Multi-agent based intermodal transportation plan generation system and method

Abstract

The application provides a multi-agent-based multimodal transport scheme generation system and method. It includes a task analysis module for generating a task context; a network modeling module for obtaining data related to the transport network and operating status and constructing an environmental situation representation; a candidate path generation module for generating a set of candidate paths that meet the constraints on the multimodal transport network; a resource checking module for checking the resource availability and timing of the candidate path set; a decision module for generating an evaluation vector for the checked candidate path set and generating a structured explanation; an output module for generating a scheme document and machine-readable instructions and providing them to an external execution system. The application can realize the unification of environmental situation representation of multi-source dynamic data, automatic generation of candidate path set based on constraints, resource and timing checking, target scheme selection and structured explanation generation under the cooperation of intelligent agents, event-triggered incremental re-planning, and machine-readable instruction output.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, and in particular, to a multi-agent based multimodal transport scheme generation system and method. BACKGROUND

[0002] Multimodal transport completes cross-regional transportation through coordination of sea transportation, railway, highway, etc., and is widely used in international logistics of bulk commodities. Its planning covers mode selection, node and path connection, schedule and time window matching, transport capacity and loading and unloading resource allocation, and is affected by dynamic factors such as weather, port operation, traffic and rate policy, and has large-scale decision-making and frequent updates.

[0003] The existing technology mainly includes the following two methods: first, a method based on operational research optimization, which models the problem as a mathematical programming and combines heuristic solution, suitable for static or weak dynamic scenarios; second, an agent method based on reinforcement learning, which regards path and scheduling as time sequence decision-making, and learns strategy based on historical or simulation data to respond to changes.

[0004] However, the existing technology has the following problems: the operational research method relies on static assumptions, and it is difficult to describe uncertainties such as weather and congestion, new nodes or channels are slow to adapt, and the cost of rolling re-planning is high; the reinforcement learning method has large data requirements, reward and hyperparameter sensitivity, and is insufficient in constraint feasibility and multi-objective trade-off, and scene migration and interpretability are poor. Overall, there is still a lack of a scheme that can integrate multi-source real-time information and make coordinated reasoning under strong constraints. SUMMARY

[0005] Therefore, the embodiments of the present application provide a multi-agent based multimodal transport scheme generation system and method to solve the problems of the prior art that multi-source dynamic information is difficult to construct a unified environmental situation representation, candidate path generation and resource time sequence checking under strong constraints are insufficient, and there is a lack of coordinated decision-making and event-driven re-planning mechanism.

[0006] The first aspect of the embodiment of the application provides a multi-agent-based multimodal transport scheme generation system, comprising: a task analysis module, configured to receive a task description for multimodal transport, extract constraint parameters related to transport demand, and form a task context; an agent coordination framework, configured to perform subtask decomposition and distribution of the task context based on a scheduling graph, and coordinate execution of each functional agent and aggregation of intermediate results; a network modeling module, configured to obtain data related to a transport network and a running state and construct an environment situation representation; a candidate path generation module, configured to generate a candidate path set satisfying constraints on a multimodal transport network based on the task context and the environment situation representation; a resource checking module, configured to check resource availability and timing connection of the candidate path set according to the environment situation representation, and obtain a checked candidate path set; a decision module, configured to generate an evaluation vector for the checked candidate path set based on a preset evaluation criterion, and select a target scheme in combination with the task context and generate a structured description; and an output module, configured to generate a scheme document and machine-readable instructions according to the structured description, and provide the scheme document and the machine-readable instructions to an external execution system; wherein the candidate path generation module, the resource checking module and the decision module are implemented as functional agent nodes scheduled by the agent coordination framework, and have a tool calling interface to access a multi-source data interface.

[0007] The second aspect of the embodiment of the application provides a multi-agent-based multimodal transport scheme generation method based on the system of the first aspect, comprising: receiving a task description for multimodal transport, extracting constraint parameters related to transport demand to form a task context, and publishing the task context based on a scheduling graph by an agent coordination framework and coordinating execution of each functional agent and aggregation of intermediate results; obtaining data related to a transport network and a running state and constructing an environment situation representation, generating a candidate path set satisfying constraints on a multimodal transport network based on the task context and the environment situation representation; checking resource availability and timing connection of the candidate path set based on the environment situation representation, and obtaining a checked candidate path set; generating an evaluation vector for the checked candidate path set based on a preset evaluation criterion, and selecting a target scheme in combination with the task context by an inference unit comprising a language model, and generating a structured description; generating a scheme document and machine-readable instructions based on the structured description, and providing the scheme document and the machine-readable instructions to an external execution system through an external execution interface; when the environment situation representation is updated, triggering the candidate path generation module, the resource checking module and the decision module to perform re-planning by the agent coordination framework, outputting a replaced target scheme and a structured description, and generating a corresponding scheme document and machine-readable instructions.

[0008] The above at least one technical scheme adopted by the embodiment of the application can achieve the following beneficial effects:

[0009] The task analysis module is used for receiving a task description oriented to multimodal transport, extracting constraint parameters related to transport demand, and forming a task context; the agent cooperation framework is used for subtask decomposition and allocation of the task context based on a scheduling graph, and coordinating execution of each functional agent and aggregation of intermediate results; the network modeling module is used for obtaining data related to a transport network and a running state and constructing an environment situation representation; the candidate path generation module is used for generating a candidate path set satisfying constraints on the multimodal transport network based on the task context and the environment situation representation; the resource checking module is used for checking resource availability and timing connection of the candidate path set according to the environment situation representation, to obtain a checked candidate path set; the decision module is used for generating an evaluation vector of the checked candidate path set based on a preset evaluation criterion, and selecting a target scheme and generating a structured description in combination with the task context; and the output module is used for generating a scheme document and machine-readable instructions according to the structured description, and providing the scheme document and the machine-readable instructions to an external execution system; wherein the candidate path generation module, the resource checking module, and the decision module are implemented as functional agent nodes scheduled by the agent cooperation framework, and have a tool calling interface to access a multi-source data interface. The application can realize unified environment situation representation of multi-source dynamic data, automatic generation of a constraint-based candidate path set and resource timing checking, target scheme selection and structured description generation under agent cooperation, event-triggered incremental re-planning, and machine-readable instruction output. BRIEF DESCRIPTION OF DRAWINGS

[0010] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort on the basis of these drawings.

[0011] Figure 1 is a structural composition schematic diagram of a multimodal transport scheme generation system based on multiple agents provided by the embodiments of the present application;

[0012] Figure 2 is a flow schematic diagram of a multimodal transport scheme generation method based on multiple agents provided by the embodiments of the present application. DETAILED DESCRIPTION

[0013] In the following description, specific details such as specific system structures, techniques, etc. are presented in order to thoroughly understand the embodiments of the present application. However, it should be clear to those skilled in the art that the present application can also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to avoid unnecessary details that hinder the description of the present application.

[0014] Multimodal transport aims to transport goods by combining various modes of transport such as sea, rail, and road, and is widely used in international bulk commodity logistics (such as the transnational transport of iron ore). However, current multimodal transport planning relies heavily on manual experience, resulting in inefficiency and high costs. Furthermore, in complex and ever-changing transport environments, manual planning struggles to adapt dynamically to various constraints, leading to insufficient resilience of the current system.

[0015] Currently, cross-regional multimodal transport route selection and dispatching largely rely on manual decision-making based on historical experience and fixed operational data. For large-scale transport tasks involving numerous transport nodes and multiple modes of transport, this manual approach reveals the following limitations:

[0016] Low efficiency: Manual planning is time-consuming and makes it difficult to respond promptly to changes in the market and environment (such as sudden changes in weather conditions, port congestion, and transportation delays). This results in the current solution being unable to quickly adjust transportation plans when faced with complex and ever-changing geographical, climatic, and price fluctuations, severely impacting the system's resilience.

[0017] High costs: Due to the lack of globally optimal calculations, manual solutions often fail to achieve the best balance between cost and timeliness, potentially leading to poor transportation routes and high vehicle empty load rates. In bulk commodity logistics, these suboptimal decisions significantly increase transportation costs.

[0018] Poor adaptability: Complex multimodal transport involves numerous aspects such as sea route selection, rail freight connections, and road transport scheduling, and is affected by various factors such as weather, infrastructure, and policies. It is difficult for humans to make decisions based on real-time data and multiple constraints, and they cannot dynamically adapt to sudden changes in the transportation process, resulting in insufficient robustness of the solutions.

