Multi-objective adaptive urban traffic signal control method and device based on agent

Abstract

The present disclosure provides an agent-based multi-objective adaptive urban traffic signal control method and device. The method comprises: collecting traffic state data in real time by an environment perception module, and sending the traffic state data to a special vehicle detection module; determining, by the special vehicle detection module, whether there is a special vehicle in the current environment and whether the special vehicle is close to the intersection based on the traffic state data, and sending an activation instruction to a signal control decision module based on the determination result; and in response to the activation instruction, generating, by the activated decision path, a signal control instruction for opening a priority passage for the special vehicle according to the traffic state data, and sending the signal control instruction to the environment perception module for execution.

Description

Technical Field

[0001] This disclosure relates to the field of intelligent transportation technology, and more specifically, to an adaptive control method and apparatus for urban traffic signals based on artificial intelligence. Background Technology

[0002] With the continuous advancement of urbanization, urban transportation demand is constantly increasing, and traffic congestion is becoming increasingly prominent. Traditional traffic management methods often struggle to achieve efficient and flexible traffic control in the face of complex and ever-changing urban traffic environments. To more clearly illustrate the current state of technological development, the following description and evaluation of existing traffic signal control technologies are based on two relevant patent documents.

[0003] Chinese invention patent CN 120823717 A (publication date October 21, 2025) discloses an "Intelligent Optimization Management System for Urban Traffic Congestion Based on Artificial Intelligence." This system aims to address the problems of slow response, localized optimization, and low data utilization in existing traffic management systems. Its technical solution mainly includes a data acquisition and fusion module, a traffic state perception and prediction module, a decision optimization module, an instruction issuance and execution module, and a human-computer interaction and visualization module. Specifically, the data acquisition and fusion module integrates multi-source heterogeneous data (such as traffic sensor data, mobile terminal data, and environmental event data) to generate a comprehensive dataset; the traffic state perception and prediction module uses a graph neural network model to perform short- and medium-term traffic flow predictions to capture the spatiotemporal correlation of traffic flow; and the decision optimization module adopts a multi-agent reinforcement learning framework to generate globally optimal signal control strategies and path planning schemes. Through modular collaboration, this system achieves real-time perception and adaptive control of traffic flow.

[0004] Chinese invention patent CN 121011094 A (publication date November 25, 2025) discloses an "intelligent and interpretable traffic signal adaptive control method." This method focuses on improving the interpretability and generalization ability of traffic signal control, and includes reinforcement learning agent training based on an Actor-Critic network, decision generation guided by a large language model, and a positive and negative sample fine-tuning mechanism. Specifically, the method first learns the timing strategy of intersections through a reinforcement learning agent, then uses a large language model to generate pre-training data containing process explanations, and finally obtains the traffic control signal decision model through full parameter fine-tuning.

[0005] The aforementioned existing technologies represent the application of artificial intelligence in the field of traffic management. However, these existing technologies still have certain shortcomings in the design of traffic signal control objectives. First, in the technical solution disclosed in CN 120823717 A, the system realizes traffic signal control decisions through multi-source traffic data fusion and prediction models. Its optimization objectives mainly focus on traffic operation efficiency-related indicators, such as vehicle queue length, delay time, or traffic efficiency. This type of method can improve the overall operational efficiency of intersections to a certain extent, but in terms of control objective modeling, it usually focuses on a single traffic efficiency indicator and does not adequately consider multi-objective collaborative control in complex traffic scenarios.

[0006] Secondly, the technical solution proposed in CN 121011094 A optimizes traffic signal control strategies by introducing large language models and reinforcement learning methods. Its control objectives are also primarily designed around traffic flow efficiency, such as reducing vehicle waiting time or improving intersection capacity. However, such methods typically do not adequately consider the differentiated needs of different traffic participants in their objective function design, and lack a unified modeling mechanism for coordinating and optimizing multiple traffic operation objectives.

[0007] In actual urban traffic operations, traffic signal control often needs to simultaneously meet multiple control objectives. For example, when special vehicles (such as emergency vehicles or public service vehicles) pass through intersections, traffic signal control needs to provide them with priority passage to reduce emergency response time and improve traffic safety. At the same time, the traffic signal control system also needs to consider the overall traffic efficiency of ordinary vehicles to avoid a decline in the overall operational efficiency of the intersection due to excessive priority strategies. However, most existing technologies use a single traffic efficiency indicator as the optimization objective, failing to adequately consider the synergistic optimization between the priority passage needs of special vehicles and the traffic efficiency of ordinary vehicles, making it difficult to meet the multi-objective control requirements in complex urban traffic environments.

