A Collaborative Approach for Passenger-Free Flow Based on Multi-Agent Continuous Learning

Abstract

This invention discloses a collaborative method for customs clearance without personnel based on multi-agent continuous learning, comprising the following steps: S1, scenario modeling and constraint expression, abstracting the multi-agent system, formalizing constraints and compliance judgment conditions, and establishing a task state machine consistent with the customs clearance business process; S2, multi-agent collaborative decision-making mechanism, generating alternative paths, alternative handover points, and alternative resource allocation schemes for execution; S3, self-game scenario generation and strategy learning, constructing a collaborative-adversarial hybrid training mechanism, and introducing multi-objective learning criteria consistent with customs objectives; S4, continuous learning and incremental rule adaptation; S5, cross-agent knowledge distillation and shared memory, improving group consistency and transferability. This solution features improved customs clearance throughput and timeliness, reduced queuing and congestion, reduced costs of manual rule modification and scheduling, ensured "uninterrupted chain, no delays, and no violations," and effectively supports large-scale promotion and application.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and unmanned system scheduling and control, and in particular to a collaborative method for unmanned personnel flow based on multi-agent continuous learning. Background Technology

[0002] With the advancement of smart ports, smart customs, and intelligent logistics infrastructure, customs clearance operations are facing increasingly stringent requirements for transportation timeliness, operational safety, process traceability, and regulatory compliance. Within the port area, the process of goods unloading, temporary storage, inspection, transfer, and release often involves the parallel operation and coordinated cooperation of multiple types of equipment, such as unmanned vehicles, drones, transfer cabinets, gates, or checkpoints. Compared to common park delivery or warehouse handling, customs clearance scenarios without human or goods flow are characterized by long operational chains, dense parallel operation of multiple vehicles, frequent cross-regional relays, strong constraints on passage permissions and time windows, rapid iteration of regulatory rules, and numerous unforeseen events. Insufficient scheduling and coordination capabilities can easily lead to channel congestion, failed handovers, resource competition, delays, and even the risk of failing to meet regulatory constraints, thereby affecting overall customs clearance efficiency and regulatory safety.

[0003] Existing unmanned logistics systems primarily employ three methods in customs clearance scenarios. The first method involves manually pre-setting rules and static scheduling, such as fixed priorities, fixed routes or areas, fixed handover nodes, and threshold-triggered replanning, combined with heuristic or offline path planning and resource allocation algorithms. This method is simple to implement and highly controllable, but it heavily relies on human experience and scenario assumptions, resulting in high rule maintenance costs. When regulatory rules are adjusted, checkpoint strategies change, site layouts are modified, or business volume fluctuates, existing rules often become ineffective, requiring frequent manual intervention or reconfiguration. This makes it difficult to guarantee overall optimization and real-time performance under complex conditions, and it also lacks sufficient coverage for long-tail anomalies.

[0004] The second type of method involves multi-agent strategies trained offline in a single instance. These strategies utilize historical data or simulation environments to train multi-agent learning policies for tasks such as device allocation, path selection, or local collision avoidance. While these methods may perform well in fixed environments, the scenarios presented are subject to constantly changing rules and disturbances. Training often relies on relatively stable reward designs and environmental distribution assumptions. When new regulatory rules are introduced, inspection strategies are adjusted, temporary lockdowns occur, equipment malfunctions occur, or new risk events emerge, the strategy's generalization ability is insufficient, leading to a sharp performance drop. Furthermore, adapting to these changes typically requires re-collecting data and retraining the entire system, which is time-consuming, costly, and difficult to implement quickly in real-world operations.

[0005] The third category of methods involves online adaptive or local replanning approaches, which address disturbances through rolling optimization, local search, or online fine-tuning of the strategy. While these methods can improve response speed to some extent, they are prone to getting trapped in "local optima," leading to global coordination imbalances. For example, congestion might shift from one region to another, or queuing at handover nodes might worsen. Furthermore, without effective stabilization mechanisms, online updates can lead to policy drift, causing the forgetting of existing effective coordination strategies and resulting in performance fluctuations and uncontrollable risks. In customs clearance scenarios that emphasize compliance and auditability, such instability significantly reduces system reliability.

[0006] Therefore, existing technologies generally suffer from the following shortcomings: difficulty in continuously evolving with changes in regulatory rules and the environment; insufficient systemic coordination capabilities for multi-vehicle parallel and cross-regional relay operations; weak response to long-tail risk events; and susceptibility to forgetting or requiring costly retraining during strategy updates, making it difficult to balance long-term stable operation and rapid adaptation. Therefore, there is an urgent need to provide a human-machine interface-based collaborative method based on continuous learning of multiple agents to address these issues. Summary of the Invention

[0007] To achieve the above objectives, the inventors provide a collaborative method for human-to-human flow based on multi-agent continuous learning, comprising the following steps:

[0008] S1, Scenario Modeling and Constraint Expression: The customs area infrastructure and execution vehicles are abstracted into a multi-agent system, the customs supervision rules are formalized into computational constraints and compliance judgment conditions, and a task state machine consistent with the customs clearance business process is established to describe the state transition of arrival, handover, queuing, inspection and release.

[0009] S2, Multi-agent collaborative decision-making mechanism: Under a unified constraint framework, each agent acts jointly based on local observations and shared customs status information. When disturbances occur, online replanning is triggered to generate alternative paths, alternative handover points, and alternative resource allocation schemes for execution, ensuring that the customs clearance link remains unbroken and meets compliance constraints.

[0010] S3, Self-Game Scenario Generation and Strategy Learning: Construct a collaborative-adversarial hybrid training mechanism to automatically generate multiple types of customs clearance adversarial scenarios through self-games, so as to learn collaborative strategies under different loads and rule combinations. At the same time, introduce multi-objective learning criteria consistent with customs objectives to achieve a balance between customs clearance timeliness, resource utilization, risk control and compliance.

[0011] S4, Continuous Learning and Incremental Rule Adaptation: Incremental updates are performed using a continuous learning approach, enabling the strategy to maintain stable performance on existing level-clearing modes while adapting to new rules, and supporting the injection of new knowledge in the form of change packages or templates to achieve iteration;

[0012] S5, Cross-Agent Knowledge Distillation and Shared Memory: Establish a cross-agent knowledge sharing mechanism, write collaborative experience, decision fragments and handling cases into a shared memory bank, and synchronize the collaborative patterns in the group strategy to various vehicles and node agents through knowledge distillation, so that newly added vehicles or new customs area deployments can obtain customs clearance collaboration capabilities, and improve group consistency and portability.