[0019] To address the aforementioned issues, industry and academia have proposed several decision-making support solutions, primarily including two categories: traditional optimization algorithms based on operations research and agent-based methods based on reinforcement learning.

[0020] Operations research-based optimization methods: Some systems model multimodal transport route optimization as a mathematical programming problem, using heuristic algorithms (such as genetic algorithms and particle swarm optimization) to find approximate optimal solutions. For example, existing patents propose using adaptive genetic algorithms to optimize vehicle scheduling and route selection in multimodal transport to reduce empty load rates and travel distances, thereby lowering transportation costs. These methods can improve efficiency to some extent in static environments.

[0021] Reinforcement learning-based agent methods: Some research treats multimodal transport route planning as a reinforcement learning problem, using agents to continuously interact with the environment and learn the optimal strategy. For example, some solutions use container or transport companies as agents, employing Q-learning algorithms to progressively optimize multimodal transport route selection based on a reward function, allowing decisions to be continuously adjusted according to environmental changes during transport. This type of method has explored consideration of dynamic changes (such as carbon tax policies and the impact of third-party behaviors).

[0022] While the aforementioned existing technologies have alleviated the problems of manual planning to some extent, they still have the following shortcomings:

[0023] 1. Operations research optimization methods typically assume that the transportation network and constraints are relatively static and deterministic. In reality, uncertainties such as weather and port congestion are not included in the model. If the actual situation deviates from these assumptions, the optimization effect will be significantly reduced. Furthermore, mathematical models require pre-defining all possible nodes and routes, lacking flexibility to accommodate new options and making it difficult to respond promptly to special events.

[0024] 2. While agent-based methods such as reinforcement learning consider dynamic adjustment, the training process relies on a large amount of historical data and repeated trial and error, resulting in high training costs. However, data for multimodal transport in real-world scenarios is extremely difficult to collect, making training challenging. Furthermore, these methods typically focus on a single objective (such as minimizing carbon emissions or cost), limiting their ability to handle the balance of multiple objectives (cost, time, and risk) in real-world scenarios. Additionally, deployment requires specialized parameter tuning, limiting their applicability.

[0025] Overall, there is currently a lack of intelligent systems capable of making reasoning decisions by integrating diverse knowledge and real-time information, much like human experts. Neither traditional algorithms nor learning algorithms can fully simulate the experience, judgment, and adaptability of human dispatchers in complex multimodal transport scenarios. Therefore, it is necessary to explore new technological solutions to overcome these shortcomings.

[0026] In view of the problems existing in the prior art, this application proposes a multi-agent-based multimodal transport scheme generation system and method. The technical solution of this application mainly includes the following:

[0027] 1. Multi-Agent Cooperative Scheduling System: Construct a scheduling system composed of multiple agents and diverse MCP-driven tools. Each agent possesses the ability to autonomously perceive the environment, make independent reasoning decisions, and collaborate with each other. Through a graph scheduling network, processing nodes are adaptively allocated to sub-tasks, enabling the system to simulate human team collaboration in planning complex transportation tasks.

[0028] 2. Large Language Model as Core Intelligence: A multimodal large language model (MLLM) is introduced as the cognitive and decision-making center of the multi-agent system. The LLM possesses powerful knowledge acquisition and reasoning capabilities, enabling it to act as the "brain" of the agents, understanding complex transportation task requirements and various constraints, and guiding the agents to work collaboratively.

[0029] 3. Optimized Path Through Agent Division of Labor: Multiple agents share different functions based on transportation task objectives and cooperate closely. Taking into account the cost, time, and reliability factors of various transportation modes such as sea, rail, and road, the agents collaboratively plan the optimal multimodal transport route and scheduling scheme. Throughout the process, the system dynamically senses the external environment (such as weather conditions, real-time port and customs status) and incorporates it into decision-making, thereby ensuring that the scheme is feasible and optimal under realistic conditions.

[0030] 4. Automatic Output and Dynamic Adaptation of Optimal Solutions: The system automatically generates efficient and executable cross-modal transportation routes and scheduling strategies. Compared to manual planning, the generated solutions significantly optimize overall transportation time, economic costs, and robustness. Furthermore, the system can adjust the solutions in real time based on environmental changes during execution, ensuring safe and efficient transportation. Overall performance far surpasses traditional static planning methods that rely on manual intervention.

[0031] The following detailed description, in conjunction with the accompanying drawings and specific embodiments, outlines the framework and functions of the multi-agent-based multimodal transport scheme generation system provided in this application. Figure 1 This is a schematic diagram of the structural composition of the multi-agent-based multimodal transport scheme generation system provided in the embodiments of this application, as shown below. Figure 1 As shown, the multi-agent-based multimodal transport scheme generation system may specifically include the following modules:

[0032] Task parsing module 101 is used to receive task descriptions for multimodal transport, extract constraint parameters related to transport requirements, and form a task context.

[0033] The intelligent agent collaboration framework 102 is used to decompose and allocate task context into subtasks based on the scheduling graph, and coordinate the execution of each functional intelligent agent and the aggregation of intermediate results.

[0034] Network modeling module 103 is used to acquire data related to the transportation network and its operational status and to construct an environmental situation representation;

[0035] The candidate path generation module 104 is used to generate a set of candidate paths that meet the constraints on the multimodal transport network based on the task context and environmental situation representation.

[0036] The resource verification module 105 is used to verify the resource availability and temporal sequence consistency of the candidate path set based on the environmental situation representation, and obtain the verified candidate path set.

[0037] The decision module 106 is used to generate an evaluation vector for the verified candidate path set based on preset evaluation criteria, and to select the target solution and generate a structured description in combination with the task context.

[0038] Output module 107 is used to generate a solution document and machine-readable instructions based on the structured specification, and to provide the solution document and machine-readable instructions to an external execution system;

[0039] Among them, the candidate path generation module 104, the resource verification module 105 and the decision module 106 are implemented as functional intelligent agent nodes scheduled by the intelligent agent collaboration framework 102, and have tool call interfaces to access multi-source data interfaces.

[0040] In some embodiments, the task parsing module is used for:

[0041] The parsing unit containing the language model is used to perform intent recognition and element extraction on the task description, and the extraction results are structured into task context.

[0042] The constraint parameters in the task context are hierarchically encoded according to hard constraints and soft constraints and unified into executable constraint expressions;

[0043] Perform consistency and reachability checks on constraint expressions, and infer default values ​​for missing parameters based on the rule base and boundary conditions;

[0044] Confidence levels and source identifiers are added to constraints with uncertainties, and the resulting task context is published to the candidate path generation module and decision-making module through the intelligent agent collaboration framework.

[0045] Specifically, the task parsing module in this embodiment works collaboratively around the following technical features: including a parsing unit for a language model, task context, hierarchical encoding of constraint parameters, executable constraint expressions, consistency and reachability checks, default inference driven by a rule base and boundary conditions, confidence level and source identifier, and publication through an intelligent agent collaboration framework. The process is as follows: receiving a task description → the parsing unit performs intent recognition and element extraction → generating an initial instance of the task context → hierarchically encoding the constraint parameters using hard and soft constraints → unifying the encoding results into executable constraint expressions → performing consistency and reachability checks on the expressions → performing default inference for missing items based on the rule base and boundary conditions → adding confidence level and source identifier to constraints with uncertainties → publishing the task context as a versioned record within the intelligent agent collaboration framework for reference by the candidate path generation module and the decision-making module.

[0046] The following explains the concepts of important technical terms used in this embodiment, in detail as follows:

[0047] Task context: A structured description of a transportation task, including at least a task identifier, origin and destination information, cargo type and batch information, constraint set, preference set, and reference metadata.

[0048] Constraint parameter hierarchical encoding: Constraints are divided into two levels: hard constraints and soft constraints. Hard constraints limit the feasible region boundary, while soft constraints are used for preferences and trade-offs. Both are described by a unified key value and spatiotemporal range.

[0049] Executable constraint expressions: Without limiting the specific implementation syntax, constraints such as time window, mode switching, accessibility, and resource consumption are transformed into predicate expressions that can be directly judged and calculated by downstream agents, and reference identifiers and version numbers are supported.

