A data-driven traffic simulation method, system, and storage medium

Abstract

The present invention relates to a data-driven traffic simulation method, system and storage medium. The method includes the following steps: acquiring multi-source traffic data; defining a dynamic heterogeneous graph structure, and converting the multi-source traffic data into a traffic system graph structure, where the nodes in the traffic graph structure represent traffic participants and the edges represent the relationships between traffic participants; based on the traffic graph structure, using a dynamic graph learning algorithm to learn traffic patterns and simulate the scenarios required by users to achieve traffic simulation. Compared with the prior art, the present invention has the advantages of simple model, applicability to complex traffic environments, high simulation efficiency, etc.

Description

Technical Field

[0001] The present invention relates to the field of traffic simulation, and in particular to a data-driven traffic simulation method, system and storage medium. Background Art

[0002] Traffic simulation is an important method for realizing urban intelligent transportation. The traffic simulation system uses simulation technology to study traffic behavior, reproduce the traffic operation status of the existing traffic system, and help users master various traffic models by analyzing and interpreting the traffic phenomena displayed by the system.

[0003] The traditional traffic simulation method has the following defects:

[0004] 1. Traditional methods usually rely on physical dynamic models, which need to predefine the structure of the model based on experience or artificial assumptions, and only use data-driven methods to learn the parameters in the predefined model. However, this method may introduce over-simplification and assumptions, thus limiting the authenticity and applicability of the model.

[0005] 2. Traditional models are suitable for specific tasks, but do not have scalability or extensibility, so there are challenges in adapting to different environments, tasks and managing a large amount of complex data inputs.

[0006] 3. The inherent complexity of the traffic system is affected by various factors and agents that affect traffic behavior, making it a challenging task to truly capture the spatio-temporal variability and complexity of traffic states. Summary of the Invention

[0007] The purpose of the present invention is to provide a data-driven traffic simulation method, system and storage medium, which uses data-driven algorithms and graph computing technologies to learn traffic dynamics, automatically learn traffic patterns from real data, can generate more realistic traffic patterns, and help decision-makers formulate traffic policies and make decisions more accurately.

[0008] The purpose of the present invention can be achieved by the following technical solutions:

[0009] A data-driven traffic simulation method includes the following steps:

[0010] Obtain multi-source traffic data;

[0011] Define a dynamic heterogeneous graph structure, and convert the multi-source traffic data into a traffic graph structure, where the nodes in the traffic graph structure represent traffic participants, and the edges represent the relationships between traffic participants;

[0012] Based on the traffic graph structure, use a dynamic graph learning algorithm to learn traffic patterns and simulate the scenarios required by users to achieve traffic simulation.

[0013] Furthermore, in the dynamic heterogeneous graph, nodes represent agents, edges represent the relationships between nodes, each node corresponds to a node category, and is represented by a first unique mapping function; each edge is a directed edge and corresponds to an edge category, which is represented by a second unique mapping function.

[0014] Furthermore, the node categories include vehicles, lanes, roads, and traffic signal controllers. All participants in the traffic system can be regarded as nodes in the traffic graph network, so the system has high scalability.

[0015] Furthermore, in the dynamic heterogeneous graph, each node corresponds to a node category and is represented by a unique mapping function. Each node includes multiple node features according to different node categories:

[0016] When the node category is a vehicle, the corresponding node features include the current position, speed, acceleration, vehicle type, and vehicle length;

[0017] When the node category is a lane, the corresponding node features include the lane length, maximum allowable speed, current traffic flow, occupancy rate, lane type, and lane markings;

[0018] When the node category is a road, the corresponding node features include the road type, road length, width, and road surface condition;

[0019] When the node category is a traffic signal controller, the corresponding node features include the signal duration, signal plan, and signal light status.

[0020] Furthermore, the edge category is determined according to the node categories connected by the edge. The edge categories include the relationship between a lane and a road, the relationship between a vehicle and its lane, the relationship between lanes, the relationship between a vehicle and a traffic signal controller, the relationship between a traffic signal controller and a lane, and the relationship between vehicles.

