An intelligent decision support system and method based on longitudinal model conductive chain

Abstract

The application discloses a kind of intelligent decision support system and method based on longitudinal model conduction chain, belong to data processing system technical field, wherein, the system includes: model conduction engine, for building ordered model sequence, the output vector of each model in sequence as the input of subsequent model, form continuous decision conduction chain;Dynamic weight module, for realizing the weighted transmission of predecessor model output to subsequent model input by conduction matrix, the conduction matrix is based on model confidence error and dynamically adjusted;Closed-loop control unit, for monitoring the actual achievement situation of decision target, realizes the global optimization of longitudinal conduction chain by optimizing full-stack target function.By the application, it avoids the repeated processing of original data, realizes the ordered conduction of knowledge between models, improves the consistency, efficiency and robustness of decision.

Description

Technical Field

[0001] This invention relates to the field of data processing system technology, specifically to a decision support system and method, which is particularly suitable for scenarios requiring continuous decision-making across multiple stages, such as enterprise strategic planning, marketing, and operations management. By constructing a vertical transmission architecture between models, the system achieves the integration and optimization of decision-making logic. Background Technology

[0002] Currently widely used intelligent decision support systems generally adopt a horizontal parallel architecture of models. Under this architecture, various commonly used business management models (such as SWOT analysis, balanced scorecard, Ansoff matrix, etc.) are deployed as independent modules, lacking organic connection and data transmission between them, resulting in structural breaks in the decision chain.

[0003] From a corporate decision-making process perspective, it typically involves multiple consecutive stages, including "environmental diagnosis, strategy formulation, resource allocation, execution monitoring, and effect evaluation," each corresponding to a specific business management model. For example, the environmental diagnosis stage relies on the PESTEL and SWOT models, the strategy formulation stage uses the Ansoff matrix, the resource allocation stage uses the Boston Consulting Group (BCG) matrix, and the execution monitoring stage relies on the Balanced Scorecard. However, in traditional systems, these models are designed as parallel modules, each receiving raw data input and outputting results independently, without a pre-defined logical transmission path between them. This horizontally isolated architecture leads to severe decision-making gaps. For instance, market opportunities identified by the SWOT model cannot be automatically transmitted to the Ansoff matrix, resulting in a disconnect between strategy formulation and market diagnosis; execution deviations monitored by the Balanced Scorecard cannot be fed back to the earlier resource allocation model, making corrective measures lack specificity. Decision-makers need to manually integrate the outputs of each model, which not only increases decision-making time costs but may also lead to logical breaks due to human interpretation biases. A case study of a manufacturing company shows that because the "technological advantages" output by the SWOT model in its decision-making system were not taken into account by the Ansoff matrix, the final market development strategy it formulated ignored its own technological capabilities, resulting in insufficient product competitiveness and failure to expand the market.

[0004] Furthermore, the horizontal isolation of models and the lack of coordination and integration mechanisms for their outputs can lead to fragmented decision-making perspectives. When the outputs of different models conflict (e.g., the SWOT model recommends "expansion" while the cash flow model indicates "risk"), the lack of a conflict reconciliation mechanism in the vertical transmission makes it difficult for decision-makers to determine the root cause of the contradictions. They can only rely on experience to make trade-offs, increasing the risk of decision-making.

[0005] In traditional decision-making systems, business management models typically rely directly on raw data (such as market research data, financial statement data, and operational indicator data) as input, lacking effective utilization of the output results of precursor models. This leads to redundant data processing, low decision-making efficiency, and neglects the progressive and continuous nature that decision-making logic should possess.

[0006] In a horizontally parallel architecture, the input interface of each model is designed to directly connect to the original database. Even if there is a clear logical relationship between two models, such as Porter's Five Forces model and a pricing strategy model, the subsequent model still directly reads the original data and recalculates, rather than inheriting the analytical conclusions of the predecessor model. This homogeneous input mode has three problems:

[0007] First, there is data processing redundancy. The same batch of raw data is repeatedly read, cleaned, and transformed by multiple models, which not only consumes a lot of computing resources, but may also lead to inconsistencies in data interpretation due to different processing methods. For example, in the traditional system of a retail company, both the SWOT model and the 4P model need to process customer satisfaction data, but the former uses a "five-point scale" for transformation and the latter uses a "percentage scale" for transformation, making it difficult to directly compare and integrate the output results of the two models.

[0008] Second, there is a disconnect in decision-making logic. Key information extracted by the precursor model (such as the "policy-favorable trend" derived from the PESTEL model) could have served as high-quality input to support subsequent strategy formulation. However, traditional systems force subsequent models to re-extract information from the original data, reducing decision-making efficiency and increasing the risk of overlooking important conclusions. In the case of an internet company, the conclusion that "young users prefer short videos" output by the user profiling model was not utilized by the product development model. The latter re-analyzed the original user behavior data, ultimately leading to product design deviations due to the omission of this crucial information.

[0009] Third, decision-making efficiency is low. Processing raw data takes far longer than the structured transformation of the output of the precursor model. Especially when the data volume is large, parallel processing of raw data by multiple models will significantly prolong the decision-making cycle. A test by a financial institution showed that in a traditional system, the total time taken by five parallel models to process the same raw data is 4.2 times that of the vertical transmission architecture.

[0010] In traditional decision-making systems, there is a lack of effective confidence assessment and correction mechanisms among various business management models. The output error of the predecessor model is directly inherited by the successor model and amplified step by step, ultimately leading to a serious deviation of the decision result from reality, and it is difficult to trace the source of the error.

[0011] In actual decision-making processes, the output of each model may introduce a certain degree of error (such as bias in market forecasting models or misjudgment in risk assessment models). In a vertical decision-making chain, if the precursor model has a 10% error, the error may be amplified to 20%-30% when subsequent models calculate based on this result. After multiple stages, the final decision error may exceed 50%. For example, in the chain of "market demand forecasting - production planning - procurement planning", if the demand forecasting model overestimates market demand by 10%, the production planning model will accordingly arrange 10% more capacity, and the procurement planning model will further purchase 10% more raw materials, which may ultimately lead to more than 10% inventory backlog. Traditional systems cannot identify this accumulation process of error.

[0012] More seriously, traditional systems lack error tracing mechanisms. When a final decision deviates from its intended purpose, decision-makers cannot determine which model in which stage caused the problem, and must instead conduct a comprehensive check of all models, resulting in extremely high investigation costs. One automaker, for example, launched an investigation after new car sales fell far short of expectations. Because the traditional system could not trace the error propagation path of each model, it took three months to discover that the initial market segmentation model had a flawed target audience, leading to a chain reaction of errors in subsequent pricing and distribution models.

[0013] Furthermore, the difference in confidence levels between different models is not taken into account. For example, one model may have a historical prediction accuracy of 90%, while another model may only have 70%. However, traditional systems assign equal weights to both models. When the output of the low-confidence model is used as the input of the high-confidence model, it will significantly reduce the overall decision quality.

[0014] In summary, existing intelligent decision support systems generally adopt a model architecture of horizontal parallelism, homogeneous input, and isolated operation, which leads to problems such as broken decision-making processes, redundant data processing, amplified error accumulation, and lack of conflict reconciliation mechanisms. As a result, they are difficult to adapt to the requirements of decision-making consistency, efficiency, and accuracy in complex business environments. Summary of the Invention

[0015] To address the aforementioned technical problems, this invention proposes an intelligent decision support system and method based on a longitudinal model transmission chain.