[0050] Consistency and reachability verification: Consistency verification includes aligning coverage scope, standardizing units, and normalizing time sequence relationships and geographic identifiers; Reachability verification includes connecting origin and destination points, valid mode sequence, and existence of basic resources.

[0051] Rule base and boundary conditions: a set of knowledge and allowed range settings used to infer missing parameters, such as mapping default packaging form based on cargo type and mapping buffer time limit based on transportation mileage.

[0052] Confidence level and source identifier: Generate confidence level values ​​and source identifiers for uncertain or external references, and bind them to constraint expressions for downstream reference and re-evaluation.

[0053] Release mechanism: Within the agent collaboration framework, release is performed using a referable task context object, which includes a version number and dependency pointers, ensuring that the candidate path generation module and the decision-making module can be deterministically read.

[0054] For example, in some examples, after a natural language description of a cross-border bulk transportation task is input, the parsing unit performs intent recognition and feature extraction on the description to generate an initial instance of the task context. Hard constraints include arrival time limits, origin and destination, and minimum loading batches, while soft constraints include mode preferences and cost weights. The module standardizes time units, maps geographical names to standard location identifiers, and retains the most recent valid entry for repeated descriptions of the same parameter.

[0055] The hard and soft constraints are then transformed into executable constraint expressions, such as using time windows to describe the allowable range of shipments and arrivals, using mode switching rules to limit the combination of transport modes in adjacent segments, and using accessibility restrictions to require the connectivity of origin and destination nodes in the multimodal transport network.

[0056] For missing parameters, inferences are made based on the rule base and boundary conditions. For example, if no container type is given, the default container type and quantity range are mapped according to cargo type and batch. If no connection buffer is given, the default buffer upper limit is mapped according to mileage range. For rates or timetables referenced from external sources, the source identifier is recorded and a confidence level is added.

[0057] Finally, the module generates a task context object with a version number, publishes it to the candidate path generation module and the decision-making module through the intelligent agent collaboration framework, and registers reference relationships within the framework for subsequent replanning for location and updating.

[0058] In some embodiments, the agent collaboration framework is used for:

[0059] Within the scheduling graph, a set of subtasks with dependency and constraint annotations is established based on the task context and environmental situation representation, and the subtasks are mapped to executable units of functional agents;

[0060] Configure execution priorities and concurrency for functional agents based on dependencies and constraints, and maintain data and constraint consistency during execution through constraint propagation and conflict detection;

[0061] Intermediate results generated by each functional agent are aggregated according to a traceable referencing mechanism for reference by unfinished subtasks;

[0062] When the environmental situation is updated, a local replanning is triggered to determine the boundaries of the affected subtasks and complete the recalculation and replacement within the scheduling graph;

[0063] Configure multi-source data interfaces and computing resource access routes for functional agents by calling the tool's interface.

[0064] Specifically, the agent collaboration framework in this embodiment revolves around a scheduling graph, a set of subtasks, dependency and constraint annotations, functional agent mapping, execution priority and concurrency, constraint propagation and conflict detection, a traceable reference mechanism, local replanning, and tool call interfaces. The specific process is as follows: A set of subtasks with dependency and constraint annotations is generated within the scheduling graph based on the task context and environmental situation representation; each subtask is mapped to an executable unit of a functional agent and input / output references are registered; execution priority and concurrency are configured for the functional agents according to the dependency structure and constraint strength; during execution, constraints such as time windows, resource consumption, and mode switching are propagated along the dependency direction, and conflict detection is performed; intermediate results returned by each functional agent are versioned and aggregated according to the traceable reference mechanism, and a deterministic snapshot is provided; when the environmental situation representation is updated, the boundaries of the affected subtasks are located, and local recalculation and result replacement are performed on the relevant branches within the scheduling graph; multi-source data interfaces and computing resource access routes are configured for the functional agents through the tool call interface, and the call context is recorded.

[0065] The following explains the concepts of important technical terms used in this embodiment, in detail as follows:

[0066] Scheduling graph: A graph structure that carries subtask nodes and their directed dependencies, with constraint labels and triggering conditions attached to the edges, used to drive agent execution and dependency management.

[0067] Dependency and constraint annotation: Maps the task context and environmental situation representation into a set of constraints of nodes and edges, including at least time window, accessibility, mode switching rules and resource consumption limits, to limit the feasible execution order.

[0068] Functional agent mapping: Capabilities such as candidate path generation, resource verification and decision-making are encapsulated into schedulable functional agents. Subtasks and functional agents are bidirectionally bound, which facilitates unified orchestration and backfilling of results.

[0069] Traceable referencing mechanism: assign a unique reference identifier and version number to intermediate results, retain source and dependency pointers, and consume only stable snapshots for unfinished subtasks to ensure consistent reading.

[0070] Local replanning: Based on the update event of the environmental situation representation, the affected subgraph is segmented, and the corresponding functional agent is triggered to perform incremental recalculation and replacement on the subgraph. The unaffected part continues to be executed according to the established dependencies.

[0071] Tool call interface and access routing: Select and connect multi-source data interfaces and computing resources for functional intelligent agents, complete authentication and routing, and write call summaries into reference metadata for auditing and reuse.

[0072] For example, in some examples, after the task context and environmental situation representation of a cross-regional transportation task are generated, the collaboration framework establishes a set of subtasks within the scheduling graph, including candidate path generation nodes, resource verification nodes, and decision nodes.

[0073] The candidate path generation node references the task context and environmental status representation, and outputs a versioned snapshot of the candidate path set; the resource verification node reads the snapshot with the reference identifier, establishes a relationship between the path segment and the resource status, completes the verification of capacity availability and time window alignment, and marks the path with soft constraint conflict with a revision identifier; for the path to be revised, the framework triggers local substitution and remapping without breaking upstream dependencies, generates the revised path and updates the reference identifier.

[0074] Subsequently, the decision node generates an evaluation vector based on the verified snapshot of the candidate path set and preset evaluation criteria, selects the target solution in conjunction with the task context, and forms a structured description. Intermediate results are aggregated throughout the process using a traceable referencing mechanism; incomplete subtasks only consume deterministic snapshots, avoiding inconsistencies caused by read / write contention.

[0075] When the environmental situation is updated (e.g., changes in the congestion index of a specific port or adjustments to the weather forecast for a specific sea area), the collaborative framework locates the affected sub-task boundaries in the scheduling graph based on the update event, triggering recalculation only for the relevant path segments and resource references; the candidate path generation node performs incremental updates on the affected segments, the resource verification node only performs re-verification on the updated segments, and the decision node completes the reselection of the target scheme on the new evaluation vector set; the unaffected nodes maintain their original execution state and continue to consume the original snapshot, and the framework assigns a new version number to the replacement result and updates the reference relationships.

[0076] Through the above implementation methods, the intelligent agent collaborative framework achieves controllable concurrency and orderly arrangement of subtasks under the constraints of the scheduling graph, uses a traceable reference mechanism to ensure version consistency and deterministic reading of intermediate results, and limits environmental updates to incremental replacement within the minimum impact domain through event-triggered local replanning. Thus, it supports the continuity and stability of candidate path generation, resource verification, and decision output under multi-source dynamic information and strong constraints.

[0077] In some embodiments, the network modeling module is used for:

[0078] Heterogeneous data related to the topology and operation status of the transportation network are accessed through multi-source data interfaces, and format normalization and spatiotemporal alignment are completed in a unified mode to generate standardized records with spatiotemporal keying.

[0079] Based on validity period and update priority, standard records are subject to freshness control and conflict resolution, forming an environmental situation representation with accompanying source identification and confidence level labeling;

[0080] An incremental fusion mechanism is adopted to maintain the versioned representation of the environmental situation, and a snapshot with version identifier is output for the candidate path generation module and the resource verification module to use deterministically.

[0081] For missing or uncertain entries, rule-driven completion and uncertainty quantification are performed under preset boundary conditions, and the completion results are written into the environmental situation representation;

[0082] Based on change thresholds and triggering rules, update events are generated from the environmental situation representation, the affected network range is labeled, and the events are published through the agent collaboration framework to drive replanning.

