Big model-based three-party game constraint cross-border logistics decision method and system

Abstract

The application discloses a kind of based on big model's three-party game constraint cross-border logistics decision-making method and system, belong to cross-border e-commerce intelligent logistics decision-making technical field, including: obtaining the private observation information of each participating subject and external dynamic environment data;Utilize big language model to process generation structured expert knowledge after quality check;Accordingly, construct three-party imperfect information game model, and generate three-party game constraint;Extract hard boundary constraint from three-party game constraint, and map hard boundary constraint into action mask;Private observation information and public belief state are input into multi-agent reinforcement learning model, and each agent generates and outputs decision action under the constraint of action mask;According to the decision action of each participating subject, generate collaborative strategy.The application improves the compliance, executability and dynamic environment adaptability of decision action.

Description

Technical Field

[0001] This invention relates to the field of intelligent logistics decision-making technology for cross-border e-commerce, and in particular to a cross-border logistics decision-making method and system based on a large-scale model and constrained by a three-party game. Background Technology

[0002] Cross-border e-commerce logistics typically involves cross-border e-commerce platforms, logistics service providers, and overseas fulfillment entities. These overseas fulfillment entities can include overseas warehouses, overseas merchants, and overseas last-mile delivery partners. Compared to regular e-commerce logistics, cross-border e-commerce logistics is characterized by longer supply chains, more nodes, higher customs clearance risks, greater uncertainty in delivery times, complex return processing, rapid changes in external policies and regulations, and complex relationships among multiple stakeholders.

[0003] These business characteristics place stringent demands on decision-making systems, including multi-source heterogeneous fusion, dynamic environmental adaptation, and balancing the interests of multiple stakeholders. However, existing collaborative decision-making methods for cross-border e-commerce logistics mainly rely on human experience, rule engines, single-agent prediction models, or conventional reinforcement learning models. When facing these requirements, they still have the following shortcomings:

[0004] First, the information held by cross-border e-commerce platforms, logistics service providers, and overseas fulfillment entities differs significantly in terms of source, dimension, format, and update frequency, and there are inherent information barriers among the three parties. Specifically, cross-border e-commerce platforms typically possess information such as orders, users, complaints, subsidies, and platform revenue; logistics service providers typically possess information such as transportation routes, clearance points, transport capacity, transportation costs, and delay compensation; and overseas fulfillment entities typically possess information such as overseas warehouse inventory, shipping capacity, replenishment cycles, and return processing capabilities. The proprietary information held by each party is not transparent to the others, and existing methods lack effective information integration and sharing mechanisms, making it difficult to adequately address this problem of incomplete information.

[0005] Second, cross-border e-commerce logistics decisions are affected by the dynamic external environment, including destination country customs clearance policies, restrictions on shipments, changes in taxes and fees, carrier announcements, overseas warehouse anomalies, and platform rules. These policies and rules are frequently updated and can be sudden. Traditional models typically rely on static training samples or fixed rules for decision-making, making it difficult to promptly perceive changes in the external environment and adaptively adjust decision-making strategies.

[0006] Third, while large language models possess semantic understanding and reasoning capabilities, their direct application in logistics decision-making may lead to issues such as output illusions, policy knowledge lagging behind the latest announcements, generated actions being unexecutable in actual business operations, and unstable confidence assessments. Without structured output constraints and rigorous quality verification mechanisms, unreliable knowledge generated by large language models can be directly introduced into downstream decision-making models, posing a serious risk to the security and accuracy of the final collaborative strategy.

[0007] Fourth, the interests of the three parties involved in cross-border e-commerce logistics differ significantly: platforms focus on order revenue and user experience, logistics service providers focus on transportation profits and capacity utilization, and overseas fulfillment entities focus on inventory turnover and fulfillment revenue. Furthermore, the degree to which different entities bear risks such as customs clearance delays, inventory backlogs, and capacity overload varies, and the long-term prediction accuracy of each party's information sources differs objectively. Existing multi-agent reinforcement learning methods typically employ a unified collaborative reward allocation mechanism, lacking explicit modeling of the aforementioned differences in revenue, risk-bearing, and information credibility. This leads to an unreasonable allocation of collaborative rewards, easily triggering "free-riding" behavior or policy degradation due to risk aversion, thus affecting long-term collaborative stability.

[0008] Therefore, there is an urgent need for a collaborative decision-making method for cross-border e-commerce logistics that can combine knowledge enhancement of large language models, constraints of three-party imperfect information games, and heterogeneous multi-agent reinforcement learning to solve the problems of incomplete information, poor dynamic adaptability, unreliable output of large language models, and unreasonable distribution of collaborative rewards, thereby improving the accuracy, stability, and dynamic adaptability of collaborative decision-making in cross-border e-commerce logistics. Summary of the Invention

[0009] The purpose of this invention is to provide a cross-border logistics decision-making method and system based on a large-scale model and a three-party game constraint, in order to solve the problem that in cross-border e-commerce logistics scenarios, due to the highly heterogeneous and opaque information of the participating entities and the frequent changes in external policies and rules, existing decision-making methods cannot simultaneously take into account decision security, collaborative stability, credit allocation fairness, and dynamic environment adaptability.

[0010] To achieve the above objectives, this application adopts the following technical solution:

[0011] This application presents a cross-border logistics decision-making method based on a large-scale model and constrained by a three-party game, comprising the following steps:

[0012] The method involves acquiring private observation information corresponding to multiple participating entities in cross-border e-commerce logistics decision-making, as well as acquiring external dynamic environmental data related to cross-border logistics operations. The participating entities include at least cross-border e-commerce platforms, logistics service providers, and overseas fulfillment entities. The method uses a large language model to process the private observation information and external dynamic environmental data to generate quality-verified structured expert knowledge. The structured expert knowledge includes at least a public belief state used to characterize the shared environmental cognition of the three parties.

[0013] Based on the private observation information and the state of public beliefs, a three-party imperfect information game model is constructed, and three-party game constraints are generated based on the game model to constrain the decision-making behavior of each participating entity.

[0014] Extract hard boundary constraints from the three-party game constraints and map the hard boundary constraints to action masks;

[0015] The private observation information and public belief state are input into a pre-trained multi-agent reinforcement learning model. The multi-agent reinforcement learning model contains agents that correspond one-to-one with each participating subject. Each agent generates and outputs decision actions under the constraints of the action mask.