[0016] According to one aspect of the present invention, an intelligent decision support system based on a longitudinal model transmission chain is provided, comprising a model transmission engine, a dynamic weight module, and a closed-loop control unit. The model transmission engine is used to construct an ordered model sequence, where the output vector of each model in the sequence serves as the input of the subsequent model, forming a continuous decision transmission chain. The model transmission engine includes two sub-modules: a model sequence configuration module and a model interface adaptation unit. The model sequence configuration module supports the ordered combination of multiple types of models and automatically generates the model execution order based on user settings or a historical decision case library. The model interface adaptation unit standardizes and converts the data formats of different model outputs to ensure data compatibility between models. The dynamic weight module is used to implement weighted transfer of the output of the preceding model to the input of the subsequent model through a transmission matrix, wherein the transmission matrix is ​​dynamically adjusted based on the model confidence error. The confidence error is the deviation between the model's predicted output and the actual result, expressed in 2-general form. When the confidence error exceeds a first preset threshold, the transmission matrix adjustment process is initiated, and the element values ​​of the transmission matrix are optimized through the gradient descent algorithm until the confidence error falls back to a safe range. The closed-loop control unit is used to monitor the actual achievement of the decision objective and calculate the target offset. When the target offset exceeds a set threshold, the strategy correction mechanism is triggered. Optimization instructions are generated by optimizing the full-stack objective function and fed back to the model transmission engine to achieve global optimization of the transmission chain. The full-stack objective function includes a precision cost term and a smoothing penalty term. The precision cost term is the sum of the model's squared error after weighting by the model hierarchy weights. The smoothing penalty term is the sum of the differences between adjacent transmission matrices. The differences between adjacent transmission matrices are calculated using the F-norm form after subtracting adjacent transmission matrices.

[0017] Optionally, the dynamic weighting module implements weighted transfer from the output of the predecessor model to the input of the successor model through a transmission matrix, specifically as follows:

[0018] The output vector of the predecessor model is weighted by the transmission matrix, then a bias term is added, and finally the output vector is generated by the model-specific operator. This output vector is then used as the input to the subsequent model. The formula is as follows:

[0019] ;

[0020] in, :Model The output vector;

[0021] : Dynamic transmission matrix, elements in the matrix express The Dimension to model The The influence weight of dimensional inputs;

[0022] :Model The bias term is used to correct for systematic errors;

[0023] :Model The dedicated operators are customized according to the model's functions;

[0024] :Model The output vector is used as the successor model. The core input.

[0025] Optionally, the transmission matrix Based on dynamic adjustment of model confidence error, the model's confidence error Defined as the L2 distance between the model's predicted output and the target output, i.e. ,in, For the model The predicted output, For the model The adjustment rule for the transmission matrix is ​​as follows:

[0026] ;

[0027] In the formula, : Transmission matrix at time step;

[0028] : Transmission matrix at time step;

[0029] Learning rate controls the step size for weight updates;

[0030] :Model The predicted output;

[0031] :Model The target output;

[0032] The dynamic weighting module also adjusts the transmission matrix. The values ​​of each element in Apply boundary constraints. When an element value exceeds the set boundary during the optimization process, it is automatically truncated to the corresponding boundary value, and the values ​​of other elements in the transmission matrix are adjusted proportionally.

[0033] Optionally, the dedicated operator Based on model functionality customization, different dedicated operators are used for different types of models, specifically including:

[0034] Strategic diagnostic models employ a multi-dimensional weighted summation operator:

[0035] ;

[0036] in, This is an operator specifically for strategic diagnostic models. For the first The weights of each diagnostic dimension, For the first Input values ​​for each diagnostic dimension, d Total number of diagnostic dimensions For dimensional indexing, ;

[0037] Resource allocation models employ linear programming operators:

[0038] ;

[0039] in, This is an operator specific to resource allocation models, with the following constraints: , For the first Cost coefficient of similar resources For the first The amount of resources invested in this category It is the first The constraint on the first Resource consumption coefficient, For the first The minimum requirement value for each constraint, For the number of resource types, The number of constraints;

[0040] Performance evaluation models employ linear programming operators:

[0041] ;

[0042] in, This is an operator specifically for performance evaluation models. For the first The weights of each evaluation dimension, For the first Dimension 1 The weight of each indicator, For the first The first dimension The actual value of each indicator To evaluate the dimensional coefficients, The number of metrics for each dimension, To evaluate the dimensional index, For indexing indicators.

[0043] Optionally, the dynamic weight module also integrates a confidence correction mechanism, used to adaptively correct the vector input to the current model based on the reliability of the predecessor model's output. The correction formula is expressed as:

[0044] ;

[0045] in, Corrected model The input vector;

[0046] Model before correction The input vector, , For the model The output vector;

[0047] Hadamard product, element-wise multiplication;

[0048] : Industry sensitivity coefficient, controlling the strength of the impact of confidence level on the correction;

[0049] :Model The confidence level is calculated from the historical accuracy:

[0050] ;

[0051] in, It is a model In the The confidence error in historical decision-making, where T is the total number of historical decision samples.

[0052] Optionally, the formula for calculating the target offset in the closed-loop control unit is:

[0053] ;

[0054] In the formula, The target offset. To preset the target value for decision-making, This is the actual value achieved.

[0055] The closed-loop control unit is based on the target offset. A tiered response is achieved using dual-threshold triggering logic:

[0056] when When the value is less than or equal to the primary threshold, a risk warning will be issued and offset details will be displayed, but automatic correction will not be initiated, allowing users to make their own judgment and intervention.

[0057] When multiple consecutive decision cycles When the value exceeds the advanced threshold, an advanced response is triggered, a correction strategy is initiated, and an early warning is issued. At the same time, the warning information is pushed to the decision-maker.

[0058] Optionally, the full-stack objective function of the closed-loop control unit is defined as:

[0059] ;

[0060] in, For all transmission matrices Find the minimum value;

[0061] Model hierarchy weights reflect the importance of the model in the decision chain;

[0062] :Model The squared error, , For the model The predicted output, For the model The target output;

[0063] : Transmission continuity penalty term, used to constrain the difference between adjacent transmission matrices;

[0064] Frobenius norm, used to measure the performance of adjacent transmission matrices. and The differences between them should be considered to avoid abrupt changes in the transmission logic that could lead to a break in the decision chain.

[0065] The total number of models in the model sequence;

[0066] The full-stack objective function is optimized using gradient descent. The process is as follows: Step 1, initialize the transmission matrix based on historical data. Step 2: Calculate the objective function value. : Step 3: Solve for the objective function with respect to the transmission matrix. Partial derivatives: Step 4: Update the transmission matrix along the negative gradient direction. , The global learning rate; for Transmission matrix at time step; The updated transmission matrix is ​​obtained; steps 2-4 are repeated until... The difference between adjacent iterations is less than the second preset threshold.

[0067] Optionally, the closed-loop control unit further includes a confidence feedback mechanism to feed back the evaluation results of the terminal model to the front-end model in the form of confidence scores, thereby correcting its decision-making logic. The feedback path of the confidence feedback mechanism includes: calculating the terminal model... Actual results Compared with the prediction results Deviation between : Based on deviation Update the predecessor model according to predefined rules. Confidence level: ,in, For feedback coefficients, For the model confidence level The updated confidence level is used to correct the error. Output: , This is the industry sensitivity coefficient, used to control the strength of the impact of confidence level on the correction. based on Restart the process, initiate a new round of decision-making, and achieve dynamic correction.

[0068] Optionally, the closed-loop control unit further includes a strategy correction submodule, the execution flow of which is as follows: Monte Carlo simulation: Based on the current decision chain parameters, the Monte Carlo method is used to generate a preset number of random scenarios, with random variables based on historical data. The interval, the decision chain parameters include the transmission matrix. Model operators Probability distribution generation: Statistical simulation results are used to calculate the target offset. The probability distribution is calculated, a probability distribution histogram is generated, and the high-risk probability is calculated, which refers to the probability that the target offset exceeds a third preset threshold. A contradiction matrix is ​​invoked for judgment: when the high-risk probability exceeds the preset risk threshold, the contradiction matrix library is invoked to match the corresponding correction strategy; otherwise, the current strategy is maintained. The contradiction matrix library stores several mapping relationships between innovation principles and decision contradictions. An optimized strategy is output: multiple applicable innovation principles are matched to generate an executable strategy that includes adjustment parameters, expected effects, and implementation steps.