[0083] Specifically, the network modeling module in this embodiment is constructed around elements such as multi-source data interfaces, unified mode, spatiotemporal alignment, standardized records of spatiotemporal keying, freshness management and conflict resolution, environmental situation representation, incremental fusion mechanism, version maintenance and snapshots, rule-driven completion and uncertainty quantification, change thresholds and triggering rules, update events and affected network range.

[0084] The implementation process is as follows: Heterogeneous data related to the transportation network topology and operational status are accessed through multi-source data interfaces → Format normalization and spatiotemporal alignment are completed in a unified mode to generate standardized records with spatiotemporal keying → Freshness control and conflict resolution are implemented based on validity period and update priority to form an environmental situation representation → The environmental situation representation is versioned using an incremental fusion mechanism and a snapshot with version identifier is output for deterministic reference by the candidate path generation module and resource verification module → Rule-driven completion and uncertainty quantification are performed for missing or uncertain entries under preset boundary conditions → Update events are generated from the environmental situation representation according to change thresholds and triggering rules, the affected network range is marked, and the updates are published through the intelligent agent collaboration framework to drive replanning.

[0085] The following explains the concepts of important technical terms used in this embodiment, in detail as follows:

[0086] Standardized records with unified mode and spatiotemporal keying: Port, railway, highway and shipping related data are mapped into a unified field system, using location identifier and time identifier as keys to record topological units and operating status, ensuring cross-source comparability and cross-time period referenceability.

[0087] Freshness control and conflict resolution: For multi-source records with the same key, select or merge them based on validity period and update priority. When time overlaps or values ​​conflict, reconcile according to priority and confidence weight, and retain the source identifier and confidence level.

[0088] Environmental situation representation and incremental fusion mechanism: Standardized records are aggregated in a spatiotemporal range into a situation view that can be directly referenced for reasoning and verification. Only affected keys are updated incrementally to avoid full replacement.

[0089] Versioning and snapshots: Assign a version number to each fusion output and provide a read-only reference in the form of a snapshot to ensure that downstream modules obtain a stable and consistent data view during computation.

[0090] Rule-driven completion and uncertainty quantification: Based on the rule base and boundary conditions, missing fields are completed, and uncertainty quantification results are generated and written back to the situation representation for downstream reference when making constraint judgments and trade-offs.

[0091] Change thresholds and triggering rules: Set magnitude thresholds and duration thresholds for key fields. When the thresholds are reached, generate an update event and mark the affected network range to limit the impact domain of replanning.

[0092] For example, in some examples, in cross-border bulk transportation scenarios, the module accesses heterogeneous data such as port operation system data, liner / train schedules, road traffic restrictions, weather and sea conditions, and freight rate announcements through multi-source data interfaces. The module maps fields from different sources to a unified schema and completes spatiotemporal alignment using location and time identifiers as keys, forming a standardized record set covering port nodes, railway sections, and sea routes.

[0093] When the capacity of a berth at the same port is reported simultaneously from two sources with overlapping timeframes, the module mediates conflicts based on pre-configured update priorities and source credibility, outputting a single record with an appended source identifier and confidence level. For missing minimum transshipment time, the module generates a completion value based on boundary conditions for cargo type and operation window in the rule base, and simultaneously writes the uncertainty quantification result for downstream reference.

[0094] After forming the environmental situation representation, the module locates and replaces newly added or changed records according to the incremental fusion mechanism, generates a new version number for this fusion, and outputs a snapshot. The candidate path generation module and the resource verification module use the version identifier of this snapshot for deterministic reading to avoid inconsistencies in reading due to data changes during the calculation.

[0095] When meteorological data indicates that the wind and wave levels in a specific sea area exceed a preset change threshold, the module generates an update event according to triggering rules, marks the affected sea route range and related port nodes, and publishes it through the intelligent agent collaboration framework. Based on this, the collaboration framework triggers replanning only for the affected area, and downstream modules perform incremental updates based on the new version snapshot; unaffected paths and resource references remain unchanged.

[0096] Through the above implementation methods, the network modeling module realizes the normalized expression of multi-source heterogeneous data with a unified mode and spatiotemporal keying mechanism, ensures the consistency and referability of environmental situation representation with freshness control and conflict mediation, supports deterministic reading of downstream data with incremental fusion and versioned snapshots, and limits the release of environmental changes to the affected network range with threshold-triggered update events. This provides a stable, traceable and incrementally updatable data foundation for candidate path generation and resource verification.

[0097] In some embodiments, the candidate path generation module is used for:

[0098] Based on the environmental situation representation, the multimodal transport network is parameterized and mapped to form a unified representation for route search;

[0099] Based on the constraints of the task context, constraint-driven path expansion and pruning are performed on the unified representation. The constraints include temporal consistency, accessibility and transportation mode switching rules, generating feasible paths from the origin to the destination.

[0100] The generated feasible paths are standardized and deduplicated, and their size is controlled to retain a limited number of candidate paths for subsequent reference.

[0101] Upon receiving an environment update event published by the network modeling module, the affected path segments are located and incrementally updated, and a snapshot of the candidate path set with version identifiers is output for the resource verification module to use for deterministic reference.

[0102] Specifically, the candidate path generation module in this embodiment operates around the following elements: environmental situation representation, parameterized mapping and unified representation, constraint-driven path expansion and pruning, standardized expression and deduplication of feasible paths, scale control and versioned snapshots, and incremental updates triggered by update events. The process is as follows: based on the environmental situation representation, the multimodal transport network is parameterized to obtain a unified representation for searching; based on the unified representation, path expansion and pruning are performed according to the hard and soft constraints of the task context to generate feasible paths that meet the rules of temporal consistency, accessibility, and mode of transport switching; the feasible paths are standardized and deduplicated, and scale control is implemented to form a candidate path set; a snapshot with a version identifier is generated for the candidate path set for deterministic reference by the resource verification module; when an environmental update event is received from the network modeling module, the affected path segments are located, incremental updates are performed, and a new version snapshot is output.

[0103] The following explains the concepts of important technical terms used in this embodiment, in detail as follows:

[0104] Parametric mapping and unified representation: Port nodes, railway sections, highway corridors and their operating parameters are represented by a unified field system, and time windows, capacity, traffic restrictions, transshipment rules and other parameters are attached to edges or nodes as parameter sets to form a consistent and searchable search view across modes.

[0105] Constraint-driven path expansion and pruning: During the search process, endogenous constraints are used as expansion and truncation conditions, hard constraints (such as time windows, connectivity, and legality of mode sequences) are used as necessary conditions, and soft constraints (such as preferences and cost trade-offs) are used as priority factors to avoid generating infeasible branches.

[0106] Standardized representation and deduplication: The path is standardized and encoded using node sequence, mode sequence and segment time parameters, and duplicate or isomorphic path entries are merged using equivalence criteria.

[0107] Size control and snapshots: Limit the number of candidate entries while meeting coverage requirements, output read-only snapshots and assign version identifiers to ensure view stability during downstream reading.

[0108] Incremental update and affected segment location: Based on the affected network range carried by the update event, locate the associated edges or nodes in the unified representation, and only recalculate and replace the encoding and parameters of the corresponding path segments.

[0109] For example, in some examples, after the task context of shipping 50,000 tons of bulk cargo from Indonesia to a coastal port in China is determined, the module first receives an environmental situation representation, which includes shipping route weather, port operational capacity, rail schedules, and road traffic restrictions. The module performs a parameterized mapping of the multimodal transport network, describing sea segments, inland rail segments, and connecting roads with a unified field, and attaching minimum transshipment time, mode switching permission pairs, and time window parameters to the relevant edges.

[0110] The module then initiates a constraint-driven search on the unified representation. Using the origin-destination, arrival time, and mode sequence constraints given in the task context as hard constraints, it expands adjacent reachable nodes; branches that do not meet timing consistency or mode switching rules are pruned immediately. For branches that meet the hard constraints, the expansion priority is adjusted based on cost and reliability preferences in the soft constraints, forming multiple feasible paths for backup.

[0111] The module standardizes the generated feasible paths by encoding the node sequence, mode sequence, and time segment parameters; it deduplicates equivalent paths and implements size control based on the principle of covering different transit strategies, retaining a limited number of candidate paths. The module generates a read-only snapshot of this set with version number V1 and registers its reference for deterministic reading by the resource verification module.