[0021] Furthermore, the dynamic graph learning algorithm, based on the constructed dynamic traffic graph, uses the interaction generation module to predict the generation probability of edges according to the traffic graph information at the current moment, represents the interaction relationship between traffic participants by generating edges, and uses the state prediction module to predict the node features at the next time step based on the historical node information and the predicted edge information, obtaining the traffic participant state prediction result. Repeat the above process to complete traffic simulation.

[0022] Furthermore, the dynamic graph learning algorithm is implemented based on HGT. Assume that the input feature of node v is the node feature H (l-1) [v] at the historical moment, and the output representation of the target node v is the node feature H (l)[v], in order to aggregate information from all neighbor nodes to the target node, the vector is updated by using a multi-head attention mechanism and a message passing mechanism. Based on Predict the target node output:

[0023]

[0024]

[0025] Among them, e i =(u i ,v i ),u i ,v i ∈V n represents the agent, edge e i =(u i ,v i ),e i ∈E n represents the relationship between nodes / agents; θ is a parameter that needs to be learned during the training process and is used to aggregate feature vectors Passed to the activation function A-Linear φ(v) And add the feature vector H of the previous layer (l-1) [v] Before, (l) [v] is transformed; Attention(u,e,v) represents the attention mechanism, which determines the importance of edge e between node u and node v; Message(u,e,v) represents the message passing mechanism, which indicates the message passed from node u to node v through edge e.

[0026] A data-driven traffic simulation system, comprising:

[0027] The data layer is used to collect data inputs from different sources, obtain multi-source traffic data, and define a dynamic heterogeneous graph structure to convert the multi-source traffic data into a traffic graph structure, in which nodes represent traffic participants and edges represent the relationships between traffic participants;

[0028] The simulation layer is used to learn traffic patterns based on the traffic graph structure using a dynamic graph learning algorithm and simulate the scenarios required by users to achieve traffic simulation;

[0029] The interface layer is used to provide a UI interface for users to interact with the traffic simulation system.

[0030] A data-driven traffic simulation system comprises a memory, a processor, and a program stored in the memory, wherein the processor implements the above-mentioned method when executing the program.

[0031] A storage medium stores a program thereon, and when the program is executed, the above-mentioned method is implemented.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] (1) Data-driven simulation environment: The present invention uses data-driven algorithms and dynamic graph computing technologies to automatically learn traffic patterns without predefining complex models and parameters, generating more realistic traffic patterns.

[0034] (2) Applicable to complex traffic scenarios: The present invention can handle complex traffic behaviors and relationships, better understand the relationships and behaviors between different nodes, can handle large-scale scenarios, and remains efficient when dealing with large-scale scenarios, so it is applicable to complex traffic scenarios.

[0035] (3) Good scalability: The present invention has good scalability. As the scene scale increases, the computing time increases linearly. It can handle large-scale scenarios and remains efficient when dealing with large-scale scenarios.

[0036] (4) Parallel processing: The present invention uses a graph structure for parallel processing, which can significantly reduce the simulation computing time. At the same time, a pre-trained model is used to reduce the computing amount required during the simulation, improving the efficiency of the simulation. Description of the Drawings

[0037] Figure 1 is a flowchart of the method of the present invention;

[0038] Figure 2 is a schematic diagram of the traffic graph structure generated by the present invention;

[0039] Figure 3 is a system structure diagram of the present invention. Detailed Embodiments

[0040] The present invention will be described in detail below with reference to the drawings and specific embodiments. This embodiment is implemented on the premise of the technical solution of the present invention, and the detailed implementation manners and specific operation processes are given, but the protection scope of the present invention is not limited to the following embodiments.

[0041] Embodiment 1

[0042] This embodiment provides a data-driven traffic simulation method, as Figure 1 shown, including the following steps:

[0043] S1. Obtain multi-source traffic data;

[0044] S2. Define a dynamic heterogeneous graph structure and convert the multi-source traffic data into a traffic graph structure;

[0045] In this embodiment, a graph - structure - based method is adopted to model the traffic system. Nodes represent various elements in the traffic system, such as vehicles, intersections, streets, blocks, and cities, etc. Edges represent the relationships between these elements, such as vehicles driving on roads, connections between intersections, and connections between streets and blocks, etc. The method of dynamic heterogeneous graphs is highly open and extensible in design. Different from traditional simulation models, this method allows users to dynamically add or delete nodes and edges, as well as the features of nodes and edges, without reconstructing the entire system. In this way, the present invention can provide topological structures at multiple levels of the traffic system, enabling traffic planners to better understand the entire traffic system and make better decisions.

