Urban traffic cooperative scheduling method and system based on large model and multiple agents

Abstract

The invention discloses an urban traffic cooperative scheduling method and system based on a large model and multiple agents. The method comprises the steps that natural language task description is converted into a task semantic graph and a structured prompt; generating a plurality of strategy candidates by using a large language model, and performing language scoring and reasoning arbitration; based on the scoring result, selecting an optimal strategy for multi-agent execution; behavior execution data are collected and evaluated and fed back, and strategy closed-loop updating is achieved through parameter optimization and Prompt fine tuning; and a distributed task embedding mechanism and a lightweight migration module are combined, so that the multi-task adaptability and the training efficiency of the system are improved. The cooperative training system constructed by the invention has the capabilities of natural language interaction, strategy interpretability, behavior controllability and task migration, and is suitable for various complex urban tasks such as traffic jam dispersion, emergency response, signal lamp linkage and the like.

Description

Technical Field

[0001] This invention belongs to the field of urban traffic scheduling technology, and in particular relates to an urban traffic collaborative scheduling method and system based on a large model and multiple agents. Background Technology

[0002] With the continuous advancement of urbanization, transportation systems are increasingly becoming key infrastructure affecting urban operational efficiency, public safety, and residents' quality of life. Particularly in large cities and transportation hubs, problems such as uneven distribution of transportation resources, frequent emergencies, and concentrated traffic pressure have long existed. Traditional traffic control systems have revealed limitations in handling complex task scenarios, including slow response, lack of flexibility, and insufficient intelligence. Currently, in tasks such as traffic emergency response, cross-regional dispatching, and real-time congestion mitigation, there are common pain points such as rigid command patterns, fragmented agent behavior, and coarse-grained task allocation, making it difficult for dispatching systems to achieve efficient coordination in dynamic, changing, and high-density urban traffic scenarios.

[0003] Traditional urban traffic control relies primarily on static rules and expert experience to set signal cycles, adjust route priorities, or generate area closure plans. While this approach generally offers a degree of operability and stability, it often suffers from delayed responses, scheduling delays, and fragmented information when faced with unexpected events (such as chain reactions, overlapping congestion on multiple road segments, or traffic signal failure due to severe weather). Furthermore, the lack of unified behavioral modeling and intelligent interaction channels among different types of traffic participants (such as traffic lights, police patrol cars, and autonomous guided vehicles) prevents the system from achieving collaborative cooperation among agents, thus weakening the efficiency of emergency dispatch and the system's responsiveness.

[0004] In recent years, Multi-Agent Systems (MAS) have been gradually introduced into the fields of traffic simulation and intelligent scheduling. These systems attempt to abstract control nodes (such as traffic lights), execution nodes (such as patrol cars), and information nodes (such as traffic monitoring sensors) in road networks into distributed agents, each possessing partial perception, decision-making, and behavioral capabilities. By introducing methods such as reinforcement learning, adaptive planning, and cooperative game theory, these agents can make local adjustments in response to traffic changes. However, most MAS systems still face problems such as fragmented policies and a lack of overall scheduling logic, exhibiting significant shortcomings in tasks involving semantic reasoning, policy fusion, and large-scale collaboration. Furthermore, many systems lack high-level interaction capabilities with human users, unable to directly express task intentions or modify policy logic through natural language, reducing the system's controllability and versatility in actual scheduling.

[0005] In contrast, Large Language Models (LLMs) have achieved significant breakthroughs in natural language understanding, semantic reasoning, and structured task generation. In many general domain applications, LLMs demonstrate the ability to transform fuzzy language tasks into actionable processes and support the interpretation, partitioning, and semantic optimization of complex task logic. This capability offers a new approach for urban transportation systems: leveraging LLMs to handle task parsing and policy organization in scheduling systems, transforming natural language input from human users into structured policy prompts, thereby guiding the underlying multi-agent system to collaboratively complete various complex traffic tasks.

[0006] Currently, there have been a few attempts to combine language models with traffic simulation to generate simulation scenarios, describe task logic, or assist in generating scheduling rules. However, most of these efforts remain at the prototype stage, lacking a complete scheduling system design capable of closed-loop execution. Especially in real urban traffic environments with high density, multiple types, and multiple tasks running simultaneously, how to integrate the strategies generated by language models with multi-agent behavioral mechanisms to achieve both global consistency and local adaptation and distributed scheduling remains a key technical challenge that has not been fully resolved. In addition, current systems generally lack interpretable policy output capabilities, making it difficult for users to trace the policy logic behind the system's scheduling and to quickly correct or intervene in erroneous behaviors, thus limiting both the system's practicality and security.