[0112] When the network modeling module releases an update event indicating that the wind and wave levels in a specific sea area have increased and the congestion index of a certain port has exceeded the threshold, the candidate path generation module locates the relevant sea segments and port nodes in the unified representation based on the affected range carried by the event. It only performs incremental recalculation on the path branches of the affected segments, replaces the encoding and time parameters of the original segments, forms a new set of candidate paths, and outputs a snapshot with the version number V2. Unaffected paths retain their encoding and parameters in V1, and the resource verification module can consume the changed segments in V2 without changing the unaffected references.

[0113] This embodiment constructs a unified search view across modes through parameterized mapping, generates a candidate path set that satisfies the rules of temporal consistency, accessibility, and mode of transport switching through constraint-driven expansion and pruning, and achieves stable and referable versioned snapshots by combining standardized expressions and deduplication, and performs incremental updates with the affected fragments as boundaries when the environment is updated, thereby supporting the resource verification module to perform deterministic and continuous subsequent processing on the candidate path set.

[0114] In some embodiments, the resource verification module is used for:

[0115] Resource status representations are generated based on environmental situation representations and associated with segments of the candidate path set;

[0116] Based on the hard and soft constraints of the task context, the availability of transport capacity, time window alignment, and mode switching compatibility are checked. Paths that do not meet the hard constraints are removed, and soft constraint conflicts are marked as pending revision.

[0117] Within the dependency limits allowed by the scheduling graph, local substitution and remapping of labeled items are performed to generate revision paths;

[0118] Output a verified versioned snapshot of the candidate path set with section-level resource references and connection parameters, and perform incremental re-verification when the environment is updated for the decision-making module to reference.

[0119] Specifically, the resource verification module in this embodiment revolves around elements such as environmental situation representation, resource status representation, candidate path set, hard and soft constraints, annotations to be revised, local substitution and remapping, segment-level resource references and connection parameters, versioned snapshots, and incremental re-verification. Its implementation process is as follows: Based on the environmental situation representation, it aggregates capacity, operational capability, and time window information to generate a resource status representation; it establishes a one-to-one or many-to-one association between the resource status representation and segments of the candidate path set; it performs capacity availability verification, time window alignment verification, and mode switching compatibility verification according to the hard and soft constraints of the task context, eliminating paths that do not meet hard constraints and marking paths with soft constraint conflicts as needing revision; within the dependency range allowed by the scheduling graph, it performs local substitution and remapping on the labeled items to generate revised paths; it outputs a versioned snapshot of the verified candidate path set with segment-level resource references and connection parameters; when the environmental situation representation is updated, it only performs incremental re-verification on the affected references and generates a new version for the decision-making module to reference.

[0120] The following explains the concepts of important technical terms used in this embodiment, in detail as follows:

[0121] Resource status representation: The availability of transport capacity such as ship positions / trains / vehicles, loading and unloading and storage capacity, port and road passage time windows, minimum transshipment time required for mode switching and rules are uniformly modeled and organized with referenceable identifiers and validity periods, and segment-level binding is supported.

[0122] Capacity availability check: Match the segment’s transport volume and time period demand with the capacity and available time periods in the corresponding resource status representation. If the capacity or available time periods are insufficient, it is considered that the hard constraint is not met.

[0123] Time window alignment check: Check the inclusion or intersection relationship between the planned arrival time of the segment and the available resource time window, and allow slight alignment without changing the upstream dependencies.

[0124] Mode switching compatibility check: Check the rule compliance of adjacent segments in terms of mode sequence, change position and minimum change time.

[0125] Revision markup and local substitution: Generate revision markup for sections with soft constraint conflicts, and perform local substitution and path remapping within the dependency boundary of the scheduling graph using similar resources or adjacent windows, while keeping upstream references unchanged.

[0126] Segment-level resource reference and connection parameters: For each verified path, resource reference identifier, replacement location, minimum replacement duration and arrival / departure timing parameters are written at the segment level to facilitate deterministic reading and verification downstream.

[0127] Versioned snapshots and incremental re-verification: Output a read-only snapshot of the verified candidate path set and assign a version number. When the environment is updated, only the affected references are re-verified and replaced.

[0128] For example, in some cases, after a transnational bulk transportation task is parsed into a task context and a candidate path generation module outputs a candidate path set, the resource verification module first constructs a resource status representation based on the environmental situation representation. This representation covers the ship's position and expected berthing window for maritime segments, the loading, unloading, and storage capacity of transshipment ports, the available train schedules and quotas for inland railway segments, and the restricted traffic periods and vehicle availability for connecting highways. The module then associates each segment of the candidate path set with its corresponding resource status, forming an initial mapping from segment to resource reference.

[0129] The module then performs hard constraint checks. For a feasible route via transit, if the transshipment port's storage capacity is insufficient on the planned date or the minimum transshipment time between adjacent sections cannot be met, the module determines that the hard constraint is not met and removes the route. If the sea voyage has available space but the railway section has a tight quota within the target time window, the module retains the route and marks the railway section as pending revision, proceeding to the subsequent partial substitution process.

[0130] In local substitution and remapping, the module is limited to retrieving adjacent available resources and time windows without disrupting upstream dependencies. For example, it might select adjacent train time slots on the same route or change loading / unloading time slots within the same node range. If the substitution is successful, the module generates a revised path and updates the reference relationships between segments and resources, while maintaining the reference identifiers and order of the remaining segments of the path. For branches where substitution fails, the module either rolls them back to the candidate list or removes them based on the scheduling graph strategy.

[0131] After resource verification is completed, the module writes segment-level resource references and connection parameters for the verified paths, generates a versioned snapshot V1 of the verified candidate path set, and registers the references. If the network modeling module publishes an environmental update event during the reading of V1, such as short-term congestion at a transshipment port or weather changes in a specific sea area, the resource verification module locates the resource references of the corresponding segment based on the affected range carried by the event, and only performs incremental re-verification on these references; if the replacement resource and time window are available, a new version snapshot V2 is generated and the references of unaffected paths to V1 are retained unchanged, ensuring that the decision module can complete subsequent processing under a consistent view.

[0132] Through the above implementation methods, the resource verification module realizes segment-level feasibility determination and fine-grained revision of the candidate path set under a unified resource status representation. It completes the minimum-range path remapping within the scheduling graph boundary by using the unrevised annotation and local substitution mechanism, and ensures the deterministic reading and continuous reference of the verified candidate path set by downstream users through versioned snapshots and incremental re-verification.

[0133] In some embodiments, the decision module is used for:

[0134] Based on preset evaluation criteria and task context, a configurable multi-dimensional evaluation template is constructed to generate standardized evaluation vectors for each of the verified candidate path sets.

[0135] Dominance relationships of candidate paths are established based on evaluation vectors to form a feasible frontier set, and candidate paths not included in the frontier set are marked as objects to be eliminated.

[0136] By using an inference unit that includes a language model, the target solution is selected from the set of feasible frontiers, while satisfying the constraints of the task context and the consistency of the environmental situation representation. The output is a structured description containing path identifiers, segment-level resource references and temporal connection parameters.

[0137] Specifically, the decision-making module in this embodiment operates around elements such as a configurable multidimensional evaluation template, evaluation vectors, dominance relationships and feasible frontier sets, an inference unit containing a language model, and structured descriptions. The process is as follows: a multidimensional evaluation template is constructed based on preset evaluation criteria and task context, and constraints and weight boundaries are instantiated; standardized evaluation vectors are calculated and generated for each of the verified candidate path sets; dominance relationships are determined within the evaluation vector space, and a feasible frontier set is constructed, while non-frontier candidates are marked as objects to be eliminated; the inference unit containing a language model performs constraint consistency verification and preference parsing on the frontier set, under the premise of satisfying the consistency of task context and environmental situation representation, and outputs the target solution and corresponding structured descriptions.

[0138] The following explains the concepts of important technical terms used in this embodiment, in detail as follows:

[0139] Configurable multidimensional evaluation templates: Evaluation dimensions and value ranges are loaded based on the task context. Dimensions can include time cost, economic cost, reliability measurement, constraint satisfaction measurement, etc. A unified scale is used and weight range and priority description are supported.

[0140] Evaluation vector: A fixed sequence vector formed by summarizing and mapping the segment parameters of a single verified candidate path according to the template, with path identifier and version number attached for easy traceability and reference.