[0069] According to another aspect of the present invention, an intelligent decision support method based on a vertical model transmission chain is provided. Based on the aforementioned intelligent decision support system based on a vertical model transmission chain, the method includes: model sequence initialization: loading a model sequence according to a preset decision logic, detecting the compatibility of each model interface, and if a format conflict is detected, initiating a format conversion tool to convert the format and unify the output into a compatible data format; vertical data transmission: weighting the output of the predecessor model through a transmission matrix, adding a bias term, and then calling a dedicated operator to generate an output vector, which is input to the successor model; closed-loop dynamic optimization: calculating the target offset, and if the target offset exceeds a set threshold, initiating full-stack objective function optimization, adjusting the transmission matrix, and using the updated transmission matrix in the next round of transmission to reduce the target offset. The full-stack objective function includes a precision cost term and a smoothing penalty term. The precision cost term is the sum of the model squared errors weighted by model hierarchy weights, and the smoothing penalty term is the sum of the differences between adjacent transmission matrices. The differences between adjacent transmission matrices are calculated using the F-norm form obtained by subtracting adjacent transmission matrices.

[0070] This invention presents an intelligent decision support system based on a vertical model transmission chain. Employing a vertical model overlay architecture, it utilizes a three-tiered component structure—model transmission engine, dynamic weight module, and closed-loop control unit—to achieve continuous decision logic and end-to-end quality control. The core innovative framework uses "predecessor model output → successor model input" as its central logic, forming a closed-loop decision chain of "diagnosis-decision-execution-evaluation-correction." Model sequence configuration can be user-defined or generated based on a case library. The system also supports interface adaptation, converting outputs of different formats into numerical vectors through standardized protocols to ensure data compatibility between models.

[0071] This invention integrates fragmented business management models into an organic whole by constructing an ordered transmission chain, dynamic weight transfer, and confidence correction mechanisms. This ensures that the decision-making logic remains continuous from the starting point to the end point. At the same time, by using dynamic weights and confidence correction, the quality of decision-making at each stage is ensured, achieving orderly continuity of decision-making logic and full-process quality control, thus significantly improving the overall decision-making performance of the system.

[0072] Compared with traditional architectures, the intelligent decision support system based on the vertical model transmission chain of this invention exhibits significant advantages in several key dimensions, specifically:

[0073] Regarding the integrity of the decision chain, traditional architectures suffer from logical incoherence and structural breaks due to the use of parallel and isolated models. This invention uses a vertical transmission mechanism to drive subsequent inputs with the output of the predecessor, ensuring logical consistency and continuous flow, thereby avoiding decision-making gaps.

[0074] In terms of error control, traditional architectures lack correction mechanisms, resulting in significant error accumulation. This invention combines confidence correction and closed-loop optimization, using dynamic weight adjustment to offset system errors and employing confidence to reduce deviations.

[0075] In terms of data utilization efficiency, traditional architectures require each model to repeatedly read the original data; this invention uses the predecessor output as the core input and realizes the structured reuse of information through the transmission matrix, which greatly reduces the dependence on the original data.

[0076] In terms of decision-making efficiency, traditional architectures take a long time on average to process raw data in parallel; this invention reduces redundant calculations through a vertical transmission mechanism, and the processing results can be directly reused between models, thus shortening the processing time.

[0077] In terms of environmental adaptability, traditional architectures have fixed weights, making it difficult to cope with environmental changes. This invention uses time-varying weights and adjustable sensitivity coefficients to achieve weight self-adaptation, which can flexibly respond to market fluctuations and industry differences, demonstrating higher adaptability. Attached Figure Description

[0078] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0079] Figure 1 This is a schematic diagram of the structure of an intelligent decision support system based on a vertical model transmission chain according to an embodiment of the present invention;

[0080] Figure 2 This is a flowchart illustrating an intelligent decision support method based on a vertical model transmission chain according to an embodiment of the present invention. Detailed Implementation

[0081] To enable those skilled in the art to clearly understand the technical solutions of this application, the technical solutions in the embodiments of this application are described in a complete and rigorous manner below with reference to the accompanying drawings and examples. Obviously, the described embodiments are merely some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other.

[0082] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0083] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, product, or apparatus that comprises a series of steps or units is not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus. The term "module," as used below, can refer to a combination of software and / or hardware that performs a predetermined function. Although the apparatus described in the following embodiments is preferably implemented in software, hardware implementations, or a combination of software and hardware, are also possible and contemplated.

[0084] Example 1: This example is designed based on a real business scenario. The data and logic form a closed loop with the claims and background technology. Features that do not conflict in this example can be combined with each other.

[0085] Reference Figure 1 , Figure 1 This is a schematic diagram of the intelligent decision support system based on a vertical model transmission chain according to an embodiment of the present invention. The system solves the core pain points of traditional decision systems, namely "model isolation, data redundancy, and error out-of-control", through a collaborative architecture of "model transmission engine - dynamic weight module - closed-loop control unit". The technical details of each module are as follows:

[0086] (1) Model transfer engine:

[0087] Core functionality: Constructing logically coherent and ordered sequences of models. Each model ( The output vector of ) As a successor model The core input source, namely, ( For the model The inputs are used to form a continuous decision transmission chain and can achieve cross-model data format compatibility, ensuring the orderly transmission of decision information from front-end diagnosis to back-end execution.

[0088] in, It represents the entire ordered sequence of models. It is an abstract general term that refers to the complete chain of models constructed for a specific decision-making scenario (such as corporate strategic planning or marketing strategy).

[0089] It is an abstract reference to a single model in the sequence, indicated by the subscript number. This indicates the order and position of the model in the transmission chain.

[0090] The model representing the starting point of the transmission chain is responsible for the analysis or diagnosis at the very front end (such as strategic diagnostic models or market environment analysis models).

[0091] It represents any intermediate model in the sequence.

[0092] The terminal model representing the transmission chain is responsible for the final evaluation or execution (such as performance evaluation model, risk warning model).

[0093] The model transfer engine consists of two sub-modules:

[0094] ① The model sequence configuration module is used to generate a complete model chain of "diagnosis-decision-execution-evaluation" based on the core objectives of the decision-making scenario. It supports two modes: "custom configuration" and "case library recommendation".

[0095] Custom configuration: Users can drag and drop model components according to business logic or decision-making scenarios to customize the model order. For example, in a new product launch decision-making scenario, the model sequence can be configured as: "Regional Market Research Model ( → Channel Priority Ranking Model → Promotional resource allocation model ( → Sales Performance Tracking Model )”;

[0096] Case Study Recommendation: Based on a historical case study database of enterprise decisions (containing at least 3 years of industry decision data, with a sample size of ≥100), the optimal sequence is recommended using cosine similarity matching. Taking an enterprise strategic decision-making scenario as an example, the recommended sequence is: Macro-environmental analysis model (…). → Competitive Situation Assessment Model → Strategic Goal Decomposition Model → Resource matching model → Execute monitoring model ( The functions of each model are as follows:

[0097] Macro-environmental analysis model Input policy, economic, and social data, and output dimensions such as "policy benefit coefficient" and "economic cycle matching degree";

[0098] Competitive Situation Assessment Model Input competitor market share and product differentiation data, and output "competitive intensity level" and "differentiation advantage value";

[0099] Strategic Goal Decomposition Model :enter , Output "Revenue target breakdown value" and "Market share target value";

[0100] Resource matching model Input the resource requirements of each business line and output the "funds / human resources allocation ratio";

[0101] Execution monitoring model Input monthly execution data and output "target achievement progress" and "deviation warning signal".

[0102] ② The model interface adaptation unit eliminates the differences in output formats between different models through standardized protocols, ensuring that the output of the predecessor model can be directly used as the input of the successor model. The core process is "format recognition → rule conversion → compatibility verification".

[0103] Format recognition: Automatically identifies the model output type, including structured data (such as Excel spreadsheets), matrix data (such as 3×3 competition intensity matrix), and text labels (such as "high / medium / low chance").

