A Store Operation Decision-Making Method and Device Based on Multi-Agent Collaboration

Abstract

This invention relates to the interdisciplinary field of artificial intelligence and enterprise management, and discloses a method and apparatus for store operation decision-making based on multi-agent collaboration. The method includes: constructing a digital twin environment through data perception and environment construction agents; identifying operational problems through diagnostic analysis agents using business rules and anomaly detection models; generating multiple strategy proposals corresponding to the operational problems through a strategy generation agent based on a pre-built reinforcement learning strategy library and business knowledge graph; injecting the multiple strategy proposals into the digital twin environment through a strategy simulation agent, and performing simulation deduction based on an environmental dynamics model and a customer response model to generate a simulation evaluation report; and generating a final decision recommendation based on the simulation evaluation report through a decision arbitration agent using a multi-objective optimization algorithm. This invention solves the problems of existing store operation decision-making processes that rely on personal experience, have delayed responses, and are subject to uncontrollable risks.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and enterprise management, specifically to a method and apparatus for store operation decision-making based on multi-agent collaboration. Background Technology

[0002] In the retail industry, the quality of a store manager's daily operational decisions directly impacts a store's performance and survival. However, the current store operational decision-making process faces numerous challenges: 1. Decision-making relies on personal experience: Store managers often make decisions based on personal experience and intuition, such as ordering, scheduling, and promotions. This approach lacks data support, is difficult to standardize and replicate, and is prone to performance fluctuations due to personnel changes.

[0003] 2. Information overload and difficulty in gaining insights: Stores generate massive amounts of data (sales, customer traffic, inventory, membership, etc.), but this data is scattered across different systems and lacks effective integration. Store managers struggle to quickly pinpoint core issues and understand their root causes from the data.

[0004] 3. Uncontrollable decision-making risks: Any change in business strategy (such as significant price reductions and promotions) may lead to unpredictable results (such as declining profits and inventory buildup). Due to the lack of pre-emptive simulation and evaluation methods, decision-making is like "flying blindly," with high trial-and-error costs.

[0005] 4. Slow response speed: The market environment and customer needs are constantly changing, but the cycle from problem discovery to analysis and decision-making to execution is long, often missing the best opportunity to respond.

[0006] Existing store management tools (such as BI (Business Intelligence) reporting systems and ERP (Enterprise Resource Planning) modules) have failed to effectively solve the above problems, for example: 1. Passive analysis: Existing tools mainly provide historical data queries and report displays, which is a "hindsight" approach and cannot provide proactive decision-making suggestions and early warnings.

[0007] 2. Functional silos: Modules such as inventory management, personnel management, and marketing management are independent of each other, data is not shared, and it is difficult to weigh multiple factors when making decisions, making it impossible to achieve global optimization.

[0008] 3. Lack of cognitive and reasoning abilities: Traditional software cannot perform causal analysis, strategy deduction, and simulation evaluation like human experts. Its function is limited to information presentation rather than decision support. Summary of the Invention

[0009] This invention provides a store operation decision-making method and apparatus based on multi-agent collaboration to solve the problems of existing store operation decision-making processes that rely on personal experience, have slow response times, and are subject to uncontrollable risks.

[0010] In a first aspect, the present invention provides a store operation decision-making method based on multi-agent collaboration, the method comprising: By using data perception and environmental construction to build intelligent agents, multi-source heterogeneous data of stores can be obtained, and a digital twin environment can be built based on the multi-source heterogeneous data; By analyzing the data of the digital twin environment through the diagnostic agent, the system uses preset business rules and preset anomaly detection models to diagnose the data of the digital twin environment and generate a diagnostic report containing operational issues. The strategy-generating agent generates multiple strategy proposals corresponding to the operational issues in the diagnostic report based on a pre-built reinforcement learning strategy library and business knowledge graph. By injecting various strategy proposals into the digital twin environment through a strategy simulation intelligent agent, and performing simulations based on pre-built environmental dynamics models and customer response models, the execution effects of each strategy proposal are predicted and a simulation evaluation report is generated. Based on the simulation evaluation report, the decision-arbitration agent uses a multi-objective optimization algorithm to rank various strategy proposals and generate a final decision recommendation.

[0011] The above technical solution addresses several key aspects. First, by integrating multi-source heterogeneous data and constructing a dynamic digital twin environment, it solves the problems of information silos and delayed state perception. Second, by utilizing rules and models for automatic diagnosis, it changes the subjective decision-making model that relies on personal experience. Third, by employing reinforcement learning and knowledge graph generation strategies, it ensures the diversity and professionalism of the strategies. Crucially, by using simulation to predict the effectiveness of the strategies, it achieves pre-emptive quantitative assessment and controllability of decision-making risks. Finally, by generating final recommendations through multi-objective optimization arbitration, it enables decisions to systematically weigh multiple business objectives, thereby improving the timeliness, scientific rigor, and reliability of decision-making overall. By leveraging the collaborative work of multiple agents, it solves the problems of existing technologies that rely on personal experience, have delayed responses, and suffer from uncontrollable risks in store operation decision-making processes.

[0012] In one alternative implementation, the digital twin environment is dynamically updated; the multi-source heterogeneous data includes time-series sales data, product data, and external environment data. Building a digital twin environment based on multi-source heterogeneous data includes: The time-series sales data is encoded using a time-series Transformer model; Based on product data, a graph neural network is used to model the relationships between products; External environmental data is used as environmental variables, and the encoded time-series sales data and the relationships between products are mapped to the dynamic state vector of the store to form a digital twin environment for strategy simulation.

[0013] In the above technical solution, the sales trend and cyclical pattern are accurately encoded by the time-series Transformer model, the relationship between products is deeply mined by the graph neural network, and the external environment data is dynamically integrated into environmental variables. Finally, a high-fidelity, dynamically updatable store state vector is generated, which realizes the efficient fusion and semantic representation of multi-source heterogeneous data, and significantly improves the accuracy and completeness of store state description.

[0014] In one optional implementation, a diagnostic analysis agent, based on the digital twin environment, uses preset business rules and a preset anomaly detection model to diagnose the data in the digital twin environment, generating a diagnostic report containing operational issues, including: The system uses diagnostic analytics to obtain operational status data for stores from the digital twin environment. Based on the status data, a preset business rule library is called to make logical judgments, and a preset anomaly detection model is run to perform pattern recognition, so as to obtain the logical judgment results and the pattern recognition results respectively. By fusing and cross-validating the results of logical judgments and pattern recognition, operational problems are identified, and these problems are correlated with potential business causes to generate a diagnostic report containing the operational problems.

[0015] In the aforementioned technical solution, a diagnostic analysis agent utilizes a pre-set business rule base and anomaly detection model in parallel to perform dual analysis and cross-validation of state data in the digital twin environment, achieving proactive, accurate, and interpretable identification of store operation problems. This method deeply integrates rule-based judgment based on expert experience with data-driven anomaly detection, not only improving the timeliness and coverage of problem discovery but also effectively revealing the underlying business roots of problems through result fusion and causal correlation. This transforms the traditional passive and fragmented diagnosis relying on personal experience into a systematic, standardized, and insightful automated diagnostic process, significantly improving the responsiveness and depth of understanding in business decision-making.

[0016] In one alternative implementation, the reinforcement learning policy library is constructed in the following manner: Determine the state space, action space, and reward function for reinforcement learning, and obtain historical operating data of the stores; Based on the state space, action space, and reward function, reinforcement learning algorithms are used to perform offline reinforcement learning on historical business data, or online training is conducted through a simulated environment to obtain the optimal learning strategy. The optimal learning strategy is stored in a preset data format to generate a reinforcement learning strategy library.