[0013] In a preferred embodiment of the present invention, step S1 includes the following steps:

[0014] S101 takes as input infrastructure information and the real-time status of execution vehicles. Through heterogeneous graph construction, different types of elements are mapped to nodes, and reachability relationships, operational relationships, handover relationships, queuing relationships, or service relationships are mapped to edges with types. Key attributes are encoded as node or edge features. The output is a unified heterogeneous graph representation of the customs area and its set of state vectors used for subsequent decision-making. The expression is:

[0015] ;

[0016] in, This represents a heterogeneous map of customs districts or a unified digital twin representation. Represents a set of nodes. Denotes the set of edges. Represents the type mapping function of a node. Represents the mapping function for the relation type of edges. Represents the set of node features. Represents a node The original properties, This represents the encoding function related to the node type. Represents a node The encoded state vector, Represents type mapping functions, This means that this operation should be performed on any node;

[0017] S102, Input Customs Supervision Rules and Operational Constraints Simultaneously input the current observation status of the customs area. With candidate scheduling actions Each rule is formalized into a constraint function. The hard compliance judgment and soft feasibility score are combined to output the compliance and feasibility of candidate solutions for screening, ranking, and subsequent optimization. The expression is:

[0018] ;

[0019] ;

[0020] in, Indicates the observation status of the customs area. Indicates candidate scheduling actions, Indicates the number of rule or constraint entries. Indicates the first The constraint function corresponding to each regulatory rule, when the rule is satisfied When the rules are violated , This represents a vector consisting of all constraint functions. This is an indicator function; it returns 1 if the condition is true, and 0 otherwise. This indicates a hard compliance determination; a value of 1 indicates that all rules are met, and a value of 0 indicates that there is a violation.

[0021] S103, Input customs clearance business process definition, input real-time event flow. Input context state Through the nominal transfer function Describe the process logic and use gating functions. Embed regulatory rules and resource availability into state transitions, and output the next state of the task. The expression is:

[0022] ;

[0023] in, Indicates the time of the task The process status, Indicates time The events that occurred, Indicates time The context state, Represents the nominal transfer function, Represents the gate function. Indicates the time of the task The output status.

[0024] In a preferred embodiment of the present invention, in step S101, the infrastructure information includes the customs area road network, operational zones, handover nodes, checkpoints or gates, inspection positions and / or transfer cabinets; the real-time status of the executing vehicle includes its location, battery level or range, load and / or task occupancy; in step S102, the operational constraints include access permissions, time windows and service windows, inspection priorities, sealing status, checkpoint release strategies, prohibited and / or restricted areas; the customs area observation status includes the vehicle's location, load and range, The checkpoint flow control status, queue length, inspection position occupancy and / or remaining occupancy time window; the candidate scheduling actions include equipment allocation, task assignment, path selection, handover node selection, enqueueing or dequeueing and / or triggering replanning; in step S103, the real-time event stream includes arrival at the handover point, handover completion, entry into the queue, turn for inspection, receipt of release signal and / or triggering risk warning; the context status includes whether resources are idle, whether the checkpoint is flow-controlled, access permission, sealing status and / or remaining task time window.

[0025] In a preferred embodiment of the present invention, step S2 includes the following steps:

[0026] S201: Map each type of decision-making object to an agent, uniformly define its observations, actions, and local state, and input the local observations of each agent. and shared customs area status summary By employing a centralized training and distributed execution mechanism with constraint gating, the joint action is decomposed into the local actions of each agent, and a joint scheduling decision is generated collaboratively, outputting a joint action vector. The expression is:

[0027] ;

[0028] ;

[0029] in, Indicates the number of agents. Indicates the index of the decision moment. Indicates the first Local observations of individual agents This represents the set of state vectors of the gate nodes. Represents customs area nodes At any moment The state vector, This represents the set of nodes in the customs area map. Represents the aggregation operator, Indicates a shared customs area status summary. Indicates the first Local policy functions of each agent Indicates the first The local actions output by an agent, Represents the joint action vector;

[0030] S202, Input shared customs area status summary Joint action with candidates Under a unified constraint framework, hard compliance is determined. If the rule is not met, minimal modifications are made, and the output is an executable joint action. To ensure the constraints are satisfied, the expression is:

[0031] ;

[0032] in, Represents the set of constraint functions. Indicates the first Constraint functions, This represents the feasible region projection operator. This indicates the modified executable joint action. Represents the L2 norm distance;

[0033] S203, Input Disturbance Event When the disturbance intensity exceeds the threshold, online replanning is triggered, the original execution action is partially replaced, alternative paths, alternative handover points and alternative resource allocations are generated, and the replanned joint action is output. And satisfying uniform constraints, the expression is:

[0034] ;

[0035] in, Indicates a disturbance event. This represents the disturbance intensity assessment function. Indicates the trigger threshold. Indicates an indicator function, Indicates whether to trigger replanning. Indicates the online replanning operator, This indicates joint actions following a replanning process.

[0036] In a preferred embodiment of the present invention, in step S201, the shared customs area status summary includes checkpoint flow restriction status, inspection position occupancy, key channel congestion and / or restricted / prohibited status; in step S203, the disturbance events include checkpoint congestion, inspection queue jumping, channel blocking and / or equipment failure.

[0037] In a preferred embodiment of the present invention, step S3 includes the following steps:

[0038] S301: Input the basic customs district model and regulatory rule set; set up a collaborating agent to complete customs clearance and delivery tasks; set up an adversary scenario generator to apply controllable perturbations and rule combinations; through alternating optimization by the collaborating and adversary agents, the adversary agent automatically generates customs clearance adversarial scenarios that can expose system vulnerabilities; output scenario distribution. This is used for subsequent policy learning and robustness improvement, and its expression is:

[0039] ;

[0040] in, Indicates scenario parameters, Represents the scenario parameter space, This indicates the number of iterations in the self-game. Indicates the first Round-the-round cooperation strategy, This represents the adversary's objective function, used to measure the situation parameters. The degree of pressure or failure exerted on the collaborating parties. Represents the scenario generation operator, Indicates the situational parameters The generated training scenario distribution;

[0041] S302, Input the training scenario distribution generated by the adversary. And the trajectory data obtained by collaborators within the customs area, from each round of training. Sample the scenario and run the collaborator strategy, update the collaborator strategy based on the total trajectory reward, and output the updated collaboration strategy. The expression is:

[0042] ;

[0043] in, Represents the set of collaborating strategies. Indicates the first The scenario parameters generated by the opposing side in the round-robin format. Indicates the first The distribution of scenarios used in round training This represents the interaction trajectory obtained by executing the collaborating party's strategy in a specific scenario. Indicates the total reward of the trajectory. Represents the mathematical expectation;

[0044] S303, Input the trajectory indicators obtained by the collaborating party in the scenario, and calculate the customs clearance time, equipment and resource utilization, risk control effectiveness, and compliance indicators respectively; Train the strategy through a multi-objective function to achieve a stable trade-off under different business preferences and regulatory intensities, and output the final strategy that satisfies the objective constraints, the expression of which is:

[0045] ;

[0046] in, Indicates the total reward of the trajectory. This indicates the reward component related to customs clearance timeliness. This indicates the return component related to resource utilization. This indicates the risk control-related return component. This indicates compliance and non-compliance penalties. This indicates the weight of each objective.