[0104] Rule Conversion: Based on industry-standard conversion rules or historical enterprise rules, non-numerical formats are converted into numerical vectors of uniform dimensions (the number of dimensions is determined according to the input requirements of subsequent models, with a default of 3-5 dimensions). For example, based on the label-result mapping relationship of similar decisions in the past two years, the text label "Market Opportunity Level" is set to "Low Opportunity = 0.3, Medium Opportunity = 0.6, High Opportunity = 0.8"; the matrix data "Competitive Intensity Matrix" is compressed into a vector by taking the mean of each row, such as converting a 3×3 matrix into a vector of [0.5, 0.7, 0.4].

[0105] Verify compatibility: Verify whether the dimension and numerical range (default 0~1) of the transformed vector are consistent with the input requirements of the subsequent model. If they do not match, trigger secondary processing of "dimensional compression / expansion" or "numerical normalization".

[0106] (2) Dynamic weight module:

[0107] Core Functionality: This module enables weighted transfer of data from the predecessor model's output to the successor model's input via a transfer matrix. It dynamically adjusts the transfer matrix based on the confidence error to prevent transfer failure caused by fixed weights. When the confidence error exceeds a first preset threshold (default 0.15), the transfer matrix adjustment process is initiated, optimizing the element values ​​using a gradient descent algorithm until the confidence error falls back to a safe range (≤0.008). The dynamic weight module comprises five sub-mechanisms:

[0108] ① Vertical transmission mathematical model:

[0109] The vertical transmission mathematical model is the core tool for the dynamic weight module to achieve ordered transmission between models. It ensures that decision information is neither distorted nor misrepresented during transmission, while still meeting the computational needs of the subsequent model, by quantifying the mapping relationship between the output of the predecessor model and the input of the successor model. This is achieved through the transmission matrix. The output of the predecessor model is weighted, a bias term is added, and a dedicated operator is used to generate an output vector, which is then used as the input to the successor model. This achieves weighted transfer of the predecessor model's output to the successor model's input. The longitudinal transmission mathematical model is expressed as:

[0110] ;

[0111] The above equation is the core equation for longitudinal transmission, and the meanings of each parameter are as follows:

[0112] :Model The output vector represents the model. The decision outcome, each component of which corresponds to a specific decision characteristic, for example, in a market diagnostic model... Output The components can be used to represent quantitative values ​​of characteristics such as market size, growth rate, and competitive intensity, in turn. , For the model The input vector;

[0113] The dynamic transmission matrix, whose element values ​​represent the influence weights of each dimension of the predecessor model's output on the successor model's input. Each element in the matrix... The precursor model was quantified. The j-th feature is used to construct subsequent models The relative weight of the i-th input feature can be initially obtained by linear regression fitting of historical decision data from the past 3 years;

[0114] :Model The bias term is used to correct inherent systematic errors caused by industry characteristics, data deviations, etc. during the transmission process, such as differences in resource conversion rates caused by industry characteristics.

[0115] :Model The dedicated operators are customized according to the model's functions. Different types of models use different operational logics. For example, the strategic diagnosis model uses a weighted summation operator, while the resource allocation model uses a linear programming operator.

[0116] :Model The output vector is used as the next level model. The input source.

[0117] The core logic of this vertical transmission mathematical model equation is that the output of the successor model depends not only on its own operation rules, but also on the result of the output of the predecessor model after being transformed by the transmission matrix, thereby ensuring the vertical coherence of the decision-making logic at the mathematical level.

[0118] Taking the transmission from "strategic diagnosis to resource allocation" as an example, the formula is: ;

[0119] in, ( The output is the "market opportunity vector". This includes "market size (0.8), growth rate (0.6), and competitive intensity (0.3)" (value range 0~1).

[0120] The (transmission matrix) is a 2×3 matrix, representing the transformation relationship of "market opportunity → resource input": The first row corresponds to the "capital input coefficient", and the second row corresponds to the "human resource input coefficient".

[0121] (Bias term) is This is used to compensate for the industry's basic resource needs;

[0122] but = + = ;

[0123] The linear programming operator for the resource allocation model (is a mapping) takes the input vector after processing by the transmission matrix and bias terms. Transform into an output vector (Resource allocation plan).

[0124] Through this transmission process, the assessment results of market opportunities are directly transformed into the basis for resource allocation decisions, avoiding the problem of resource allocation being disconnected from market diagnosis in traditional systems. This vertical transmission mathematical model explicitly translates the precursor output... As a core variable, it replaces the " ( The general form of the original data is used to mathematically lock in the vertical transmission relationship.

[0125] ② Adaptive adjustment of the transmission matrix:

[0126] The adaptive adjustment of the transmission matrix (i.e., weights) enables the transmission relationship between models to be optimized in real time according to the decision effect and environmental changes, avoiding the problem of "fixed weights causing transmission failure" in traditional systems.

[0127] Transmission matrix for A real matrix of dimension ( To provide dimensions for subsequent models, (output dimension of the predecessor model), transmission matrix elements in The output vector of the predecessor model is represented by the first... The first dimension inputs the successor model. The influence weights of each dimension are initially obtained through training with historical decision data. For example, in the transmission of "market opportunity → resource input", =0.7 indicates that the "market size" dimension has a weight of 70% over the "capital input" dimension.

[0128] Transmission matrix Based on confidence error Dynamic adjustments are made, modifying matrix element values ​​in real time. Confidence error. Defined as a model The deviation between the predicted output and the target output is expressed in L2 form, i.e., (2-norm), where yes The predicted output, yes The target output. When Exceeding the first preset threshold, such as When the decision accuracy requirements need to be reset, the dynamic weight module initiates the matrix adjustment process, optimizing the matrix using the gradient descent algorithm. The element values ​​are counted until the error falls back to the safe range.

[0129] Transmission matrix The update rules are as follows:

[0130] ;

[0131] in, For the first Transmission matrix at time step; For the first Transmission matrix at time step; The learning rate (default 0.01, range 0.001~0.1, can be adjusted according to industry characteristics) is used to control the update step size; For the model The predicted output; For the model Target output; partial derivative terms For the squared loss function pair The gradient indicates the direction of weight adjustment.

[0132] Gradient direction: when When the gradient is positive, the corresponding Increase, for example, when the model The predicted output of the (resource allocation model) (capital input 0.65, i.e.) When the gradient is lower than the target output (0.7) (=0.65), the gradient is positive. The corresponding weights for the "market size" dimension (e.g.) The value will increase from 0.7 to 0.75, enhancing the impact of this dimension on resource input in subsequent transmission and gradually narrowing the gap between the forecast and the target.

[0133] Convergence condition: when (When the difference between the F-norm and the matrix is ​​less than the set convergence threshold, the iteration stops.)

[0134] ③ Weight boundary control:

[0135] The dynamic weight module also applies to the transmission matrix. Boundary constraints are applied to each element value. When an element value exceeds a set boundary during optimization, it is automatically truncated to the corresponding boundary value. This prevents the model from becoming overly reliant on a single factor due to excessively high weights in a single dimension, or from missing information due to excessively low weights. For example, the element value range is set to... ,like >0.8, automatically truncated to 0.8. If the value is less than 0.2, it will automatically be increased to 0.2.

[0136] The weights of the excess portions are allocated according to the original proportions of the remaining dimensions, ensuring that the total weight of that row remains unchanged. For example, if the weight of the "Market Opportunities" dimension... =0.85 (0.05 over the limit), remaining dimensions =0.3, =0.2 (original ratio 3:2), then after adjustment , After adjustment, the total weight of this row remains 0.8 + 0.33 + 0.22 = 1.35 (consistent with the previous weight). This control mechanism makes decision-making more robust and avoids the disruption of the transmission chain by extreme values.

[0137] ④ Model-specific operator design:

[0138] Model Dedicated operators Depending on the model's functionality, different types of business management models have different computational logics, such as:

[0139] Strategic diagnostic models ( A multi-dimensional weighted summation operator is used to perform weighted synthesis on multiple diagnostic dimensions:

[0140] ;

[0141] Operators specific to strategic diagnostic models;

[0142] : No. Actual input values ​​for each diagnostic dimension (e.g., policy support strength score 0.8, market demand growth rate 0.6).