[0141] Domination Relationships and Feasible Frontier Sets: Identify candidate sets that cannot be dominated by others in a multidimensional space based on criteria that are not inferior to and at least one dimension superior, as a candidate pool for further selection.

[0142] The reasoning unit, which includes a language model, parses the preferences and exceptions in the task context, performs interpretable selection and ranking of the frontier set, and generates the target solution and structured description, provided that the formal constraints pass the verification.

[0143] Structured description: Includes path identifiers, section-level resource references and timing connection parameters, reference version information and dependency pointers, which are used by the output module to directly generate solution documents and machine-readable instructions.

[0144] For example, in some examples, in a transportation task from Indonesia to Lianyungang, the task context provides an arrival time limit and cost ceiling, and prefers a transit sequence via Singapore. The decision module first instantiates a multi-dimensional evaluation template based on the task context, mapping time cost, economic cost, segment-level constraint satisfaction, and reliability metrics to a unified scale, while setting a higher priority for time cost.

[0145] The verified candidate route set includes direct sea routes, routes transiting through Singapore, and sea-rail combined transport routes. The decision-making module reads a versioned snapshot, calculates an evaluation vector for each candidate route, and binds the route identifier with segment-level resource references. Then, it determines the dominance relationship within the evaluation vector space. Direct routes are dominated by transit routes in the reliability dimension, while sea-rail combined routes are inferior to other candidates in the economic cost dimension and are marked as candidates to be eliminated. The feasible frontier set consists of the Singapore transit route and one alternative direct route.

[0146] The inference unit, which includes a language model, performs constraint consistency verification and preference parsing on the frontier set. This unit combines mode sequence preferences and arrival time limits within the task context to select one of two frontier candidates, choosing the transit route via Singapore as the target solution. It then outputs a structured description, including the path identifier, resource references for each segment, arrival and departure time parameters, and the dependent snapshot version number. This allows the output module to directly generate a solution document and machine-readable instructions.

[0147] After the environmental situation is updated to reflect rising winds and waves in a specific sea area, the decision-making module receives a new snapshot of the evaluation vector. The dominance relations and frontier set are then recalculated. The inference unit, which includes the language model, performs a reselection within the new frontier set. If the target solution remains unchanged, only the referenced version in the structured description is refreshed; if a replacement occurs, the new target solution is output and the dependency pointers are updated.

[0148] This embodiment establishes a unified decision space through configurable multidimensional evaluation templates and standardized evaluation vectors. It constructs a feasible frontier set based on dominance relationships to narrow down the candidate range. The reasoning unit, which includes a language model, completes the selection and explanatory output of the frontier candidates under the constraint of consistency. Finally, it forms a structured description that can be directly consumed by subsequent modules and supports rapid reselection under environmental updates.

[0149] In some embodiments, the output module is used for:

[0150] Based on the structured specification call document and instruction template library, generate a scheme document containing path identifiers, segment-level resource references and timing connection parameters, and construct the corresponding machine-readable instruction set;

[0151] Deterministic serialization and version identification are performed on the solution documents and machine-readable instructions, and the reference relationships and verification information are written into the status and model storage unit;

[0152] Machine-readable instructions are published to the external execution system through the interface, and protocol mapping and parameter verification are completed.

[0153] When the environmental situation is presented or the target solution is updated, the instruction change set is output and the version is replaced to provide the update to the external execution system.

[0154] Specifically, the output module in this embodiment operates around elements such as structured descriptions, document and instruction template libraries, solution documents, machine-readable instructions, deterministic serialization and version identification, reference relationships and verification information, state and model storage units, external execution interfaces, protocol mapping and parameter verification, instruction change sets, and version replacement. The process is as follows: receiving structured descriptions → generating solution documents containing path identifiers, segment-level resource references, and timing connection parameters based on the template library and constructing machine-readable instruction sets → performing deterministic serialization on the documents and instructions and assigning version identifiers → writing reference relationships and verification information into the state and model storage units → publishing machine-readable instructions through the external execution interface and completing protocol mapping and parameter verification → outputting instruction change sets and completing version replacement when the environmental situation representation or target solution is updated.

[0155] The following explains the concepts of important technical terms used in this embodiment, in detail as follows:

[0156] Document and instruction template library: Defines fields such as path identifier, section-level resource reference, arrival / departure sequence, change location and minimum change duration with parameterized placeholders, and supports template version management and field constraint validation.

[0157] Deterministic serialization and version identification: Generate repeatable binary or text representations by using fixed field order and standardized time / location encoding, and identify version snapshots by combining version number and timestamp.

[0158] Reference relationships and verification information: Establish pointer relationships for path identifiers, resource references, and snapshot versions referenced in documents and instructions, attach integrity verification values ​​and source identifiers, and write them to the status and model storage units for auditing and playback.

[0159] External execution interface and protocol mapping: Mapping internal fields to data contracts of external interfaces such as booking system and transportation scheduling system, and completing field semantic alignment, unit standardization and verification of required fields.

[0160] Instruction Changesets and Version Replacement: Calculates differences between versions, generates minimal changesets for add, change, and undo instructions, and atomically replaces old version references in external systems.

[0161] For example, in some examples, after the decision-making module produces a structured description, the output module selects a template corresponding to the task type from the template library, injects path identifiers, segment-level resource references, and timing parameters, and generates a solution document; simultaneously, it constructs a machine-readable instruction set covering booking requests, port operation reservations, and train / vehicle dispatch instructions, etc. The module then performs deterministic serialization, fixing the field order and unifying time and location encoding, generating a document and instruction snapshot with a version identifier V1.

[0162] The output module writes the reference relationships in documents and instructions into the status and model storage unit, records the path identifier and resource reference to the pointer of snapshot version V1, and writes the verification information and source identifier. Then, it publishes machine-readable instructions to the corresponding system through the external execution interface, completes field mapping and parameter verification, such as mapping the segment time sequence to the departure / arrival time window of the booking, mapping resource references to external resource identifiers, and verifying the required fields and value ranges.

[0163] When the network modeling module publishes an environment update event and triggers replanning to update the structured specification to V2, the output module compares V1 and V2 to generate an instruction change set, containing only the additions, changes, and removals of affected sections. The module completes the version replacement at the external execution interface, ensuring that unaffected instructions retain references to V1, affected instructions switch to V2, and new reference relationships and verification information are registered in the state and model storage unit.

[0164] Through the above implementation methods, the output module ensures the consistent generation of solution documents and machine-readable instructions through templated and deterministic serialization, achieves traceable and auditable external release through version identification and reference relationships, and completes atomic version replacement after environment updates with a minimal set of instruction changes, thereby providing a stable, standardized and incrementally updatable interface for external execution systems.

[0165] In some embodiments, the system further includes:

[0166] The replanning module is used to receive updated environmental situation information during the execution of the external execution system, so as to trigger the intelligent agent collaborative framework to call the candidate path generation module, resource verification module and decision module to perform replanning and output the replacement target solution.

[0167] Specifically, the replanning module in this embodiment operates collaboratively around the following elements: environmental situation representation update events, affected network range localization, affected subgraph identification within the scheduling graph, incremental calls to the candidate path generation module and resource verification module, target scheme reselection by the decision module, version replacement and change set output, and atomic switching with the external execution interface. The process is as follows: During the operation of the external execution system, an environmental situation representation update event published by the network modeling module is received → the affected subgraph is located within the scheduling graph based on the affected network range and confidence level carried by the event → the intelligent agent collaboration framework is triggered to call the candidate path generation module and resource verification module to perform incremental recalculation on the affected subgraph → the verified candidate path set is submitted to the decision module to generate a new structured description → a change set at the instruction level is generated and version replacement is completed → atomically published through the external execution interface when the safe switching point conditions are met.

[0168] The following explains the concepts of important technical terms used in this embodiment, in detail as follows:

[0169] Update event model: Encapsulate version number, time range, affected network range and confidence level label for a single environment change, which is used to drive the recalculation of the minimum impact domain.

[0170] Affected subgraph location: Based on the reference relationships in the scheduling graph and the binding of segments to resources, the subgraph is extracted along the dependency boundary to ensure that only nodes directly related to the affected resources or time windows are recalculated.

[0171] Safe switch point: refers to the boundary of a segment that has not been started or is within a reversible window. The replanning module only completes the scheme replacement and freezes the references of the executed segment at this boundary.