[0007] Therefore, there is an urgent need for a new traffic scheduling method that integrates natural language understanding and intelligent behavior coordination capabilities. This method should be able to quickly parse task semantics, generate multiple policy candidates, and realize task allocation and behavioral linkage among multiple types of intelligent agents in highly complex traffic scenarios. It should also have scalable, interpretable, and sustainable optimization capabilities to adapt to the ever-changing urban operation needs and traffic challenges. Summary of the Invention

[0008] This invention provides a method and system for urban traffic collaborative scheduling based on large models and multiple agents, which can achieve higher global efficiency and lower operating costs in core urban traffic tasks such as congestion relief, accident response, signal linkage, fleet guidance and cross-regional collaboration.

[0009] A method for coordinated urban traffic scheduling based on large models and multi-agent systems includes the following steps: (1) Receive natural language scheduling instructions input by the dispatcher and organize them into a task semantic graph; and summarize the road network structure, real-time traffic, event alarms, control rules and historical execution records and organize them into structured prompts; (2) Based on the task semantic graph and combined with structured prompts, multiple text-based candidate solutions and multiple structured candidate solutions are generated using a large language model and a procedural method, and all the obtained solutions are merged into the candidate pool. (3) Convert each solution in the candidate pool into a unified execution format, including sub-tasks, completion conditions and an execution template; among them, sub-tasks and completion conditions are used for reinforcement learning recognition of the three intelligent agents: traffic light control, vehicle guidance and area scheduling. (4) Use a large language model to perform semantic scoring on each solution in the candidate pool, obtain the semantic score, confidence and constraints of each solution, and output the explanation of the solution; (6) Incorporate the semantic score into the total reward so that the total reward includes the environmental reward and the semantic score, and automatically adjust the weight of the semantic score according to the confidence level, task stage and historical error; at the same time, use the confidence level to correct the value baseline and convert the constraint into a consistency penalty and add it to the total reward. (6) Select an optimal solution based on semantic score, confidence level and total reward after adding consistency penalty; or, use a fusion strategy to combine the advantageous segments of different candidates to obtain the optimal solution and output the corresponding explanatory text. (7) Using the CTDE algorithm, the traffic light control agent, vehicle guidance agent and area scheduling agent are trained in a centralized manner and executed separately. The strategy is implemented according to the execution template of the optimal solution, and the status, actions and rewards of the agents are recorded. (8) Collect training data in parallel in multiple environments, perform distributed training and continuous learning, and update parameters in a unified manner to accelerate the training process of the CTDE algorithm; (9) During training and execution, the optimal solution output is evaluated in real time, and the structured prompts, semantic score weights and consistency penalty weights are optimized based on the evaluation results and problem fragments; and the final control instructions are output.

[0010] In step (2), the procedural approach includes rule-based planning, finding similar solutions from historical cases and modifying parameters, and using a lightweight strategy to provide a draft time sequence.

[0011] In step (3), the execution template includes execution order, conditions and node parameters. The node parameters include at least signal period and phase sequence, green wave synchronization window, lane release weight, temporary blockade duration, detour priority and bus priority; and correspond to each action space of each agent.

[0012] In step (5), the formula for the total return is as follows: ; In the formula, For total return, In return for the environment, For semantic scoring, The weights of the semantic score are automatically adjusted based on confidence level, task stage, and historical error; the task stage includes congestion relief, incident response, and activity evacuation; historical error is the deviation between the target and the actual effect in the historical execution record.

[0013] In step (5), converting constraints into consistency penalties includes: adding penalties for phase snatching, route blocking, and boundary crossing.

[0014] In step (6), the fusion strategy includes (a) Time-segmented splicing: Different candidate advantageous segments are used in different time windows; (b) Hierarchical merging: Merging sub-target hierarchies without conflict; (c) Voting weighting: Weighting is performed using consistency score or confidence level as the weight.

[0015] In step (7), a centralized training and separate execution approach is adopted, specifically as follows: During training, each agent shares or partially shares backbone parameters to utilize more comprehensive information for evaluation; during execution, each agent makes independent decisions to balance global consistency and local response speed.