[0047] In a preferred embodiment of the present invention, step S4 includes the following steps:

[0048] S401: Input the newly added change information, encapsulate the change information into a rule change package, layout change package, or event template package, and use a unified parsing and mapping module to transform the rule change package, layout change package, or event template package into constraint updates and environment parameter updates for calculation. Output the incrementally updated constraint set and environment configuration to trigger subsequent continuous learning. The expression is:

[0049] ;

[0050] in, Indicates the first Version index adapted by the second incremental step Indicates the first This rule change package. Indicates the first Secondary layout change package Indicates the first Sub-event template package This indicates a change in the parsing operator. This represents the set of incremental changes obtained after analysis. This represents the set of constraints before the change. This represents the set of constraints after the change. Indicates the incremental merging operator;

[0051] S402, Input the constraint set after incremental change. Incremental data acquired during online operation or simulation Simultaneously, constraint playback data is extracted from historical memory. To cover existing clearance patterns, the old and new data are reweighted, making the learning process focus more on the regions affected by rule changes and long-tail risk event samples, and outputting updated policy parameters. The expression is:

[0052] ;

[0053] in, Indicates the strategy parameters, Indicates the first The policy parameters after the next incremental update. Indicates the observation status of the customs area. Indicates an action, Indicates in constraints Define the training loss. Represents the sample weight function;

[0054] S403, Input the policy parameters from the previous version. With the current incremental training objective, a parameter regularization mechanism is used to constrain the new parameters to not deviate from the effective solution space of the old parameters in key dimensions. This allows the system to adapt to new rules while maintaining stable performance on the existing level-clearing pattern, and outputs update policy parameters with forgetting suppression. The expression is:

[0055] ;

[0056] in, Indicates the first The primary training objective is updated incrementally. This indicates the strategy parameters of the previous version. This represents the L2 norm.

[0057] In a preferred embodiment of the present invention, step S5 includes the following steps:

[0058] S501: Input trajectory fragments generated during multi-agent collaborative execution; perform quality assessment and compression encoding on the trajectory fragments; filter out collaborative experience, decision fragments, and handling cases, and write them into the shared memory; during policy execution or learning, perform a similarity search on the shared memory based on the current shared customs area state summary and task context; output reusable collaborative evidence or experience fragments for each agent to use for decision-making reference or training replay, as expressed in the following expression:

[0059] ;

[0060] ;

[0061] in, Indicates a shared memory bank. Indicates the first These candidate experience fragments are derived from key decision windows extracted from collaborative trajectories. This represents an empirical coding function used to compress empirical fragments into vectors or structured representations. This represents the experience quality score. Indicates experience tags. The write operator writes the encoded result and metadata to the shared memory and returns a shared memory unit. , Indicates a shared memory unit. Indicates time Shared customs area status summary This represents a retrieval operator used to return a set of memories related to the current context from a shared memory bank. This represents the relevant memory set retrieved.

[0062] S502, Input group collaboration strategy as teacher strategy And the student strategy set of various vehicles and node intelligent agents. Shared customs district status summary within the same customs district The teacher's strategy generates a joint action distribution and extracts effective cooperative patterns from it. Student strategies learn to approximate the teacher's decision distribution based on their own observations. Through distillation training, different types of agents acquire consistent cooperative preferences, enabling newly added vehicles or deployments in new checkpoints to inherit the cooperative clearance capabilities. The output is the set of student strategy parameters after distillation, expressed as:

[0063]

[0064] in, Represent the group teacher strategy, and output the joint action distribution under the condition of shared gate state summary. Represents intelligent agents individual student strategies Indicates student strategy parameters, This represents the optimal parameters after distillation. Indicates time Shared customs area status summary Represents intelligent agents Local observations, This represents the Kullback-Leibler divergence, used to measure the difference in how the student distribution approximates the teacher distribution. Represents an intelligent agent.

[0065] Unlike existing technologies, the above technical solution achieves the following beneficial effects:

[0066] (1) This method provides a multi-agent collaborative decision-making method for customs port operations without passenger flow, which is used for joint scheduling and collaborative control of unmanned vehicles, drones or transfer containers and customs checkpoint equipment in the process of "consolidation - transfer - inspection - checkpoint release", which can effectively improve customs clearance throughput and timeliness, and reduce queuing and congestion.

[0067] (2) This method provides an adaptive mechanism that adapts to the dynamic adjustment of customs supervision rules, access rights, inspection priorities and checkpoint strategies, so that the system can still maintain compliant operation and stable operation in the event of policy updates, temporary inspections, blockade rerouting, site layout changes, etc., and effectively reduce the cost of manual rule modification and manual scheduling.

[0068] (3) This method provides a robust collaborative strategy for updating targets for long-tail risks and emergencies at ports. It can achieve rapid replanning and cross-regional relay in situations such as equipment failure, channel congestion, risk warning triggering secondary inspection, and abnormal cargo handling, ensuring "no chain break, no delay, and no violation".

[0069] (4) This method provides a continuous learning and group synchronous evolution target suitable for the migration and deployment of multiple customs districts or multiple stations. It enables the system to continue to update without forgetting the effective strategies in the past after adding new regulatory requirements, introducing new port equipment, and adjusting the operation process. It also achieves consistent improvement of multi-vehicle capabilities through cross-agent knowledge sharing, effectively supporting large-scale promotion and application. Attached Figure Description

[0070] Figure 1 The flowchart is for a specific implementation method. Detailed Implementation

[0071] To explain in detail the technical content, structural features, objectives, and effects of the technical solution, the following description is provided in conjunction with specific embodiments and accompanying drawings.

