision data. Based on the combined optimization decision data and the loss function, the training loss data is determined; The parameters of the initial solution model are updated based on the training loss data, and the training loss, validation regret value, encoder gradient state and differentiable layer output state are monitored during the training process to obtain training monitoring results. When the training monitoring results meet the preset training stop conditions, the trained solution model is determined.

7. The linear target combination optimization problem solving model training configuration method of claim 1, wherein, The step of inputting the natural language data into the trained solution model and outputting the combinatorial optimization decision result corresponding to the linear objective combinatorial optimization problem includes: The natural language data is adapted to obtain the input data. The input data is input into the encoder of the trained solution model to obtain the target function parameter prediction data; Based on the predicted data using the objective function parameters, the solution interface is called to obtain the optimal decision vector; Based on the optimal decision vector, the objective function value, constraint satisfaction result, and baseline comparison result are determined. The optimal decision vector, objective function value, constraint satisfaction result, and baseline comparison result are output as the combinatorial optimization decision result corresponding to the linear objective combinatorial optimization problem.

8. A linear target combination optimization problem solution model training configuration device, characterized in that, Implementing the training configuration method for a linear objective combinatorial optimization problem solution model as described in claim 1, comprising: The first processing module is used to perform linear validation processing on the natural language data input by the user, and determine the extended quintuple dataset corresponding to the natural language data. The extended quintuple dataset includes user query information, question type label information, learning signal identification information, objective function linear structure information, and optimization model information. The second processing module is used to perform uncertainty classification processing on the data in the extended quintuple dataset and determine the learning paradigm configuration data used for the training strategy configuration of the solution model. The first determining module is used to determine the training component configuration data of the solution model based on the extended quintuple dataset and the learning paradigm configuration data. The training component configuration data includes encoder configuration data and differentiable layer configuration data. The second determining module is used to determine the loss function for solving model training based on the learning paradigm configuration data and the training component configuration data. The construction module is used to construct and train the solution model based on the learning paradigm configuration data, the training component configuration data and the loss function, so as to obtain the trained solution model. The output module is used to input the natural language data into the trained solution model and output the combinatorial optimization decision result corresponding to the linear objective combinatorial optimization problem. 9.A computer device, comprising a memory, a processor, and computer readable instructions stored on the memory and running on the processor, wherein, When the processor executes the computer-readable instructions, it implements the training and configuration method for solving the linear objective combinatorial optimization problem as described in any one of claims 1 to 7.

10. A readable storage medium, having stored thereon computer readable instructions, characterized in that, When the computer-readable instructions are executed by the processor, they implement the training and configuration method for solving the linear objective combinatorial optimization problem as described in any one of claims 1 to 7.

Images