Vehicle speed prediction method and system used in urban scene
By performing personalized aggregation steps on the server and using a sequence-to-sequence length short-term memory network model of multi-head attention, the personalized problem of vehicle speed prediction in urban market scenarios is solved, and efficient personalized prediction and data privacy protection is achieved.
Patent Information
- Application Number
- CN202411091310.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-09
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2044-08-09
AI Technical Summary
Existing vehicle speed prediction methods cannot achieve personalized predictions, especially in urban scenarios. Traditional federated learning methods cannot effectively capture specific information from clients, resulting in limited prediction accuracy and risk of private data breaches.
A personalized federated learning method is used to perform aggregation steps on the server, personalized prediction of vehicle speed through a sequence-to-sequence length short-term memory network model based on multi-head attention, and aggregation of client local models and global at the element level using personalized aggregation weights. Model, reducing client computing and communication overhead.
It realizes the protection of driver data privacy while providing personalized vehicle speed prediction, reducing client computing and communication overhead, and improving prediction accuracy.
Smart Images

Figure CN119091646A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of vehicle speed prediction, and relates to a vehicle speed prediction method and system for use in urban scenes. Background Art
[0002] Vehicle speed prediction plays an important role in modern intelligent transportation systems and is a key technology for improving road safety, traffic efficiency, and vehicle energy efficiency. It is widely used in path planning, energy management of hybrid electric vehicles, and eco-adaptive cruise control. However, accurately predicting the speed of a single vehicle is very challenging because the vehicle speed is affected by traffic conditions, vehicle type, road conditions, and driver behavior. Traditional methods for vehicle speed prediction include Markov chains, Bayesian networks, etc. Compared with traditional prediction methods, recurrent neural networks using vehicle-to-vehicle and vehicle-to-infrastructure communications have been shown to be more effective in vehicle speed prediction. However, previous studies have not taken into account the significant differences in driver driving styles and vehicle types, which greatly affect the speed prediction of the target vehicle, and therefore cannot achieve personalized prediction. Collaborative training of driving data collected from multiple vehicles is essential to improve prediction accuracy because one vehicle does not have enough data to train a reliable model. However, traditional distributed machine learning techniques require the private driving data of all vehicles to be concentrated on a central server (e.g., a cloud server), which poses a potential risk for private data leakage.
[0003] Federated learning is a collaborative distributed machine learning paradigm in which learning tasks are coordinated by a central server and solved jointly by a network of multiple participating devices (called clients). Each client holds its own private training dataset, which is not uploaded to the server. Instead, the local model trained locally is uploaded to the server for aggregation to obtain a global model, which effectively reduces the privacy risk of the system. However, when the traditional federated learning method is applied to vehicle speed prediction, a single global model cannot achieve personalized vehicle speed prediction.
[0004] Personalized federated learning methods can address statistical heterogeneity in federated learning and provide personalized solutions. Unlike traditional federated learning that seeks a single high-quality global model, personalized federated learning methods are proposed to prioritize training the local model of each client.
[0005] However, almost all existing personalized federated learning methods require additional steps on the client to achieve personalization, which usually increases the client's computing or communication overhead. Considering the high real-time requirements of driving scenarios and the limited computing and communication resources of vehicles, it is undoubtedly more appropriate to implement the personalization steps on a powerful server. Therefore, there is an urgent need for a personalized federated learning method on the server to achieve personalized vehicle speed prediction. Summary of the invention
[0006] In view of this, an object of the present invention is to provide a vehicle speed prediction method and system for urban scenarios, which captures client-specific information by implementing a personalized aggregation step on a server and performs personalized prediction of vehicle speed through a new prediction model.
[0007] In order to achieve the above object, the present invention provides the following technical solutions:
[0008] A vehicle speed prediction method for urban scenarios, in which a personalized federated learning method is used to implement a personalized aggregation step on a server to capture client-specific information; the acquired data information is used to perform personalized prediction of vehicle speed through a sequence-to-sequence long short-term memory network model based on multi-head attention.