[0016] Based on the decision-making actions of each participating entity, a collaborative strategy is generated to execute cross-border e-commerce logistics operations.

[0017] A cross-border logistics decision-making system based on a large-scale model and constrained by a three-party game includes:

[0018] The data acquisition module is used to acquire private observation information corresponding to multiple participating entities in cross-border e-commerce logistics decision-making, as well as external dynamic environmental data related to cross-border logistics operations. The participating entities include at least cross-border e-commerce platforms, logistics service providers, and overseas fulfillment entities.

[0019] An enhanced processing module is used to process the private observation information and external dynamic environment data using a large language model to generate quality-verified structured expert knowledge, wherein the structured expert knowledge includes at least a public belief state for characterizing the shared environmental cognition of the three parties.

[0020] The game generation module is used to construct a three-party imperfect information game model based on the private observation information and the public belief state, and to generate three-party game constraints based on the game model to constrain the decision-making behavior of each participating subject.

[0021] The constraint mapping module is used to extract hard boundary constraints from the constraints of the three-party game and map the hard boundary constraints into action masks.

[0022] The reinforcement learning module is used to input the private observation information and public belief state into a pre-trained multi-agent reinforcement learning model. The multi-agent reinforcement learning model contains agents that correspond one-to-one with each participating subject. Each agent generates and outputs decision actions under the constraints of the action mask.

[0023] The strategy output module is used to generate collaborative strategies for executing cross-border e-commerce logistics operations based on the decision-making actions of each participating entity.

[0024] The present invention has the following beneficial effects:

[0025] 1) Local decision knowledge is generated by using a large language model corresponding to the roles of each participating entity, and quality evaluation, correction and cross-entity fusion are performed by using a large language model of domain experts. This generates a public belief state without directly sharing the private observation details of each entity, which improves the reliability of fusion of multi-source heterogeneous logistics data under imperfect information conditions.

[0026] 2) By using structured prompt templates, field format validation, field type validation, valid action set validation, and confidence threshold validation, the probability of format errors, unexecutable actions, and distorted risk labels in the output of large language models is reduced, thereby improving the structuring and usability of input data for downstream decision-making models.

[0027] 3) By mapping prohibition and restriction rules, customs clearance risk thresholds, inventory safety levels, carrier capacity constraints, and route availability constraints into action masks, candidate actions that do not meet strong risk safety constraints are directly masked during the inference stage, thereby improving the compliance and feasibility of cross-border logistics decision-making actions.

[0028] 4) By mapping the cooperative stability constraint, the unilateral deviation payoff constraint, and the weak risk safety constraint into a soft penalty term in the reward function, the multi-agent reinforcement learning model can learn to avoid high-risk and unstable actions during the training phase, thereby improving the stability of policy output in dynamic environments.

[0029] 5) By introducing the action preference distribution into the teacher-guided reward channel and combining it with private benefit reward, collaborative benefit reward and game constraint penalty channels for joint training, the convergence efficiency and policy interpretability of the multi-agent reinforcement learning model in complex cross-border logistics scenarios are improved.

[0030] 6) By updating the public belief fusion parameters, tripartite game constraint parameters, and reward function weights based on actual performance feedback, the system can perform closed-loop corrections based on customs clearance results, delays, inventory changes, returns, and complaint feedback, thereby improving the adaptability of the cross-border logistics decision-making system to changes in external rules and performance anomalies. Attached Figure Description

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

[0032] Figure 1 This is a flowchart illustrating the overall process of a cross-border logistics decision-making method based on a large-scale model and constrained by a three-party game, as provided in an embodiment of this application.

[0033] Figure 2 This is another flowchart of a cross-border logistics decision-making method based on a large model and constrained by a three-party game, provided in an embodiment of this application;

[0034] Figure 3 This is a flowchart of the structured expert knowledge generation method provided in the embodiments of this application;

[0035] Figure 4 This is a schematic diagram of the large language model enhancement processing mechanism provided in the embodiments of this application;

[0036] Figure 5 This is a schematic diagram of the structure of three-party game-constrained reinforcement learning and reward allocation provided in the embodiments of this application;

[0037] Figure 6 This is a schematic diagram of the structure of a cross-border logistics decision-making system based on a large model and constrained by a three-party game, provided in an embodiment of this application. Detailed Implementation

[0038] To make the technical solution of this application clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The terms "first," "second," etc., in the claims and specification 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 terms can be used interchangeably where appropriate. This is merely a way of distinguishing objects with the same attributes in the embodiments of this application. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, so that a process, method, system, product, or apparatus that comprises a series of units is not necessarily limited to those units, but may include other units not explicitly listed or inherent to these processes, methods, products, or apparatuses.

[0039] Example 1

[0040] like Figure 1 and Figure 2 As shown, a cross-border logistics decision-making method based on a large model and constrained by a three-party game includes steps S110-S160.

[0041] S110. Obtain private observation information corresponding to multiple participating entities in cross-border e-commerce logistics decision-making, and obtain external dynamic environmental data related to cross-border logistics operations. Among them, the participating entities include at least cross-border e-commerce platforms, logistics service providers, and overseas fulfillment entities.

[0042] In this embodiment, the first step is to start from the cross-border e-commerce platform. Logistics service providers and overseas contract fulfillment entities Private observation information is collected within the business system. It should be noted that private observation information is heterogeneous across different participating entities, specifically manifested in differences in information sources, data formats, data dimensions, and update frequencies. Therefore, private observation information can also be referred to as heterogeneous observation data.

[0043] Specifically, the private observation information of cross-border e-commerce platforms This includes, but is not limited to: order amount, destination country, user timeliness preferences, user complaint information, platform subsidy budget, platform commission, fulfillment score, and traffic allocation information. Among these, user timeliness preferences characterize users' preferred choices regarding delivery time, estimated delivery time, tolerance for delays, and expedited delivery needs; traffic allocation information characterizes the rules and results of cross-border e-commerce platforms allocating page display traffic, recommendation traffic, or search traffic to different products, merchants, or fulfillment channels based on product category, fulfillment capability, inventory status, marketing activities, and user profiles.

[0044] Private observation information of logistics service providers This includes, but is not limited to: logistics routes, international trunk line status, clearance point status, transportation costs, capacity load, historical on-time rate, delay compensation records, and carrier announcements.

[0045] Overseas compliance entities' private observation information This includes, but is not limited to: overseas warehouse inventory, inventory synchronization time, delivery capacity, replenishment cycle, inventory turnover rate, return processing capacity, out-of-stock records, and overseas fulfillment score.