[0172] Versioning replacement and dual-use strategy: The structured descriptions and instructions of the old and new versions coexist for a short period of time. If the new version meets the verification, it atomically replaces the old version reference; if it fails, it rolls back.

[0173] Change set generation: Using path identifiers and section-level resource references as keys, calculate three types of differences: addition, modification, and revocation, for output modules and external execution interfaces to perform minimal updates.

[0174] Throttling and debouncing: Aggregate continuously updated events by time and range to prevent high-frequency jitter from causing repeated recalculations.

[0175] For example, in some cases, during the execution of a cross-regional multimodal transport mission according to target scheme V1, the network modeling module publishes an update event indicating that the loading and unloading capacity of a transshipment port will decrease within the next 48 hours, accompanied by an increase in the weather rating of the shipping segment. After receiving the event, the replanning module reads the affected network range, locates the affected subgraph formed by the corresponding port node and adjacent sea segments in the scheduling graph, and marks the associated segment-level resource references.

[0176] Subsequently, the replanning module triggers the agent collaboration framework to perform incremental recalculation on the affected subgraph: the candidate path generation module performs only substitution search and encoding update on the affected segments in the unified representation to form revised path branches; the resource verification module performs availability and time window alignment verification on the revised branches under the new resource state representation, and outputs an incremental snapshot of the verified candidate path set; the decision module generates an evaluation vector on the incremental snapshot based on the current evaluation template and task context, and completes the reselection of the target scheme within the feasible frontier set to form a structured description V2.

[0177] The replanning module performs difference calculations on V1 and V2, generating an instruction change set. This set includes resource reference replacements and timing parameter adjustments for the changed sections, along with the old and new version numbers and dependency pointers. To avoid affecting already running sections, the replanning module selects the nearest safe switch point as the replacement boundary. Before this boundary, V1 references remain unchanged; after this boundary, a nuclear switch to V2 is performed. The output module then publishes a minimal change set, which takes effect after the external execution interface completes protocol mapping and parameter verification.

[0178] If the external system returns partial change rejection or resource locking failure, the replanning module executes the rollback strategy based on the receipt: undoes ineffective changes, restores V1 reference relationships, and records the reason for failure and resource usage conflict; at the same time, it waits for the next aggregated update event according to the throttling strategy before initiating recalculation to avoid jitter.

[0179] Through the above implementation, the replanning module triggers incremental recalculation with the affected subgraph as the boundary when the environmental situation representation is updated. It combines a safe switching point and a versioning replacement mechanism to complete the minimum range of scheme updates, ensuring that the reference relationship during execution is stable, external releases are atomic and rollback is possible, and the replanning frequency is controlled by throttling and debouncing strategies. This provides consistent, traceable and controllable replanning capabilities for candidate path generation, resource verification and decision reselection in continuous operation scenarios.

[0180] The above embodiments have described in detail the specific modules and functions of the multi-agent-based multimodal transport scheme generation system of this application. The implementation process of the multi-agent-based multimodal transport scheme generation method of this application will be described in detail below with reference to specific embodiments. Figure 2 This is a flowchart illustrating the method for generating multimodal transport schemes based on multi-agent systems provided in this application. Figure 2 As shown, the method for generating multimodal transport schemes based on multi-agent systems may specifically include the following steps:

[0181] S201: Receives a task description for multimodal transport, extracts constraint parameters related to transport demand to form a task context, and the intelligent agent collaboration framework publishes the task context based on the scheduling graph and coordinates the execution of each functional intelligent agent and the aggregation of intermediate results.

[0182] S202: Acquire data related to the transportation network and its operational status and construct an environmental situation representation. Based on the task context and the environmental situation representation, generate a set of candidate routes that meet the constraints on the multimodal transport network.

[0183] S203, Based on the environmental situation representation, the resource availability and temporal sequence connection of the candidate path set are checked to obtain the checked candidate path set;

[0184] S204: Based on preset evaluation criteria, an evaluation vector is generated from the verified candidate path set, and the reasoning unit containing the language model selects the target solution in combination with the task context, generating a structured description.

[0185] S205 generates solution documents and machine-readable instructions based on structured specifications and provides them to external execution systems through an external execution interface;

[0186] S206, when the environmental situation representation is updated, the agent collaboration framework triggers the candidate path generation module, resource verification module and decision module to perform replanning, output the replacement target solution and structured description, and generate corresponding solution documents and machine-readable instructions.

[0187] Specifically, the multi-agent-based multimodal transport scheme generation method in this embodiment operates under the scheduling graph of the agent collaboration framework, forming a closed loop around data objects such as task context, environmental situation representation, candidate path set, verified candidate path set, evaluation vector, structured description, scheme document and machine-readable instructions.

[0188] The operation flow of this embodiment is as follows: In S201, a task description is received and a task context is formed. The intelligent agent collaboration framework publishes and orchestrates intelligent agents for candidate path generation, resource verification, and decision-making within the scheduling graph. In S202, the network modeling module generates an environmental situation representation and drives the candidate path generation module to produce a candidate path set. In S203, the resource verification module verifies the availability and temporal sequence of the candidate path set based on the environmental situation representation and outputs the verified candidate path set. In S204, the decision-making module generates an evaluation vector based on a multi-dimensional evaluation template, and the inference unit containing a language model selects the target solution in conjunction with the task context, generating a structured description. In S205, the output module generates a solution document and machine-readable instructions based on this and publishes them through an external execution interface. In S206, when the environmental situation representation is updated, the collaboration framework triggers a partial recalculation of candidate path generation, resource verification, and decision-making, outputs the replaced target solution and corresponding document / instruction, and completes the incremental update.

[0189] The following section provides a detailed explanation of the operation flow of the multi-agent-based multimodal transport scheme generation method of this application, with examples. The specific content is as follows:

[0190] S201 Mission Context Release and Orchestration

[0191] In a cross-border transportation task, the input is "Transport 50,000 tons of bulk cargo from a mining area in Indonesia to Lianyungang, China, with a delivery time of 6 weeks, and priority given to transiting through Singapore." The task parsing module extracts the task context from the description. Hard constraints include the delivery time and origin / destination, while soft constraints include mode preferences and cost weights. The agent collaboration framework publishes the task context as the root input to the scheduling graph, establishes a dependency chain of candidate path generation → resource verification → decision → output, and configures concurrently executable branches for the path generation and resource verification nodes.

[0192] S202 Environmental Situation Representation and Candidate Path Set Generation

[0193] The network modeling module integrates data such as port operations, liner / train schedules, road traffic restrictions, and weather and sea conditions to achieve spatiotemporal alignment and conflict resolution, outputting a snapshot V1 representing the environmental situation. The candidate path generation module performs constraint-driven expansion and pruning on a unified representation, generating several feasible paths that satisfy temporal consistency, accessibility, and mode switching rules. These paths are then standardized, deduplicated, and scaled to form a candidate path set snapshot P1, which is then registered and referenced.

[0194] S203 Resource Availability and Timing Alignment Verification

[0195] The resource verification module generates a resource status representation based on the environmental situation representation V1 and establishes a correlation between the path segments in P1 and the resource status. It performs capacity availability, time window alignment, and mode switching compatibility checks on each path. Paths that do not meet hard constraints are eliminated; conflicting branches with soft constraints are marked as needing revision and local substitution and remapping are performed within the dependency limits allowed by the scheduling graph, generating revised paths and updating segment-level resource references. A snapshot C1 of the verified candidate path set is output, recording segment-level resource references and connection parameters.

[0196] S204 Evaluation Vector Calculation and Target Scheme Selection

[0197] The decision-making module instantiates a multidimensional evaluation template based on the task context, generates standardized evaluation vectors for each C1, and constructs a feasible frontier set. The inference unit, which includes a language model, selects target solution B from the frontier set under the premise of satisfying the task context constraints and the consistency of the environmental situation representation, and generates a structured description D1, which includes path identifiers, segment-level resource references, temporal connection parameters, and dependent snapshot versions.

[0198] S205 Solution Documentation and Machine-Readable Instruction Output

[0199] The output module generates solution documents based on the template library called by D1, and constructs a machine-readable instruction set covering instructions such as booking, port operation reservation, and train / vehicle dispatching; it performs deterministic serialization of documents and instructions and assigns version number O1, writes reference relationships and verification information to storage; and publishes them to the external execution system after completing protocol mapping and parameter verification through the external execution interface.