[0009] Furthermore, the personalized federated learning method comprises the following steps:
[0010] S1. The server sends the initialized global model to all clients and initializes the local model of each client. The client is numbered as i, and the number of iterations of the current personalized federated learning method is t.
[0011] S2, client i executes the tth iteration, trains the local model by gradient descent, and uploads the trained local model to the server;
[0012] S3: The server performs global aggregation on the uploaded local model to obtain a global model.
[0013] S4. The server determines whether it is necessary to calculate the personalized aggregation weight according to whether the current iteration round number t is greater than the set threshold r; if not, the server sends the global model to the selected client as its initialization local model and returns to step S2; if greater, proceed to step S5;
[0014] S5. The server calculates the weighted mean square error between the client local model and the global model parameters and normalizes it to obtain the personalized aggregation weight for this iteration.
[0015] S6, the server obtains the client-customized aggregate model based on the personalized aggregation weight by aggregating the highest p-layer network of the client local model and the global model at the element level, while the remaining lower-layer networks adopt the corresponding parts of the global model;
[0016] S7, the server sends the client-customized aggregation model to the selected client as its initialization local model;
[0017] S8. Repeat steps S2 to S7, perform T rounds of iterations, t∈[0, T], until the local model training of each client converges.
[0018] Further, in step S1, the global objective function of federated learning is defined as:
[0019]
[0020] in, and and is the local objective function of the i-th client; D i is the dataset of the i-th client; |D i | is the dataset size of the i-th client; Initialize the local model on the i-th client.
[0021] Further, in step S2, for the tth round of iterative training, the server randomly selects a subset S of N clients t Train and send the global model or aggregate model obtained in the previous iteration to the set S t The client in;
[0022] Client i performs the tth iteration and trains the local model by gradient descent. The specific method is as follows:
[0023]
[0024] Where η is the learning rate for local model training.
[0025] Further, in step S3, the server performs global aggregation on the uploaded local model to obtain a global model, specifically in the following manner:
[0026]
[0027] Further, in step S5, the client subset S is calculated t All local models and the p higher layers of the global model: With Θ t,p The weighted mean square error of the parameters is used to obtain the model parameter difference measure M t,p, which reflects the degree of difference in model parameters between different clients, and this difference is weighted based on the number of local data samples, as follows:
[0028]
[0029] M t,p Each layer After normalization, the aggregation weight W is obtained t,p , so that The specific method is as follows:
[0030]
[0031] Further, in step S6, it includes:
[0032] The server is based on the personalized aggregation weight W t,p , by aggregating client-side local models at the element level and the global model Θ t The highest p-layer network is used, while the remaining lower-layer networks adopt the corresponding parts of the global model to obtain the client-customized aggregation model, which is as follows:
[0033]
[0034] Where ⊙ is the Hadamard product, which means that the elements in the same position of two matrices are multiplied; L(Θ i )yes The number of layers, and The lower layers of The element is 0; the aggregation weight W t,p Same shape as the remaining p higher layers.
[0035] Furthermore, the sequence-to-sequence long short-term memory network model based on multi-head attention specifically includes: predicting the vehicle speed in a future period of time based on a period of historical input feature data; the model adopts an encoder-decoder architecture; in addition, the model adopts a long short-term memory network and a multi-head attention mechanism; specifically, the model first processes the input sequence through a two-layer long short-term memory network as an encoder to extract a high-dimensional representation of the time series data; then, through a custom multi-head attention layer, the sequence dependencies are captured in multiple subspaces in parallel to enhance the understanding of the complex patterns of the time series; next, the output of the multi-head attention module is sent to a two-layer long short-term memory network as a decoder to gradually build a prediction of the future vehicle speed; finally, the output of the decoder is mapped to the predicted vehicle speed through a fully connected layer.