[0046] The traffic system has its unique characteristics and behaviors, such as vehicle behavior, spatio - temporal traffic flow changes, traffic control rules, etc. How to appropriately transform and encode these characteristics into a graph structure for use by graph - learning algorithms is a core technical challenge.

[0047] In one embodiment, the traffic system is defined as a dynamic heterogeneous graph G n (V n , E n , O n , R n ), where nodes V n represent agents, edges E n represent the relationships between nodes / agents, O n represents the set of node categories, and R n represents the set of edge categories. Each node corresponds to a node category and can be represented by a unique mapping function φ:

[0048] φ: V n → O n , o i ∈ O n

[0049] Each node v i can be mapped to a specific category o i according to the function φ. For example, in the traffic system, these node categories may include "vehicle", "pedestrian", or "traffic signal controller".

[0050] Similarly, each edge e i is a directed edge and is associated with an edge type r i and can be represented by a unique mapping ψ function:

[0051] ψ: E n → R n , r i ∈ R n , e i = (ui , v i )

[0052] The above formula indicates that each edge e can be mapped to a specific category r according to the function ψ i . For example, the types of edges may be "front", "rear". For instance, vehicle A is "in front of" or "behind" vehicle B, and the relationship between road 1 and road 2 is "connected", etc. i

[0053] In a specific embodiment, the node categories include vehicles, lanes, roads, and traffic signal controllers. In another embodiment, it may also include other traffic participants in the traffic system, such as intersections, pedestrians, etc. All traffic participants can be modeled as a node in the traffic graph structure through feature extraction.

[0054] In the dynamic heterogeneous graph of the traffic system, each node (agent) has associated features, which are represented as node features. Taking the vehicle agent as an example, its node features may include its current position, speed, and acceleration. This means that when the simulation system views a specific vehicle node, it can not only know that it is a vehicle (based on its node type), but also understand its specific motion state, such as which position on a certain road it is at, what speed it is traveling at, and whether its acceleration is positive or negative. A lane is also an agent, and its node features may include the length, width, maximum allowable speed, current traffic flow, and whether it is a dedicated lane (such as a bus lane or a bicycle lane), etc. For example, a specific lane node may represent an asphalt lane that is 500 meters long, 3.5 meters wide, has a maximum allowable speed of 60 km / h, and a current traffic flow of 20 vehicles per minute. These features provide detailed information about the physical properties and current usage status of the lane. In the dynamic graph structure, the attributes of the agent can be represented as the features of the nodes in the graph. Assuming the node has a feature dimension then the node features can be represented as:

[0055]

[0056] In the traffic graph structure, each node may include multiple node features according to different node categories. The nodes include roads, traffic signal controllers, lanes, and vehicles, etc., and they have their own characteristics, such as name, length, speed, position, etc., as shown in Table 1 specifically.

[0057] Table 1 Node Categories and Their Corresponding Node Features

[0058]

[0059] ​

[0060] Edges in the graph structure represent various relationships between entities. The edge categories are determined according to the node categories connected by the edges, including the relationship between lanes and roads (lane_phy / to_road), the relationship between vehicles and the lanes they are in (veh_phy / to_lane), the connections between lanes and other lanes (lane_phy / to_lane and lane_phy / from_lane), the relationship between vehicles and traffic signal controllers (veh_phy / to_tlc), the relationship between traffic signal controllers and lanes (tlc_phy / to_lane), and the front - rear relationship between vehicles (veh_phy / behind_veh and veh_phy / ahead_veh), etc.

[0061] The nodes and edges set in this embodiment are to meet the basic simulation task requirements. Users can quickly add new types of nodes and edges on the existing framework, and can also introduce new attributes and characteristics for existing entities to make it closer to the required research needs. And edge features are not a necessary parameter for traffic simulation, so this embodiment does not describe it further. In one embodiment, edge features can represent weights or information flows, etc., and are added as a later extension to improve the performance of the traffic simulation system.