[0046] The system also acquires dynamic external environment data related to cross-border e-commerce logistics decisions by monitoring and collecting data from external data sources. This includes, but is not limited to: destination country customs clearance policies, prohibited or restricted shipping rules, tax and fee rules, exchange rate fluctuations, platform rules, logistics industry dynamics, overseas warehouse anomalies, and historical fulfillment feedback information. Overseas warehouse anomalies include one or more of the following: overseas warehouse overload, discrepancies between inventory records and actual stock, inventory freezes, picking delays, reduced shipping capacity, return backlogs, warehouse shutdowns, last-mile delivery anomalies, and fulfillment anomalies caused by local policies, weather, strikes, or sudden public events.

[0047] After obtaining the aforementioned private observation information and external dynamic environment data, the data is processed including but not limited to field cleaning, format unification, missing value marking, and structured serialization to generate structured input data for each participating entity, which can then be used in subsequent large language model enhancement processing steps.

[0048] S120. Process private observation information and external dynamic environmental data using a large language model to generate quality-verified structured expert knowledge, which includes at least a public belief state used to characterize the shared environmental cognition of the three parties.

[0049] In step S110, the system has acquired the private observation information of each participating entity and the external dynamic environment data related to cross-border logistics operations. After field cleaning, format standardization, missing value marking, and structured serialization, structured input data corresponding to each participating entity is generated. In this step, the system further utilizes a large language model to enhance the structured input data and the external dynamic environment data, generating quality-verified structured expert knowledge.

[0050] In one specific embodiment, such as Figure 3 As shown, the generated quality-verified structured expert knowledge includes:

[0051] S1201. Generate local decision knowledge from the perspective of each subject using the role-based large language model corresponding to each participating subject. The local decision knowledge includes at least one of decision information and output quality measurement. The decision information includes at least one of implicit state estimation, risk label, candidate action ranking, action preference distribution and reward adjustment suggestion. The output quality measurement includes at least one of confidence and uncertainty.

[0052] S1202. The local decision knowledge is evaluated for quality using a domain expert big language model. When the quality evaluation does not meet the preset conditions, correction information is generated and the correction information is returned to the corresponding role big language model to regenerate the local decision knowledge.

[0053] S1203. Perform cross-subject fusion on the local decision-making knowledge obtained through the quality evaluation to generate a public belief state.

[0054] like Figure 4 As shown, the process involves four interconnected stages: role knowledge generation, expert review and correction, integration of public belief states, and global reflection and updating.

[0055] I. Character Knowledge Generation

[0056] Based on the structured inputs from each participating entity, the system generates platform-side local decision-making knowledge through a platform role big language model, logistics-side local decision-making knowledge through a logistics role big language model, and overseas fulfillment-side local decision-making knowledge through an overseas fulfillment role big language model. Each role's big language model performs differentiated inference based on its received private observation information to generate local decision-making knowledge that matches the decision-making needs of the corresponding entity. This local decision-making knowledge includes at least one of two categories: decision information and output quality metrics. Specifically, the decision information category includes one or more of implicit state estimation, risk labeling, candidate action ranking, action preference distribution, and reward adjustment suggestions; the output quality metrics category includes one or more of confidence and uncertainty.

[0057] During the generation process, the system first fills the pre-defined prompt template with the structured input data of each participating subject, the summary of the external dynamic environment, the candidate action set, and the constraints, and then inputs the corresponding role's large language model. The pre-defined prompt template includes subject role fields, private observation fields, public environment fields, candidate action fields, constraint field, and structured output fields. The main role field is used to define the platform role, logistics role, or overseas fulfillment role corresponding to the current large language model; the private observation field is used to fill in the order, route, inventory, cost, transportation capacity, complaint, or fulfillment capability data that the corresponding subject can directly obtain; the public environment field is used to fill in the destination country's policies, prohibition and restriction rules, tax rules, exchange rate changes, logistics announcements, and historical fulfillment feedback summaries; the candidate action field is used to list the currently available order allocation, logistics route, clearance point, overseas warehouse transfer, platform subsidies, return processing, or abnormal warning actions; the constraint field is used to fill in the customs clearance risk threshold, prohibition and restriction rules, inventory safety level, carrier capacity constraints, subsidy budget constraints, and return risk threshold; and the structured output field is used to define one or more of the following: output field name, field type, value range, set of legal actions, and confidence level requirements.

[0058] Each role's large language model outputs local decision knowledge in structured field format. The system performs structured output constraint and validation processing on the model output, including one or more of the following: preset field format validation, field type validation, enumeration value validation, numerical range validation, valid action set validation, and confidence threshold validation. When the output has missing fields, incorrect field types, actions that do not belong to the valid action set, unnormalized action preference distribution, or confidence scores below the preset threshold, the system performs corresponding processing such as regeneration, field completion, action filtering, normalization, weight reduction, or discarding.

[0059] II. Expert review and revision

[0060] After obtaining the local decision knowledge output by the large language model of each role, the system evaluates the quality of the local decision knowledge through the domain expert's large language model.

[0061] The quality assessment results include a factual consistency score. Policy consistency score Performance Execution Score Risk coverage integrity score Structured field integrity score Action legality score and confidence level reasonableness score One or more of the following: Factual Consistency Score, used to evaluate whether the output content is consistent with the input observation data; Policy Consistency Score, used to evaluate whether the output content complies with the destination country's customs clearance policies, restrictions on smuggling, and platform rules; Fulfillment Feasibility Score, used to evaluate whether the suggested actions meet the constraints of inventory, transportation capacity, customs clearance points, and fulfillment capabilities; Risk Coverage Completeness Score, used to evaluate whether the risks of customs clearance, delays, shortages, returns, and complaints are covered; Structured Field Completeness Score, used to evaluate whether the output fields are complete and correctly formatted; Action Legality Score, used to evaluate whether the candidate action belongs to the set of legal actions; Confidence Reasonableness Score, used to evaluate whether the confidence level matches the data basis and historical errors.

[0062] The system performs a weighted sum of the above scores to obtain a comprehensive quality score. :

[0063]

[0064] in, Let be the weighting coefficients for each quality evaluation dimension, and , When the overall quality score Below the preset quality threshold If any key dimension score falls below the corresponding minimum threshold, the system determines that the local decision knowledge does not meet the quality requirements. The domain expert's large language model then generates a correction question and modification suggestions, which are returned to the corresponding role's large language model to regenerate the local decision knowledge. The system repeats the above quality evaluation and regeneration steps until the quality evaluation result meets the preset quality threshold or the preset maximum number of corrections is reached.