[0017] The above technical solution abstracts store operation decisions into a reinforcement learning problem, clearly defines the state, action and reward function, and uses historical data to train the optimal strategy through reinforcement learning algorithm, and finally builds a reusable strategy library, realizing the transformation of decision-making paradigm from experience-driven to data-driven, and improving the objectivity and scientific nature of the strategy.

[0018] In one alternative implementation, a policy-generating agent generates multiple policy proposals corresponding to the operational issues in the diagnostic report, based on a pre-built reinforcement learning policy library and a business knowledge graph. These proposals include: The strategy generates an intelligent agent that analyzes the type, severity, and context of operational issues in the diagnostic report. Based on the type, severity, and context of the business problem, retrieve relevant candidate strategies from the business knowledge graph; Input the current store status data into the reinforcement learning strategy library and calculate the expected value of each candidate strategy; Based on expected value and actual business constraints, various strategy proposals are generated to address the operational issues identified in the diagnostic report.

[0019] In the above technical solution, firstly, knowledge graph-based retrieval ensures that the generated strategies are highly relevant to the problem and conform to business logic, inheriting expert experience and historical data; secondly, reinforcement learning is used to quantify the expected value of candidate strategies based on the current state, introducing data-driven effect prediction and making strategy ranking more objective; finally, combining the advantages of both and generating diverse strategy proposals under business constraints ensures both the feasibility and professionalism of the strategies, while providing ample space for comparison and selection.

[0020] In one optional implementation, multiple policy proposals are injected into the digital twin environment through a policy simulation agent, and simulations are performed based on a pre-built environmental dynamics model and a customer response model to predict the execution effects of each policy proposal and generate a simulation evaluation report, including: The executable parameters of each policy proposal are extracted by a policy simulation agent, and the executable parameters of each policy proposal are injected into the digital twin environment. Based on pre-built environmental dynamics models and customer response models, multiple rounds and time steps of simulation are performed in a digital twin environment after injecting executable parameters. This simulates the dynamic interaction and evolution of the market, customer and store status after the implementation of each strategy proposal, and obtains simulation results. From the simulation results, extract the sequence of key business indicators, and based on the sequence of key business indicators, predict the final effect of the strategy proposal and assess the uncertainty. Based on the final results and uncertainties, a structured simulation evaluation report is generated for each strategy proposal.

[0021] In the above technical solution, by injecting strategy parameters into the dynamic environment and performing multi-step deduction based on environmental dynamics and customer response models, a high degree of fidelity simulation of the complex interaction process after strategy execution is achieved, significantly improving the realism of the prediction. Furthermore, by extracting key indicator sequences and using ensemble methods for effect prediction and uncertainty assessment, not only is a quantitative effect expectation provided, but also a clear risk insight for decision-makers through confidence intervals and risk levels. That is, by using a strategy simulation agent to perform high-fidelity, multi-dimensional pre-simulation of strategy proposals in a digital twin environment, a quantitative prediction of decision risk and a scientific estimate of the effect are achieved.

[0022] In one alternative implementation, the method further includes: Once the policy simulation agent has completed the evaluation of all policy proposals, it will publish all simulation evaluation reports to the communication management center.

[0023] In the above technical solution, the centralized release and asynchronous transmission of simulation evaluation results are realized through the communication management center, which effectively decouples the direct dependencies between various intelligent agents in the system and improves the collaborative efficiency of the decision-making process and the scalability of the system architecture.

[0024] In one alternative implementation, a decision-arbitration agent, based on a simulation evaluation report, uses a multi-objective optimization algorithm to rank various strategy proposals and generate a final decision recommendation, including: The decision-arbitration intelligent agent extracts the sequence of key business indicators from the simulation evaluation report and assigns corresponding weights to each business indicator based on the preset business objective priority. A multi-objective optimization algorithm is used to calculate the comprehensive score of each strategy proposal based on each business indicator and its corresponding weight. The strategies are ranked based on their overall scores, and the preferred strategy proposal and alternative strategy proposals are determined based on the ranking results. Generate an executable list of instructions for the preferred strategy proposal, add conditional descriptions for alternative strategy proposals, and generate a final decision recommendation.

[0025] In the above technical solution, by pre-setting target priorities and weight allocation, complex business objectives are transformed into computable optimization problems, ensuring the consistency between decision-making and overall strategy; by using multi-objective optimization algorithms to comprehensively score and rank strategies, rational trade-offs and choices under multiple constraints are realized; finally, decision suggestions containing executable instructions and alternative explanations are generated, that is, by using a decision arbitration agent based on multi-objective optimization algorithms to make objective and quantitative comprehensive judgments on simulation evaluation results, efficient transformation from multi-strategy simulation to executable decisions is achieved.

[0026] In one alternative implementation, the method further includes: The final decision recommendations are sent to the execution terminal, and the actual effect data after execution is collected; Based on actual performance data, optimize the pre-built environmental dynamics model and reinforcement learning strategy library.

[0027] In the above technical solution, by feeding back the execution results of the decision suggestions to the model and strategy library for continuous optimization, the system is upgraded from static assistance to dynamic self-learning, forming a complete intelligent closed loop of decision-making-execution-feedback-optimization. Secondly, this invention provides a store operation decision-making device based on multi-agent collaboration, the device comprising: The digital twin environment construction module is used to acquire multi-source heterogeneous data of stores through data perception and environment construction of intelligent agents, and to build a digital twin environment based on the multi-source heterogeneous data; The diagnostic report generation module is used to diagnose the data in the digital twin environment by using a diagnostic analysis agent based on the digital twin environment, and adopting preset business rules and preset anomaly detection models to generate a diagnostic report containing operational issues. The strategy proposal generation module is used to generate various strategy proposals corresponding to the operational issues in the diagnostic report by a strategy generation agent based on a pre-built reinforcement learning strategy library and a pre-built business knowledge graph. The simulation and deduction module is used to inject multiple policy proposals into the digital twin environment through a policy simulation agent, and perform simulation and deduction based on a pre-built environmental dynamics model to predict the execution effect of each policy proposal and generate a simulation evaluation report. The decision recommendation generation module is used by a decision arbitration agent to rank various strategy proposals based on simulation evaluation reports and employ a multi-objective optimization algorithm to generate final decision recommendations.

[0028] Thirdly, the present invention provides an electronic device, comprising: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the store operation decision-making method based on multi-agent collaboration described in the first aspect or any corresponding embodiment.

[0029] Fourthly, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to execute the multi-agent collaborative store operation decision-making method of the first aspect or any corresponding embodiment described above.

[0030] Fifthly, the present invention provides a computer program product, including computer instructions, which are used to cause a computer to execute the multi-agent collaborative store operation decision-making method described in the first aspect or any corresponding embodiment above. Attached Figure Description

[0031] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0032] Figure 1 This is a schematic diagram of an application scenario according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the first type of store operation decision-making method based on multi-agent collaboration according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the second process of a store operation decision-making method based on multi-agent collaboration according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the third process of the store operation decision-making method based on multi-agent collaboration according to an embodiment of the present invention; Figure 5 This is an interactive diagram illustrating the multi-agent collaboration in a store operation decision-making method based on multi-agent collaboration according to an embodiment of the present invention. Figure 6 This is a schematic diagram of the internal working principle of the strategy simulation agent in the store operation decision-making method based on multi-agent collaboration according to an embodiment of the present invention. Figure 7 This is a schematic diagram of the fourth process of the store operation decision-making method based on multi-agent collaboration according to an embodiment of the present invention; Figure 8 This is a structural block diagram of a store operation decision-making device based on multi-agent collaboration according to an embodiment of the present invention; Figure 9 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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 embodiments of the present invention, not all embodiments. 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.