[0016] In step (9), the optimal solution is evaluated based on indicators such as congestion relief, average delay, green wave quality, strategy consistency, response delay, resource efficiency, and anomaly rate. The final control commands include signal phase / green wave parameters, vehicle guidance paths, and lockdown / release strategies.

[0017] A city traffic cooperative scheduling system based on a large model and multiple agents, characterized in that it includes a memory and one or more processors, wherein the memory stores executable code, and the one or more processors execute the executable code to implement the city traffic cooperative scheduling method according to any one of claims 1-8.

[0018] Compared with the prior art, the present invention has the following beneficial effects: This invention uses a large language model as a high-level semantic hub to parse the dispatcher's natural language intent into structured task semantics and high-level policies; it uses multi-agent reinforcement learning as the execution and learning framework to perform collaborative decision-making and optimization in sub-tasks such as signal phase control, convoy guidance, and area release / blockade; and it achieves interpretable, controllable, and stable convergence of policy improvement through deep coupling of language arbitration signals and reinforcement learning objective functions (reward shaping / value correction / consistency constraints); at the same time, relying on the distributed training and continuous learning mechanism of Actor-Learner decoupling, it achieves rapid migration and low-cost deployment across time periods, events, and regions. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of an urban traffic collaborative scheduling method based on a large model and multiple agents, according to an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] It should be noted that, unless otherwise specified, the features in the following embodiments and implementation methods can be combined with each other.

[0023] like Figure 1 As shown, a collaborative scheduling method for urban traffic based on a large model and multiple agents includes the following steps: S1. Obtain Task and Context: Receive natural language scheduling instructions and summarize road network structure, real-time traffic, event alarms, control rules and historical execution records as subsequent inputs.

[0024] S2. Semantic parsing and structured representation: The above information is organized into a task semantic graph and structured prompts (including target area, time limit, priority, available resources, constraints and precautions) using a large language model, providing a unified foundation for subsequent automatic solution generation.

[0025] S3. Generate multiple candidate strategies at once: Under the same instruction, generate multiple candidates simultaneously using two methods. (1) Generate multiple textual candidates in the large language model by "rewriting prompts + changing sampling parameters"; (2) Generate multiple structured candidates by calling simple procedural methods (such as rule-based planning, finding similar solutions from historical cases and changing parameters, and using lightweight strategies to provide a time sequence draft); the two are combined to form a "candidate pool".

[0026] The generation of multiple candidates must include at least: (a) Rewrite the prompts (rephrase the objectives and constraints or rearrange their order); (b) Change the sampling parameters (temperature, kernel probability, etc.); (c) Align with historical nearest neighbor schemes (find similar scenarios from the case library and replace critical roads and time windows); (d) Provide a time / phase draft using simple rules or lightweight strategies; thereby ensuring that the candidates are both diverse and executable.

[0027] S4. Candidates are unified into an executable format: Each solution in the candidate pool is converted into a unified execution format, including: subtasks and completion conditions that can be recognized by reinforcement learning (corresponding to "options" and "termination conditions"), and an execution template (which can be expressed by a flowchart / state machine: the triggering conditions, actions to be performed, and when to end for each step).

[0028] The unified executable format includes: (a) Write the sub-goals and constraints as “sub-tasks + completion conditions”; (b) Write the execution order and conditions as an “execution template” (a flowchart / state machine can be used). The node parameters (the key control quantities required for each step of the template) should include at least the signal period and phase sequence, green wave synchronization window, lane release weight, temporary blockade duration, detour priority, and bus priority; and correspond to each action space of each agent.

[0029] S5. Score each candidate and provide reasons: Use a large language model to perform language scoring, checking for consistency, compliance with task objectives and traffic rules; rank the candidates based on historical preferences (trained from human feedback or historical results); obtain three things: semantic score ( The semantic score measures whether the solution is reasonable and aligned with the goal, while the confidence score measures how reliable the solution is.

[0030] S6. Transform scoring into numerical data usable by reinforcement learning: Incorporate semantic scores into the total reward, enabling reinforcement learning to "understand" the language-level evaluation; the calculation formula is as follows: ; in, In return for the environment, Score the language. The weights are automatically adjusted based on confidence level, task stage (referring to the stage of the response, such as congestion relief / accident response / event evacuation), and historical error (referring to the recently statistical "deviation between the target and the actual effect"), so that effective strategies can still be learned quickly even when environmental rewards are sparse or delayed.