[0065] It should be noted here that the preset quality threshold Used to determine whether local decision-making knowledge meets the minimum quality requirements for inclusion in public belief fusion and downstream reinforcement learning training; maximum number of corrections. This is used to limit the number of interaction corrections between the role-based large language model and the domain expert's large language model in the same round of decision-making. When the quality evaluation result is lower than... And the number of corrections has not been reached. When the number of corrections reaches a certain threshold, the system triggers a regeneration; when the number of corrections reaches a certain threshold, the system will regenerate. Still not satisfied At that time, the system will perform a process of downgrading, discarding, or adopting a conservative default strategy for the corresponding local decision knowledge.

[0066] When conflicts arise in the local beliefs, risk labels, or action preferences within the local decision-making knowledge output by different role-based big language models, the domain expert big language model executes conflict resolution based on data freshness, historical prediction accuracy, quality evaluation results, dynamic environmental update information, and constraint priority. For conflicts involving restrictions on transportation, customs clearance policies, inventory safety levels, and carrier unavailability, external dynamic environmental data and strong risk constraints take priority; for conflicts involving revenue prediction, fulfillment capabilities, or user experience, the side with higher historical prediction accuracy and quality evaluation results takes priority; for conflicts where priority cannot be determined, the system retains an uncertainty marker and reduces the weight of the corresponding field in public belief fusion and teacher-guided rewards.

[0067] When multiple modification suggestions generated by the domain expert's large language model conflict with each other, the system resolves them according to the priority of strong risk constraints, policy and rule constraints, legality constraints of actions, enforceability constraints of performance, and benefit optimization suggestions. If the priority still cannot be determined, the corresponding field is marked as a conflict field, its confidence is reduced, and its weight is reduced in the calculation of public belief fusion or teacher guidance reward.

[0068] III. Integration of Public Belief States

[0069] The system performs cross-agent fusion of local decision-making knowledge obtained through quality evaluation to generate a public belief state. The public belief state is used to characterize the implicit state estimates shared by the three parties under imperfect information conditions, including one or more of the following: customs clearance risk belief, logistics capacity belief, overseas warehouse inventory credibility belief, return risk belief, destination country policy risk belief, and the entity's performance capability belief.

[0070] Specifically, for the first There are three latent state variables, and the three roles in the large language model each output local belief values. and confidence level ,in The system determines the fusion score based on quality evaluation results, historical prediction accuracy, data freshness, and dynamic environmental update information. Based on the fusion score, the fusion weights of the large language models for each role for this latent state variable are calculated. :

[0071] ,

[0072] ,

[0073] in, Indicates the first The first role's large language model is for the first Quality evaluation score of each hidden state variable; This indicates its prediction accuracy for similar latent state variables within a historical window; This represents the freshness score of the corresponding input data; This represents the domain relevance score between the role and the z-th latent state variable; This indicates the degree to which dynamic environment update information supports this implicit state variable; It is a non-negative adjustment coefficient; Indicates the first The first role's large language model is for the first The fusion weights of the hidden state variables.

[0074] When the latent state variable is a categorical risk label, the system uses a weighted voting method to determine the public belief state. Specifically, it statistically analyzes the category labels of each role's large language model, uses the fusion weight as the voting weight, and selects the category with the highest weighted votes as the public belief state. The calculation formula is as follows:

[0075] ,

[0076] in, Indicates the first A set of candidate categories for each latent state variable. This is an indicator function. When strong risk constraints are involved, the system prioritizes the outcome with the higher risk level as the conservative public belief.

[0077] IV. Overall Reflection and Update

[0078] Throughout the entire process of role knowledge generation, expert review and correction, and public belief state integration, the system also generates dynamic environment update information based on external dynamic environment data through a global reflective big language model. Specifically, the global reflective big language model collects and timestamps external dynamic environment data such as destination country policies, embargo and restriction rules, customs clearance notices, carrier announcements, platform rules, exchange rate changes, overseas warehouse anomalies, and historical performance feedback. It then uses entity recognition, relation extraction, and conflict detection to determine timeliness conflicts, rule conflicts, or scope of application conflicts between newly collected information and historical knowledge, thereby generating dynamic environment update information. Dynamic environment update information It is used to update one or more of the following: cue constraints, quality evaluation criteria, public belief fusion weights, three-party game constraint parameters, reinforcement learning reward weights, and optional action space.

[0079] The global reflection big language model updates continuously throughout the decision-making process. Updates are triggered when external dynamic environment data changes, and the updated parameters are fed back in real time or at preset cycles to each stage of role knowledge generation, expert review and correction, public belief fusion, and subsequent three-party game constraint generation and multi-agent reinforcement learning inference, thereby achieving real-time linkage response between decision-making strategies and dynamic environments.

[0080] Through the synergistic effects of role knowledge generation, expert review and correction, public belief state fusion, and global reflection and updating, the system ultimately generates quality-verified structured expert knowledge. This structured expert knowledge includes at least a public belief state, and also one or more of the following: risk labels, candidate action ranking, action preference distribution, and reward adjustment suggestions. For example, for cross-border orders destined for a target country, the system can encapsulate the following: public belief states include medium customs clearance risk, high logistics capacity reliability, and low overseas warehouse inventory reliability; risk labels include customs clearance delay risk and inventory synchronization delay risk; candidate action ranking includes prioritizing logistics route A, followed by logistics route B; action preference distribution includes 0.55 for route A, 0.30 for route B, and 0.15 for route C; and reward adjustment suggestions include increasing the penalty weight for customs clearance risk and decreasing the guidance weight for low-confidence teachers. This structured expert knowledge will be used in the subsequent three-party game constraint generation step, serving as one of the inputs to the three-party imperfect information game model, participating in the generation and mapping of three-party game constraints, and ultimately guiding the reasoning and decision-making of the multi-agent reinforcement learning model.

[0081] S130. Based on private observation information and public belief state, construct a three-party imperfect information game model, and generate three-party game constraints based on the game model to constrain the decision-making behavior of each participating entity.

[0082] Specifically, the system treats the cross-border e-commerce platform, logistics service provider, and overseas fulfillment entity as three parties in a game. The state of the three-party imperfect information game is represented as follows:

[0083] ,

[0084] in, It indicates the state of public belief.