[0143] Weights of each diagnostic dimension (e.g., policy factors 0.3, market factors 0.4, technology factors 0.3);

[0144] Dimension index ( );

[0145] Total number of diagnostic dimensions (determined according to industry standards, such as 3-5 dimensions by default for strategic diagnostics);

[0146] Resource allocation model ( ), using linear programming operators:

[0147] ;

[0148] The constraints are ;

[0149] in, Operators specific to resource allocation models;

[0150] : is the first The amount of resources invested (such as capital investment and human resources investment).

[0151] : is the cost coefficient, representing the first The unit cost of such resources (e.g., capital cost of 0.08 million yuan / unit, human resource cost of 0.05 million yuan / person);

[0152] Technical coefficient, representing the unit of resources Constraints Contribution / consumption (e.g., producing 1 unit of product requires 2 units of capital and 1 person of manpower);

[0153] : No. Minimum demand values ​​for each constraint (e.g., minimum output of 1000 units, minimum number of service personnel of 500).

[0154] p: Number of resource types (e.g., funds and human resources).

[0155] q: Number of constraints (e.g., output, number of people served, 2);

[0156] : is an operator in mathematical optimization, which means finding the solution (or solutions) that minimize the objective function from the set of solutions that satisfy all constraints.

[0157] It is the objective function, representing the total cost to be minimized.

[0158] It is a set of linear inequality constraints that ensure that the resource allocation scheme can meet the minimum requirements.

[0159] The non-negative constraint ensures that the amount of resources invested is non-negative.

[0160] therefore The output is the optimal resource allocation vector that minimizes the total cost while simultaneously satisfying the above constraints. ,..., .

[0161] Performance evaluation models ( ), using the analytic hierarchy process (AHP) operator:

[0162] ;

[0163] Operators specific to performance evaluation models;

[0164] : No. The actual value of the l-th indicator under each evaluation dimension (e.g., “Financial Dimension - Profit Margin 15%”);

[0165] : No. The weights of each evaluation dimension (determined based on corporate strategic goals, such as a financial dimension weight of 0.4 and a customer dimension weight of 0.3).

[0166] : No. The weight of the l-th indicator under the dimension (e.g., "profit margin 0.6, cash flow 0.4" under the financial dimension);

[0167] m: Number of evaluation dimensions (default 3~4, such as financial, customer, internal processes, learning and growth);

[0168] Number of metrics for each dimension (default 2-3, to avoid metric redundancy);

[0169] Dimension index, ;

[0170] Indicator Index .

[0171] The design of dedicated operators ensures that the input data can be efficiently processed by subsequent models, while preserving the model's own professional decision-making logic.

[0172] ⑤ Confidence-based correction transmission:

[0173] Since the output of the predecessor model may contain errors, the dynamic weight module also includes a confidence correction mechanism to adjust the input to the model. The input vector is modified as follows, primarily by adjusting the input vector based on the reliability of the predecessor model to prevent low-quality output from affecting subsequent decisions:

[0174] ;

[0175] in,

[0176] : Corrected The input vector;

[0177] Before revision The input vector (i.e., the predecessor model) The output vector , = );

[0178] Hadamard product (element-level multiplication);

[0179] Industry Sensitivity Coefficient: Used to control the strength of the impact of confidence level on correction. Different industries can be set differently. The default value is 0.3. The financial industry has high risk sensitivity and can be set to 0.5 (to amplify the impact of confidence level). The manufacturing industry focuses on stability and can be set to 0.2 (to reduce volatility).

[0180] Precursor Model confidence level ( ), calculated from historical accuracy:

[0181] ;

[0182] For the model In the Confidence error in historical decision-making The number of historical decision samples (at least 50, default 50~100 to ensure statistical validity).

[0183] The confidence adjustment mechanism dynamically adjusts the influence of the predecessor model's output on the successor model based on the reliability of the predecessor model's output: if the predecessor model has a high confidence level, its output has a stronger impact on the successor model, thus more significantly influencing subsequent decisions. (Historical accuracy rate 90%) , The formula was re-optimized as follows:

[0184] ;

[0185] The value 1.016 is automatically truncated to 1.0 (ensuring the value is within the range of 0 to 1); similarly, when the confidence level is low, its influence is weakened to avoid decision-making bias caused by low-quality input.

[0186] (3) Closed-loop control unit:

[0187] The closed-loop control unit is used to monitor the actual achievement of the decision-making objectives, and when the objective deviates... ( When the threshold is set, a policy correction mechanism is triggered. Through the optimization of the full-stack objective function, the global optimization of the vertical transmission chain is achieved, ensuring that the output of each model not only meets its own decision-making needs, but also synergistically improves the overall decision-making quality, avoiding the problem of "local optimum but global suboptimal".

[0188] Among them, target offset The calculation formula is: “ "Assuming a pre-set decision target value (e.g., a market share growth rate of 10%)," "This represents the actual value achieved." User-defined settings are supported; the default value is 15%, and different industries can have different settings (such as the financial industry). =10%, Manufacturing =20%).

[0189] ① Dual threshold triggering logic:

[0190] The closed-loop control unit uses dual threshold triggering logic to detect the target offset. Implement a graded response:

[0191] If the value is less than or equal to the primary threshold, the system will issue a risk warning but will not initiate automatic correction. For example, when... At this time, the system only marks the risk hotspot in yellow on the decision dashboard and displays the deviation details (such as "actual market share growth of 8%, lower than the target of 10%, deviation of 2%), without initiating active correction, leaving it to the user to decide whether to intervene.

[0192] Multiple consecutive decision cycles If the threshold is exceeded, automatic policy correction will be triggered, for example, if the conditions are met for three consecutive decision cycles. When the system automatically triggers an advanced response, it forces the Monte Carlo simulation to generate a set of corrective strategies, which is then displayed at the top of the decision dashboard with a red warning icon, and simultaneously pushed to the decision-maker via email.

[0193] The advantage of the dual threshold design is that it can avoid overreacting to small fluctuations (such as a 1-2% offset caused by short-term market fluctuations) and reduce the waste of system resources; at the same time, it ensures timely intervention for persistent deviations (such as failure to reach the target for three consecutive months) to prevent the accumulation and expansion of errors.

[0194] ② Full-stack objective function:

[0195] Closed-loop optimization achieves coordination among models at each stage of the decision-making chain and improves overall decision quality through systematic optimization of the full-stack objective function. The full-stack objective function is defined as follows:

[0196] ;

[0197] The meanings of the parameters in the formula are as follows:

[0198] For all transmission matrices { Find the minimum value;

[0199] Model hierarchy weights reflect the importance of the model at different stages in the decision-making chain (e.g., strategic level). =0.6, Execution layer =0.3, Evaluation layer =0.1, the strategic layer weight is greater than the execution layer, and greater than the evaluation layer), which satisfies the normalization condition that the sum = 1. The specific value can be adjusted according to the actual decision-making scenario.

[0200] :Model The squared error, This is used to ensure the decision-making accuracy of a single model;

[0201] : Transmission continuity penalty term (default 0.2), used to constrain the difference between adjacent transmission matrices;

[0202] : F-norm (where Frobenius norm) Used for quantizing adjacent conduction matrices and The overall degree of difference (the smaller the value, the smoother the parameter transfer between models and the more coherent the decision logic).

[0203] The total number of models in the model sequence (determined based on decision complexity, default is 3-5);

[0204] First item (accuracy cost): This requires that the output of each model in the sequence be as accurate as possible. Weights This reflects the relative importance of different levels of models in decision-making;

[0205] Second item (smoothing penalty): This project requires that the transmission logic (i.e., the transmission matrix) between adjacent models does not change drastically, thus ensuring the continuity of the decision chain. Penalty coefficient. The importance of smoothness was controlled.