[0034] It is understood that before using the technical solutions disclosed in the various embodiments of the present invention, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in the present invention and their authorization should be obtained in accordance with relevant laws and regulations through appropriate means.

[0035] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0036] As an optional application scenario of this invention, such as Figure 1 The diagram shown illustrates the architecture of a multi-agent system provided by this invention. The multi-agent system includes: The data perception and environment construction intelligent agent 101 is used to collect multi-source heterogeneous data from stores in real time and to build and maintain a dynamically updated digital twin environment.

[0037] The digital twin environment 102 serves as a virtual mapping of the real store, providing the system with a consistent and emulated view of data and a representation of state.

[0038] The diagnostic analysis agent 103 is used to identify operational problems and generate diagnostic reports based on business rules and anomaly detection models in a digital twin environment.

[0039] The strategy generation agent 104 receives diagnostic reports and combines them with a reinforcement learning policy library and a business knowledge graph to generate various response strategy proposals.

[0040] The strategy simulation agent 105 is used to inject strategy proposals into the digital twin environment, perform simulations based on the environmental dynamics model and the customer response model, evaluate the effectiveness of the strategy, and generate simulation reports.

[0041] The decision arbitration agent 106 is used to comprehensively evaluate and rank various strategies based on the simulation evaluation report using a multi-objective optimization algorithm, and generate the final decision recommendation.

[0042] The communication management hub 107 serves as the system's shared information bus, supporting asynchronous and decoupled data exchange and collaborative work among various intelligent agents.

[0043] The decision suggestion output module 108 pushes the final decision suggestion to the execution terminal or relevant business system, completing the closed loop from analysis to execution.

[0044] A schematic diagram of multi-agent collaborative interaction is shown below. Figure 5As shown, the overall process is as follows: The system constructs a digital twin environment 102 through data perception and environment construction intelligent agent 101, performs problem diagnosis by diagnostic analysis intelligent agent 103, generates strategies by strategy generation intelligent agent 104, performs simulation evaluation by strategy simulation intelligent agent 105, and completes decision arbitration by decision arbitration intelligent agent 106. The entire process is achieved through communication management center 107 to realize collaborative communication between intelligent agents, and finally outputs executable decision suggestions through decision suggestion output module 108, forming a complete intelligent decision closed loop from perception, analysis, generation, simulation to decision.

[0045] According to an embodiment of the present invention, a method for store operation decision-making based on multi-agent collaboration is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0046] This embodiment provides a store operation decision-making method based on multi-agent collaboration, which can be used in the aforementioned electronic devices or terminal devices. The electronic devices or terminal devices are equipped with a multi-agent system. Figure 2 This is a flowchart of a store operation decision-making method based on multi-agent collaboration according to an embodiment of the present invention, such as... Figure 2 As shown, the process includes the following steps: Step S201: Obtain multi-source heterogeneous data of the store through data perception and environmental construction of intelligent agents, and construct a digital twin environment based on the multi-source heterogeneous data.

[0047] Specifically, in this embodiment, the store is an offline physical store.

[0048] By using data perception and environmental construction to create intelligent agents, we can acquire multi-source heterogeneous data in real time, including sales, inventory, customer flow and external environment of stores, and build and continuously update a digital twin environment that reflects the real status of stores based on this data.

[0049] Step S202: The diagnostic analysis agent diagnoses the data in the digital twin environment based on the digital twin environment using preset business rules and preset anomaly detection models, and generates a diagnostic report containing operational issues.

[0050] Specifically, the diagnostic analysis agent analyzes the current environmental data based on the digital twin environment, calls the preset business rule base and anomaly detection model, identifies existing operational problems, and generates a structured diagnostic report accordingly.

[0051] Step S203: Based on a pre-built reinforcement learning policy library and business knowledge graph, the policy-generating agent generates multiple policy proposals corresponding to the operational issues in the diagnostic report.

[0052] Specifically, a strategy-generated intelligent agent queries a pre-built business knowledge graph based on the diagnostic report to obtain strategy clues, and evaluates candidate strategies in conjunction with a reinforcement learning strategy library, ultimately generating multiple strategy proposals corresponding to the operational issues in the diagnostic report.

[0053] Step S204: Inject multiple strategy proposals into the digital twin environment through a strategy simulation agent, and perform simulation based on a pre-built environmental dynamics model and customer response model to predict the execution effect of each strategy proposal and generate a simulation evaluation report.

[0054] Specifically, the environmental dynamics model is a model that simulates how the store's own state and the external market environment evolve over time. It defines how various endogenous variables (such as natural decline in customer traffic and natural inventory loss) and external factors (such as weather changes, holiday effects, and competitor activities) dynamically affect the overall state of the store without direct intervention from strategies.

[0055] A customer response model is a model that simulates how different individual customers or groups react to store strategies (such as promotions, price adjustments, and display changes). Based on customer attributes (such as membership tags and historical behavior) and the current context, it predicts changes in their purchasing decisions, spending amounts, and other behavioral changes.

[0056] By using a strategy simulation agent, the generated strategy proposals and their parameters are injected into the digital twin environment. Multiple rounds of simulation are conducted using a pre-built environmental dynamics model and a customer response model to simulate the dynamic process after the strategy is executed, and a simulation evaluation report containing predicted effects and risk assessments is output.

[0057] Step S205: Based on the simulation evaluation report, the decision arbitration agent sorts out various strategy proposals using a multi-objective optimization algorithm to generate the final decision recommendation.

[0058] Specifically, the decision-arbitration agent comprehensively analyzes the simulation evaluation reports of all strategy proposals, uses a multi-objective optimization algorithm to comprehensively evaluate and rank each proposal, and generates a final decision recommendation that includes the preferred solution and execution instructions.

[0059] The store operation decision-making method based on multi-agent collaboration provided in this embodiment innovatively combines temporal modeling, graph neural networks and multi-agent simulation to achieve high-fidelity simulation of complex retail environments, forming a four-stage decision-making process of diagnosis-generation-simulation-arbitration, and has a complete cognitive decision-making chain.

[0060] This embodiment provides a store operation decision-making method based on multi-agent collaboration, which can be used in the aforementioned electronic devices or terminal devices. The electronic devices or terminal devices are equipped with a multi-agent system. Figure 3 This is a flowchart of a store operation decision-making method based on multi-agent collaboration according to an embodiment of the present invention, such as... Figure 3 As shown, the process includes the following steps: Step S301: Obtain multi-source heterogeneous data of the store through data perception and environmental construction of intelligent agents, and construct a digital twin environment based on the multi-source heterogeneous data.

[0061] Specifically, the digital twin environment is dynamically updated; the multi-source heterogeneous data includes time-series sales data, product data, and external environment data; among which: Time-series sales data includes: Sales data: sales revenue, sales volume, transaction time, product category, etc. Customer flow data: customer traffic, time period distribution, heat map, dwell time, etc. Order data: order status, delivery information, return records, etc. Membership data: member attributes (age, gender), purchase history, preference tags, etc.

[0062] Product data includes: Inventory data: inventory level, inbound / outbound records, turnover rate, etc.

[0063] External environmental data: weather data, holiday information, competitor store activities, etc.

[0064] Step S301 above includes: Step S3011: Encode the time-series sales data using the time-series Transformer model.

[0065] Specifically, the temporal Transformer model is used to encode temporal sales data (such as sales data, customer flow data, membership data, etc.) to capture long-term dependencies and periodic patterns.

[0066] Step S3012: Based on the product data, a graph neural network is used to model the relationships between products.

[0067] Specifically, based on product data, graph neural networks are used to model the relationships between products (such as complementary products and substitutes) and construct a category association graph.