[0062] Assume that the edge has a feature dimension Then, similarly, the edge feature can be expressed as:

[0063]

[0064] Figure 2 shows the topological structure of the hierarchical traffic graph structure, including presenting the dynamic characteristics of individual vehicles at the vehicle level, showing traffic signal control strategies and traffic conditions at bottlenecks at the intersection level, and showing strategies and policies that have an impact on a larger scale at the street, block, and city levels. This multi - scale approach enables the simulation system of the present invention to provide a comprehensive view of the traffic system and enables traffic planners to make informed decisions based on the simulation results.

[0065] S3. Based on the traffic graph structure, use a dynamic graph learning algorithm to learn traffic patterns and simulate the scenarios required by users to achieve traffic simulation.

[0066] To achieve traffic simulation and prediction, the present invention uses a Transformer-based dynamic graph encoder, a hierarchical probabilistic interaction generation module, and an autoregressive node state prediction module. These modules work together to predict the evolution of the traffic system by learning patterns and associations in historical data. The key to this method lies in the introduction of the concept of a dynamic traffic graph (DTG), which is a comprehensive framework capable of representing the traffic system and its participants in real time. Specifically, the dynamic graph learning algorithm includes the following:

[0067] (1) Dynamic graph representation: Generate the dynamic traffic graph structure according to step S2.

[0068] (2) Interaction generation: Predict the generation probability of edges using the interaction generation module based on the traffic graph information at the current moment, and represent the interaction relationship between traffic participants by generating edges.

[0069] (3) State prediction: The states of traffic participants evolve over time. The state prediction module predicts the node features at the next time step based on historical node information and predicted edge information to obtain the traffic participant state prediction result.

[0070] (4) Repeat the above process to complete traffic simulation.

[0071] To simulate complex traffic behaviors and relationships, this embodiment uses Heterogeneous Graph Transformer (HGT) to model the evolution of the traffic system. Given a dynamic heterogeneous graph G(V, E, O, R), which represents the state of the traffic simulation system, assume that the input feature of node v is the node feature H (l-1) [v] at the historical moment, and the output of the target node v is represented as the node feature H (l) [v] to be predicted. To aggregate information from all neighbor nodes to the target node, update the vector by using the multi-head attention mechanism and the message passing mechanism and predict the target node output based on :

[0072]

[0073]

[0074] where e i =(u i , v i ), u i , v i ∈V n represents an agent, and the edge e i =(u i , v i ), e i ∈En represents the relationship between nodes / agents; θ is a parameter to be learned during the training process, which is used to transform the aggregated feature vector before passing it to the activation function A-Linear φ(v) and adding it to the feature vector H of the previous layer (l-1) [v], and transform H (l) [v]; Attention(u,e,v) represents the attention mechanism, which determines the importance of the edge e between nodes u and v; Message(u,e,v) represents the message passing mechanism, indicating the message passed from node u to node v through edge e.

[0075] The evolution process of the traffic network is regarded as a series of graph structure transformations. Each static graph represents the state of the traffic network within a specific time period, and this series of graphs describes the entire evolution process of the traffic network:

[0076] T:G0→G1…→G n

[0077] Embodiment 2

[0078] This embodiment provides a data-driven traffic simulation system (TransWorldNG). The purpose of TransWorldNG is to solve the defects existing in traditional traffic simulation methods, including over-simplification and assumptions, lack of scalability and applicability, and difficulty in capturing the time variability and complexity of traffic conditions. To this end, TransWorldNG uses data-driven algorithms and graph computing technologies to learn traffic dynamics. Specifically, TransWorldNG uses a dynamic heterogeneous graph learning algorithm to learn complex traffic behaviors and relationships, so as to better understand the relationships and behaviors between different nodes. By directly learning traffic dynamics from multi-source data (such as road sensors, traffic cameras, mobile phones, social media, etc.), a realistic traffic simulation environment can be generated. The data-driven method makes TransWorldNG a powerful traffic simulation tool, which can generate accurate and realistic simulation environments and be used for various traffic-related multi-type tasks, such as traffic flow prediction, accident management, public transportation, etc.