[0206] The objective function aims to achieve accuracy in decision-making at each stage while ensuring the smoothness and consistency of the transmission process. The first part of the objective function ( To ensure that the output error of each model is minimized, the second part ( This ensures smooth changes in the transmission matrix and maintains the consistency of decision-making logic. For example, in a strategic adjustment scenario, if... (Strategy → Resource Transmission) suddenly changed drastically, and If the (resource → performance transmission) is not adjusted synchronously, the objective function value will increase. The system will optimize accordingly. Adjustments should be made accordingly to maintain consistency in the transmission logic.

[0207] ③ Optimize the solution process:

[0208] The optimization of the full-stack objective function uses gradient descent, and the specific process is as follows:

[0209] 1. Initialize the transmission matrix (Initial values ​​fitted based on historical decision-making data, such as those in resource allocation models) (Data obtained from channel investment from 2022 to 2024)

[0210] 2. Calculate the total error (overall objective function value) of the current decision chain: ;

[0211] 3. Solve the objective function pair Partial derivatives: ;

[0212] 4. Update the weights along the negative gradient direction: ( This is the global learning rate, which defaults to 0.005. (This is the updated transmission matrix).

[0213] 5. Repeat steps 2-4 above until... Convergence occurs when the difference between adjacent iterations is less than a second preset threshold, i.e. Stop iteration; in this example, the second preset threshold is set to... Based on the accuracy requirements of numerical calculation.

[0214] This process ensures that the transmission matrix is ​​continuously optimized during iteration, so that the overall decision chain maintains both the accuracy of each link and the logical coherence.

[0215] Optimization results: Tests conducted by a manufacturing company showed that after 50 iterations, the objective function value decreased from 1.2 to 0.38 (a decrease of 68%); the decision error decreased from ±15% to within ±5% (meeting industry accuracy requirements).

[0216] ④ Confidence feedback mechanism:

[0217] The closed-loop control unit also includes a confidence feedback mechanism, which feeds back the evaluation results of the terminal model to the front-end model in the form of confidence scores to correct its decision-making logic. The feedback path of the confidence feedback mechanism includes:

[0218] 1. Calculate the terminal model error (the deviation between the actual and predicted results of the terminal model): (like Performance deviation =0.2);

[0219] 2. Based on end modulus difference Update front-wheel drive model Confidence level: ,in, It is the feedback coefficient (default 0.1, to avoid a sudden drop in confidence). For the model Confidence level, For the updated confidence level, such as ;

[0220] 3. Correct the output of the predecessor model using the updated confidence level:

[0221] ;

[0222] like (1.11 is truncated to 1.0).

[0223] 4. Restart the decision-making chain to achieve dynamic adjustment of decision-making logic. based on Rerun, and pass the output to ,like After adjusting resource allocation, The performance deviation decreased to 0.08.

[0224] For example, if The actual benefits of resource input were found to be 20% lower than predicted. =0.2), then confidence level It decreased from 0.8 to 0.78. The output "market opportunity score" will be weakened (e.g., revised from 0.9 to 0.85), prompting a more conservative allocation of resources and avoiding over-investment.

[0225] ⑤ Strategy Correction Submodule:

[0226] The strategy correction submodule includes a Monte Carlo simulation engine, a contradiction matrix library, and a strategy generator. When the correction process is triggered, it first generates the outcome distribution of 10,000 possible scenarios through Monte Carlo simulation to calculate the probability of target deviation; if the high-risk probability exceeds a preset risk threshold, such as... Then, the principle of matching optimization based on the contradiction matrix library is invoked, as shown in Table 1 below:

[0227] Table 1: Contradiction Matrix

[0228] Decision-making contradictions Applicable innovation principles Insufficient resources - overly ambitious goals Segmentation Principle - Pre-action Principle Over-input - low efficiency Integration principle, value optimization principle, etc.

[0229] The strategy generator outputs specific corrective measures (such as adjusting resource allocation ratios and correcting strategic goal decomposition coefficients).

[0230] The execution flow of the strategy correction submodule is as follows:

[0231] 1. Monte Carlo simulation: based on the parameters of the current decision chain (such as the transmission matrix). Model operators The system generates 10,000 random scenarios to simulate the distribution of target achievement values. In each simulation, the input variables are randomly perturbed (e.g., market growth rate fluctuation ±10%, resource conversion rate fluctuation ±5%), and the corresponding target offset is recorded. .

[0232] 2. Probability distribution generation: Statistical analysis of 10,000 simulations. Values ​​are used to generate a probability distribution histogram and calculate the probability of high-risk events. (That is, the percentage of scenarios with an offset exceeding 20%).

[0233] 3. Decision-making and judgment: If If the value is >0.7 (high risk probability exceeds 70%), then call the contradiction matrix library; otherwise, maintain the current strategy and only record the offset for subsequent analysis.

[0234] 4. Contradiction Matrix Library: The contradiction matrix library stores the mapping relationship between 40 innovation principles and decision-making contradictions (based on TRIZ theory). For example, the contradiction between "insufficient resources and overly ambitious goals" corresponds to the "segmentation principle" and the "preemptive action principle." Based on the cause of the deviation (such as insufficient resource investment leading to delayed market expansion), 2-3 applicable principles are matched to generate specific correction strategies.

[0235] 5. Optimize Strategy Output: The strategy generator transforms innovative principles into executable optimization strategies. Each strategy includes "adjustment parameters," "expected results," and "implementation steps." For example, based on the "segmentation principle," a strategy can be generated to "divide the national market target into three regional sub-targets, prioritizing investment in high-potential regions," which is expected to reduce the overall offset from 25% to less than 10%.

[0236] This invention constructs a vertical model transmission architecture. This architecture, with the core logic of "the output of the preceding model serving as the input of the subsequent model," connects business management models from different stages such as strategic diagnosis, resource allocation, and performance evaluation into an organic whole. It breaks through the limitations of horizontal model isolation and logical fragmentation in traditional decision-making systems. Through this vertical model overlay architecture, it avoids redundant processing of raw data, achieves orderly knowledge transmission between models, and lays the foundation for closed-loop system optimization. By constructing an orderly model transmission chain, it achieves continuous flow of decision-making logic. Furthermore, by introducing dynamic weight transfer and confidence feedback mechanisms, it forms a complete closed-loop decision-making process of "diagnosis-decision-execution-evaluation-correction," significantly improving the consistency, efficiency, and robustness of decisions.

[0237] Example 2: Based on the intelligent decision support system based on the vertical model transmission chain in Example 1 above, this embodiment of the invention provides an intelligent decision support method based on the vertical model transmission chain, referring to... Figure 2 The method includes the following steps:

[0238] S1, Model sequence initialization;

[0239] S11, Load sequence:

[0240] Load the model sequence according to the preset decision logic order. With the goal of "expanding new product sales in East China," the model sequence is as follows: Market opportunity diagnosis → Channel resource allocation → Performance evaluation.

[0241] S12, parameter settings:

[0242] : A strategic diagnostic model used to analyze the macro environment, market opportunities and competitive landscape, with three output dimensions (market size, growth potential, and competitive intensity), weighted at 0.4, 0.3 and 0.3 respectively based on the importance of the FMCG industry dimensions.

[0243] This is a resource allocation model that allocates resources such as funds and human resources based on strategic diagnostic results. It outputs dimension 2 (online and offline input coefficients) and uses linear programming constraints on "online input". 75% of the total budget;

[0244] : Performance evaluation model to monitor the actual effect of resource investment, output dimension 1 (goal achievement rate), evaluation indicators include "sales volume, customer traffic, and repurchase rate" (weights of 0.6, 0.3, and 0.1 respectively).

[0245] S2, Interface compatibility check;

[0246] If there are format conflicts in the interfaces of the various models (e.g.) Output text labels If numerical input is required, then the format conversion tool will be launched to convert it to a standardized numerical vector (e.g., convert "high chance" to 0.8, "medium chance" to 0.5). Specifically,

[0247] Inspection content: The output "high growth" is a text label, while Numeric input required;

[0248] Conflict resolution: Based on two years of growth data in East China, high growth corresponds to a certain growth rate. 15%, which was normalized to convert "high growth" to 0.8;

[0249] Verification result: Output → Format compatible, vector dimension 3→2 (first 2 dimensions compressed by mean).