[0068] Step S3013: External environment data is used as environmental variables, and the encoded time-series sales data and the relationships between products are mapped into the dynamic state vector of the store to form a digital twin environment for strategy simulation.

[0069] Specifically, external environmental data (such as weather and holidays) are integrated as environmental variables, and the weights of these factors are dynamically adjusted through an attention mechanism to reflect the impact of different time periods or seasons.

[0070] The data processed in steps S3011 and S3012 is mapped to a virtual store model. This model includes: a dynamic state representation: real-time updates of the store's state vector, such as inventory levels, customer flow density, and sales trends; and a simulation engine foundation: providing an interactive environment interface for the strategy simulation agent, supporting the injection of strategies and simulating their execution effects.

[0071] Continuous environment updates: Real-time reception of state vectors via streaming technologies (such as Apache Kafka or Flink) and triggering model recalculation ensure that the digital twin environment is synchronized with the real store status.

[0072] It should be noted that the generated digital twin environment provides data services to the outside world through a communication management hub (blackboard system). Other intelligent agents (such as diagnostic analysis agents) can subscribe to and access this environment to obtain a consistent data view. The data in the digital twin environment is referred to as multi-dimensional state data.

[0073] This embodiment innovatively combines temporal modeling, graph neural networks, and multi-agent simulation to achieve high-fidelity simulation of complex retail environments.

[0074] Step S302: The diagnostic analysis agent diagnoses the data in the digital twin environment based on the digital twin environment using preset business rules and preset anomaly detection models, and generates a diagnostic report containing operational issues.

[0075] Specifically, based on a pre-built business rule base and machine learning anomaly detection model, data in the digital twin environment is analyzed to proactively identify operational problems (such as: low sales rate of a certain product category, insufficient manpower during peak lunch hours, abnormal inventory turnover of a certain product). The above step S302 includes: Step S3021: Obtain the status data of store operation from the digital twin environment through diagnostic analysis of the intelligent agent.

[0076] The diagnostic analysis agent extracts multi-dimensional status data reflecting the store's operating status in real time from the dynamically updated digital twin environment, including key indicators such as sales, inventory, customer traffic, and manpower, providing unified and accurate data input for subsequent analysis.

[0077] Multidimensional state data specifically includes: 1. Sales data: Sales revenue, sales volume, number of transactions. Sales turnover rate (e.g., category sales turnover rate, product sales turnover rate). Sales trends (e.g., month-on-month and year-on-year changes). Promotional effectiveness data (e.g., sales increase rate during the promotion period).

[0078] 2. Inventory Data: Inventory Levels (Current Inventory Level, Safety Stock Threshold). Inventory Turnover Rate, Inventory Age Structure. Stockout Rate, Overstock Indicator. Inbound / Outbound Transaction Records.

[0079] 3. Passenger Flow Data: Passenger volume (time period distribution, average daily passenger flow). Passenger density (heat map area). Average dwell time, conversion rate (conversion of passenger flow to transactions). Peak period identification (e.g., peak passenger flow from 11:00 AM to 1:00 PM).

[0080] 4. Member Data: Member purchase frequency, average order value, repurchase rate. Member demographics (age, gender distribution). Member preference tags (e.g., preferred product categories, price sensitivity). Member activity (last purchase time, login frequency).

[0081] 5. Human Resources Data: Employee schedules, number of employees on duty. Work efficiency indicators (e.g., number of transactions per employee). Matching of labor costs with sales.

[0082] 6. External environment data: weather data (such as the impact of temperature and rainfall on customer flow), holiday markers, and information on competitor store activities (such as promotional activities).

[0083] Step S3022: Based on the status data, call the preset business rule library to perform logical judgment, and at the same time run the preset anomaly detection model to perform pattern recognition, and obtain the logical judgment result and the pattern recognition result respectively.

[0084] Specifically, based on the acquired state data, the diagnostic analysis agent synchronously calls the preset business rule base to make logical judgments based on expert experience, and runs the preset anomaly detection model to perform pattern recognition based on data statistics, thereby obtaining rule-driven judgment results and data-driven recognition results respectively.

[0085] Step S3023: The logical judgment result and the pattern recognition result are fused and cross-validated to identify the business problem, and the business problem is associated with the potential business causes to generate a diagnostic report containing the business problem.

[0086] Specifically, the diagnostic analysis agent integrates and cross-validates logical judgment results with pattern recognition results to comprehensively identify specific operational problems, associates the problems with potential operational causes, and finally generates a structured diagnostic report that includes a problem description, severity, and causal analysis (including potential causes).

[0087] Step S303: The policy-generating agent generates multiple policy proposals corresponding to the operational issues in the diagnostic report based on a pre-built reinforcement learning policy library and business knowledge graph. For details, please refer to [link to details]. Figure 2 Step S203 of the illustrated embodiment will not be described again here.

[0088] Step S304: Multiple strategy proposals are injected into the digital twin environment through a strategy simulation agent. Simulations are then performed based on a pre-built environmental dynamics model and customer response model to predict the execution effects of each strategy proposal and generate a simulation evaluation report. For details, please refer to [link to relevant documentation]. Figure 2 Step S204 of the illustrated embodiment will not be described again here.

[0089] Step S305: Based on the simulation evaluation report, the decision-arbitration agent uses a multi-objective optimization algorithm to rank various strategy proposals and generate a final decision recommendation. For details, please refer to [link to details]. Figure 2 Step S205 of the illustrated embodiment will not be described again here.

[0090] The store operation decision-making method based on multi-agent collaboration provided in this embodiment achieves proactive, accurate, and interpretable identification of store operation problems by using a pre-set business rule base and anomaly detection model in parallel with the diagnostic analysis agents to perform dual analysis and cross-validation of state data in the digital twin environment.

[0091] This embodiment provides a store operation decision-making method based on multi-agent collaboration, which can be used in the aforementioned electronic devices or terminal devices. The electronic devices or terminal devices are equipped with a multi-agent system. Figure 4 This is a flowchart of a store operation decision-making method based on multi-agent collaboration according to an embodiment of the present invention, such as... Figure 4 As shown, the process includes the following steps: Step S401: Acquire multi-source heterogeneous data from the store through data perception and environmental construction of an intelligent agent, and construct a digital twin environment based on the multi-source heterogeneous data. For details, please refer to [link to details]. Figure 3 Step S301 of the illustrated embodiment will not be described again here.

[0092] Step S402: The diagnostic analysis agent, based on the digital twin environment, uses preset business rules and a preset anomaly detection model to diagnose the data in the digital twin environment, generating a diagnostic report containing operational issues. For details, please refer to [link to relevant documentation]. Figure 2 Step S302 of the illustrated embodiment will not be described again here.

[0093] Step S403: Based on a pre-built reinforcement learning policy library and business knowledge graph, the policy-generating agent generates multiple policy proposals corresponding to the operational issues in the diagnostic report.

[0094] In an optional implementation, the reinforcement learning policy library in step S403 above is constructed in the following manner: The state space, action space, and reward function of reinforcement learning are determined, and historical operating data of stores are obtained. Based on the state space, action space, and reward function, reinforcement learning algorithms are used to perform offline reinforcement learning on the historical operating data, or online training is conducted through a simulated environment to obtain the optimal learning strategy. The optimal learning strategy is stored in a preset data format to generate a reinforcement learning strategy library.

[0095] Specifically, offline reinforcement learning training is performed using historical store data (such as state-action-reward sequences), or online training is performed using a simulation environment. The construction process is as follows: Define the state space: including store status (such as inventory level, customer traffic, and sales trends).

[0096] Define the action space: including possible business strategies (such as promotions, price adjustments, and scheduling changes).