[0072] The motivation behind this invention is to construct a decision-making system capable of maintaining collaborative efficiency and compliance under the dynamic and highly constrained environment of unmanned logistics in customs clearance, amidst rule updates, layout changes, and disturbance events. To this end, this invention proposes to uniformly model unmanned vehicles, drones, transfer lockers, and checkpoints as a multi-agent decision-making system. It generates diverse collaborative and adversarial scenarios through self-games and other methods to learn key decision-making strategies, and introduces a continuous learning mechanism to ensure the system can continuously update itself when new rules or events occur, minimizing forgetting. Simultaneously, it leverages cross-agent knowledge sharing to enhance group collaboration capabilities, thereby improving the scalability, portability, and operational stability of unmanned delivery within customs areas under complex and changing conditions.

[0073] like Figure 1 As shown, this embodiment provides a human-free flow-based collaborative method based on multi-agent continuous learning, including the following steps:

[0074] S1, Scenario Modeling and Constraint Representation: The infrastructure such as the customs area road network, operation zones, handover nodes, checkpoints or gates, inspection positions, and transfer cabinets, as well as execution vehicles such as unmanned vehicles and drones, are abstracted into a multi-agent system; customs supervision rules, such as: passage permissions, time windows, inspection priorities, sealing status, checkpoint release strategies, prohibited or restricted areas, are formalized into computable constraints and compliance judgment conditions, and a task state machine consistent with the customs clearance business process is established to describe the state transitions of key links such as arrival, handover, queuing, inspection, and release;

[0075] S2, Multi-Agent Collaborative Decision-Making Mechanism: Under a unified constraint framework, each agent takes joint actions based on local observations and shared customs status information, such as making decisions on equipment or capacity allocation, task assignment, route selection, handover node selection, and arrival timing; when disturbances such as checkpoint congestion, queue jumping, channel closure, or equipment failure occur, online replanning is triggered to generate executable alternative routes, alternative handover points, and alternative resource allocation schemes to ensure that the customs clearance link remains unbroken and meets compliance constraints;

[0076] S3, Self-Game Scenario Generation and Strategy Learning: Constructing a collaborative-adversarial hybrid training mechanism, which automatically generates multiple types of customs clearance adversarial scenarios through self-games, such as peak congestion, increased proportion of temporary inspections, checkpoint flow control, severe weather restricting drones, and risk warnings triggering secondary inspections, to learn collaborative strategies under different loads and rule combinations; at the same time, introducing multi-objective learning criteria consistent with customs objectives, so that the strategy achieves a balance between customs clearance timeliness, resource utilization, risk control and compliance;

[0077] S4, Continuous Learning and Incremental Rule Adaptation: When new regulatory rules, site layout modifications, checkpoint strategy updates, or new long-tail risk events occur, the system adopts a continuous learning approach for incremental updates. Through mechanisms such as constraint replay, parameter regularization, and sample reweighting, catastrophic forgetting is suppressed, enabling the strategy to adapt to new rules while maintaining stable performance on existing clearance modes. It also supports the rapid injection of new knowledge in the form of "rule change packages, layout change packages, or event templates" to achieve low-cost iteration.

[0078] S5, Cross-Agent Knowledge Distillation and Shared Memory: Establish a cross-agent knowledge sharing mechanism to write high-quality collaborative experiences, key decision fragments, and violation or risk handling cases into a shared memory bank. Through knowledge distillation, effective collaborative patterns in the group strategy are synchronized to various vehicle and node agents, enabling newly added vehicles or new customs area deployments to quickly acquire customs clearance collaboration capabilities, thereby improving group consistency and portability.

[0079] The specific implementation of step S1 in the above embodiments includes the following steps:

[0080] S101 inputs information on infrastructure such as the customs district's road network, operational zones, handover nodes, checkpoints or gates, inspection locations, and transfer lockers, as well as the real-time status of vehicles such as unmanned vehicles or drones, including their location, battery level or remaining range, payload, and task occupancy. Through heterogeneous graph construction, different types of elements are mapped to nodes, and reachability relationships, operational relationships, handover relationships, queuing or service relationships are mapped to edges with types. Key attributes are encoded as node or edge features. The output is a unified heterogeneous graph representation of the customs district and its set of state vectors that can be used for subsequent decision-making, expressed as:

[0081] ;

[0082] in, Representing heterogeneous maps or unified digital twin representations of customs districts; This represents a set of nodes, including road nodes, work area nodes, handover nodes, checkpoint or gate nodes, inspection station nodes, transfer cabinet nodes, and vehicle nodes. This represents a set of edges, used to represent relationships such as road connectivity, cross-regional reachability, entry checkpoint, entry queue, handover, loading / unloading, or storage / retrieval. Represents the type mapping function for nodes; Represents the mapping function for the relation type of edges; Represents the set of node features; Represents a node Original attributes, such as: road length or speed limit, checkpoint flow restriction threshold, inspection station service rate, transfer cabinet capacity, vehicle battery power or load or mission status, etc. This represents the encoding function associated with the node type, used to map raw attributes to a uniform dimension vector; Represents a node The encoded state vector, where ϕ represents the type mapping function. This means that this operation is performed on any node.

[0083] S102, Input Customs Supervision Rules and Operational Constraints This includes access permissions, time windows and service windows, inspection priorities, sealing status, checkpoint release policies, and restricted and prohibited areas; it also requires inputting the current customs area observation status. With candidate scheduling actions This includes equipment allocation, task assignment, path selection, handover point selection, whether to enqueue, and whether to trigger replanning; each rule is formalized as a constraint function. The hard compliance determination and soft feasibility score are combined to obtain the compliance and feasibility scores of candidate solutions, which are then used for screening, ranking, and subsequent optimization. The formal expression is as follows:

[0084] ;

[0085] ;

[0086] in, It indicates the status of customs area observation, including vehicle position and load and endurance, checkpoint flow control status, queue length, inspection position occupancy, and remaining occupancy time window; This indicates candidate scheduling actions, including device allocation, task assignment, path selection, handover node selection, enqueueing or dequeueing, and triggering replanning. Indicates the number of rule or constraint entries; Indicates the first The constraint function corresponding to each regulatory rule, when the rule is satisfied When the rules are violated ; This represents a vector consisting of all constraint functions; This indicates an indicator function; it returns 1 if the condition is true, and 0 otherwise. This indicates a hard compliance determination; a value of 1 indicates that all rules are met, while a value of 0 indicates that a violation exists.