[0085] The actions of the platform intelligent agent, logistics intelligent agent, and overseas fulfillment intelligent agent are respectively denoted as Joint actions are recorded as The set of optional actions for each participating entity includes one or more of the following: order allocation, logistics route selection, clearance point selection, overseas warehouse transfer, platform subsidy, return processing, and anomaly warning.

[0086] Construct platform revenue functions, logistics revenue functions, and overseas fulfillment revenue functions respectively:

[0087] ,

[0088] Among them, the platform revenue function Related to one or more of the following: platform profits, user satisfaction, complaint costs, and subsidy costs; logistics revenue function Related to one or more of the following: carrier revenue, transportation costs, customs clearance anomaly costs, and delay compensation; overseas performance revenue function It is related to one or more of the following: fulfillment revenue, inventory holding costs, stockout losses, and return processing costs.

[0089] Based on the aforementioned payoff function, the system generates three-way game constraints to govern the decision-making behavior of each participating entity. These constraints include one or more of the following: individual rationality constraints, risk and safety constraints, collaborative stability constraints, and unilateral deviation payoff constraints.

[0090] Specifically, individual rationality constraints ensure that the benefits each participant receives from collaborative decision-making are no less than the minimum benefits they could obtain through independent decision-making, thus guaranteeing their willingness to participate in collaboration. Risk safety constraints limit the decision-making actions of each participant from exceeding pre-set risk tolerance boundaries. These include strong and weak risk safety constraints. Strong risk safety constraints involve inviolable mandatory bottom lines such as customs clearance risk thresholds, prohibited or restricted transport rules, inventory safety levels, logistics route availability, carrier capacity constraints, and return risk thresholds. Weak risk safety constraints refer to risk constraints with lower risk levels that do not involve mandatory compliance bottom lines; their violation will not lead to unacceptable legal or security consequences, but will affect the overall system performance. Collaborative stability constraints ensure the balance of benefit distribution and the continuity of decision-making strategies among the three parties in long-term collaboration, preventing the breakdown of the collaborative relationship due to short-term fluctuations in interests.

[0091] Unilateral deviation payoff constraints are used to constrain the additional payoff that any participant can obtain by unilaterally changing its own action, given that the actions of other participants remain unchanged. For candidate joint actions... , No. The maximum one-sided deviation payoff for an agent is:

[0092] ,

[0093] in, This represents the current state of the imperfect information game among the three parties. The system determines the [number]th [player's] payoff based on the difference between the maximum single-player deviation payoff and the current payoff. Deviation from the stimulus value of each agent:

[0094] .

[0095] When any agent deviates from the excitation value When the deviation exceeds the preset threshold, the system determines that the current candidate joint action has a risk of unilateral deviation and performs one or more of the following actions on the candidate joint action: action de-weighting, action blocking, reward / penalty, or compensation parameter adjustment, in order to reduce the motivation of each participating entity to unilaterally deviate from the collaborative decision and ensure the stability of the three-party collaboration.

[0096] By constructing the above-mentioned three-party imperfect information game model and generating the three-party game model, the system organically combines the private observation information and public belief state of each participating subject to form a complete constraint system on the decision-making behavior of the three parties.

[0097] S140. Extract hard boundary constraints from the three-party game constraints and map these hard boundary constraints as action masks.

[0098] Specifically, hard boundary constraints are strong-risk security constraints within the three-way game constraints. For such constraints, the system maps them to action masks, which are used to directly mask candidate actions that do not meet the conditions during the inference process of the trained multi-agent reinforcement learning model. The action mask is implemented in binary vector form, with a vector length equal to the size of the action space. Each bit corresponds to a candidate action, where a value of 1 indicates that the action can be executed, and a value of 0 indicates that the action is masked. During the model inference phase, the agent selects only from the set of candidate actions marked as available by the action mask, thereby ensuring that the output decision action always meets the mandatory safety and compliance requirements.

[0099] The remaining non-hard constraints in the three-party game are called soft trade-off constraints. Soft trade-off constraints include at least individual rationality constraints, cooperative stability constraints, stable payoff constraints, unilateral deviation payoff constraints, and weak risk constraints in the three-party game. For these constraints, the system maps them to soft penalties in reward calculation, which are used to apply corresponding penalty terms to the reward function of the trained multi-agent reinforcement learning model based on the degree of constraint violation.

[0100] No. The game constraint penalty for an individual agent can be expressed as:

[0101] ,

[0102] in, Indicates the degree to which an individual's rational constraints are violated. This indicates a penalty item for weak risk constraints. Indicates deviation from the threshold. Indicates the degree of violation of the cooperative stability constraint. These are non-negative weighting coefficients.

[0103] Specifically, the system calculates penalty values ​​based on the degree to which each agent's decision-making actions violate various soft trade-off constraints. The higher the degree of violation, the larger the penalty value. Thus, during the training phase, the reward function guides each agent to gradually learn to optimize decisions while satisfying soft constraints. Since this penalty mechanism is embedded in the reward function and completes policy optimization during the model training phase, the trained model inherently avoids actions that trigger soft penalties when outputting character actions during the actual inference phase.

[0104] It's important to note that action masks and soft penalties operate at different times in reinforcement learning models: action masks act during the inference phase, filtering candidate actions in real-time, a "hard filtering" mechanism; soft penalties act during the training phase, guiding the agent's policy learning direction by setting a penalty term in the reward function. After training, this guidance is internalized into the model parameters, a "soft guidance" mechanism. Through this hierarchical mapping, the system embeds different types of three-way game constraints into different stages of the reinforcement learning model according to their degree of enforceability—hard boundary constraints, through action masks, ensure the bottom-line safety of decisions during the inference phase, while soft trade-off constraints, through reward shaping, guide the agent to gradually learn to seek equilibrium among multiple objectives during the training phase, thus balancing decision safety and policy flexibility. Figure 5 As shown, Figure 5 A schematic diagram of the structure of three-party game-constrained reinforcement learning and reward allocation is provided.