[0031] At the same time, confidence is used to correct the value baseline / advantage estimate (a type of quantity used in reinforcement learning to reduce training variance / fluctuation and provide an "average level" reference to judge whether the current action is better, commonly implemented as GAE or TD), and the "situations that should not happen" (constraints given in step S5) are transformed into consistency penalties and added to the loss function, so that the training converges in a more stable and compliant direction.

[0032] S7. Select or merge a final strategy and provide an explanation: Based on the scoring and the results of the above numerical coupling ("scoring" refers to the semantic score and confidence in step S5; "results of numerical coupling" refers to the calculation results after incorporating the language signal into the reward / value and adding the constraint penalty in step S6), either directly select the optimal strategy or combine the advantageous segments of different candidates (e.g., using different candidates in different time periods, or merging sub-goals when there is no conflict), and output the corresponding explanatory text.

[0033] The strategy fusion adopts at least one method: (a) time-segmented splicing (using different candidate advantage segments in different time windows); (b) hierarchical merging (merging sub-target hierarchies without conflict); (c) voting weighting (weighting with consistency score or confidence level as weight); and automatically generates corresponding explanatory text for recording and auditing.

[0034] S8. Multi-agent execution and data collection: Under the "centralized training, separate execution" approach ("centralized training" means that global information or a centralized value function can be used during training; "separate execution" means that each agent makes independent decisions based solely on local observations during online execution) (more global information is available during training, and each intersection / vehicle makes quick decisions using its own local information during execution), agents such as traffic light control, vehicle guidance, and area scheduling implement strategies according to the execution template and record data such as status, actions, rewards, and advantages.

[0035] S9. Distributed Training and Continuous Learning: Adopting a "separate sampling and learning" architecture, data is collected in parallel across multiple environments (Actors), and parameters are uniformly updated by a central learner (Learner) to accelerate the training of the CTDE algorithm in step S8 (here, "distributed" refers only to the computational method of parallel sampling and centralized updating, not "training three agents"; it belongs to a different dimension from CTDE in step S8 and can be used simultaneously); a task code tells the model which region, time period, and task it is currently in; important parameters are protected, and labeled historical data is replayed to avoid forgetting the basics when changing tasks; when a new task is launched, only a few minor parameter modules need to be fine-tuned for adaptation.

[0036] Distributed training specifically includes: (a) The sampling end and the learning end are separated, and multiple processes collect data in parallel, while the parameters are updated centrally. (b) Use a task code to tell the model "where, when, and what" so as to enable robust migration when switching between different regions / time periods / tasks; (c) Protect important parameters (do not easily change parameters that have been proven to be important to historical tasks), and replay tagged historical segments in layers according to task / time period / region to avoid forgetting; (d) Quickly adapt to new tasks by fine-tuning small parameter modules (adjusting only a small number of new parameters or unfreezing only a few layers).

[0037] S10. Evaluate and Feedback to Form a Closed Loop: Evaluate based on indicators such as congestion mitigation, average delay, green wave quality, strategy consistency, response latency, resource efficiency, and anomaly rate; feed back the evaluation results and "problem segments" (time segments in the execution trajectory that are ineffective, violate constraints, or cause anomalies) to steps S2, S5, and S6, automatically or semi-automatically optimizing the prompt structure and scoring weights. The penalty weights are adjusted until stability is achieved. The output control commands should at least cover: signal phase and phase sequence, green wave linkage parameters, variable speed limits / variable lanes, vehicle guidance paths, temporary closures / releases, bus priority and emergency lanes, and should be integrated with urban traffic control or simulation systems through a unified interface.

[0038] In this invention, to ensure traceability, each candidate is recorded in a log in groups from "generation source → unified mapping → scoring and constraints → selection / fusion → execution effect", which can be used to reproduce experiments and perform offline playback analysis.

[0039] Candidate generation (step S3), scoring (step S5), numerical coupling (step S6), and selection / fusion (step S7) correspond one-to-one: the source, score, confidence level, and constraint basis of each executed or spliced ​​content can be traced in the log.

[0040] This invention provides three entry points for manual intervention: (a) modify prompts (directly edit structured prompts or constraint text and apply immediately); (b) select strategies (specify adoption / exclusion from multiple candidates); and (c) correct the trajectory (provide scores or corrective suggestions for problematic segments during execution). All interventions are recorded and used to update historical preferences and constraint templates.