[0079] As Figure 3 shown, TransWorldNG includes:

[0080] (1) Data layer, which is used to collect data inputs from different sources (such as sensors, GPS devices, etc.), obtain multi-source traffic data, and define a dynamic heterogeneous graph structure to convert the multi-source traffic data into a traffic graph structure. Nodes in the traffic graph structure represent traffic participants, and edges represent the relationships between traffic participants; the purpose of this layer is to convert the original traffic data into a structured format for subsequent analysis and simulation. This process involves the definition of the dynamic heterogeneous graph structure and the extraction of node features described above.

[0081] (2) The simulation layer is responsible for learning traffic patterns using dynamic graph learning algorithms based on the traffic graph structure and simulating the scenarios required by users to achieve traffic simulation. This layer includes a controller, SimCore, and an analysis module. The controller is responsible for managing the entire simulation process, including initializing the simulation, controlling the progress of the simulation, and ending the simulation. SimCore is the core component of the simulation, responsible for implementing the dynamic graph learning algorithm and simulating traffic scenarios based on the learned patterns. The analysis module is responsible for analyzing the results of the simulation, such as calculating metrics like traffic flow, speed, and density. This process involves the implementation of the dynamic graph learning algorithm and the learning of traffic patterns described above.

[0082] (3) The interface layer is responsible for providing a UI interface for users to interact with the traffic simulation system, so that users can conveniently use the system. This interface can include some basic controls, such as buttons, sliders, and text boxes, as well as some advanced functions, such as map display, traffic flow visualization, and simulation result analysis.

[0083] TransWorldNG is applicable to the following application scenarios:

[0084] (1) Urban traffic planning: Urban traffic planning needs to consider factors such as complex road network structures, driving behaviors of different types of vehicles, and population flow. TransWorldNG uses a dynamic heterogeneous graph learning model that can simulate these complex traffic behaviors and relationships, providing more accurate predictions and optimization solutions.

[0085] (2) Vehicle decision-making and planning: TransWorldNG uses real data to train the model and predict traffic flow, optimize routes, and plan strategies such as public transportation. For example, historical traffic data can be used to predict future traffic conditions, and then optimize the vehicle's route planning and driving strategies.

[0086] (3) Signal light control: TransWorldNG can be used to develop an intelligent traffic signal light control system to optimize the timing and control strategies of intersection signal lights, thereby improving the traffic efficiency and safety of vehicles.

[0087] (4) Development of Intelligent Transportation Systems: Intelligent transportation systems need to respond quickly to different traffic scenarios and levels of abstraction. TransWorldNG's model-free approach can adapt to different traffic scenarios and levels of abstraction without extensive model development and calibration, offering higher flexibility.

[0088] TransWorldNG is a traffic simulator that can be used in the field of traffic management and control. It uses data-driven algorithms and graph computing techniques to learn traffic dynamics and automatically learns traffic patterns from real data. This simulator can generate more realistic traffic patterns, helping decision-makers formulate traffic policies and make decisions more accurately. In addition, TransWorldNG can also be applied to emerging technology areas including mobility services and autonomous vehicles.

[0089] Embodiment 3

[0090] This embodiment provides a data-driven traffic simulation system, including a memory, a processor, and a program stored in the memory. When the processor executes the program, it implements the method described in Embodiment 1 above.

[0091] Embodiment 4

[0092] This embodiment provides a storage medium with a program stored thereon. When the program is executed, it implements the method described in Embodiment 1 above.

[0093] If the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory, or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD), or other optical storage, magnetic cassette tapes, magnetic tape magnetic disk storage, or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media do not include transitory media such as modulated data signals and carrier waves.

[0094] The preferred specific embodiments of the present invention have been described in detail above. It should be understood that those of ordinary skill in the art can make many modifications and variations based on the concept of the present invention without creative efforts. Therefore, all technical solutions that can be obtained by those skilled in the art in the technical field based on the concept of the present invention through logical analysis, reasoning, or limited experiments on the basis of the prior art should fall within the protection scope determined by the claims.