[0097] Define the reward function: based on business objectives (such as profit, sales, customer satisfaction).

[0098] Training algorithm: Deep Q-Network (DQN) and Policy Gradient (PG) are used to learn the optimal policy.

[0099] Output: The reinforcement learning policy library is stored as a Q-value table or neural network parameters, mapping the value of states to actions.

[0100] The method for constructing a business knowledge graph is as follows: The graph structure is constructed by extracting knowledge from industry experts, historical cases, and business rules, including: Nodes: Entities such as products, categories, promotional methods, customer groups, and problem types. Edges: Relationships such as "Product-belongs to-category", "Problem-solvable-strategy", and "Strategy-impact-metric". Attributes: Node-attached attributes (such as strategy application conditions and historical effects).

[0101] Example: For the problem of "low sales rate", the knowledge graph may include related strategies: "cross marketing", "coupon push", "display adjustment", and indicate the applicable scenarios and constraints of these strategies.

[0102] Specifically, step S403 includes: Step S4031: Generate an intelligent agent through a strategy to analyze the type, severity, and context of operational problems in the diagnostic report.

[0103] Specifically, the strategy-generating agent receives diagnostic reports (such as "the sales rate of product category A is lower than the threshold of 30%), and analyzes the type, severity, and context of the operational problem (such as the affected product category and time period).

[0104] Step S4032: Based on the type, severity, and context of the business problem, retrieve relevant candidate strategies from the business knowledge graph.

[0105] Specifically, the strategy nodes related to the problem are queried from the business knowledge graph. For example, using "low sales rate" as the query, the set of related strategies {"cross-marketing", "coupon push", "display adjustment"} is retrieved. At the same time, the preset parameters and constraints of the strategies are retrieved (such as "cross-marketing" requires specifying the related product category, and "coupon push" requires setting the discount rate).

[0106] Step S4033: Input the current store status data into the reinforcement learning strategy library and calculate the expected value of each candidate strategy.

[0107] Specifically, the current store status (obtained from the digital twin environment) is input into the reinforcement learning policy library, and the Q-value or expected reward for each possible action (policy) is calculated.

[0108] For example, for the current state, the strategy library might output: "Cross-marketing" Q value = 0.8, "Coupon push" Q value = 0.6, "Display adjustment" Q value = 0.7.

[0109] Step S4034: Based on expected value and actual business constraints, generate multiple strategy proposals corresponding to the operational issues in the diagnostic report.

[0110] Specifically, based on Q-score ranking and knowledge graph constraints, multiple policy proposals are generated: High Q-value strategies are preferred, but diversity (such as different strategy types) should be ensured.

[0111] Adjust strategy parameters: Generate variants in the parameter space (such as discount rate, promotion duration) through random sampling or optimization algorithms (such as Bayesian optimization).

[0112] Example output: Strategy 1: In-store cross-selling (related category B, 10% discount).

[0113] Strategy 2: Targeted coupon push (for members, 15% discount).

[0114] Strategy 3: Adjust the display location (move Category A to the entrance).

[0115] It should be noted that this embodiment also performs the following feasibility filtering: Use a business rule base (such as cost constraints and resource availability) to filter out unrealistic strategies and ensure that the strategies are executable.

[0116] Finally, a policy proposal is output, and the generated policy proposal and its parameters are published to the communication management center for further evaluation by the policy simulation agent.

[0117] Step S404: Inject multiple strategy proposals into the digital twin environment through a strategy simulation agent, and perform simulation based on a pre-built environmental dynamics model and customer response model to predict the execution effect of each strategy proposal and generate a simulation evaluation report.

[0118] Specifically, the internal working principle diagram of the strategy simulation intelligent agent is as follows: Figure 6 As shown, the strategy simulation agent is the core agent of the multi-agent system. This agent injects strategy proposals into the digital twin environment and simulates the effects of strategy execution based on historical data and learned environmental dynamics models. The simulation considers the mutual influence between strategies and market uncertainty, predicting their combined impact on multiple indicators such as sales revenue, gross profit margin, and inventory levels.

[0119] The final output is a simulation evaluation report for each strategy, including predicted values ​​of key indicators, confidence intervals, and risk warnings.

[0120] Step S404 above includes: Step S4041: Extract the executable parameters of each policy proposal through the policy simulation agent, and inject the executable parameters of each policy proposal into the digital twin environment.

[0121] Specifically, executable parameters are a set of key quantitative input values ​​extracted by the strategy simulation agent from the "strategy proposal" provided by the strategy generation agent. These parameters are used to define and initiate a specific sales strategy simulation in the digital twin environment. Executable parameters may include sales-related parameters and non-sales-related parameters, such as scheduling adjustments and display optimizations.

[0122] For example, the executable parameters for a "targeted coupon push" strategy proposal might include: Coupon discount rate: for example, 15%; Applicable product categories: such as grains, oils, and non-staple foods; Target customer group: such as silver card members who have not made a purchase in the past 30 days; Validity period: for example, 7 days; Distribution channels: such as the store's APP message center.

[0123] For a strategy proposal of a "limited-time discount promotion", the executable parameters may include: Discount amount: for example, 20% off; Participating product SKU list: for example, [Product A, Product B]; Promotional period: For example, from October 1, 2023 to October 7, 2023; Purchase limit per person: for example, 2 items.

[0124] In this step, the policy simulation agent parses the executable parameters from each policy proposal and loads these parameters as input conditions into the digital twin environment, thereby specifically characterizing the execution settings of each policy in the simulation.

[0125] Step S4042: Based on the pre-built environmental dynamics model and customer response model, perform multi-round, multi-time-step simulations in the digital twin environment after injecting executable parameters to simulate the dynamic interaction and evolution of the market, customer and store status after the implementation of each strategy proposal, and obtain simulation results.

[0126] Specifically, the environmental dynamics model is obtained by an environmental simulation agent based on an LSTM-Markov hybrid model to predict the evolution of the environmental state, and external factors are taken into account.

[0127] Customer Response Model: In response to the high proportion of elderly customers in community-based stores, the system has established a specialized customer response model: 1. A quantitative system for the characteristics of middle-aged and elderly customers, including: (1) Based on the age tags and shopping basket data of the membership system, construct a profile of the middle-aged and elderly customer group, including: Price sensitivity index: calculated based on the price elasticity coefficient, middle-aged and elderly customers usually show high price sensitivity (0.8-0.9); Brand loyalty score: calculated based on repeat purchase rate and brand switching frequency; New product acceptance threshold: assessed based on the first purchase time of new products and the response rate to trial promotions.

[0128] (2) Shopping behavior feature modeling: Time period preference coefficient: The weight of the morning period (9-11 am) is increased by 30%. Category attention matrix: The attention of health category is increased by 40%, and the attention of daily necessities is increased by 25%.

[0129] 2. The simulation optimization mechanism for age-friendly strategies includes: (1) Adaptive adjustment of marketing strategy parameters: Promotion cycle extension algorithm: Based on the decision cycle data of middle-aged and elderly customers, the promotion duration is automatically extended by 25-50%; Discount range optimization: A tiered discount trigger mechanism is set for price-sensitive characteristics.

[0130] (2) Service strategy effect simulation: Evaluation of the benefits of manual service channels: Quantify the effect of increasing manual services on the average transaction value and customer satisfaction; Shopping convenience improvement simulation: Simulate the impact of improvement measures such as display optimization and channel widening on the retention rate of middle-aged and elderly customers.

[0131] In this step, based on the pre-built environmental dynamics model and customer response model, the strategy simulation agent performs multi-round, multi-time-step simulations in the digital twin environment after the parameters are injected. It fully reproduces the time-progressive interactive evolution process between the market, customers and store status after the strategy is executed, and outputs the simulation results.