[0008] In summary, existing traffic signal control methods tend to focus on a single traffic efficiency index in terms of control objective design, and pay insufficient attention to the multi-objective coordinated optimization problem between priority passage for special vehicles and the overall passage efficiency of ordinary vehicles. Summary of the Invention

[0009] In view of the aforementioned defects or deficiencies in the existing technology, this disclosure aims to provide a multi-objective adaptive urban traffic signal control method and device based on intelligent agents. This invention constructs a complete technical chain of "environmental perception - scene detection - decision-making and traffic diversion - instruction execution." By introducing a special vehicle detector and its two-layer judgment mechanism, it achieves dynamic recognition of scene priorities. Through the specialized division of labor and seamless collaboration between reinforcement learning and a large language model, it ensures absolute priority passage for special vehicles while minimizing the impact on regular traffic efficiency, fundamentally solving the long-standing problem of multi-objective cooperative control in the existing technology.

[0010] To achieve one or more of the above objectives, this disclosure provides the following aspects.

[0011] According to a first aspect of this disclosure, a multi-objective adaptive urban traffic signal control method based on intelligent agents is provided, comprising: an environmental perception module collecting traffic state data in real time and sending the traffic state data to a special vehicle detection module; the special vehicle detection module determining whether a special vehicle exists in the current environment and whether the special vehicle is approaching an intersection based on the traffic state data, and sending an activation command to a signal control decision module based on the determination result; and in response to the activation command, the activated decision path generating a signal control command for opening a priority lane for the special vehicle based on the traffic state data, and sending it to the environmental perception module for execution.

[0012] According to an exemplary embodiment, when a special vehicle is present and approaching the intersection, the Large Language Model (LLM) decision path in the signal control decision module is activated, with the LLM responsible for flexible decision-making in emergency scenarios. In all other cases, the Reinforcement Learning (RL) decision path in the signal control decision module is activated, with the RL model responsible for efficiency optimization in normal scenarios. This division of labor fully leverages the respective advantages of the two technologies: the RL model, trained on massive amounts of data, can stably and efficiently handle normal traffic flow, achieving continuous optimization of overall traffic efficiency; while the LLM, with its powerful logical reasoning and common-sense understanding capabilities, can comprehensively assess sudden special vehicle scenarios and generate customized control strategies that balance priority passage needs with minimizing traffic disturbance.

[0013] According to an exemplary embodiment, determining whether a special vehicle exists in the current environment includes: determining whether a special vehicle exists in the current environment by means of a vehicle ID, a special identifier, or a preset vehicle type rule. Special vehicles include ambulances, fire trucks, and police cars.

[0014] According to an exemplary embodiment, determining whether a special vehicle is approaching an intersection includes: judging whether the distance between the special vehicle and the stop line of the intersection is less than a preset distance, or determining whether the estimated arrival time calculated based on the speed and distance of the special vehicle is less than a predetermined time.

[0015] According to an exemplary embodiment, when the LLM decision path is activated, the signal control decision module integrates the current traffic state and the decision request into a prompt message and sends it to the large language model. The decision request includes a request to generate a signal control instruction to open a priority lane for special vehicles approaching the intersection.

[0016] According to an exemplary embodiment, when the LLM decision path is activated, the large language model performs comprehensive reasoning based on the current traffic state, not only considering the passage demand of special vehicles, but also evaluating the impact of different control strategies on traffic flow in other directions, generating natural language instructions containing priority passage strategies, and parsing the natural language instructions into one or more signal control commands including interrupting the current phase, inserting a special phase, or extending the green light time in a specific direction.

[0017] According to an exemplary embodiment, when the RL decision path is activated, a pre-trained reinforcement learning model is invoked to automatically select a signal phase switching scheme with the goal of optimizing overall traffic efficiency based on the real-time traffic conditions.

[0018] According to an exemplary embodiment, the traffic state data includes at least the real-time number of vehicles in each traffic flow direction, queue length, currently active signal phase ID, and the location and speed information of all vehicles near the intersection.

[0019] According to an exemplary embodiment, the environmental perception module collects queue lengths in real time by deploying lane detectors.

[0020] According to a second aspect of this disclosure, a multi-objective adaptive urban traffic signal control device based on intelligent agents is provided, comprising: an environment perception module configured to collect traffic state data in real time and send the traffic state data to a special vehicle detection module; a special vehicle detection module configured to determine, based on the traffic state data, whether a special vehicle exists in the current environment and whether the special vehicle is approaching an intersection, and to send an activation command to a signal control decision module based on the determination result; and a signal control decision module configured to activate a decision path in response to the activation command, and to generate a signal control command for opening a priority lane for special vehicles based on the traffic state data through the activated decision path, and send it to the environment perception module for execution.