[0250] S3, Vertical data transmission;

[0251] S31, first-stage conduction ( → ):

[0252] ( Output);

[0253] For example, a 2×3 matrix. The first row corresponds to the weight of the "capital input" dimension, and the second row corresponds to the weight of the "human resource input" dimension.

[0254] This is a bias term, used to compensate for the industry's basic resource needs;

[0255] calculate Input vector: = ;

[0256] Execute dedicated operators (Linear programming), generate output (Capital input coefficient 72%, human resource input coefficient 65%).

[0257] S32, Second-stage conduction ( → ):

[0258] For a 1×2 matrix, example: =[0.7,0.3] (70% weighting for capital investment, 30% weighting for human resource investment);

[0259] =0.06;

[0260] calculate Input vector: ;

[0261] Execute the weighted summation operator → Output (Target achievement rate).

[0262] S4, closed-loop dynamic optimization;

[0263] S41, Offset Calculation:

[0264] Real-time computing Target offset: (Actual = 500 × 0.75 = 3.75 million yuan) (Assuming the target achievement rate is 100%).

[0265] S42, Corrected triggering and execution:

[0266] like (like ; If the error occurs, the reverse correction process will be initiated:

[0267] (1) Monte Carlo simulation: Call the contradiction matrix;

[0268] (2) Matching and segmentation principle: The East China region is divided into "Shanghai, Jiangsu and Zhejiang", with priority given to Shanghai;

[0269] (3) Adjustment The weight of the Shanghai region changed from 0.6 to 0.7.

[0270] (4) Verification results: New , , (A decrease of 7 percentage points);

[0271] (5) Iterative optimization: Repeat S3~S4 until .

[0272] In this embodiment of the invention, the model The output result is obtained through the transmission matrix. After processing, it is passed to the model as the core input. ;Model The output is transmitted through the transmission matrix After processing, the driving model Run; Model The evaluation results are then fed back into the model in the form of confidence feedback. This enables dynamic optimization of the entire decision-making chain. This vertical transmission mechanism ensures that the output of each model can be efficiently utilized by subsequent models, avoiding the problems of "repeated input and logical breakage" in traditional systems, and maintaining consistency in decision logic from the starting point to the end point.

[0273] Example 3: Taking a fast-moving consumer goods company's "new product market expansion decision" as an example, the invention will be further explained. The system operation flow is as follows:

[0274] (1) Model initialization:

[0275] load (Market Opportunity Diagnosis) → (Channel resource allocation) → (Marketing performance evaluation) model sequence;

[0276] Format compatibility handling: The initial output contains text labels "high / medium / low" (e.g., "competition intensity: low"). The system converts the text labels into numerical vectors using a label-value mapping rule (low=0.3, medium=0.6, high=0.9) to ensure consistency with numerical values. It is compatible with numerical input requirements.

[0277] Input "Population density of East China (850 people / km²), per capita disposable income (78,000 yuan / year), and competitor coverage (35%)", and output a standardized numerical vector through multi-dimensional weighted calculation. ,in:

[0278] First dimension: Market size coefficient (0.9, high population density and strong consumption power);

[0279] Second dimension: Growth potential coefficient (0.7, the penetration rate of nut gift boxes still has 15% room for improvement);

[0280] The third dimension: Competition intensity coefficient (0.4, indicating that there are only 3 similar competing products in the region, resulting in low competitive pressure).

[0281] (2) Vertical transmission (core logic verification):

[0282] ① (Market Opportunity Diagnosis) → (Channel resource allocation);

[0283] Transmission matrix Definition: Based on the company's channel spending data for 12 months from 2022 to 2023 (including online sales on Tmall / JD.com and offline sales in supermarkets / community stores), a 2×3 matrix is ​​obtained through linear regression fitting. Output two dimensions: "online / offline investment coefficient". Output 3 dimensions):

[0284] → : ;

[0285] The meaning of the matrix elements is as follows:

[0286] First row (online channel weight): Market size has the greatest impact on online investment (0.6), while competition intensity has a negative impact (-0.2, the stronger the competition, the higher the online traffic cost).

[0287] The second row (offline channel weight): growth potential has the greatest impact on offline investment (0.5, and offline stores in regions with high growth potential have better reach).

[0288] Bias term Settings: Based on the characteristics of channel operating costs in the FMCG industry, set up... ,in,

[0289] Online bias 0.08 (covering platform promotion fees and logistics and warehousing costs);

[0290] Offline bias 0.05 (covering basic costs such as store rent and staff commission).

[0291] calculate ;

[0292] Output generate: Using linear programming operators (Objective: Maximize channel ROI; Constraints: Online investment ≤ 75% of total budget, Offline investment ≥ 25% of total budget), the final output is... This indicates that online channel investment accounts for 72% of the total budget (with a focus on the Tmall flagship store), and offline channel investment accounts for 65% (with a focus on covering community stores in East China) (Note: The investment coefficient is slightly greater than 1 due to the existence of channel synergy investment, such as joint promotion expenses).

[0293] ② (Channel resource allocation) → (Marketing performance evaluation);

[0294] Transmission matrix Definition: Based on the company's performance data from January to June 2023 (online conversion rate 12%, offline conversion rate 18%), a 1×2 matrix is ​​fitted. Output one dimension: "goal achievement rate" Meaning: Online sales account for a higher proportion (70%), contributing more to the achievement of the target.

[0295] Bias term Setting: Referencing the historical execution performance of the East China regional marketing team (105% target achievement rate in the first half of 2023), a positive bias is set. This demonstrates how team capabilities contribute to performance gains.

[0296] Dedicated Operator: A weighted summation operator is used, with a value of 1. This indicates a target achievement rate of 76%, meaning actual sales of 3.8 million yuan (500 × 0.76). The target was not met, and the target deviation was [missing information]. .

[0297] This invention addresses the pain points of traditional decision-making systems—"isolated models, redundant data, and uncontrolled errors"—through a vertical model transmission architecture, dynamic transmission mechanism, and closed-loop optimization system. It achieves continuous and dynamic optimization of decision-making logic, providing efficient and accurate intelligent support for complex decision-making in various organizations. Its technical solution is at the forefront in terms of mathematical rigor, universality of implementation, and unique innovation, possessing broad industrial application value. It is suitable for intelligent decision-making across the entire domain of enterprise management. In the context of multi-model vertical linkage (based on goal management), it can optimize organizational strategy, conduct organizational design based on quantitative goals, perform human resource planning, operational management planning, provide financial management advice, and design marketing strategies.

[0298] Specific examples in this embodiment can be found in the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0299] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0300] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0301] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The system embodiments described above are merely illustrative; for example, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection of units or modules may be electrical or other forms.

[0302] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0303] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0304] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0305] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. An intelligent decision support system based on a vertical model transmission chain, characterized in that, It includes a model propagation engine, a dynamic weight module, and a closed-loop control unit, among which, The model transmission engine is used to construct an ordered sequence of models, where the output vector of each model in the sequence serves as the input of the subsequent model, thus forming a continuous decision transmission chain. The model transmission engine includes two sub-modules: a model sequence configuration module and a model interface adaptation unit. The model sequence configuration module supports the ordered combination of multiple models and automatically generates the model execution order based on user settings or a historical decision case library. The model interface adaptation unit standardizes and converts the data formats output by different models to ensure data compatibility between models. The dynamic weighting module is used to transfer the weighted values ​​from the output of the predecessor model to the input of the successor model through a transfer matrix. The transfer matrix is ​​dynamically adjusted based on the model confidence error. The confidence error is the deviation between the model's predicted output and the actual result, expressed in 2-norm form. When the confidence error exceeds a first preset threshold, the transfer matrix adjustment process is initiated, and the element values ​​of the transfer matrix are optimized through a gradient descent algorithm until the confidence error falls back to a safe range. The closed-loop control unit is used to monitor the actual achievement of the decision-making objective and calculate the objective offset. When the objective offset exceeds the set threshold, the strategy correction mechanism is triggered. The optimization instruction is generated by optimizing the full-stack objective function and fed back to the model transmission engine to achieve global optimization of the transmission chain. The full-stack objective function of the closed-loop control unit is defined as follows: ； in, For all transmission matrices Find the minimum value; Model hierarchy weights reflect the importance of the model in the decision chain; :Model The squared error, , For the model The predicted output, For the model The actual output; : Transmission continuity penalty term, used to constrain the difference between adjacent transmission matrices; Frobenius norm, used to measure the performance of adjacent transmission matrices. and The differences between them should be considered to avoid abrupt changes in the transmission logic that could lead to a break in the decision chain. The total number of models in the model sequence.