[0087] S103, Input the customs clearance business process definition, including the state set and nominal transition relationship; Input the real-time event flow. This includes arriving at the handover point, completing the handover, entering the queue, being checked, receiving the release signal, and triggering a risk warning; inputting the context status. This includes whether resources are available, whether checkpoints are flow-limited, access permissions, block status, and remaining task time window; through a nominal transfer function. Describe the process logic and use gating functions. Embed regulatory rules and resource availability into state transitions; output the next state of the task. And ensure that each transfer is auditable, the expression is:

[0088] ;

[0089] in, Indicates the time of the task The process status; Indicates time The events that occurred; Indicates time The context state; Indicates the nominal transfer function; Represents the gate function; Indicates the time of the task The output status.

[0090] The specific implementation of step S2 in the above embodiments includes the following steps:

[0091] S201 represents the set of decision-making objects within the control area, including execution vehicles and key operational nodes. Each type of decision-making object is mapped to an agent, with its observations, actions, and local state uniformly defined. Input is the local observations of each agent. and shared customs area status summary , The system includes local observations, such as vehicle location, payload, range, current task, and neighboring queue information. The shared customs area state summary includes checkpoint flow control status, inspection station occupancy, key channel congestion, and restricted / prohibited status. Through a centralized training and distributed execution mechanism with constraint gating, the joint action is decomposed into local actions of each agent and collaboratively generates a joint scheduling decision. The output is a joint action vector. It includes equipment or capacity allocation, task assignment, route selection, handover node selection, and arrival sequence, and its expression is:

[0092] ;

[0093] ;

[0094] in, This indicates the number of intelligent agents, which correspond to unmanned vehicles, drones, transfer cabinet node agents, checkpoint agents, or inspection station agents. Indicates the index of the decision-making time; Indicates the first Local observations of individual agents; Represents the set of state vectors of gate nodes; Represents customs area nodes At any moment The state vector; The set of nodes representing the customs area map; Represents an aggregation operator, used to extract from... Calculate the shared gate area state summary ; Indicates a summary of the shared customs area status; Indicates the first Local policy functions of each agent; Indicates the first Local actions output by an agent; This represents the joint action vector.

[0095] S202, Input shared customs area status summary Joint action with candidates Perform hard compliance checks within a unified constraint framework; if non-compliance is found, implement minimal modifications; output executable joint actions. To ensure compliance with constraints such as access permissions, time windows, checkpoint policies, and traffic restrictions, the expression is:

[0096] ;

[0097] in, Represents the set of constraint functions; Indicates the first Constraint functions; Represents the feasible region projection operator; This indicates the modified executable joint action; This represents the L2 norm distance.

[0098] S203, Input Disturbance Event Examples of disturbances include checkpoint congestion, queue jumping during inspections, channel closures, and equipment malfunctions. When the disturbance intensity exceeds a threshold, online replanning is triggered, partially replacing the original executable actions, generating alternative paths, alternative handover points, and alternative resource allocations; the replanned joint actions are then output. And still satisfying the uniform constraints, the expression is:

[0099] ;

[0100] in, Indicates a disturbance event; This represents the disturbance intensity assessment function; Indicates the trigger threshold; Indicates an indicator function; Indicates whether replanning has been triggered; Indicates the online replanning operator; This indicates the joint action following the replanning; the meanings of the other symbols are the same as before.

[0101] The specific implementation of step S3 in the above embodiments includes the following steps:

[0102] S301: Input the basic customs district model and regulatory rule set; set up a collaborating agent to complete customs clearance and delivery tasks; set up an adversary scenario generator to apply controllable perturbations and rule combinations; through alternating optimization between the collaborating and adversary agents, the adversary agent automatically generates customs clearance adversarial scenarios that maximize the exposure of system vulnerabilities; output scenario distribution. This is used for subsequent policy learning and robustness improvement, and its expression is:

[0103] ;

[0104] in, These parameters represent scenario parameters, including congestion intensity, temporary inspection ratio, checkpoint flow restriction threshold, drone no-fly or no-fly constraints, and probability of triggering risk warnings. Represents the scenario parameter space; Indicates the number of iterations in the self-game; Indicates the first Round-robin cooperation strategy; This represents the adversary's objective function, used to measure the situation parameters. The degree of pressure or failure exerted on the collaborating parties; Represents the scenario generation operator; Indicates the situational parameters The generated training scenario distribution.

[0105] S302, Input the training scenario distribution generated by the adversary. And the trajectory data obtained by collaborators within the customs area; from each round of training... Sample scenarios and run collaborator policies, updating the collaborator policies based on the total trajectory reward, enabling them to learn stable and executable collaborative decisions under different loads and rule combinations; output the updated collaborative policies. The expression is:

[0106] ;

[0107] in, Represents the set of collaborating strategies; Indicates the first Scenario parameters generated by the opposing side in a round; Indicates the first The distribution of scenarios used in round training; This represents the interaction trajectory obtained by the execution of the collaborating party's strategy in a certain scenario; It represents the total trajectory reward, used to measure the overall effectiveness of customs clearance timeliness, resource consumption, violation penalties, and risk event handling; Indicates the training round index. It represents the mathematical expectation.

[0108] S303, Input the trajectory indicators obtained by the collaborating party in the scenario, and calculate the customs clearance time, equipment and resource utilization, risk control effectiveness, and compliance indicators respectively; Train the strategy through a weighted or adjustable multi-objective function to achieve a stable trade-off under different business preferences and regulatory intensities; Output the final strategy that satisfies the multi-objective constraints, with the expression:

[0109] ;

[0110] in, Indicates the total reward of the trajectory; This represents a negative value for the reward component related to customs clearance time, such as completion time or average delay. This represents the return component related to resource utilization, such as throughput per unit time or equipment utilization. This indicates the reward component related to risk control, such as the reward for successfully transferring to a secondary inspection and completing the disposal after a risk warning is issued; This indicates compliance and violation penalties, such as violations of permissions, time windows, restrictions on movement, or checkpoint policies. This indicates the weight of each objective.