[0021] This disclosure aims to provide a multi-objective adaptive urban traffic signal control method and apparatus based on intelligent agents. The technical solution provided by this disclosure has the following beneficial effects: This disclosure constructs a complete technical chain of "perception-detection-diversion-decision-execution," achieving multi-objective collaborative optimization and adaptive traffic signal control in complex urban traffic environments. It fundamentally solves the technical challenge of balancing priority passage for special vehicles with overall traffic efficiency for ordinary vehicles in existing technologies. The system can dynamically perceive and adaptively decide on traffic conditions, ensuring absolute priority passage for special vehicles and shortening emergency response time while minimizing the impact on regular traffic flow and maintaining overall intersection efficiency at a reasonable level.

[0022] Specifically, this disclosure constructs a hybrid decision-making traffic signal control system with environmental perception capabilities by integrating a large language model and a reinforcement learning model, achieving intelligent coordination of multiple objectives, including traffic efficiency and emergency priority passage. Unlike existing technologies that use the large language model merely as an auxiliary training tool, this disclosure introduces a "special vehicle detector" and a "two-layer judgment" mechanism, endowing the large language model with independent emergency decision-making capabilities. This allows for dynamic identification of scene priorities: under normal circumstances, the pre-trained reinforcement learning model is invoked to optimize traffic efficiency; when an approaching special vehicle is detected, the system immediately switches to the large language model's decision path, generating customized priority passage instructions. This architecture overcomes the limitations of traditional systems with a single objective, achieving multi-objective adaptive decision-making.

[0023] At the system architecture level, the modular "environmental perception-decision triage" design enhances the system's robustness and scalability. Each module functions independently, enabling the system to withstand the risk of single-model failure and facilitating the integration of new sensors or algorithms to adapt to the control needs of intersections of different sizes.

[0024] At the human-computer interaction level, the system significantly lowers the barrier to entry by leveraging the natural language processing capabilities of the large language model. Managers can interact with the system through high-level instructions, while the large language model can generate clear and reasonable natural language explanations for each control action (especially emergency instructions), enabling visualization and dialogue of the decision-making process and enhancing the system's credibility and operability.

[0025] Furthermore, the effects of this disclosure include not only those set forth herein, but also other effects that will be apparent to those skilled in the art upon reference to the claims, the specification, and the accompanying drawings. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is a diagram showing a traffic light-controlled intersection scenario.

[0028] Figure 2 This is a flowchart illustrating an agent-based multi-objective adaptive urban traffic signal control method according to an embodiment.

[0029] Figure 3 This is a block diagram illustrating an agent-based multi-objective adaptive urban traffic signal control device according to an embodiment. Detailed Implementation

[0030] The technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. The components of the embodiments of this disclosure described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this disclosure can also be combined with each other.

[0031] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. In the description of this disclosure, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0032] To address the problem that existing traffic signal control methods often focus on a single traffic efficiency index as the optimization objective, making it difficult to simultaneously consider the priority passage needs of special vehicles and the overall traffic efficiency of ordinary vehicles, this disclosure aims to provide a multi-objective adaptive urban traffic signal control method and device based on intelligent agents. This method and device construct an agent-driven traffic signal control mechanism to dynamically perceive and adaptively decide on traffic operation states. While ensuring priority passage for special vehicles, it also improves the overall traffic efficiency of ordinary vehicles, thereby achieving multi-objective collaborative optimization and adaptive traffic signal control in complex urban traffic environments.

[0033] This disclosure is based on a dynamic traffic environment constructed using the microscopic traffic simulation software SUMO (Simulation of Urban Mobility). This disclosure introduces a "special vehicle detector," which performs real-time identification and status assessment of vehicles in the simulated or real environment, and intelligently switches between different decision paths based on the assessment results. Specifically, when a special vehicle (such as an ambulance or fire truck) is detected near an intersection, the system triggers an immediate reasoning and decision path based on a large language model (LLM decision path), generating priority control commands such as "open green lane." Under normal circumstances, the system calls a pre-trained reinforcement learning model for efficient and stable optimization control. Figure 3 As shown, this hybrid architecture of "detection-judgment-diversion" achieves a balance between optimizing routine traffic and prioritizing emergency passage.

[0034] 1) Basic Concepts To facilitate understanding of the traffic signal control method disclosed herein, several basic concepts in the traffic light control intersection scenario will first be explained.

[0035] Figure 1 This is a diagram illustrating a traffic light-controlled intersection scenario. See below for reference. Figure 1 Define the relevant elements in traffic intersections.

[0036] Lane: In the context of a traffic intersection, a lane refers to the road space where vehicles travel in a predetermined direction. Lanes can be classified according to the direction of vehicle travel and their function. Generally, lanes can be divided into two categories: entry lanes and exit lanes. Entry lanes guide vehicles from the road into the intersection area, while exit lanes are used by vehicles to leave the intersection area after passing through it. Entry lanes and exit lanes together constitute the basic traffic flow structure of an intersection.