[0036] The present invention also provides a vehicle speed prediction system for urban scenes.
[0037] The beneficial effects of the present invention are:
[0038] The present invention provides a vehicle speed prediction method and system for urban scenarios, including a personalized federated learning method based on personalized aggregation weights, which performs personalized aggregation on a server to capture client-specific information; the method sends a customized aggregation model to the client to achieve personalization without bringing additional computing and communication overhead to the client; in addition, it also includes a sequence-to-sequence long short-term memory network model based on multi-head attention, and an input feature construction method thereof, which are used to replace a traditional vehicle speed prediction model; the present invention can achieve personalized vehicle speed prediction while protecting driver data privacy by providing a new personalized federated learning method and a new vehicle speed prediction model.
[0039] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:
[0041] Figure 1 A flowchart of a personalized federated learning method based on personalized aggregation weights provided in the first embodiment of the present invention;
[0042] Figure 2 A system framework diagram of a personalized federated learning method based on personalized aggregation weights provided in the first embodiment of the present invention;
[0043] Figure 3 A scene diagram for predicting vehicle speed through vehicle-to-vehicle and vehicle-to-infrastructure communications provided in Embodiment 2 of the present invention;
[0044] Figure 4 A schematic diagram of the structure of a sequence-to-sequence long short-term memory network model based on multi-head attention provided in Example 2 of the present invention. DETAILED DESCRIPTION
[0045] The technical solution of the present invention is described in detail below with reference to the accompanying drawings.
[0046] The technical solution of the present invention includes a personalized federated learning method based on personalized aggregation weights according to a first aspect, and the method includes the following steps:
[0047] Step 1: The server sends the initialized global model to all clients and initializes the local model of each client; where i is the client number and t is the number of iterations of the current personalized federated learning method;
[0048] Step 2: Client i executes the tth iteration, trains the local model by gradient descent, and uploads the trained local model to the server;
[0049] Step 3: The server performs global aggregation on the uploaded local model to obtain a global model;
[0050] Step 4: The server determines whether it is necessary to calculate the personalized aggregation weight according to whether the current iteration round number t is greater than the set threshold r; if not, the server sends the global model to the selected client as its initialization local model and returns to step 2; if greater, proceed to step 5;
[0051] Step 5: The server calculates the weighted mean square error between the client local model and the global model parameters and normalizes it to obtain the personalized aggregation weight for this iteration;
[0052] Step 6: Based on the personalized aggregation weight, the server obtains the client-customized aggregation model by aggregating the highest p-layer network of the client local model and the global model at the element level, while the remaining lower-layer networks adopt the corresponding parts of the global model;
[0053] Step 7: The server sends the client-customized aggregation model to the selected client as its initialization local model;
[0054] Step 8: Repeat steps 2 to 7 for T rounds of iterations until the local model training of each client converges; where t∈[0, T].