[0041] This invention sets safety boundaries and degradation mechanisms, such as minimum phase switching intervals, permissible deviations for green wave synchronization, maximum vehicle speed and minimum turning radius, and minimum / maximum duration of roadblocks. In the event of communication anomalies, sensor failures, or secondary accidents, it immediately selects the safety-priority candidate or triggers a degradation strategy, and automatically reduces the risk level. Increase consistency penalties.

[0042] Based on the same inventive principle, this invention also provides an urban traffic collaborative scheduling system based on a large model and multiple agents, including a memory and one or more processors. The memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the urban traffic collaborative scheduling method mentioned in the above embodiments.

[0043] Based on the same inventive principle, this invention also provides a city traffic collaborative scheduling system based on a large model and multiple agents. To achieve the objectives of the above steps, the system provided by this invention consists of seven functional modules: a natural language task input and parsing module; a high-level policy generation and option mapping module; a language arbitration and reinforcement learning fusion module; a multi-agent reinforcement learning execution and collaborative decision-making module; a distributed training and multi-task transfer module; a behavior recording, evaluation, and closed-loop optimization module; and a platform, interface, and human-in-the-loop module. These modules are interconnected in terms of data and control flow, jointly supporting the deep integration and engineering implementation of LLM×MARL.

[0044] Module 1: Natural Language Task Input and Parsing Module. This module addresses natural language commands from dispatchers or upper-level business systems. It combines context such as road network topology, real-time traffic, and accident alarms to construct structured prompts and task semantic graphs. The parsing process standardizes and extracts elements such as "target area, timeliness requirements, priority, resource constraints, and control rules," forming a graph representation composed of task nodes, resource nodes, and constraint edges. This graph serves as a stable intermediate layer between LLM and RL. On one hand, the semantic graph provides a computable structural context for subsequent high-level planning in LLM; on the other hand, constraints in the graph can be directly mapped to the conditions and termination rules of behavior trees and state machines, reducing the exploration cost and uncertainty of RL in the transformation from complex semantics to control commands. Through this module, the first cross-domain bridge from natural language to "executable semantics" is achieved, providing a unified and auditable task baseline for subsequent modules.

[0045] Module Two: High-Level Policy Generation and Option Mapping Module. Based on the task semantic graph and structured prompt, LLM generates multiple candidate high-level policies, including sub-objective breakdown, role allocation (signal control unit, guidance / patrol vehicles, etc.), execution order, collaborative constraints, and expected risk descriptions. To ensure that high-level language policies can be stably consumed by RL, this module maps the LLM output to "options" and termination conditions in hierarchical reinforcement learning (HRL / Options), and simultaneously produces behavior tree / state machine templates. Thus, "natural language → high-level sub-objectives / options → executable templates" forms a continuous link, preserving interpretability at the language level while providing MARL with a clearly structured goal prior and an implementable interface. In the context of urban traffic, these options can correspond to "forming L3–L4 green wave zones," "prioritizing the clearing of the East Second Ring Road entrance ramps," and "implementing temporary closures and guiding detours for accident-affected sections," etc.

[0046] Module 3: Language Arbitration and Reinforcement Learning Integration Module. This module performs language scoring and CoT consistency tests on the multi-candidate high-level policies output by Module 2, and introduces Human Preference Learning (RLHF) to obtain a preference model for traffic governance. Unlike traditional LLM that only provides hints, this invention directly injects the arbitration results into the objective function of reinforcement learning: firstly, by constructing semantic reward shaping terms. Environmental returns according to First, synthesis significantly alleviates the exploration difficulties under sparse or delayed rewards. Second, the confidence / risk assessment of language arbitration is used for value baseline correction, entering into advantage estimation or TD objectives, reducing policy variance and accelerating convergence. Third, the causal chain and semantic constraints of the thought chain are transformed into consistency regularization terms, imposing penalties on trajectory segments that violate traffic rules or conflict with collaborative objectives. These three paths elevate LLM from a "describer / scriptwriter" to a differentiable / semi-differentiable policy evaluator and constraint generator, enabling the "interpretation and preference" of the language layer to numerically enter the optimization process of RL, achieving interpretable and controllable collaborative learning.