[0037] Traffic flow: Traffic flow refers to the directional movement of vehicles from the entering lane to the exit lane, reflecting the specific travel path of vehicles at an intersection. In a typical four-way intersection, there are usually four main incoming paths: east, west, south, and north. For each incoming path, vehicles can generally perform two main exit movements: left turn (denoted as l) and straight ahead (denoted as s). In the traffic control framework considered in this disclosure, based on the assumption that "right-turning vehicles under the right-hand rule can generally pass without strict control by traffic lights," right-turn movement is not modeled for the time being. Therefore, in this intersection scenario, the traffic signal control system mainly manages eight different traffic movements, labeled m1 to m8, each movement corresponding to a specific vehicle travel path.

[0038] Traffic lights: Traffic lights are used to direct and manage traffic flow in all directions within an intersection. Different colored lights indicate different traffic control states; a green light means vehicles are allowed to proceed in that direction, while a red light means vehicles must stop and wait. By setting corresponding light states for different traffic flows, effective management of vehicle traffic order within the intersection can be achieved. For example, ... Figure 1 As shown, in a standard four-way intersection, when moving signals m4 and m8 are green, it means that the corresponding traffic movement (i.e., left-turning vehicles from the west and east) is allowed to proceed, while traffic movement in other directions must remain stopped.

[0039] Phase: Traffic signal phase refers to a combination of traffic signals that simultaneously permit passage within the same time period. By rationally designing traffic signal phases, multiple non-conflicting traffic flows can proceed simultaneously while ensuring traffic safety, thereby improving the overall traffic efficiency of intersections. For example... Figure 1 As shown, this intersection operates using four different signal phases, labeled P1 to P4. Each phase corresponds to a set of non-conflicting traffic movements. For example, during phase P4, the green light signals for movements m4 and m8 are activated simultaneously, allowing vehicles to safely make left turns from the westbound and eastbound entrance lanes respectively without interfering with traffic from other directions.

[0040] 2) Simulation platform and core environmental parameters Simulation Engine: The system uses SUMO (Simulation of Urban Mobility) software as its core environmental simulation platform.

[0041] Data Acquisition: Virtual lane area detectors are deployed in the simulation environment to collect data such as the number of vehicles and queue length in each lane in real time. The data acquisition range is centered on the intersection, with a maximum distance of, for example, 150 meters, to simulate the typical constraints of sensor monitoring range in a real city.

[0042] Traffic light sequence: Following the common traffic light change sequence in the city, the standard phase change rule is set as: green light phase → 3-second yellow light → red light phase.

[0043] 3) Vehicle behavior and dynamic parameters Speed ​​limit: The maximum speed limit for vehicles on the road section can be set at 13.9 m / s (i.e. 50 km / h) to conform to the typical speed limit standards for urban roads.

[0044] Safe following distance: The minimum safe distance between vehicles can be set to 2.5 meters to simulate typical following behavior in an urban environment.

[0045] Speed ​​distribution: The vehicle speed is not a fixed value, but is modeled using a Gaussian distribution to reflect the variability in driver behavior. The mean of this distribution can be set to 10 m / s, and the standard deviation to 1.73 m / s.

[0046] Figure 2 This is a flowchart illustrating an agent-based multi-objective adaptive urban traffic signal control method according to an embodiment.

[0047] According to a first aspect of this disclosure, a multi-objective adaptive urban traffic signal control method based on intelligent agents is provided, comprising: an environmental perception module collecting traffic state data in real time and sending the traffic state data to a special vehicle detection module (S1); the special vehicle detection module determining whether a special vehicle exists in the current environment and whether the special vehicle is approaching an intersection based on the traffic state data, and sending an activation command to a signal control decision module based on the determination result (S2); and in response to the activation command, the activated decision path generating a signal control command for opening a priority lane for the special vehicle based on the traffic state data, and sending it to the environmental perception module for execution (S3).

[0048] According to an exemplary embodiment, when a special vehicle is present and the special vehicle is approaching the intersection, the Large Language Model (LLM) decision path in the signal control decision module is activated; otherwise, the Reinforcement Learning (RL) decision path in the signal control decision module is activated.

[0049] According to an exemplary embodiment, determining whether a special vehicle exists in the current environment includes: determining whether a special vehicle exists in the current environment through vehicle ID, special identifier, or preset vehicle type rules. Special vehicles include emergency vehicles, public service vehicles, and specially designated ordinary vehicles. Special vehicles include ambulances, fire trucks, and police cars.