[0132] This embodiment establishes a quantitative behavioral characteristic model for specific customer groups such as the middle-aged and elderly, which significantly improves the simulation accuracy in specific scenarios.

[0133] Step S4043: Extract the key business indicator sequence from the simulation results, and based on the key business indicator sequence, predict the final effect of the strategy proposal and assess the uncertainty.

[0134] Specifically, the strategy simulation agent extracts the sequence data of key business indicators changing over time from the inference results, and uses an ensemble learning method to predict the final execution effect of the strategy based on the sequence. At the same time, it calculates the confidence interval and risk level of the predicted value through statistical methods to quantify the uncertainty of the effect.

[0135] Step S4044: Based on the final effect and uncertainty, generate a structured simulation evaluation report for each strategy proposal.

[0136] Specifically, the strategy simulation agent integrates the aforementioned predicted values, confidence intervals, and risk analysis to generate a structured simulation evaluation report for each strategy proposal, which includes quantitative predictions, effect comparisons, and risk warnings.

[0137] This embodiment provides decision-makers with comprehensive risk insights and decision-making basis through multi-model integration and uncertainty quantification.

[0138] Step S405: After the policy simulation agent completes the evaluation of all policy proposals, it publishes all simulation evaluation reports to the communication management center.

[0139] Specifically, the communication management hub provides a shared workspace and communication bus for all agents. Agents publish their results (such as diagnostic reports and policy proposals) to the "blackboard," and other agents can subscribe to and consume this information, enabling an asynchronous and decoupled collaborative working mode.

[0140] Step S406: Based on the simulation evaluation report, the decision arbitration agent sorts out various strategy proposals using a multi-objective optimization algorithm to generate the final decision recommendation.

[0141] Specifically, the decision-making arbitration agent receives all strategies and their simulation reports, and comprehensively evaluates and ranks the strategies based on multi-objective optimization algorithms (such as balancing revenue, profit, and customer satisfaction) or predefined business rules.

[0142] Output: Final decision recommendations, including preferred strategy proposals, alternative strategy proposals, and a detailed list of execution instructions.

[0143] Step S406 above includes: Step S4061: Extract the sequence of key business indicators from the simulation evaluation report through the decision arbitration agent, and assign corresponding weights to each business indicator based on the preset business target priority.

[0144] Specifically, the decision-making arbitration agent automatically reads the predefined key business indicator sequence by parsing the structured data fields (such as JSON or specific protocol format) of the simulation evaluation report; then, it queries the preset business target configuration table, dynamically retrieves and loads the weight coefficients corresponding to each indicator according to the current decision scenario (such as profit priority or turnover priority), and completes the weight allocation.

[0145] Step S4062: Using a multi-objective optimization algorithm, calculate the comprehensive score for each strategy proposal based on each business indicator and its corresponding weight.

[0146] Specifically, the decision-arbitration agent inputs the extracted index sequence and assigned weights into a built-in multi-objective optimization algorithm (such as a linear weighted method or a Pareto front-based solver). The algorithm first normalizes the heterogeneous indices, then calculates the weighted utility value of each policy proposal based on the weights, and finally outputs a comparable score that represents the overall performance.

[0147] Step S4063: Rank the various strategies based on the comprehensive score, and determine the preferred strategy proposal and alternative strategy proposal based on the ranking results.

[0148] Specifically, all strategy proposals are sorted in descending order based on the calculated comprehensive score. On this basis, the decision arbitration agent calls the preset screening rules (such as score threshold and minimum differentiation requirement) to verify the sorting results, automatically marks the first-ranked and rule-compliant preferred strategy proposal as the final recommended solution, and selects several proposals that follow closely behind as alternative solutions recognized by the system.

[0149] Step S4064: Generate an executable instruction list for the preferred strategy proposal, attach conditional descriptions to the alternative strategy proposals, and generate a final decision recommendation.

[0150] Specifically, the decision-arbitration intelligent agent matches the corresponding operation template from the instruction template library based on the type and parameters of the preferred strategy proposal, and fills the template with specific parameters (such as promotion range and execution period) to generate an operation instruction list that can be directly issued to the business system. At the same time, it attaches explanatory text consisting of its triggering conditions (such as "activate when the preferred solution is blocked") to each alternative solution, and finally packages it into a structured decision suggestion document.

[0151] In step S407, the final decision recommendation is sent to the execution terminal, and the actual effect data after execution is collected; based on the actual effect data, the pre-built environmental dynamics model and reinforcement learning strategy library are optimized.

[0152] The store operation decision-making method based on multi-agent collaboration provided in this embodiment achieves an efficient transformation from multi-strategy simulation to executable decisions by using a decision arbitration agent to make objective and quantitative comprehensive judgments on simulation evaluation results based on a multi-objective optimization algorithm. From problem discovery to strategy execution and effect feedback, a complete automated decision-making chain is formed, realizing the continuous self-optimization of the system.

[0153] As one or more specific application embodiments of the present invention, combined with Figure 7 The present invention provides a more detailed description of the store operation decision-making method based on multi-agent collaboration, such as... Figure 7 As shown, the specific process is as follows: 1. Data synchronization and environment construction: The data-aware intelligent agent runs continuously to build and update the digital twin environment.

[0154] 2. Problem Triggering and Diagnosis: The diagnostic analysis agent periodically scans the digital twin environment, or is triggered by a preset threshold event to start the diagnostic process and publish the diagnostic report to the communication center.

[0155] 3. Strategy Generation: The strategy generation agent subscribes to diagnostic reports, generates strategy proposals based on the problem type, and publishes them to the communication management center.

[0156] 4. Policy Simulation: The policy simulation agent obtains policy proposals, conducts multiple rounds of simulation in the digital twin environment, and publishes the evaluation report back to the communication management center.

[0157] 5. Arbitration and Decision-Making: The decision-making arbitration agent collects all information, conducts arbitration, and generates a final decision recommendation.

[0158] 6. Execution and Feedback: Decision recommendations are pushed to the store manager's app or relevant business systems (such as marketing platforms and ERP) via an interface for execution. The system simultaneously monitors the actual effects after execution to optimize and calibrate the simulation model, forming a closed loop.

[0159] The store operation decision-making method based on multi-agent collaboration provided in this embodiment has the following core technical points: I. A Multi-Agent Interaction Simulation Method Based on Dynamic Environment Modeling: Unlike static rule-based simulations in related technologies, this invention proposes a dynamic environment modeling mechanism: 1. Dynamic update mechanism for environmental status: Build an environmental state encoder based on temporal Transformer to fuse multi-source heterogeneous data (sales data, customer flow data, weather data, competitive intelligence, etc.) in real time. Graph neural networks are used to model product relationships and capture the cross-influence effects between product categories. An attention mechanism is introduced to dynamically adjust the weights of environmental parameters, adapting to changes in business characteristics at different times and in different seasons.

[0160] 2. Multi-agent cooperative simulation architecture: Environmental simulation agent: predicts environmental state evolution based on LSTM-Markov hybrid model, taking into account external factors; Customer Response Agent Cluster: The customer decision-making process is modeled using a deep Q-network, with each agent having independent attribute characteristics and behavioral strategies; Strategy execution agent: Simulates employee execution efficiency and trains an execution effect prediction model based on historical execution data.

[0161] II. Differentiated Response Modeling for Middle-Aged and Elderly Customers: In response to the high proportion of elderly customers in community-based stores, a specialized customer response model has been established.