[0200] Local replanning triggered by the S206 update event

[0201] During execution, the network modeling module detected an increase in wind and wave levels in a specific sea area and that the congestion index of a certain transshipment port exceeded the threshold, generating update event E1 and marking the affected network range. The collaborative framework located the affected subgraph within the scheduling graph, triggering incremental recalculation of candidate path generation and resource verification, forming a new candidate path set snapshot P2 and a verified snapshot C2; the decision module generated a structured description D2 on the new evaluation vector. The output module generated an instruction change set based on the differences between D2 and O1 and implemented version replacement at the nearest safe switch point, releasing the new version O2; unaffected segments continued to use the old version, ensuring atomic switching and rollback capability of external interfaces.

[0202] This embodiment, driven by a scheduling graph and involving multi-agent collaboration, completes an end-to-end method flow from task context release, environmental situation representation construction, candidate path generation and resource verification, evaluation vector calculation and target scheme selection to document / instruction output and event-triggered replanning. Deterministic transfer of data objects is achieved between each step through versioned snapshots and traceable references, and incremental updates are implemented with the affected subgraph as the boundary, supporting method-level closed-loop operation in continuous execution scenarios.

[0203] It should be understood that the sequence number of each step in the above method embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0204] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although the technical solutions of this application have been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A multi-agent based intermodal transportation scheme generation system, characterized by, The method comprises the following steps: a task analysis module is used to receive a task description oriented to multimodal transport, extract constraint parameters related to transport demand, and form a task context; an agent collaboration framework is used to decompose and allocate the task context based on a scheduling graph, and coordinate the execution of each functional agent and the aggregation of intermediate results; a network modeling module is used to obtain data related to the transport network and operating status and construct an environmental situation representation; a candidate path generation module is used to generate a set of candidate paths that meet the constraints on the multimodal transport network based on the task context and the environmental situation representation; a resource checking module is used to check the resource availability and timing connection of the set of candidate paths based on the environmental situation representation, and obtain a set of checked candidate paths; a decision module is used to generate an evaluation vector for the set of checked candidate paths based on a preset evaluation criterion, select a target scheme in combination with the task context, and generate a structured explanation; an output module is used to generate a scheme document and machine-readable instructions based on the structured explanation, and provide the scheme document and machine-readable instructions to an external execution system; wherein the candidate path generation module, resource checking module, and decision module are implemented as functional agent nodes scheduled by the agent collaboration framework, and have a tool calling interface to access a multi-source data interface; wherein the task analysis module is used to: use a parsing unit containing a language model to perform intent recognition and element extraction on the task description, and structure the extraction results as the task context; hierarchically encode the constraint parameters in the task context as hard constraints and soft constraints, and unify them into executable constraint expressions; perform consistency and reachability checks on the constraint expressions, and perform default inference on missing parameters based on a rule base and boundary conditions; add confidence and source identification to the constraints with uncertainty, and publish the formed task context to the candidate path generation module and the decision module through the agent collaboration framework.

2. The system of claim 1, wherein, The agent collaboration framework is used to: establish a set of subtasks with dependency and constraint annotations based on the task context and environmental situation representation within a scheduling graph, and map the subtasks to executable units of functional agents; configure execution priority and concurrency for functional agents according to the dependencies and constraints, and maintain data and constraint consistency through constraint propagation and conflict detection during execution; aggregate intermediate results generated by each functional agent according to a traceable reference mechanism for reference by incomplete subtasks; trigger local re-planning when the environmental situation representation is updated, determine the affected subtask boundaries, and complete recalculation and replacement within the scheduling graph; configure access routes for functional agents to multi-source data interfaces and computing resources through a tool calling interface.

3. The system of claim 1, wherein, The network modeling module is used to: access heterogeneous data related to the topology and operating status of the transport network through a multi-source data interface, and complete format normalization and space-time alignment in a unified mode to generate standardized records keyed by space-time; The freshness control and conflict mediation are performed on the standardized record based on the validity period and the update priority, an environmental situation representation with source identification and confidence annotation is formed; An incremental fusion mechanism is adopted to versionize the environmental situation representation, and a snapshot with version identification is output for deterministic reference of the candidate path generation module and the resource checking module; Rule-driven completion and uncertainty quantification are performed on missing or uncertain entries under preset boundary conditions, and the completion results are written into the environmental situation representation; According to the change threshold and the trigger rule, an update event is generated from the environmental situation representation, the affected network range is marked, and the intelligent agent cooperation framework is released to drive re-planning.

4. The system of claim 1, wherein, The candidate path generation module is configured to: Parameterize mapping of the multimodal transport network based on the environmental situation representation, and form a unified representation for path search; Perform constraint-driven path expansion and pruning on the unified representation according to the constraints of the task context, including timing consistency, reachability and transport mode switching rules, to generate feasible paths from the starting point to the ending point; Standardize the generated feasible paths and remove duplicates, and implement scale control, and retain a limited number of candidate path sets for subsequent reference; When receiving the environmental update event released by the network modeling module, the affected path segments are located and updated incrementally, and a candidate path set snapshot with version identification is output for deterministic reference of the resource checking module.

5. The system of claim 1, wherein, The resource checking module is configured to: Form a resource state representation based on the environmental situation representation and associate it with the segments of the candidate path set; Perform capacity availability, time window alignment and mode switching compatibility checking according to the hard and soft constraints of the task context, and remove paths that do not meet the hard constraints and mark soft constraint conflicts for revision; Locally replace and remap the marked items within the dependency range allowed by the scheduling graph to generate revised paths; Output a versioned snapshot of the checked candidate path set with segment-level resource references and connection parameters, and perform incremental re-checking when the environment is updated for reference by the decision module.

6. The system of claim 1, wherein, The decision module is configured to: Based on the preset evaluation criteria and the task context, a configurable multi-dimensional evaluation template is constructed, and a standardized evaluation vector is generated for each of the checked candidate path set; According to the evaluation vector, the dominance relationship of the candidate paths is established and a feasible frontier set is formed, and the candidate paths not contained in the frontier set are marked as objects to be selected; Under the premise of meeting the constraints of the task context and the consistency of the environmental situation representation, the inference unit containing the language model selects the target scheme from the feasible frontier set, and outputs a structured explanation containing path identification, segment-level resource reference and timing connection parameters.

7. The system of claim 1, wherein, The output module is configured to: Invoke a document and instruction template library based on the structured explanation, generate a scheme document containing path identification, segment-level resource reference and timing connection parameters, and construct a corresponding machine-readable instruction set; Determine the sequence and version of the scheme document and machine-readable instructions, and write the reference relationship and verification information into the state and model storage unit; The machine readable instructions are published to an external execution system through an external execution interface, and protocol mapping and parameter checking are completed; When the environment situation representation or target scheme is updated, an instruction change set is output and version replacement is performed, and an update is provided to the external execution system.

8. The system of claim 7, wherein, The system further comprises: A re-planning module configured to receive updated information of the environment situation representation during execution of the external execution system, to trigger the agent collaborative framework to call the candidate path generation module, resource checking module and decision module to perform re-planning, and to output a replaced target scheme. 9.A multi-agent based intermodal transportation scheme generation method based on the system according to any one of claims 1 to 8, characterized in that, Comprise: Receiving a task description oriented to multimodal transport, extracting constraint parameters related to transportation demand to form a task context, and publishing the task context based on a scheduling graph and coordinating execution and intermediate result aggregation of each functional agent by the agent collaborative framework; Obtaining data related to the transportation network and running state and constructing an environment situation representation, and generating a set of candidate paths that meet the constraints on the multimodal transport network based on the task context and the environment situation representation; Based on the environment situation representation, the set of candidate paths is checked for resource availability and timing connection, and a set of checked candidate paths is obtained; Based on a preset evaluation criterion, an evaluation vector is generated for the set of checked candidate paths, and a target scheme is selected by an inference unit comprising a language model in combination with the task context to generate a structured explanation; Based on the structured explanation, a scheme document and machine readable instructions are generated and provided to an external execution system through an external execution interface; When the environment situation representation is updated, the agent collaborative framework triggers the candidate path generation module, resource checking module and decision module to perform re-planning, outputs a replaced target scheme and structured explanation, and generates a corresponding scheme document and machine readable instructions.

Images