[0047] Module Four: Multi-Agent Reinforcement Learning Execution and Collaborative Decision-Making Module. This module uses Multi-Agent Reinforcement Learning (MARL) as the core framework for execution and learning, employing a centralized training / distributed execution (CTDE) paradigm. It supports algorithm families such as MAPPO, QMIX, and VDN to adapt to different coupling strengths and observability levels. High-level options and constraints from Modules Two and Three are implemented as executable control signals such as signal phase timing, green wave linkage parameters, vehicle guidance paths, and area release / lockdown commands through behavior trees / state machines; and language arbitration... Value correction and consistency regularization are incorporated into the policy improvement process, becoming crucial components of the loss function or advantage estimation. During training, agents share or partially share the policy backbone to enhance collaborative consistency; during execution, each agent makes independent decisions based on local observations, relying on higher-level options to achieve global coordination. This module ensures stable collaboration and global efficiency in highly dynamic and uncertain traffic environments.

[0048] Module 5: Distributed Training and Multi-Task Transfer Module. To support large-scale simulations and rapid deployment, this module adopts a distributed training system with decoupled Actor-Learner architecture. Multiple Actors sample in parallel, multiple environments run asynchronously, and a central Learner aggregates and updates the data. Task embedding vectors are introduced to dynamically modulate the policy network, enabling the model to distinguish between different time periods (morning / evening rush hours), different regions (main / secondary arterial roads), and different task types (diversion / lockdown / public transport priority, etc.). Elastic Weight Retention (EWC) and labeled experience replay mitigate catastrophic forgetting, allowing the reuse of old knowledge when introducing new tasks. Combined with efficient transfer strategies such as LoRA / module freezing / partial weight inheritance, adaptation to new regions / events is completed within small samples and short time windows. This module ensures the system possesses engineering capabilities for continuous learning, low-cost transfer, and stable convergence.

[0049] Module Six: Behavior Recording, Evaluation, and Closed-Loop Optimization Module. This module records the trajectory of state-action-reward-advantage in a structured manner, establishing a multi-dimensional indicator system for traffic governance, including congestion relief rate, coordination consistency, response latency, resource efficiency, and signal switching smoothness, and provides playback and visualization capabilities. Evaluation results are used to adjust hyperparameter adaptation (such as learning rate, λ coefficient, and constraint weights), and are also fed back into prompt orchestration and language arbitration to update the prompt structure, constraint templates, and preference model. This enables the system to achieve a self-consistent closed loop of semantics-policy-execution-evaluation during continuous execution, improving long-term steady-state performance and auditability.

[0050] Module Seven: Platform, Interfaces, and Human-in-the-Loop Module. The platform layer provides plug-and-play integration with simulated or real-world interfaces such as SUMO, CityFlow, and CARLA. It defines unified communication protocols and action interfaces, and uses behavior tree / state machine templates to standardize the timing and conditions of the execution side. It supports hybrid cloud-edge-device deployment and microservice management to meet the requirements of large-scale concurrency and low latency. Regarding the human-in-the-loop, the platform offers three types of interfaces: prompt-level intervention, strategy selection intervention, and trajectory correction. Dispatchers can edit constraints, specify candidate solutions, or correct deviations during execution. The system also retains evidence chains such as thought chains, semantic scores, and constraint sources to meet the safety compliance and audit traceability requirements in traffic management. Through this module, the system strives for adaptive intelligence while ensuring explainability, controllability, and feasibility.

[0051] The system of this invention achieves a deep integration of large language models and multi-agent reinforcement learning through the organic synergy of the aforementioned seven modules: it bridges the "language-control" structure with task semantic graphs and option mappings; it transforms "interpretation and preference" into optimizable numerical objectives by incorporating language arbitration signals into rewards, values, and constraints; and it supports cross-task and cross-regional engineering deployment through distributed continuous learning and efficient parameter transfer. Compared with existing technologies, this invention achieves significant improvements in convergence stability, collaborative consistency, transfer efficiency, interpretability, and controllability. It can achieve higher global efficiency and lower operating costs in core urban traffic tasks such as congestion mitigation, accident response, signal coordination, fleet guidance, and cross-regional collaboration, demonstrating broad application prospects and industrial transformation value.