2. The intelligent decision support system according to claim 1, characterized in that, The dynamic weighting module uses a transmission matrix to achieve weighted transfer from the output of the predecessor model to the input of the successor model, specifically: The output vector of the predecessor model is weighted by the transmission matrix, then a bias term is added, and finally the output vector is generated by the model-specific operator. This output vector is then used as the input to the subsequent model. The formula is as follows: ； in, :Model The output vector; : Dynamic transmission matrix, elements in the matrix express The Dimension to model The The influence weight of dimensional inputs; :Model The bias term is used to correct for systematic errors; :Model The dedicated operators are customized according to the model's functions; :Model The output vector of is used as the successor model. The core input.

3. The intelligent decision support system according to claim 2, characterized in that, The transmission matrix Based on dynamic adjustment of model confidence error, the model's confidence error Defined as the L2 distance between the model's predicted output and the target output, i.e. ,in, For the model The predicted output, For the model The adjustment rule for the transmission matrix is ​​as follows: ； In the formula, : Transmission matrix at time step; : Transmission matrix at time step; Learning rate controls the step size for weight updates; :Model The predicted output; :Model The target output; The dynamic weighting module also adjusts the transmission matrix. The values ​​of each element in Apply boundary constraints. When an element value exceeds the set boundary during the optimization process, it is automatically truncated to the corresponding boundary value, and the values ​​of other elements in the transmission matrix are adjusted proportionally.

4. The intelligent decision support system according to claim 2, characterized in that, The dedicated operator Based on model functionality customization, different dedicated operators are used for different types of models, specifically including: Strategic diagnostic models employ a multi-dimensional weighted summation operator: ； in, This is an operator specifically for strategic diagnostic models. For the first The weights of each diagnostic dimension, For the first Input values ​​for each diagnostic dimension This represents the total number of diagnostic dimensions. For dimensional indexing, ; Resource allocation models employ linear programming operators: ； in, This is an operator specific to resource allocation models, with the following constraints: , For the first Cost coefficient of similar resources For the first The amount of resources invested in this category It is the first The constraint on the first Resource consumption coefficient, For the first The minimum requirement value for each constraint, For the number of resource types, The number of constraints; Performance evaluation models employ linear programming operators: ； in, This is an operator specifically for performance evaluation models. For the first The weights of each evaluation dimension, For the first Dimension 1 The weight of each indicator, For the first The first dimension The actual value of each indicator To evaluate the dimensional coefficients, The number of metrics for each dimension, To evaluate the dimensional index, For indexing indicators.

5. The intelligent decision support system according to claim 1, characterized in that, The dynamic weight module also integrates a confidence correction mechanism, which adaptively corrects the vector input to the current model based on the reliability of the predecessor model's output. The correction formula is expressed as: ; in, Corrected model The input vector; Model before correction The input vector, , For the model The output vector; Hadamard product, element-wise multiplication; : Industry sensitivity coefficient, controlling the strength of the impact of confidence level on correction; :Model The confidence level is calculated from the historical accuracy: ； in, It is a model In the Confidence error in historical decision-making This represents the total number of historical decision-making samples.

6. The intelligent decision support system according to claim 1, characterized in that, The formula for calculating the target offset in the closed-loop control unit is: ； In the formula, The target offset. To preset the target value for decision-making, This is the actual value achieved. The closed-loop control unit is based on the target offset. A tiered response is achieved using dual-threshold triggering logic: when When the value is less than or equal to the primary threshold, a risk warning will be issued and offset details will be displayed, but automatic correction will not be initiated, allowing users to make their own judgment and intervention. When multiple consecutive decision cycles When the value exceeds the advanced threshold, an advanced response is triggered, a correction strategy is initiated, and an early warning is issued. At the same time, the warning information is pushed to the decision-maker.

7. The intelligent decision support system according to claim 1, characterized in that, The full-stack objective function is optimized using gradient descent, and the process is as follows: Step 1: Initialize the transmission matrix based on historical data ; Step 2, calculate the objective function value : ; Step 3: Solve for the objective function with respect to the transmission matrix. Partial derivatives: ; Step 4, update the transmission matrix along the negative gradient direction: , The global learning rate; for Transmission matrix at time step; This is the updated transmission matrix; Repeat steps 2-4 until... The difference between adjacent iterations is less than the second preset threshold.

8. The intelligent decision support system according to claim 7, characterized in that, The closed-loop control unit also includes a confidence feedback mechanism, used to feed back the evaluation results of the terminal model to the front-end model in the form of confidence scores, thereby correcting its decision-making logic. The feedback path of the confidence feedback mechanism includes: Computational end model Actual results Compared with the prediction results Deviation between : ; Based on deviation Update the predecessor model according to predefined rules. Confidence level: ,in, For feedback coefficients, For the model confidence level The updated confidence level; Correct using the updated confidence level Output: , This is the industry sensitivity coefficient, used to control the strength of the impact of confidence level on the correction. based on Restart the process, initiate a new round of decision-making, and achieve dynamic correction.

9. The intelligent decision support system according to claim 8, characterized in that, The closed-loop control unit further includes a strategy correction submodule, the execution flow of which is as follows: Monte Carlo simulation: Based on the current decision chain parameters, the Monte Carlo method is used to generate a preset number of random scenarios, with random variables based on historical data. The interval, the decision chain parameters include the transmission matrix. Model operators ; Probability distribution generation: Statistical simulation results are used to calculate the target offset. The probability distribution is calculated, a probability distribution histogram is generated, and the high-risk probability is calculated. The high-risk probability refers to the probability that the target offset exceeds a third preset threshold. Contradiction matrix call judgment: When the high risk probability exceeds the preset risk threshold, the contradiction matrix library is called to match the corresponding correction strategy; otherwise, the current strategy is maintained. The contradiction matrix library stores several mapping relationships between innovation principles and decision contradictions. Optimize strategy output: Match multiple applicable innovation principles to generate an executable strategy that includes parameter adjustments, expected results, and implementation steps.

10. An intelligent decision support method based on a vertical model transmission chain, based on the intelligent decision support system according to any one of claims 1 to 9, characterized in that, include: Model sequence initialization: Load the model sequence according to the preset decision logic, check the compatibility of each model interface, and if a format conflict is detected, start the format conversion tool to convert the format and unify the output into a compatible data format; Vertical data transmission: The output of the predecessor model is weighted by the transmission matrix, and after the bias term is added, the output vector is generated by calling the special operator and input into the successor model; Closed-loop dynamic optimization: Calculate the target offset. If the target offset exceeds the set threshold, start full-stack objective function optimization, adjust the transmission matrix, and use the updated transmission matrix in the next round of transmission to reduce the target offset. The full-stack objective function is defined as follows: ； in, For all transmission matrices Find the minimum value; Model hierarchy weights reflect the importance of the model in the decision chain; :Model The squared error, , For the model The predicted output, For the model The actual output; : Transmission continuity penalty term, used to constrain the difference between adjacent transmission matrices; Frobenius norm, used to measure the performance of adjacent transmission matrices. and The differences between them should be considered to avoid abrupt changes in the transmission logic that could lead to a break in the decision chain. The total number of models in the model sequence.

Images