[0055] The second aspect of the present invention provides a sequence-to-sequence long short-term memory network model based on multi-head attention, and the input features and model structure of the vehicle speed prediction model include:
[0056] For a discrete time setting with an interval of 1s, the vehicle speed prediction model can predict the speed scalar of the target vehicle within a certain range in the future through some input features. Through vehicle-to-vehicle and vehicle-to-infrastructure communication, the input features mainly come from the target vehicle, the preceding vehicle and the traffic lights, which are the target vehicle speed v T , front vehicle speed v P , Is there a preceding vehicle? P 、Is there a traffic light? TL , the relative distance d between the target vehicle and the preceding vehicle P , the relative distance d between the target vehicle and the stop line of the traffic light TL, the current state of the traffic light signal TL , the future state of the traffic light signal TL,k+1 , ..., s TL,k+H , target vehicle throttle, target vehicle brake, target vehicle steering angle steer; where the preceding vehicle refers to the nearest preceding vehicle in the same lane, H is the predicted time range, and k represents the current discrete time scale;
[0057] The multi-head attention-based sequence-to-sequence long short-term memory network model predicts the vehicle speed in the future based on a period of historical input feature data, that is, the input features are organized into the input sequence of the model in chronological order, and the future speed of the target vehicle is used as the output sequence of the model; the model adopts an encoder-decoder architecture, which has been proven to be effective in various sequence modeling tasks; in addition, the model adopts a long short-term memory network and a multi-head attention mechanism, the former is good at capturing time dependencies, and the latter can capture complex time dynamics and long-distance dependencies in the sequence; specifically, the model first processes the input sequence through a two-layer long short-term memory network as an encoder to extract a high-dimensional representation of the time series data; then, through a custom multi-head attention layer, the sequence dependencies are captured in multiple subspaces in parallel to enhance the understanding of the complex patterns of the time series; next, the output of the multi-head attention module is sent to a two-layer long short-term memory network as a decoder to gradually build a prediction of the future vehicle speed; finally, the output of the decoder is mapped to the predicted vehicle speed through a fully connected layer.
[0058] Embodiment 1:
[0059] Reference Figure 1 , which is a flow chart of a personalized federated learning method based on personalized aggregation weights provided in the first embodiment of the present invention;
[0060] The personalized federated learning method based on personalized aggregation weights provided in the first embodiment of the present invention is applied to vehicle speed prediction of intelligent connected vehicles, and includes the following steps:
[0061] Step 1: The server sends the initialized global model to all clients and initializes the local model of each client; where i is the client number and t is the number of iterations of the current personalized federated learning method;
[0062] Specifically, step 1 includes:
[0063] Step 1.1: Number the N intelligent connected vehicles (i.e., clients) participating in federated learning from 1 to N; record the dataset of the i-th client as D i , the size of the data set is denoted as |D i|, and these data sets are not independent and identically distributed; the local model on the i-th client is denoted as Θ i ; The number of layers of the local model is recorded as L(Θ i ); the initialized local model on the i-th client is recorded as The global model is denoted as Θ; the personalized aggregation weight is denoted as W p , where p means that personalized aggregation is performed only in the last p layers of the local model; the loss function is recorded as The local objective function of the i-th client is recorded as
[0064] Step 1.2: Define the global objective function of federated learning as:
[0065]
[0066] in, and and is the local objective function of the i-th client.
[0067] Step 1.3: The server sends the initialized global model Θ 0 For all clients, initialize the client's local model.
[0068] Step 2: Client i executes the tth iteration, trains the local model by gradient descent, and uploads the trained local model to the server;
[0069] Specifically, step 2 includes:
[0070] Step 2.1: For the tth round of iterative training, the server randomly selects a subset S of N clients t Train and send the global model or aggregate model obtained in the previous iteration to the set S t The client in
[0071] Step 2.2: Client i performs the tth iteration and trains the local model by gradient descent. The specific method is as follows:
[0072]
[0073] Where η is the learning rate of local model training;
[0074] Step 3: The server performs global aggregation on the uploaded local model to obtain the global model. The specific method is as follows:
[0075]
[0076] Step 4: The server determines whether it is necessary to calculate the personalized aggregation weight according to whether the current iteration round number t is greater than the set threshold r; if not, the server sends the global model to the selected client as its initialization local model and returns to step 2; if greater, proceed to step 5;
[0077] Step 5: The server calculates the weighted mean square error between the client local model and the global model parameters and normalizes it to obtain the personalized aggregation weight for this iteration;
[0078] Specifically, step 5 includes:
[0079] Step 5.1: Calculate the client subset S t All local models and the p higher layers of the global model: With Θ t,p The weighted mean square error of the parameters is used to obtain the model parameter difference measure M t,p , which reflects the degree of difference in model parameters between different clients, and this difference is weighted based on the number of local data samples, as follows:
[0080]
[0081] Step 5.2, M t,p Each layer After normalization, the aggregation weight M is obtained t,p , so that The specific method is as follows:
[0082]
[0083] Step 6: Based on the personalized aggregation weight, the server obtains the client-customized aggregation model by aggregating the highest p-layer network of the client local model and the global model at the element level, while the remaining lower-layer networks adopt the corresponding parts of the global model. The specific method is as follows:
[0084]
[0085] Where ⊙ is the Hadamard product, which means that the elements in the same position of two matrices are multiplied; L(Θ i )yes The number of layers, and The lower layers of The element is 0; the aggregation weight W t,p Same shape as the remaining p higher layers.