[0050] According to an exemplary embodiment, determining whether a special vehicle is approaching an intersection includes: judging whether the distance between the special vehicle and the stop line of the intersection is less than a preset distance, or determining whether the estimated arrival time calculated based on its speed and distance is less than a predetermined time.

[0051] According to an exemplary embodiment, when the LLM decision path is activated, the signal control decision module integrates the current traffic state and the decision request into a prompt message and sends it to the large language model. The decision request includes a request to generate a signal control instruction to open a priority lane for special vehicles approaching the intersection.

[0052] According to an exemplary embodiment, when the LLM decision path is activated, the large language model performs comprehensive reasoning based on the current traffic state, generates natural language instructions containing priority passage strategies, and parses the natural language instructions into one or more signal control commands, including interrupting the current phase, inserting a special phase, or extending the green light time for a specific direction.

[0053] According to an exemplary embodiment, when the RL decision path is activated, a pre-trained reinforcement learning model is invoked to automatically select a signal phase switching scheme with the goal of optimizing overall traffic efficiency based on the real-time traffic conditions.

[0054] According to an exemplary embodiment, the traffic state data includes at least the real-time number of vehicles in each traffic flow direction, queue length, currently active signal phase ID, and the location and speed information of all vehicles near the intersection.

[0055] According to an exemplary embodiment, the environmental perception module collects queue lengths in real time by deploying lane detectors.

[0056] Figure 3 This is a block diagram illustrating an agent-based multi-objective adaptive urban traffic signal control device according to an embodiment.

[0057] According to a second aspect of this disclosure, a multi-objective adaptive urban traffic signal control device based on intelligent agents is provided, comprising: an environment perception module configured to collect traffic state data in real time and send the traffic state data to a special vehicle detection module; a special vehicle detection module configured to determine, based on the traffic state data, whether a special vehicle exists in the current environment and whether the special vehicle is approaching an intersection, and to send an activation command to a signal control decision module based on the determination result; and a signal control decision module configured to activate a decision path in response to the activation command, and to generate a signal control command for opening a priority lane for special vehicles based on the traffic state data through the activated decision path, and send it to the environment perception module for execution.

[0058] The device or system disclosed herein mainly consists of three modules: an environmental perception module (or an environmental perception and simulation module), a special vehicle detection module, and a signal control decision module. These modules are closely integrated and work collaboratively to form a complete "perception-judgment-decision-execution" technology chain. The specific composition, functions, and interaction flow are as follows: 1) Environmental Perception and Simulation Module This module is responsible for building and maintaining the system's operating environment, which is the foundation for interacting with the physical or virtual transportation world. It consists of three parts: the initial environment, environmental observation, and the SUMO environment.

[0059] Initial Environment: The system startup component is responsible for loading and configuring the initial state parameters of the traffic simulation. These parameters include at least: road network topology (e.g., a standard four-way intersection), initial traffic light phase scheme, vehicle generation rate and behavior parameters (e.g., maximum speed, safe following distance, speed distribution model), and preset information for special vehicles. This component provides reproducible initial conditions for the entire simulation run.

[0060] Environmental observation: During simulation, this part acts as the system's "sensors," responsible for collecting real-time observation data from the "SUMO environment." The collected data includes at least: real-time vehicle numbers in each lane (or traffic flow direction), queue lengths, current signal phase states, vehicle speeds, and other intersection traffic state information, providing accurate environmental situation information for subsequent decision-making.

[0061] SUMO Environment: A dynamic traffic system model driven by the SUMO simulation kernel, simulating the real-time evolution of vehicle behavior and traffic flow. It receives control signals from the signal control decision module, executes traffic light phase changes, and updates the state of the entire traffic environment based on rules.

[0062] 2) Special vehicle detection module This module is the decision path selection hub of this disclosure and a core component for achieving dynamic multi-objective control. It possesses a two-layer progressive logical judgment capability, continuously receiving "real-time observation data" streams from "environmental observation," and dynamically guiding the system to different advanced decision paths accordingly.

[0063] Special Vehicle Detector: This component performs real-time analysis of vehicle information in the observed data, executing a two-layer decision-making process to select the appropriate path. The decision logic is as follows: The first layer of judgment (whether there are special vehicles): Analyze whether there are vehicles with priority right-of-way in the current environment, such as ambulances, fire trucks, police cars, etc. Identification can be achieved through vehicle ID, special identifiers, or preset vehicle type rules.

[0064] The second layer of judgment (whether it is close to the intersection): If the first layer judgment is "yes", then the spatial relationship between the special vehicle and the target intersection is further analyzed. This judgment is usually based on whether the real-time distance between the vehicle and the stop line of the intersection is less than a preset threshold, or on the estimated arrival time calculated based on its speed and position.