[0111] The specific implementation of step S4 in the above embodiments includes the following steps:

[0112] S401: Input information on changes such as new regulatory rules, site layout modifications, checkpoint strategy updates, or new long-tail risk events; encapsulate the change information into rule change packages, layout change packages, or event template packages; transform the change packages into computable constraint updates and environmental parameter updates through a unified parsing and mapping module; output the incrementally updated constraint set and environmental configuration to trigger subsequent continuous learning, expressed as:

[0113] ;

[0114] in, Indicates the first Version index adapted by incremental step; Indicates the first This rule change package; Indicates the first Secondary layout change package; Indicates the first Sub-event template package; Indicates the change of parsing operator; This represents the set of incremental changes obtained after analysis; Represents the set of constraints before the change; Represents the set of constraints after the change; This represents the incremental merging operator.

[0115] S402, Input the constraint set after incremental change. Incremental data acquired during online operation or simulation Simultaneously, constraint playback data is extracted from historical memory. To cover existing clearance patterns; to reweight old and new data samples, making the learning process focus more on the areas affected by rule changes and long-tail risk event samples; and to output updated policy parameters. The expression is:

[0116] ;

[0117] in, Indicates the strategy parameters; Indicates the first The strategy parameters after the next incremental update; Indicates the observation status of the customs area; Indicates an action; Indicates in constraints Define the training loss; This represents the sample weighting function.

[0118] S403, Input the policy parameters from the previous version. The training objective is to adapt to the current incremental training goals; by using parameter regularization, the new parameters are constrained not to deviate from the effective solution space of the old parameters in key dimensions, thereby maintaining stable performance on the existing level-clearing mode while adapting to the new rules; the updated strategy parameters with forgetting suppression are output. The expression is:

[0119] ;

[0120] in, Indicates the first The main training objective is updated incrementally. Indicates the strategy parameters of the previous version; This represents the L2 norm.

[0121] The specific implementation of step S5 in the above embodiments includes the following steps:

[0122] S501, Input the trajectory fragments generated during the multi-agent collaborative execution process. These trajectory fragments include local observations at key decision moments, shared checkpoint state summaries, action sequences, clearance results, and violation or risk handling labels. Perform quality assessment and compression encoding on the trajectory fragments, selecting high-quality collaborative experiences, key decision fragments, and violation or risk handling cases, and write them into the shared memory. During policy execution or learning, perform a similarity search on the shared memory based on the current shared checkpoint state summary and task context, outputting reusable collaborative evidence or experience fragments for decision-making reference or training replay by each agent. The expression is:

[0123] ;

[0124] ;

[0125] in, Indicates a shared memory; Indicates the first The candidate experience fragments are derived from key decision windows extracted from the collaborative trajectory; This represents an empirical coding function used to compress empirical fragments into vectors or structured representations. The experience quality score reflects the overall performance of customs clearance timeliness, compliance, risk management effectiveness, and resource expenditure. The experience tags include violation type, risk event type, checkpoint congestion level, and inspection trigger mode; The write operator writes the encoded result and metadata to the shared memory and returns the memory unit. ; Indicates a shared memory unit; Indicates time Shared customs area status summary; This represents a retrieval operator used to return the set of memories most relevant to the current context from a shared memory bank; This represents the set of relevant memories retrieved.

[0126] S502, Input group collaboration strategy as teacher strategy And student policy sets for various vehicles and node agents. Shared customs district status summary within the same customs district The teacher's strategy generates a joint action distribution and extracts effective cooperative patterns from it. Student strategies learn to approximate the teacher's decision distribution based on their own observations. Through distillation training, different types of agents acquire consistent cooperative preferences, enabling newly added vehicles or those deployed in new checkpoints to quickly inherit cooperative clearance capabilities. The output is the set of student policy parameters after distillation, expressed as:

[0127] ;

[0128] in, Represent the group teacher strategy and output the joint action distribution under the condition of shared gate state summary; Indicates the first Student strategies for individual agents; Indicates student strategy parameters, Indicates the optimal parameters after distillation; Indicates time Shared customs area status summary; Represents intelligent agents Local observations; This represents the Kullback–Leibler divergence, used to measure the difference in how the student distribution approximates the teacher distribution; Represents an intelligent agent.

[0129] To demonstrate the effectiveness of the above embodiments, the following embodiments were verified using a public dataset. Specifically, the public dataset used in this experiment is Amazon LMRRC, which comes from real last-mile delivery operations and contains 9,184 historical delivery routes collected from five large metropolitan areas. It provides route-level, station-level, and package-level features, such as route attributes, station service information, and package statistics. It also provides the travel time between stations and the actual stop order, which facilitates the evaluation of the feasibility, cost, and timeliness of route planning and scheduling strategies under real operational data distribution. To meet the requirements of public use and privacy protection, the dataset does not contain personally identifiable information, route and package identifiers have been re-randomized, and related location data has been anonymized or obfuscated.

[0130] The experimental results in Table 1 show that, compared with local optimization, heuristic routing, strategy optimization, and tabu search, this invention achieves the highest constraint satisfaction rate (99.0%) and on-time rate (95.4%) under the condition of deployable computation time (0.95s). Furthermore, it reduces the average route cost to 0.88 and the number of devices used to 10.4. This demonstrates that in customs scenarios with strong constraints such as passage permissions, checkpoint strategies, and time windows, and under the requirement of cross-regional relay collaboration, it can output a more stable and executable scheduling scheme and effectively reduce queuing and waiting costs. The invention avoids the risk of timeouts; at the same time, compared with tabu search, the invention achieves higher feasibility and timeliness, as well as lower cost and resource consumption, while significantly reducing the solution time (3.60s→0.95s). This shows that it not only improves the quality of the solution, but is also more suitable for online replanning under high-frequency rolling scheduling and sudden disturbances at ports, such as congestion, temporary inspections, and equipment failures. In summary, this invention verifies that it can take into account "compliance and controllability, timeliness and efficiency, resource saving and real-time response" in customs clearance operations with continuously changing rules and environments, forming a scalable collaborative decision-making capability for actual port deployment.

[0131] Table 1 Comparison of experimental results

[0132] method Feasibility (constraint satisfaction rate) Average route cost (distance / time) On-time performance (time window / service window) Number of devices used Average calculation time (seconds) Local optimization algorithm 92.4% 0.97 85.1% 12.3 0.80 Heuristic route algorithm 88.6% 1.00 80.4% 13.1 0.18 Policy optimization algorithm 94.1% 0.95 87.3% 12.0 0.42 Tabu search algorithm 97.2% 0.92 90.2% 11.2 3.60 This method 99.0% 0.88 95.4% 10.4 0.95

[0133] It should be noted that although the above embodiments have been described herein, this does not limit the scope of patent protection of the present invention. Therefore, any changes and modifications made to the embodiments described herein based on the innovative concept of the present invention, or equivalent structural or procedural transformations made using the content of the present invention's specification and drawings, directly or indirectly applying the above technical solutions to other related technical fields, are all included within the scope of patent protection of the present invention.