[0105] In a specific example, for a cross-border order destined for a certain country, the strong-risk security constraints in the three-party game constraints generated by the system include: the destination country's prohibition and restriction rules (prohibiting the transport of certain types of goods), customs clearance risk threshold constraints (the customs clearance risk score cannot exceed a preset threshold), and carrier capacity constraints (the carrier's current remaining capacity is insufficient). The system maps these strong-risk security constraints to action masks. When the logistics route selection action involves a prohibited route or a route with a customs clearance risk exceeding the threshold, the action is masked by the action mask, and the agent cannot select that route. At the same time, the system maps the collaborative stability constraint (penalty is triggered when the three-party revenue distribution ratio deviates from the historical average by more than a preset range), the unilateral deviation revenue constraint (penalty is triggered when the platform unilaterally lowers logistics prices to obtain additional revenue), and the weak-risk security constraint (minor penalty is triggered when the historical delay rate of a certain route is high but does not reach the prohibition standard) to soft penalties. During the model inference phase, the agent selects from the set of available actions limited by the action mask. Since the model has already completed policy optimization during the training phase through a reward function that includes soft penalties, the decision actions output by the model naturally tend to avoid triggering or trigger as few soft penalty actions as possible, thereby achieving synergistic optimization of the interests of all three parties while ensuring security.

[0106] S150. Input private observation information and public belief state into a pre-trained multi-agent reinforcement learning model. The multi-agent reinforcement learning model contains agents that correspond one-to-one with each participating subject. Each agent generates and outputs decision actions under the constraint of action mask.

[0107] The multi-agent reinforcement learning model used in this embodiment is pre-trained, including a platform agent, a logistics agent, and an overseas fulfillment agent. This pre-trained multi-agent reinforcement learning model is trained using a centralized training, distributed execution structure. During the training phase, a centralized value evaluation network is set up. This network can access each agent's private observation information, common belief state, joint actions, and three-way game constraint parameters during training. This network is used to centrally evaluate the value of joint actions and guide the policy network updates of each agent. Each agent maintains its own independent policy network, generating actions solely based on its own private observation information and common belief state.

[0108] During training, the reward function of each agent comprises four channels: a private benefit reward channel, a collaborative benefit reward channel, a game-constraint penalty channel, and a teacher-guided reward channel. The private benefit reward channel is calculated based on each agent's own benefit function, reflecting the individual decision-making gains of each participant. The collaborative benefit reward channel is calculated based on the overall gain generated by the three-way collaboration, used to incentivize decision-making behaviors that contribute to improving the overall system efficiency. The game-constraint penalty channel is calculated based on soft penalties, applying negative incentives to decision-making actions that violate collaborative stability constraints, benefit stability class constraints, unilateral deviation benefit constraints, and weak risk safety constraints. The teacher-guided reward channel is calculated based on the action preference distribution generated by the large language model enhancement processing steps, used to guide the action output direction of each agent and accelerate policy convergence.

[0109] Specifically, no. The reward function for each agent is:

[0110] ,

[0111] in, Indicates the first Private revenue rewards for each intelligent agent. This indicates a reward for collaboration among the three parties. This represents the game's constraint and penalty. This indicates a reward for teacher guidance. Indicates the collaborative reward allocation coefficient. and This represents the corresponding weighting coefficient.

[0112] Furthermore, teacher-guided rewards are determined based on the consistency between the agent's actions and the action preference distribution or candidate action ranking that passed the quality evaluation. When the action preference distribution fails the quality evaluation, has a confidence level below a preset threshold, or conflicts with dynamically updated information from the environment, rewards will be given. Reduce to a preset low weight value or set to zero.

[0113] When allocating collaborative reward rewards to each agent, the rewards are distributed according to each agent's collaborative reward allocation weight. This weight is determined based on each agent's counterfactual contribution, risk-taking level, and information credibility. Specifically, the counterfactual contribution is calculated based on the marginal increment of collaborative rewards resulting from each agent's actual actions relative to the baseline actions; the risk-taking level is calculated based on a weighted average of each agent's risk or cost exposure across multiple risk categories; and the information credibility is calculated based on the historical prediction errors of each agent's corresponding information source.

[0114] Specifically, the weighting of collaborative reward allocation ,

[0115] in, This represents the counterfactual contribution after normalization. This indicates the normalized risk tolerance. This indicates the credibility of the information after normalization. The non-negative adjustment coefficient, used to adjust the impact of counterfactual contribution, risk-taking, and information credibility on the distribution of collaborative rewards, can be preset manually or adaptively updated based on historical performance feedback, fluctuations in tripartite revenue, and the stability of the collaborative strategy; the collaborative reward distribution coefficient The result, obtained through normalization calculation, satisfies: .

[0116] No. The counterfactual contribution of each agent is: ,in, Indicates the first The baseline action of an agent.

[0117] No. The risk tolerance of each agent is: ,in, Indicates the first The agent in the th... The amount of risk or cost to be borne. Indicates the first Weights of risk or cost categories To prevent constants with a denominator of zero.

[0118] No. The credibility of information from each agent is: ,in, Indicates the first The moving average of the prediction error of each agent relative to the information source within a preset time window. (The remaining text appears to be incomplete and requires further context.) , and After normalization, we get , and .

[0119] During the execution phase (i.e., the inference phase), the system uses the private observation information and public belief state of each participating entity acquired at the current decision-making moment as input, and feeds them into the policy networks of the platform agent, logistics agent, and overseas fulfillment agent, respectively. Each agent's policy network independently generates candidate actions based on its own private observation information and public belief state. Based on this, the system applies action masks to filter the candidate actions generated by each agent before outputting the final decision action. Since soft penalties have been embedded in the reward function and optimized during the training phase, each agent's policy network has intrinsically learned to avoid decision behaviors that trigger soft penalties after training. Therefore, when outputting decision actions during the inference phase, it naturally tends to avoid triggering soft penalties or trigger them as little as possible.

[0120] The decision actions output by the pre-trained multi-agent reinforcement learning model include the action choices made by each participating entity at the current decision moment, such as the platform's order allocation and subsidy actions, the logistics service provider's logistics route selection and clearance point selection actions, and the overseas fulfillment entity's overseas warehouse allocation and return processing actions. These decision actions will be passed to the subsequent collaborative strategy generation step to generate a complete collaborative strategy that can be executed for cross-border e-commerce logistics operations.

[0121] Through a centralized training and decentralized execution architecture, each agent can independently complete decision-making and reasoning by relying only on its own private observation information and shared public belief state during the inference phase, without needing to obtain the private observation details of other participating entities. This ensures the quality of decision-making while protecting the data privacy of each participating entity.

[0122] S160. Generate collaborative strategies for executing cross-border e-commerce logistics operations based on the decision-making actions of each participating entity.