[0065] Decision path selection: Based on the results of the two-level judgment, the system precisely activates the corresponding decision path in the decision control module to achieve path switching. Triggering LLM decision path: The system immediately switches to the path if and only if both layers of judgment results are "yes" (i.e., there is a special vehicle and it is close to the intersection).

[0066] Triggering RL decision path: In all other cases (no special vehicle, or special vehicle but not yet approaching the intersection), the system invokes the RL decision path.

[0067] 3) Signal control decision module This module integrates two different advanced intelligent decision-making algorithms, which are selectively activated according to the instructions of the special vehicle detection module, and is the "brain" that generates specific control commands.

[0068] LLM Decision Path: When activated, the large language model in this path performs comprehensive reasoning based on the current traffic conditions or complex real-time scenarios (including the precise location of special vehicles, traffic flow status in other directions, intersection phase structure, etc.). Leveraging its powerful natural language understanding and logical reasoning capabilities, LLM quickly generates customized and interpretable priority passage strategies. Its typical output instruction is "open green lane," which means immediately providing right-of-way to the direction of the special vehicle through a series of operations (such as interrupting the current phase, inserting a special phase, or extending a specific green light time).

[0069] RL Decision Path: Under normal circumstances, this path invokes a pre-trained reinforcement learning model. This model has been trained on massive amounts of historical traffic data or simulation scenarios, mastering a general control strategy aimed at optimizing overall traffic efficiency (such as minimizing total waiting time and maximizing throughput). It can efficiently and stably handle regular traffic flow, automatically selecting the optimal signal phase switching scheme based on "real-time observation data".

[0070] The internal workflow of an agent-based multi-objective adaptive urban traffic signal control device is as follows: Step 1: The system starts up, builds the "initial environment" according to the preset "initial state parameters", and initializes the "SUMO environment".

[0071] Step 2: At each time step or decision point, the "Environment Observation" component obtains the latest "Real-time Observation Data" from the "SUMO Environment".

[0072] Step 3: The "real-time observation data" is transmitted to the special vehicle detection module. The special vehicle detector performs the core "two-layer judgment" and selects either the LLM or reinforcement learning (RL) decision path in the activation signal control decision module based on the result.

[0073] Step 4: The final control command generated by the "Signal Control Decision Module" is sent to the "SUMO Environment" as a "control signal" for execution, changing the state of the traffic lights.

[0074] Step 5: After the control signal is executed, the "SUMO environment" updates its internal state based on the new signal state and vehicle behavior model. The updated state becomes the data source for the next round of "environmental observation", forming a continuous feedback closed loop.

[0075] This disclosure uses a typical urban intersection during peak hours to illustrate an embodiment of the present disclosure. The entire system's operation strictly follows the system architecture. Figure 3 The logic and data flow shown cover a complete closed loop from environment initialization, real-time perception, intelligent decision-making to signal execution.

[0076] Example: Emergency priority traffic control at intersections during peak hours I. System Initialization and Environment Setup This embodiment uses a standard four-way intersection as the control object. When the system starts, the initial environment component in the environment perception and simulation module constructs a simulation environment based on preset initial state parameters.

[0077] (1) Initialization of road network and signal phase A typical four-way intersection network was loaded into the SUMO simulation platform. This network includes four approach directions (east, south, west, and north), each with dedicated lanes for straight-through and left-turn traffic (a total of 8 lanes), corresponding to 8 traffic flows (m1 to m8). Signal control employs a fixed four-phase scheme (P1 to P4): P1 allows north-south straight-through traffic, P2 allows north-south left-turn traffic, P3 allows east-west straight-through traffic, and P4 allows east-west left-turn traffic. Each phase follows a fixed sequence of "green light - 3-second yellow light - red light." Virtual loop detectors were deployed approximately 150 meters from the stop line at the intersection to collect real-time lane occupancy and queue length data.

[0078] (2) Initialization of traffic flow and vehicle parameters Configure peak-hour traffic demand using a SUMO routing file. Set a high vehicle arrival rate (e.g., 800 vehicles / hour) for north-south straight-through flows (m1, m5) to simulate tidal traffic flow. All vehicle types are set to "cars," with parameters configured as follows: maximum speed 13.9 m / s, minimum safe following distance 2.5 m, and speed following a Gaussian distribution with a mean of 10 m / s and a standard deviation of 1.73 m / s. A pre-defined ambulance, identified as "Ambulance_01," is designated as a special vehicle, with its vehicle type (vClass) marked as "Emergency." This ambulance enters from the west entrance some time after the simulation begins, with its destination being the east exit.