Claims

1. A collaborative method for people-free flow based on multi-agent continuous learning, characterized in that, Including the following steps: S1, Scenario Modeling and Constraint Expression: The customs area infrastructure and execution vehicles are abstracted into a multi-agent system, the customs supervision rules are formalized into computational constraints and compliance judgment conditions, and a task state machine consistent with the customs clearance business process is established to describe the state transition of arrival, handover, queuing, inspection and release. S2, Multi-agent collaborative decision-making mechanism: Under a unified constraint framework, each agent acts jointly based on local observations and shared customs status information. When disturbances occur, online replanning is triggered to generate alternative paths, alternative handover points, and alternative resource allocation schemes for execution, ensuring that the customs clearance link remains unbroken and meets compliance constraints. S3, Self-Game Scenario Generation and Strategy Learning: Construct a collaborative-adversarial hybrid training mechanism to automatically generate multiple types of customs clearance adversarial scenarios through self-games, so as to learn collaborative strategies under different loads and rule combinations. At the same time, introduce multi-objective learning criteria consistent with customs objectives to achieve a balance between customs clearance timeliness, resource utilization, risk control and compliance. S4, Continuous Learning and Incremental Rule Adaptation: Incremental updates are performed using a continuous learning approach, enabling the strategy to maintain stable performance on existing level-clearing modes while adapting to new rules, and supporting the injection of new knowledge in the form of change packages or templates to achieve iteration; S5, Cross-Agent Knowledge Distillation and Shared Memory: Establish a cross-agent knowledge sharing mechanism, write collaborative experience, decision fragments and handling cases into the shared memory bank, and synchronize the collaborative patterns in the group strategy to various vehicles and node agents through knowledge distillation, so that newly added vehicles or new customs area deployments can obtain customs clearance collaboration capabilities, and improve group consistency and portability. Step S3 includes the following steps: S301: Input the basic customs district model and regulatory rule set; set up a collaborating agent to complete customs clearance and delivery tasks; set up an adversary scenario generator to apply controllable perturbations and rule combinations; through alternating optimization by the collaborating and adversary agents, the adversary agent automatically generates customs clearance adversarial scenarios that can expose system vulnerabilities; output scenario distribution. This is used for subsequent policy learning and robustness improvement, and its expression is: ； in, Indicates scenario parameters, Represents the scenario parameter space, This indicates the number of iterations in the self-game. Indicates the first Round-the-round cooperation strategy, This represents the adversary's objective function, used to measure the situation parameters. The degree of pressure or failure exerted on the collaborating parties. Represents the scenario generation operator, Indicates the situational parameters The generated training scenario distribution; S302, Input the training scenario distribution generated by the adversary. And the trajectory data obtained by collaborators within the customs area, from each round of training. Sample the scenario and run the collaborator strategy, update the collaborator strategy based on the total trajectory reward, and output the updated collaboration strategy. The expression is: ； in, Represents the set of collaborating strategies. Indicates the first The scenario parameters generated by the opposing side in the round-robin format. Indicates the first The distribution of scenarios used in round training This represents the interaction trajectory obtained by executing the collaborating party's strategy in a specific scenario. Indicates the total reward of the trajectory. Represents the mathematical expectation; S303, Input the trajectory indicators obtained by the collaborating party in the scenario, and calculate the customs clearance time, equipment and resource utilization, risk control effectiveness, and compliance indicators respectively; Train the strategy through a multi-objective function to achieve a stable trade-off under different business preferences and regulatory intensities, and output the final strategy that satisfies the objective constraints, the expression of which is: ； in, Indicates the total reward of the trajectory. This indicates the reward component related to customs clearance timeliness. This indicates the return component related to resource utilization. This indicates the risk control-related return component. This indicates the penalties for compliance and non-compliance. This represents the weight of each objective.

2. The collaborative method for people-free flow based on multi-agent continuous learning according to claim 1, characterized in that, Step S1 includes the following steps: S101 takes as input infrastructure information and the real-time status of execution vehicles. Through heterogeneous graph construction, different types of elements are mapped to nodes, and reachability relationships, operational relationships, handover relationships, queuing relationships, or service relationships are mapped to edges with types. Key attributes are encoded as node or edge features. The output is a unified heterogeneous graph representation of the customs area and its set of state vectors used for subsequent decision-making. The expression is: ； in, This represents a heterogeneous map of customs districts or a unified digital twin representation. Represents a set of nodes. Denotes the set of edges. Represents the type mapping function of a node. Represents the mapping function for the relational type of edges. Represents the set of node features. Represents a node The original properties, This represents the encoding function related to the node type. Represents a node The encoded state vector, Represents type mapping functions, This means that such an operation should be performed on any node; S102, Input Customs Supervision Rules and Operational Constraints Simultaneously input the current observation status of the customs area. With candidate scheduling actions Each rule is formalized into a constraint function. The hard compliance judgment and soft feasibility score are combined to output the compliance and feasibility of candidate solutions for screening, ranking, and subsequent optimization. The expression is: ； ； in, Indicates the observation status of the customs area. Indicates candidate scheduling actions, Indicates the number of rule or constraint entries. Indicates the first The constraint function corresponding to each regulatory rule, when the rule is satisfied When the rules are violated , This represents a vector consisting of all constraint functions. This is an indicator function; it returns 1 if the condition is true, and 0 otherwise. This indicates a hard compliance determination; a value of 1 indicates that all rules are met, and a value of 0 indicates that there is a violation. S103, Input customs clearance business process definition, input real-time event flow. Input context state Through the nominal transfer function Describe the process logic and use gating functions. Embed regulatory rules and resource availability into state transitions, and output the next state of the task. The expression is: ； in, Indicates the time of the task The process status, Indicates time The events that occurred, Indicates time The context state, Represents the nominal transfer function, Represents the gate function. Indicates the time of the task The output status.