[0086] Step 7: The server sends the client-customized aggregation model Sent to the selected client as its initial local model;
[0087] Step 8: Repeat steps 2 to 7 for T rounds of iterations until the local model training of each client converges; where t∈[0,T].
[0088] Reference Figure 2 , which is a system framework diagram of the personalized federated learning method based on personalized aggregation weights provided in the first embodiment of the present invention; specifically, in the t-th round of iterative training:
[0089] (1) First, the server randomly selects some intelligent connected vehicles (i.e., clients) for personalized aggregation, and then the clients use their trained local models Upload to the server;
[0090] (2) The server then averages and aggregates these client local models to obtain a global model Θ t ; Then, the server calculates the personalized aggregation weight W t , aggregate the client local model with the global model at the element level to obtain a customized aggregate model
[0091] (3) Finally, the server distributes these customized aggregation models to each client for local training to facilitate personalized vehicle speed prediction for intelligent connected vehicles.
[0092] Embodiment 2:
[0093] Reference Figure 3 , which is a scene diagram of vehicle speed prediction through vehicle-to-vehicle and vehicle-to-infrastructure communication provided by the second embodiment of the present invention. Specifically, through vehicle-to-vehicle and vehicle-to-infrastructure communication, the input features mainly come from the target vehicle, the preceding vehicle and the traffic light, which are: target vehicle speed v T , front vehicle speed v P , Is there a preceding vehicle? P 、Is there a traffic light? TL , the relative distance d between the target vehicle and the preceding vehicle P , the relative distance d between the target vehicle and the stop line of the traffic light TL , the current state of the traffic light signal TL , the future state of the traffic light signal TL,k+1 , ..., s TL,k+H , target vehicle throttle, target vehicle brake, target vehicle steering angle steer; the preceding vehicle refers to the nearest preceding vehicle in the same lane, H is the predicted time range, and k represents the current discrete time scale.
[0094] Reference Figure 4, which is a schematic diagram of the structure of a sequence-to-sequence long short-term memory network model based on multi-head attention provided by the second embodiment of the present invention. Specifically, the model predicts the vehicle speed in the future based on a period of historical input feature data, that is, the input features are organized into the input sequence x of the model in chronological order. k-M+1 , ..., x k , and the future speed of the target vehicle is taken as the output sequence v of the model t,k+1 , ..., v t,k+H , where k represents the current discrete time scale, M represents the past time range, and H represents the predicted time range; the model adopts an encoder-decoder architecture, which has been proven to be effective in various sequence modeling tasks; in addition, the model adopts a long short-term memory network and a multi-head attention mechanism, the former is good at capturing time dependencies, and the latter can capture complex time dynamics and long-distance dependencies in the sequence; specifically, the model first processes the input sequence through a two-layer long short-term memory network (Long Short-Term Memory, LSTM) as an encoder to extract a high-dimensional representation of the time series data; then, through a custom multi-head attention (Multi-Head Attention, MHA) layer, the sequence dependencies are captured in multiple subspaces in parallel to enhance the understanding of the complex patterns of the time series; next, the output of the multi-head attention layer is fed into a two-layer long short-term memory network as a decoder to gradually build a prediction of the future vehicle speed; finally, a fully connected (Fully Connected, FC) layer is used to map the output of the decoder to the predicted vehicle speed.