[0123] Specifically, the system aggregates and integrates platform decision-making actions output by the platform intelligent agent, logistics decision-making actions output by the logistics intelligent agent, and overseas fulfillment decision-making actions output by the overseas fulfillment intelligent agent to generate a cross-border e-commerce logistics collaboration strategy that includes instructions for collaborative execution by all participating entities. This collaboration strategy includes one or more of the following: order allocation strategy, logistics route strategy, clearance point strategy, overseas warehouse allocation strategy, platform subsidy strategy, return processing strategy, and anomaly warning strategy.

[0124] The aforementioned collaborative strategy is output in the form of structured instructions, which can be directly sent to the business execution systems of each participating entity, such as the order management system of the cross-border e-commerce platform, the transportation management system of the logistics service provider, and the warehouse management system of the overseas fulfillment entity. Each system then executes specific logistics operations based on the received instructions.

[0125] In a specific example, for a batch of cross-border orders destined for a certain country, the system generates a complete collaborative strategy by aggregating the decision-making actions output by various agents: the platform agent outputs order allocation and subsidy plans, the logistics agent outputs logistics routes and clearance point plans, and the overseas fulfillment agent outputs overseas warehouse allocation and return processing plans. After integrating these decision-making actions, the system forms a complete collaborative strategy encompassing order allocation, logistics routes, clearance point plans, overseas warehouse allocation, subsidy plans, and return processing plans, which is then distributed to the business execution systems of each participating entity for execution. For orders where changes in customs clearance policies may affect normal fulfillment, the system simultaneously generates alternative customs clearance plans as a supplement, based on an anomaly warning strategy.

[0126] After the aforementioned collaborative strategy is issued and implemented, the system obtains actual performance feedback information. This feedback includes one or more of the following: delivery time, customs clearance results, delays, returns, complaints, subsidy payments, logistics costs, inventory changes, and actual benefits for all three parties.

[0127] Specifically, the system obtains actual performance results through operational data returned from the business execution systems of each participating entity: it obtains the actual receipt time, customs clearance results, whether delays occurred, and the duration of delays through the logistics service provider's transportation management system; it obtains the return processing results, actual inventory consumption, and replenishment status through the overseas fulfillment entity's warehouse management system; it obtains user complaint records and actual subsidy expenditures through the platform's order management system; and it obtains logistics costs and actual revenue data for all three parties through the financial systems of each participating entity.

[0128] Based on the aforementioned actual performance feedback information, the system adaptively updates the decision-making model, specifically including: 1) storing the input data, intermediate results, and actual performance feedback of this decision as new historical samples in the sample library, for subsequent fine-tuning of instructions or expansion of context learning samples for the role-based large language model and the domain expert large language model, so that the large language model can refer to historical experience when dealing with similar decision-making scenarios; 2) updating the customs clearance risk parameters, transportation capacity credibility parameters, and return parameters in the three-party imperfect information game model according to the deviation between the actual customs clearance results and the predicted customs clearance risks, the deviation between the actual delays and the capacity assessment, etc. Risk threshold parameters enable the dynamic perception of the game model to more accurately reflect the characteristics of the real business environment; 3) Based on the deviation between the actual signing time and the expected time, changes in the complaint rate, and the actual distribution of the three parties' benefits, adjust the weight coefficients of the private benefit reward channel, collaborative benefit reward channel, game constraint penalty channel, and teacher guidance reward channel in the multi-agent reinforcement learning model, so that the reward signal can more accurately reflect the real business goal orientation; 4) Based on the actual benefits and actual risk-bearing of each participating entity, update the weight parameters of each factor in the collaborative reward allocation coefficient, so that the credit allocation mechanism is more fair and reasonable.

[0129] The aforementioned update operation is triggered after each decision is executed, and the updated parameters are used for processing in the corresponding stage of the next decision-making process. Through the closed-loop update driven by the above performance feedback, the system continuously optimizes itself during continuous operation. For example, the knowledge base of the large language model is constantly expanded, the risk parameters of the game model become increasingly accurate, the reward signals of the reinforcement learning model become closer to actual business, and the credit allocation mechanism becomes increasingly fair and reasonable. This enables the cross-border e-commerce logistics collaborative decision-making system to achieve adaptive evolution and continuous optimization in a real business environment. The above updates can be executed in full or in part according to the system configuration, and the triggering frequency of each update item can be flexibly adjusted according to actual business needs. For example, game constraint parameters and reward weights can be updated once every N decision cycles using a sliding window method, rather than triggering a full update for every decision, thereby achieving a balance between system performance and update frequency.

[0130] Through the closed-loop update of the above-mentioned collaborative strategy generation, issuance and execution, and performance feedback, the system realizes a complete closed loop from data acquisition, LLM enhancement processing, game constraint generation, multi-agent reasoning, collaborative strategy output to actual execution feedback. Each participating entity executes specific logistics operations according to the received collaborative strategy, and the system continuously optimizes based on the execution results, thereby realizing the dynamic adaptive evolution of cross-border e-commerce logistics collaborative decision-making.

[0131] Example 2

[0132] like Figure 6 As shown, a cross-border logistics decision-making system 100 based on a large-scale model and constrained by a three-party game includes:

[0133] The data acquisition module 101 is used to acquire private observation information corresponding to multiple participating entities in cross-border e-commerce logistics decision-making, as well as external dynamic environment data related to cross-border logistics operations. Among them, the participating entities include at least cross-border e-commerce platforms, logistics service providers and overseas fulfillment entities.

[0134] The enhanced processing module 102 is used to process private observation information and external dynamic environmental data using a large language model to generate quality-verified structured expert knowledge, which includes at least a public belief state used to characterize the shared environmental cognition of the three parties.

[0135] The game generation module 103 is used to construct a three-party imperfect information game model based on private observation information and public belief state, and to generate three-party game constraints based on the game model to constrain the decision-making behavior of each participating subject.

[0136] The constraint mapping module 104 is used to extract hard boundary constraints from the constraints of the three-party game and map the hard boundary constraints into action masks.

[0137] The reinforcement learning module 105 is used to input private observation information and public belief state into a pre-trained multi-agent reinforcement learning model. The multi-agent reinforcement learning model contains agents that correspond one-to-one with each participating subject. Each agent generates and outputs decision actions under the constraint of action mask.

[0138] The strategy output module 106 is used to generate collaborative strategies for executing cross-border e-commerce logistics operations based on the decision-making actions of each participating entity.