3. The multi-agent continuous learning-based human-machine interface collaborative method according to claim 2, characterized in that: In step S101, the infrastructure information includes the customs area road network, operation zone division, handover nodes, checkpoints or gates, inspection positions and / or transfer cabinets; the real-time status of the execution vehicle includes location, battery level or remaining range, load and / or task occupancy. In step S102, the operational constraints include access permissions, time windows and service windows, inspection priorities, sealing status, checkpoint release policies, and prohibited and / or restricted areas; the customs area observation status includes vehicle location, load and endurance, checkpoint flow control status, queue length, inspection position occupancy and / or remaining occupancy time window; the candidate scheduling actions include equipment allocation, task assignment, path selection, handover node selection, enqueueing or dequeueing, and / or triggering replanning. In step S103, the real-time event stream includes arrival at the handover point, handover completion, entry into the queue, turn for inspection, receipt of a release signal, and / or triggering a risk warning; the context state includes whether resources are idle, whether the checkpoint is flow-limited, access permissions, sealing status, and / or remaining task time window.

4. The multi-agent continuous learning-based collaborative method for people-free flow of goods according to claim 2, characterized in that, Step S2 includes the following steps: S201: Map each type of decision-making object to an agent, uniformly define its observations, actions, and local state, and input the local observations of each agent. and shared customs area status summary By employing a centralized training and distributed execution mechanism with constraint gating, the joint action is decomposed into the local actions of each agent, and a joint scheduling decision is generated collaboratively, outputting a joint action vector. The expression is: ； ； in, Indicates the number of agents. Indicates the index of the decision moment. Indicates the first Local observations of individual agents This represents the set of state vectors of the gate nodes. Represents customs area nodes At any moment The state vector, This represents the set of nodes in the customs area map. Represents the aggregation operator, Indicates a shared customs area status summary. Indicates the first Local policy functions of each agent Indicates the first The local actions output by an agent, Represents the joint action vector; S202, Input shared customs area status summary Joint action with candidates Under a unified constraint framework, hard compliance is determined. If the rule is not met, minimal modifications are made, and the output is an executable joint action. To ensure the constraints are satisfied, the expression is: ； in, Represents the set of constraint functions. Indicates the first Constraint functions, This represents the feasible region projection operator. This indicates the modified executable joint action. Represents the L2 norm distance; S203, Input Disturbance Event When the disturbance intensity exceeds the threshold, online replanning is triggered, the original execution action is partially replaced, alternative paths, alternative handover points and alternative resource allocations are generated, and the replanned joint action is output. And satisfying uniform constraints, the expression is: ； in, Indicates a disturbance event. This represents the disturbance intensity assessment function. Indicates the trigger threshold. Indicates an indicator function, Indicates whether to trigger replanning. Indicates the online replanning operator, This indicates joint actions following a replanning process.

5. The multi-agent continuous learning-based collaborative method for people-free flow of goods according to claim 4, characterized in that: In step S201, the shared customs area status summary includes checkpoint flow control status, inspection station occupancy, key channel congestion, and / or restricted / prohibited status. In step S203, the disturbance events include checkpoint congestion, queue jumping during inspections, channel closures, and / or equipment malfunctions.

6. The multi-agent continuous learning-based human-free flow coordination method according to claim 4, characterized in that, Step S4 includes the following steps: S401: Input the newly added change information, encapsulate the change information into a rule change package, layout change package, or event template package, and use a unified parsing and mapping module to transform the rule change package, layout change package, or event template package into constraint updates and environment parameter updates for calculation. Output the incrementally updated constraint set and environment configuration to trigger subsequent continuous learning. The expression is: ； in, Indicates the first Version index adapted by the second incremental step Indicates the first This rule change package. Indicates the first Secondary layout change package Indicates the first Sub-event template package This indicates a change in the parsing operator. This represents the set of incremental changes obtained after analysis. This represents the set of constraints before the change. This represents the set of constraints after the change. Indicates the incremental merging operator; S402, Input the constraint set after incremental change. Incremental data acquired during online operation or simulation Simultaneously, constraint playback data is extracted from historical memory. To cover existing clearance patterns, the old and new data are reweighted, making the learning process focus more on the regions affected by rule changes and long-tail risk event samples, and outputting updated policy parameters. The expression is: ； in, Indicates the strategy parameters, Indicates the first The policy parameters after the next incremental update. Indicates the observation status of the customs area. Indicates an action, Indicates in constraints Define the training loss. Represents the sample weight function; S403, Input the policy parameters from the previous version. With the current incremental training objective, a parameter regularization mechanism is used to constrain the new parameters to not deviate from the effective solution space of the old parameters in key dimensions. This allows the system to adapt to new rules while maintaining stable performance on the existing level-clearing pattern, and outputs update policy parameters with forgetting suppression. The expression is: ； in, Indicates the first The primary training objective is updated incrementally. This indicates the strategy parameters of the previous version. This represents the L2 norm.

7. The multi-agent continuous learning-based human-machine interface collaborative method according to claim 6, characterized in that, Step S5 includes the following steps: S501: Input trajectory fragments generated during multi-agent collaborative execution; perform quality assessment and compression encoding on the trajectory fragments; filter out collaborative experience, decision fragments, and handling cases, and write them into the shared memory; during policy execution or learning, perform a similarity search on the shared memory based on the current shared customs area state summary and task context; output reusable collaborative evidence or experience fragments for each agent to use for decision-making reference or training replay, as expressed in the following expression: ； ； in, Indicates a shared memory bank. Indicates the first These candidate experience fragments are derived from key decision windows extracted from collaborative trajectories. This represents an empirical coding function used to compress empirical fragments into vectors or structured representations. This represents the experience quality score. Indicates experience tags. The write operator writes the encoded result and metadata to the shared memory and returns a shared memory unit. , Indicates a shared memory unit. Indicates time Shared customs area status summary This represents a retrieval operator used to return a set of memories related to the current context from a shared memory bank. This represents the set of relevant memories retrieved. S502, Input group collaboration strategy as teacher strategy And the student strategy set of various vehicles and node intelligent agents. Shared customs district status summary within the same customs district The teacher's strategy generates a joint action distribution and extracts effective cooperative patterns from it. Student strategies learn to approximate the teacher's decision distribution based on their own observations. Through distillation training, different types of agents acquire consistent cooperative preferences, enabling newly added vehicles or deployments in new checkpoints to inherit the cooperative clearance capabilities. The output is the set of student strategy parameters after distillation, expressed as: in, Represent the group teacher strategy, and output the joint action distribution under the condition of shared gate state summary. Represents intelligent agents Student strategies, Indicates student strategy parameters, This represents the optimal parameters after distillation. Indicates time Shared customs area status summary Represents intelligent agents Local observations, This represents the Kullback-Leibler divergence, used to measure the difference in how the student distribution approximates the teacher distribution. Represents an intelligent agent.

Images