[0162] III. Uncertainty Quantification and Risk Assessment Based on Ensemble Learning: By employing advanced ensemble learning methods to construct simulation prediction models, prediction accuracy is significantly improved. 1. Multi-model ensemble prediction framework: Base model combination: XGBoost (for handling structured features), Temporal Fusion Transformer (for time series prediction), Graph Neural Network (for modeling relationships). Dynamic weight allocation: Based on the model's performance on recent data, dynamically adjust the weights of each model in the integration; Group prediction optimization: Establish specialized model parameter sets for different customer group structures (such as the proportion of middle-aged and elderly people).

[0163] 2. Risk assessment and decision support: Confidence interval calculation: Calculate the 90% confidence interval of the predicted value using Bootstrap sampling; Risk level classification: Based on the forecast variance and the degree of business impact, the strategy risk is classified into three levels: high, medium and low. Scenario Analysis: Provides detailed forecast reports for three scenarios: best-case, expected, and worst-case.

[0164] The multi-agent system architecture of the present invention has the following improvements: 1. Dynamic closed-loop decision-making mechanism: Real-time data-driven environment updates, rather than static rule bases; Based on reinforcement learning-based policy adaptive optimization, the system can continuously learn from execution feedback; The simulation model has an online calibration mechanism that continuously corrects prediction biases using real business data.

[0165] 2. Multi-level intelligent agent collaborative architecture: A four-stage decision-making process—diagnosis, generation, simulation, and arbitration—forms a complete cognitive decision-making chain. An asynchronous communication mechanism based on a publish-subscribe pattern supports the dynamic expansion and replacement of agents; The digital twin environment serves as a unified source of facts, ensuring consistency in the decision-making basis of each agent.

[0166] 3. Specialized design for retail scenarios: We design specialized solutions to address unique challenges in store operations, such as product sales, customer conversion, and inventory turnover. Considering the differences in behavioral characteristics among different customer groups (especially middle-aged and elderly customers), it provides accurate simulation predictions; Integrate with the actual store operation systems (ERP, CRM, etc.) to achieve seamless connection from decision-making to execution.

[0167] This embodiment also provides a store operation decision-making device based on multi-agent collaboration. This device is used to implement the above embodiments and preferred embodiments, and details already described will not be repeated. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0168] This embodiment provides a store operation decision-making device based on multi-agent collaboration, such as... Figure 8 As shown, it includes: The digital twin environment construction module 801 is used to acquire multi-source heterogeneous data of stores through data perception and environment construction intelligent agents, and to construct a digital twin environment based on the multi-source heterogeneous data.

[0169] The diagnostic report generation module 802 is used to diagnose the data of the digital twin environment based on the digital twin environment by using a diagnostic analysis agent, and adopting preset business rules and preset anomaly detection models to generate a diagnostic report containing business problems.

[0170] The strategy proposal generation module 803 is used to generate various strategy proposals corresponding to the operational issues in the diagnostic report by a strategy generation agent based on a pre-built reinforcement learning strategy library and a pre-built business knowledge graph.

[0171] The simulation module 804 is used to inject multiple policy proposals into the digital twin environment through a policy simulation agent, and perform simulation based on a pre-built environmental dynamics model to predict the execution effect of each policy proposal and generate a simulation evaluation report.

[0172] The decision recommendation generation module 805 is used to rank various strategy proposals based on the simulation evaluation report by a decision arbitration agent and generate the final decision recommendation.

[0173] In some optional implementations, the digital twin environment is dynamically updated; the multi-source heterogeneous data includes time-series sales data, product data, and external environment data; the digital twin environment construction module 801 includes: The time-series modeling unit is used to encode time-series sales data using a time-series Transformer model.

[0174] The graph neural network modeling unit is used to model the relationships between products based on product data using graph neural networks.

[0175] The digital twin environment generation unit is used to use external environment data as environmental variables and map the encoded time-series sales data and the relationships between products into dynamic state vectors of stores, forming a digital twin environment for strategy simulation.

[0176] In some optional implementations, the diagnostic report generation module 802 includes: The status data acquisition unit is used to acquire store operation status data from the digital twin environment through diagnostic analysis agents.

[0177] The pattern recognition unit is used to make logical judgments based on state data by calling a preset business rule library, and at the same time run a preset anomaly detection model to perform pattern recognition, thereby obtaining logical judgment results and pattern recognition results respectively.

[0178] The diagnostic report generation unit is used to fuse and cross-validate the logical judgment results and pattern recognition results, identify operational problems, associate the operational problems with potential business causes, and generate a diagnostic report containing the operational problems.

[0179] In some alternative implementations, the reinforcement learning policy library is constructed in the following manner: Determine the state space, action space, and reward function for reinforcement learning, and obtain historical operating data of the stores; Based on the state space, action space, and reward function, reinforcement learning algorithms are used to perform offline reinforcement learning on historical business data, or online training is conducted through a simulated environment to obtain the optimal learning strategy. The optimal learning strategy is stored in a preset data format to generate a reinforcement learning strategy library.

[0180] In some optional implementations, the strategy proposal generation module 803 includes: The parsing unit is used to generate an agent through a strategy to parse the type, severity, and context of operational problems in the diagnostic report.

[0181] The strategy retrieval unit is used to retrieve relevant candidate strategies from the business knowledge graph based on the type, severity, and context of the business problem.

[0182] The reinforcement learning unit is used to input the current store status data into the reinforcement learning strategy library and calculate the expected value of each candidate strategy.

[0183] The strategy proposal generation unit is used to generate various strategy proposals corresponding to the operational issues in the diagnostic report, based on expected value and actual business constraints.

[0184] In some alternative implementations, the simulation module 804 includes: The parameter injection unit is used to extract the executable parameters of each policy proposal through the policy simulation agent and inject the executable parameters of each policy proposal into the digital twin environment.

[0185] The simulation and deduction unit is used to perform multi-round, multi-time-step deductions in a digital twin environment after injecting executable parameters, based on pre-built environmental dynamics models and customer response models. It simulates the dynamic interaction and evolution of the market, customer and store status after the implementation of each strategy proposal, and obtains simulation and deduction results.

[0186] The results evaluation unit is used to extract key business indicator sequences from simulation results, and based on the key business indicator sequences, predict the final effect of the strategy proposal and assess the uncertainty.

[0187] The simulation evaluation report generation unit is used to generate a structured simulation evaluation report for each strategy proposal based on the final results and uncertainties.

[0188] In some alternative embodiments, the device further includes: The communication hub shared unit is used to publish all simulation evaluation reports to the communication management hub after the policy simulation agent has completed the evaluation of all policy proposals.

[0189] In some alternative implementations, the decision suggestion generation module 805 includes: The indicator extraction and weight assignment unit is used to extract the sequence of key business indicators from the simulation evaluation report through the decision arbitration agent, and assign corresponding weights to each business indicator based on the preset business target priority.

[0190] The comprehensive score calculation unit is used to calculate the comprehensive score of each strategy proposal based on each business indicator and its corresponding weight using a multi-objective optimization algorithm.

[0191] The preferred alternative strategy proposal determination unit is used to rank multiple strategies based on comprehensive scores and determine the preferred strategy proposal and alternative strategy proposals based on the ranking results.

[0192] The final decision recommendation generation unit is used to generate an executable list of instructions for the preferred strategy proposal and to attach conditional descriptions for the alternative strategy proposals, thereby generating a final decision recommendation.

[0193] In one alternative embodiment, the device further includes: The practical application module is used to send the final decision recommendations to the execution terminal and collect the actual effect data after execution; based on the actual effect data, it optimizes the pre-built environmental dynamics model and reinforcement learning strategy library.

[0194] The store operation decision-making device based on multi-agent collaboration provided in this embodiment of the invention can execute the store operation decision-making method based on multi-agent collaboration provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method execution. Further functional descriptions of the above modules and units are the same as those in the corresponding embodiments described above, and will not be repeated here.