Claims

1. A data-driven traffic simulation method, characterized in that, It includes the following steps: Obtain multi-source traffic data; Define a dynamic heterogeneous graph structure, and convert the multi-source traffic data into a traffic graph structure, where the nodes in the traffic graph structure represent traffic participants and the edges represent the relationships between traffic participants; Based on the traffic graph structure, use a dynamic graph learning algorithm to learn traffic patterns and simulate the scenarios required by users to achieve traffic simulation; The described dynamic graph learning algorithm is implemented based on HGT, assuming that the input features of node v are the node features at historical moments , and the output representation of the target node is the node feature to be predicted . To aggregate information from all neighbor nodes to the target node, vectors are updated by using the multi-head attention mechanism and the message passing mechanism , and the output of the target node is predicted based on : ， ， Among them, represents an agent, and the edge represents the relationship between nodes / agents; is a parameter to be learned during training, and is used to transform before passing the aggregated feature vector to the activation function and adding it to the feature vector of the previous layer ; represents the attention mechanism, which determines the importance of the edge between node and node ; represents the message passing mechanism, indicating the message passed from node to node through the edge .

2. The data-driven traffic simulation method according to claim 1, characterized in that In the dynamic heterogeneous graph, the nodes represent agents, the edges represent the relationships between the nodes, each node corresponds to a node category, and is represented by a first unique mapping function; each edge is a directed edge and corresponds to an edge category, which is represented by a second unique mapping function.

3. A data-driven traffic simulation method according to claim 2, characterized in that, The node categories include vehicles, lanes, roads, and traffic signal controllers.

4. A data-driven traffic simulation method according to claim 3, characterized in that In the dynamic heterogeneous graph, each node corresponds to a node category and is represented by a unique mapping function. Each node includes multiple node features according to different node categories: When the node category is a vehicle, the corresponding node features include the current position, speed, acceleration, vehicle type, and vehicle length; When the node category is a lane, the corresponding node features include the lane length, maximum allowable speed, current traffic flow, occupancy rate, lane type, and lane markings; When the node category is a road, the corresponding node features include the road type, road length, width, and road surface condition; When the node category is a traffic signal controller, the corresponding node features include the signal duration, signal plan, and signal light status.

5. A data-driven traffic simulation method according to claim 2, characterized in that The edge category is determined according to the node categories connected by the edge. The edge categories include the relationship between a lane and a road, the relationship between a vehicle and its lane, the relationship between lanes, the relationship between a vehicle and a traffic signal controller, the relationship between a traffic signal controller and a lane, and the relationship between vehicles.

6. A data-driven traffic simulation method according to claim 1, characterized in that The dynamic graph learning algorithm, according to the constructed dynamic traffic graph, uses an interaction generation module to predict the generation probability of edges based on the traffic graph information at the current moment, represents the interaction relationship between traffic participants by generating edges, and uses a state prediction module to predict the node features at the next time step based on the historical node information and the predicted edge information to obtain the traffic participant state prediction result. Repeat the above process to complete traffic simulation.

7. A data-driven traffic simulation system, characterized in that, It includes: A data layer for collecting data inputs from different sources, obtaining multi-source traffic data, and defining a dynamic heterogeneous graph structure to convert the multi-source traffic data into a traffic graph structure, where the nodes in the traffic graph structure represent traffic participants and the edges represent the relationships between traffic participants; A simulation layer for learning traffic patterns based on the traffic graph structure using a dynamic graph learning algorithm and simulating the scenarios required by users to achieve traffic simulation; An interface layer for providing a UI interface for users to interact with the traffic simulation system; The dynamic graph learning algorithm is implemented based on HGT, assuming that the input features of node v are the node features at historical moments , and the output representation of the target node is the node feature to be predicted . To aggregate information from all neighbor nodes to the target node, vectors are updated by using the multi-head attention mechanism and the message passing mechanism , and the output of the target node is predicted based on : ， ， Among them, represents an agent, and the edge represents the relationship between nodes / agents; is a parameter that needs to be learned during training and is used to transform before passing the aggregated feature vector to the activation function and adding it to the feature vector of the previous layer ; represents the attention mechanism, which determines the importance of the edge between node and node ; represents the message passing mechanism, indicating the message passed from node to node through the edge ; 8. A data-driven traffic simulation system, comprising a memory, a processor, and a program stored in the memory, characterized in that, When the processor executes the program, it implements the method described in any one of claims 1-7.

9. A storage medium, on which a program is stored, characterized in that, When the program is executed, it implements the method described in any one of claims 1-7.

Images