[0079] (3) Initialization of decision model RL decision path: Load a pre-trained reinforcement learning model (such as one trained based on DQN or PPO algorithms). The model takes the queue length and occupancy rate of each lane as state inputs and the phase switching action as output. The optimization objective is to minimize the long-term average vehicle delay.

[0080] LLM Decision Path: Configure a large language model, provide system prompts to clearly define its role as "traffic signal control AI", and give it the ability to call virtual tools such as "get current phase", "get queue length of each flow direction", "get special vehicle position", and "set next phase to Px" to perform reasoning and decision-making.

[0081] After the above initialization is completed, the SUMO environment starts running, and the system enters a continuous "perception-judgment-decision-execution" cycle.

[0082] II. Execution of Closed-Loop Control Throughout the Entire Process Simulation time stepping, the system according to Figure 3 The numbers shown work together in sequence.

[0083] Step 2: Environmental Observation (Data Acquisition) The environmental observation component collects real-time observation data from the SUMO environment at each decision cycle (e.g., once per second), including: Real-time queue counts of vehicles in 8 traffic flows (m1-m8) (e.g., m1: 5 vehicles, m2: 3 vehicles, ...). The currently active signal phase ID (e.g., P3, east-west straight line is green light). Location and speed information of all vehicles The vehicle identified as "Ambulance_01" was located on the west entrance lane, 80 meters from the stop line at the intersection, and was approaching the intersection at a speed of 12 meters per second.

[0084] Step 3: Special Vehicle Detection and Decision Path Selection (Two-Level Judgment) Real-time observation data is immediately sent to the special vehicle detection module. The special vehicle detector then initiates its core "two-layer judgment" logic: First-level judgment (whether there is a special vehicle): The detector scans the vehicle list and identifies the vehicle "Ambulance_01" tagged "Emergency", and the judgment result is "Yes".

[0085] The second layer of judgment (whether it is close to the intersection): The detector reads the location information of "Ambulance_01" (80 meters away from the stop line) and compares it with the preset "priority response distance threshold" (e.g., 100 meters). If 80 meters < 100 meters, the judgment result is "yes".

[0086] Based on the fact that both judgment results are "yes", the detector outputs a decision instruction: trigger the LLM decision path. This instruction is sent to the signal control decision module.

[0087] Step 4: Intelligent Decision Making and Signal Execution The signal control decision module receives the "trigger LLM decision path" instruction and activates the LLM decision path. The system integrates the current environmental state (observation data from step 2) with the decision request ("An ambulance is approaching the intersection from west to east, please provide signal control instructions") into a prompt message and sends it to the GPT-4 Turbo model.

[0088] LLM uses its internal knowledge and understanding of traffic rules to reason and generate the following analysis and instructions: "The current phase is P3 (green light for east-west straight traffic). The ambulance is in the west entrance lane and needs to proceed straight through the intersection. The current phase has already granted the ambulance the right of way, and there is no need to switch phases. To ensure the ambulance's safe and efficient passage, the green light time of the current P3 phase should be immediately extended by 15 seconds to cover the time required for the ambulance to pass through the intersection, and the following traffic should be cleared. After the extension, the traffic should switch to the next phase, P4 (green light for east-west left turns), to allow left-turning vehicles to pass and avoid intersection congestion. Instructions: 1. Immediately extend the green light time of the current P3 phase by 15 seconds; 2. After the extension, switch to the P4 phase." The natural language instruction is parsed by the system into a specific sequence of control commands (such as "extend the current P3 phase by 15 seconds", "next phase = P4") and sent back to the SUMO environment.

[0089] Step 5: Environment Status Update and Cycle Upon receiving the control signal, the SUMO environment immediately executes the following: the current P3 phase green light duration is extended by 15 seconds. After the extension, the traffic light switches to the P4 phase (east-west left turn green light). The signal extension causes an update to the vehicle operation status: the east-west straight-ahead direction remains green, allowing ambulances to pass through the intersection smoothly, while north-south traffic remains in a red light waiting state. The environmental status is updated.

[0090] In the next decision cycle, the process returns to step 2, where the environmental observation component collects new real-time observation data (such as the ambulance having passed and the reduction in east-west queues), initiating a new round of closed-loop control. At this point, the special vehicle detector's judgment becomes "no," and the system automatically switches back to the RL decision path, where the reinforcement learning model continues to optimize the regular traffic flow.