[0195] Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.

[0196] The following is a detailed reference. Figure 9 This diagram illustrates a structural schematic suitable for implementing an electronic device according to embodiments of the present invention. The electronic device may include a processor (e.g., a central processing unit, graphics processor, etc.) 901, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 902 or a program loaded from memory 908 into random access memory (RAM) 903. The RAM 903 also stores various programs and data required for the operation of the electronic device. The processor 901, ROM 902, and RAM 903 are interconnected via a bus 904. An input / output (I / O) interface 905 is also connected to the bus 904.

[0197] Typically, the following devices can be connected to I / O interface 905: input devices 906 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 907 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; memory devices 908 including, for example, magnetic tapes, hard disks, etc.; and communication devices 909. Communication device 909 allows electronic devices to exchange data via wireless or wired communication with other devices. Although Figure 9 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown, and more or fewer devices may be implemented or have instead.

[0198] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 909, or installed from a memory 908, or installed from a ROM 902. When the computer program is executed by a processor 901, it performs the functions defined in the multi-agent collaborative store operation decision-making method of the embodiments of the present invention.

[0199] Figure 9 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0200] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the multi-agent collaborative store operation decision-making method shown in the above embodiments is implemented.

[0201] A portion of this invention can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to the invention through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0202] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A store operation decision-making method based on multi-agent collaboration, characterized in that, The method includes: By using data perception and environmental construction to build an intelligent agent, multi-source heterogeneous data of stores is obtained, and a digital twin environment is constructed based on the multi-source heterogeneous data; The diagnostic analysis agent diagnoses the data in the digital twin environment based on the digital twin environment using preset business rules and preset anomaly detection models, and generates a diagnostic report containing operational issues. The strategy-generating agent generates multiple strategy proposals corresponding to the operational issues in the diagnostic report based on a pre-built reinforcement learning strategy library and business knowledge graph. Multiple strategy proposals are injected into the digital twin environment through a strategy simulation agent, and simulations are performed based on a pre-built environmental dynamics model and customer response model to predict the execution effect of each strategy proposal and generate a simulation evaluation report. Based on the simulation evaluation report, the decision-arbitration agent uses a multi-objective optimization algorithm to rank various strategy proposals and generate a final decision recommendation.

2. The method according to claim 1, characterized in that, The digital twin environment is dynamically updated; the multi-source heterogeneous data includes time-series sales data, product data, and external environment data. Constructing a digital twin environment based on the aforementioned multi-source heterogeneous data includes: The time-series sales data is encoded using a time-series Transformer model; Based on the product data, a graph neural network is used to model the relationships between products; The external environment data is used as environmental variables, and the encoded time-series sales data and the relationships between products are mapped to the dynamic state vector of the store to form a digital twin environment for strategy simulation.

3. The method according to claim 1, characterized in that, Based on the digital twin environment, the diagnostic analysis agent uses preset business rules and a preset anomaly detection model to diagnose the data in the digital twin environment, generating a diagnostic report containing operational issues, including: The diagnostic analysis agent obtains the store operation status data from the digital twin environment; Based on the state data, a preset business rule library is invoked to perform logical judgment, and a preset anomaly detection model is run to perform pattern recognition, so as to obtain the logical judgment result and the pattern recognition result respectively. The logical judgment results and pattern recognition results are fused and cross-validated to identify operational problems. These operational problems are then associated with potential business causes to generate a diagnostic report containing the operational problems.

4. The method according to claim 1, characterized in that, The reinforcement learning policy library is constructed in the following manner: Determine the state space, action space, and reward function for reinforcement learning, and obtain historical operating data of the stores; Based on the state space, action space, and reward function, the historical operating data is subjected to offline reinforcement learning using a reinforcement learning algorithm, or online training is performed through a simulated environment to obtain the optimal learning strategy. The optimal learning strategy is stored in a preset data format to generate a reinforcement learning strategy library.

5. The method according to claim 1, characterized in that, The policy-generating agent, based on a pre-built reinforcement learning policy library and business knowledge graph, generates multiple policy proposals corresponding to the operational issues in the diagnostic report, including: The strategy generates an intelligent agent to analyze the type, severity, and context of operational problems in the diagnostic report; Based on the type, severity, and context of the business problem, relevant candidate strategies are retrieved from the business knowledge graph; Input the current store status data into the reinforcement learning strategy library and calculate the expected value of each candidate strategy; Based on the expected value and actual business constraints, various strategy proposals are generated to correspond to the operational issues in the diagnostic report.

6. The method according to claim 1, characterized in that, Multiple strategy proposals are injected into the digital twin environment through a strategy simulation agent. Simulations are then performed based on pre-built environmental dynamics and customer response models to predict the execution effects of each strategy proposal and generate a simulation evaluation report, including: The executable parameters of each policy proposal are extracted by a policy simulation agent, and the executable parameters of each policy proposal are injected into the digital twin environment. Based on pre-built environmental dynamics models and customer response models, multiple rounds and time steps of simulation are performed in a digital twin environment after injecting executable parameters. This simulates the dynamic interaction and evolution of the market, customer and store status after the implementation of each strategy proposal, and obtains simulation results. From the simulation results, a sequence of key business indicators is extracted, and based on the sequence of key business indicators, the final effect of the strategy proposal is predicted and the uncertainty is assessed. Based on the final results and uncertainties, a structured simulation evaluation report is generated for each strategy proposal.

7. The method according to claim 6, characterized in that, The method further includes: Once the policy simulation agent has completed the evaluation of all policy proposals, it will publish all simulation evaluation reports to the communication management center.

8. The method according to claim 1, characterized in that, Based on the simulation evaluation report, the decision-arbitration agent uses a multi-objective optimization algorithm to rank various strategy proposals and generate a final decision recommendation, including: The decision-arbitration intelligent agent extracts a sequence of key business indicators from the simulation evaluation report and assigns corresponding weights to each business indicator based on the preset business target priority. A multi-objective optimization algorithm is used to calculate the comprehensive score of each strategy proposal based on each business indicator and its corresponding weight. The strategies are ranked based on the comprehensive score, and the preferred strategy proposal and alternative strategy proposals are determined based on the ranking results. Generate an executable list of instructions for the preferred strategy proposal, add conditional descriptions for alternative strategy proposals, and generate a final decision recommendation.

9. The method according to claim 1, characterized in that, The method further includes: The final decision recommendation is sent to the execution terminal, and the actual effect data after execution is collected; Based on the actual performance data, the pre-built environmental dynamics model and reinforcement learning strategy library are optimized.

10. A store operation decision-making device based on multi-agent collaboration, characterized in that, The device includes: The digital twin environment construction module is used to acquire multi-source heterogeneous data of stores through data perception and environment construction intelligent agents, and to construct a digital twin environment based on the multi-source heterogeneous data; The diagnostic report generation module is used to diagnose the data of the digital twin environment based on the digital twin environment by a diagnostic analysis agent, using preset business rules and preset anomaly detection models, and generate a diagnostic report containing operational issues. The strategy proposal generation module is used to generate multiple strategy proposals corresponding to the operational issues in the diagnostic report by a strategy generation agent based on a pre-built reinforcement learning strategy library and a pre-built business knowledge graph. The simulation and deduction module is used to inject multiple policy proposals into the digital twin environment through a policy simulation agent, and perform simulation and deduction based on a pre-built environmental dynamics model to predict the execution effect of each policy proposal and generate a simulation evaluation report. The decision recommendation generation module is used by a decision arbitration agent to rank various strategy proposals based on the simulation evaluation report and generate final decision recommendations using a multi-objective optimization algorithm.

Images