[0139] This embodiment is used to implement the method provided in the above embodiments and has the corresponding beneficial effects of the method. Technical details not described in detail in this embodiment can be found in the methods provided in all the foregoing embodiments of this invention.

[0140] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A cross-border logistics decision-making method based on a large-scale model and constrained by a three-party game, characterized in that, Includes the following steps: The system acquires private observation information corresponding to multiple participating entities in cross-border e-commerce logistics decision-making, as well as external dynamic environmental data related to cross-border logistics operations. The participating entities include at least cross-border e-commerce platforms, logistics service providers, and overseas fulfillment entities. The private observation information and external dynamic environment data are processed using a large language model to generate quality-verified structured expert knowledge. The structured expert knowledge includes at least a public belief state used to characterize the shared environmental cognition of the three parties. Based on the private observation information and the state of public beliefs, a three-party imperfect information game model is constructed, and three-party game constraints are generated based on the game model to constrain the decision-making behavior of each participating entity. Extract hard boundary constraints from the three-party game constraints and map the hard boundary constraints to action masks; The private observation information and public belief state are input into a pre-trained multi-agent reinforcement learning model. The multi-agent reinforcement learning model contains agents that correspond one-to-one with each participating subject. Each agent generates and outputs decision actions under the constraints of the action mask. Based on the decision-making actions of each participating entity, a collaborative strategy is generated to execute cross-border e-commerce logistics operations.

2. The cross-border logistics decision-making method based on a large-scale model and constrained by a three-party game, as described in claim 1, is characterized in that... The generation of quality-verified structured expert knowledge includes: Local decision knowledge from the perspective of each subject is generated by using the role-based large language model corresponding to each participating subject. The local decision knowledge includes at least one of decision information class and output quality metric class. The decision information class includes at least one of implicit state estimation, risk label, candidate action ranking, action preference distribution and reward adjustment suggestion. The output quality metric class includes at least one of confidence and uncertainty. The local decision knowledge is evaluated using a domain expert big language model. When the quality evaluation does not meet the preset conditions, correction information is generated and returned to the corresponding role big language model to regenerate the local decision knowledge. The local decision-making knowledge obtained through the quality assessment is fused across subjects to generate a public belief state.

3. The cross-border logistics decision-making method based on a large-scale model and constrained by a three-party game, as described in claim 1, is characterized in that... When the large language model generates structured expert knowledge, a preset prompt template containing structured output fields is used to constrain the output format of the large language model. The structured output fields are used to limit the output format and include at least one of the following: field name, field type, value range, set of legal actions, and confidence requirement.

4. The cross-border logistics decision-making method based on a large-scale model and constrained by a three-party game, as described in claim 1, is characterized in that... Also includes: The global reflective large language model generates dynamic environment update information based on the external dynamic environment data. The dynamic environment update information is used to update at least one of the following: the fusion weight of the public belief state, the constraint parameters of the three-party game constraint, and the reward weight of the multi-agent reinforcement learning model.

5. The cross-border logistics decision-making method based on a large-scale model and constrained by a three-party game, as described in claim 1, is characterized in that... The hard boundary constraint is the high-risk security constraint in the three-party game constraint.

6. The cross-border logistics decision-making method based on a large-scale model and constrained by a three-party game, as described in claim 5, is characterized in that... The unilateral deviation payout constraint includes: Given that the actions of other participating entities remain unchanged, calculate the deviation incentive value between the maximum unilateral deviation gain of any participating entity and the gain of that participating entity under the current joint action; when the deviation incentive value is greater than a preset deviation threshold, perform at least one of the following on the current candidate joint action: action deweighting, action blocking, reward / penalty, or compensation parameter adjustment.

7. The cross-border logistics decision-making method based on a large-scale model and constrained by a three-party game, as described in claim 1, is characterized in that... The pre-trained multi-agent reinforcement learning model is trained using a centralized training and decentralized execution structure. During the training phase, a centralized value evaluation network is set up to evaluate the value based on each agent's private observation information, public belief state, actions of each agent, and three-party game constraint parameters. During the execution phase, each agent performs action reasoning based on its own private observation information and the public belief state.

8. The cross-border logistics decision-making method based on a large-scale model and constrained by a three-party game, as described in claim 1, is characterized in that... The structured expert knowledge also includes action preference distribution; During the training process, the incentive function of each agent in the pre-trained multi-agent reinforcement learning model includes a private reward channel, a collaborative reward channel, a game constraint penalty channel based on the soft penalty, and a teacher-guided reward channel that guides the agent's actions based on the action preference distribution. When the collaborative reward channel is allocated to each agent, the collaborative reward is distributed according to the collaborative reward allocation weight of each agent.

9. A cross-border logistics decision-making method based on a large-scale model and constrained by a three-party game, as described in claim 8, is characterized in that... The weighting of the collaborative reward allocation is determined based on the counterfactual contribution, risk-taking, and information credibility of each agent. The counterfactual contribution is calculated based on the marginal increment of each agent's actual action relative to the baseline action on the collaborative benefit. The risk-taking is calculated based on the weighted average of each agent's risk or cost borne by multiple types of risks or costs. The information credibility is calculated based on the historical prediction error of each agent's corresponding information source.

10. A cross-border logistics decision-making system based on a large-scale model and constrained by a three-party game, characterized in that, include: The data acquisition module is used to acquire private observation information corresponding to multiple participating entities in cross-border e-commerce logistics decision-making, as well as external dynamic environmental data related to cross-border logistics operations. The participating entities include at least cross-border e-commerce platforms, logistics service providers, and overseas fulfillment entities. An enhanced processing module is used to process the private observation information and external dynamic environment data using a large language model to generate quality-verified structured expert knowledge, wherein the structured expert knowledge includes at least a public belief state for characterizing the shared environmental cognition of the three parties. The game generation module is used to construct a three-party imperfect information game model based on the private observation information and the public belief state, and to generate three-party game constraints based on the game model to constrain the decision-making behavior of each participating subject. The constraint mapping module is used to extract hard boundary constraints from the constraints of the three-party game and map the hard boundary constraints into action masks. The reinforcement learning module is used to input the private observation information and public belief state into a pre-trained multi-agent reinforcement learning model. The multi-agent reinforcement learning model contains agents that correspond one-to-one with each participating subject. Each agent generates and outputs decision actions under the constraints of the action mask. The strategy output module is used to generate collaborative strategies for executing cross-border e-commerce logistics operations based on the decision-making actions of each participating entity.

Images