[0052] The embodiments described above provide a detailed explanation of the technical solutions and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for coordinated urban traffic scheduling based on a large model and multiple agents, characterized in that, Includes the following steps: (1) Receive natural language scheduling instructions input by the dispatcher and organize them into a task semantic graph; and summarize the road network structure, real-time traffic, event alarms, control rules and historical execution records and organize them into structured prompts; (2) Based on the task semantic graph and combined with structured prompts, multiple text-based candidate solutions and multiple structured candidate solutions are generated using a large language model and a procedural method, and all the obtained solutions are merged into the candidate pool. (3) Convert each solution in the candidate pool into a unified execution format, including sub-tasks, completion conditions and an execution template; among them, sub-tasks and completion conditions are used for reinforcement learning recognition of the three intelligent agents: traffic light control, vehicle guidance and area scheduling. (4) Use a large language model to perform semantic scoring on each solution in the candidate pool, obtain the semantic score, confidence and constraints of each solution, and output the explanation of the solution; (6) Incorporate the semantic score into the total reward so that the total reward includes the environmental reward and the semantic score, and automatically adjust the weight of the semantic score according to the confidence level, task stage and historical error; at the same time, use the confidence level to correct the value baseline and convert the constraint into a consistency penalty and add it to the total reward. (6) Select an optimal solution based on semantic score, confidence level and total reward after adding consistency penalty; or, use a fusion strategy to combine the advantageous segments of different candidates to obtain the optimal solution and output the corresponding explanatory text. (7) Using the CTDE algorithm, the traffic light control agent, vehicle guidance agent and area scheduling agent are trained in a centralized manner and executed separately. The strategy is implemented according to the execution template of the optimal solution, and the status, actions and rewards of the agents are recorded. (8) Collect training data in parallel in multiple environments, perform distributed training and continuous learning, and update parameters in a unified manner to accelerate the training process of the CTDE algorithm; (9) During training and execution, the optimal solution output is evaluated in real time, and the structured prompts, semantic score weights and consistency penalty weights are optimized based on the evaluation results and problem fragments; and the final control instructions are output.

2. The urban traffic cooperative scheduling method based on a large model and multiple agents according to claim 1, characterized in that, In step (2), the procedural approach includes rule-based planning, finding similar solutions from historical cases and modifying parameters, and using a lightweight strategy to provide a draft time sequence.

3. The urban traffic collaborative scheduling method based on a large model and multiple agents according to claim 1, characterized in that, In step (3), the execution template includes execution order, conditions and node parameters. The node parameters include at least signal period and phase sequence, green wave synchronization window, lane release weight, temporary blockade duration, detour priority and bus priority; and correspond to each action space of each agent.

4. The urban traffic cooperative scheduling method based on a large model and multiple agents according to claim 1, characterized in that, In step (5), the formula for the total return is as follows: ； In the formula, For total return, In return for the environment, For semantic scoring, The weights of the semantic score are automatically adjusted based on confidence level, task stage, and historical error. The task phases include congestion mitigation, incident response, and activity evacuation; historical error is the deviation between the target and the actual results in historical execution records.

5. The urban traffic cooperative scheduling method based on a large model and multiple agents according to claim 1, characterized in that, In step (5), converting constraints into consistency penalties includes: adding penalties for phase snatching, route blocking, and boundary crossing.

6. The urban traffic cooperative scheduling method based on a large model and multiple agents according to claim 1, characterized in that, In step (6), the fusion strategy includes (a) Time-segmented splicing: Different candidate advantageous segments are used in different time windows; (b) Hierarchical merging: Merging sub-target hierarchies without conflict; (c) Voting weighting: Weighting is performed using consistency score or confidence level as the weight.

7. The urban traffic cooperative scheduling method based on a large model and multiple agents according to claim 1, characterized in that, In step (7), a centralized training and separate execution approach is adopted, specifically as follows: During training, each agent shares or partially shares backbone parameters to utilize more comprehensive information for evaluation; during execution, each agent makes independent decisions to balance global consistency and local response speed.

8. The urban traffic cooperative scheduling method based on a large model and multiple agents according to claim 1, characterized in that, In step (9), the optimal solution is evaluated based on indicators such as congestion relief, average delay, green wave quality, strategy consistency, response delay, resource efficiency, and anomaly rate. The final control commands include signal phase / green wave parameters, vehicle guidance paths, and lockdown / release strategies.

9. A city traffic cooperative scheduling system based on a large model and multiple agents, characterized in that, The system includes a memory and one or more processors, wherein the memory stores executable code, and the one or more processors execute the executable code to implement the urban traffic collaborative scheduling method according to any one of claims 1-8.

Images