[0091] III. System Output The output of this invention is a real-time dynamic control command sequence for traffic lights and the resulting changes in traffic conditions. This can be visually observed through the SUMO-GUI interface in a simulation environment. Under normal circumstances, traffic lights are controlled by an RL model, resulting in smooth phase switching and superior performance in terms of total waiting time compared to fixed timing. As the ambulance approaches the intersection, the system responds quickly, and the traffic lights intervene according to the LLM generation strategy to create a green channel for the ambulance. After the ambulance passes, the system seamlessly switches back to RL model control and resumes normal traffic optimization. All decision-making logic (especially the LLM reasoning process) can be recorded and output to form an interpretable decision log, such as: "Ambulance 'Ambulance_01' has been detected 80 meters to the west, triggering priority passage. Decision: The green light time of the current P3 phase should be extended by 15 seconds immediately to cover the time required for the ambulance to pass through the intersection and clear the following traffic. After the extension, the next phase P4 (east-west left turn green light) should be switched to allow left-turning vehicles to pass and avoid intersection congestion." The above embodiments fully demonstrate the entire process of the present invention from environmental initialization, real-time perception, intelligent path switching based on two-layer judgment, to hybrid decision execution, and verify the effectiveness and advancement of the system in balancing normal efficiency and emergency response.

[0092] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles used, and is not intended to limit the scope of the claimed disclosure, but only to illustrate preferred embodiments of this disclosure. Those skilled in the art should understand that the scope of the invention involved in this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalent features without departing from the inventive concept. All other embodiments obtained by those skilled in the art based on the embodiments of this disclosure without inventive effort are within the scope of protection of this disclosure.

Claims

1. A multi-objective adaptive urban traffic signal control method based on intelligent agents, characterized in that, include: The environmental perception module collects traffic status data in real time and sends the traffic status data to the special vehicle detection module; The special vehicle detection module determines whether there are special vehicles in the current environment and whether the special vehicles are close to the intersection based on the traffic state data, and sends an activation command to the signal control decision module based on the determination result; as well as In response to the activation command, the activated decision path generates a signal control command to open a priority lane for the special vehicle based on the traffic state data, and sends it to the environmental perception module for execution.

2. The agent-based multi-objective adaptive urban traffic signal control method according to claim 1, characterized in that, When the special vehicle is present and it is close to the intersection, the Large Language Model (LLM) decision path in the signal control decision module is activated; otherwise, the Reinforcement Learning (RL) decision path in the signal control decision module is activated.

3. The agent-based multi-objective adaptive urban traffic signal control method according to claim 2, characterized in that, Determining whether special vehicles exist in the current environment includes: using vehicle ID, special identifiers, or preset vehicle type rules to determine whether special vehicles exist in the current environment, including ambulances, fire trucks, and police cars.

4. The agent-based multi-objective adaptive urban traffic signal control method according to claim 3, characterized in that, Determining whether a special vehicle is approaching an intersection includes: determining whether the distance between the special vehicle and the stop line at the intersection is less than a preset distance, or determining whether the estimated arrival time calculated based on the speed of the special vehicle and the distance is less than a predetermined time.

5. The agent-based multi-objective adaptive urban traffic signal control method according to claim 2, characterized in that, When the LLM decision path is activated, the signal control decision module integrates the current traffic status and decision request into a prompt message and sends it to the large language model. The decision request includes a request to generate a signal control instruction to open a priority lane for special vehicles approaching the intersection.

6. The agent-based multi-objective adaptive urban traffic signal control method according to claim 5, characterized in that, When the LLM decision path is activated, the large language model performs comprehensive reasoning based on the current traffic state, generates natural language instructions containing priority passage strategies, and parses the natural language instructions into one or more signal control commands, including interrupting the current phase, inserting a special phase, or extending the green light time for a specific direction.

7. The agent-based multi-objective adaptive urban traffic signal control method according to claim 6, characterized in that, When the RL decision path is activated, a pre-trained reinforcement learning model is invoked to automatically select a signal phase switching scheme with the goal of optimizing overall traffic efficiency based on the real-time traffic conditions.

8. The agent-based multi-objective adaptive urban traffic signal control method according to claim 1, characterized in that, The traffic status data includes at least the real-time number of vehicles in each traffic flow direction, queue length, currently active signal phase ID, and the position and speed information of all vehicles near the intersection.

9. The agent-based multi-objective adaptive urban traffic signal control method according to claim 8, characterized in that, The environmental perception module collects the queue length in real time by deploying lane detectors.

10. A multi-objective adaptive urban traffic signal control device based on intelligent agents, characterized in that, include: The environmental perception module is configured to collect traffic status data in real time and send the traffic status data to the special vehicle detection module; The special vehicle detection module is configured to determine whether there are special vehicles in the current environment and whether the special vehicles are close to the intersection based on the traffic state data, and to send an activation command to the signal control decision module based on the determination result. as well as The signal control decision module is configured to activate a decision path in response to the activation command, and generate a signal control command for opening a priority lane for the special vehicle based on the traffic state data through the activated decision path, and send it to the environmental perception module for execution.

Images