[0095] In summary, the present invention provides a vehicle speed prediction method and system for urban scenarios, including a personalized federated learning method based on personalized aggregation weights, which performs personalized aggregation on a server to capture client-specific information; the method sends a customized aggregation model to the client to achieve personalization without bringing additional computing and communication overhead to the client; in addition, it also includes a sequence-to-sequence long short-term memory network model based on multi-head attention, and an input feature construction method thereof, which are used to replace the traditional vehicle speed prediction model; the present invention can achieve personalized vehicle speed prediction while protecting the driver's data privacy by providing a new personalized federated learning method and a new vehicle speed prediction model.
[0096] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.
Claims
1. A vehicle speed prediction method for urban scenes, characterized in that: In this method, a personalized federated learning method is used to implement a personalized aggregation step on the server to capture client-specific information; the acquired data information is used to make a personalized prediction of the vehicle speed through a sequence-to-sequence long short-term memory network model based on multi-head attention.
2. A vehicle speed prediction method for urban scenes according to claim 1, characterized in that: The personalized federated learning method comprises the following steps: S1. The server sends the initialized global model to all clients and initializes the local model of each client. The client is numbered as i, and the number of iterations of the current personalized federated learning method is t. S2, client i executes the tth iteration, trains the local model by gradient descent, and uploads the trained local model to the server; S3: The server performs global aggregation on the uploaded local model to obtain a global model. S4. The server determines whether it is necessary to calculate the personalized aggregation weight according to whether the current iteration round number t is greater than the set threshold r; if not, the server sends the global model to the selected client as its initialization local model and returns to step S2; if greater, proceed to step S5; S5. The server calculates the weighted mean square error between the client local model and the global model parameters and normalizes it to obtain the personalized aggregation weight for this iteration. S6, the server obtains the client-customized aggregate model based on the personalized aggregation weight by aggregating the highest p-layer network of the client local model and the global model at the element level, while the remaining lower-layer networks adopt the corresponding parts of the global model; S7, the server sends the aggregate model customized by the client to the selected client as its initialization local model; S8. Repeat steps S2 to S7, perform T rounds of iterations, t∈[0,T], until the local model training of each client converges.
3. A vehicle speed prediction method for urban scenes according to claim 2, characterized in that: In step S1, the global objective function of federated learning is defined as: in, and and is the local objective function of the i-th client; D i is the dataset of the i-th client; |D i | is the dataset size of the i-th client; Initialize the local model on the i-th client.
4. The vehicle speed prediction method for urban scenes according to claim 3, characterized in that: In step S2, for the tth round of training iteration, the server randomly selects a subset S of N clients t Train and send the global model or aggregate model obtained in the previous iteration to the set S t The client in Client i performs the tth iteration and trains the local model by gradient descent. The specific method is as follows: Where η is the learning rate for local model training.
5. The vehicle speed prediction method for urban scenes according to claim 4, characterized in that: In step S3, the server performs global aggregation on the uploaded local model to obtain a global model, as follows:
6. The vehicle speed prediction method for urban scenes according to claim 5, characterized in that: In step S5, the client subset S is calculated t All local models and the p higher layers of the global model: With Θ t,p The weighted mean square error of the parameters is used to obtain the model parameter difference measure M t,p , which reflects the degree of difference in model parameters between different clients, and this difference is weighted based on the number of local data samples, as follows: M t,p Each layer After normalization, the aggregation weight W is obtained t,p , so that The specific method is as follows:
7. The vehicle speed prediction method for urban scenes according to claim 6, characterized in that: In step S6, it includes: The server is based on the personalized aggregation weight W t,p , by aggregating client-side local models at the element level and the global model Θ t The highest p-layer network is used, while the remaining lower-layer networks adopt the corresponding parts of the global model to obtain the client-customized aggregation model, which is as follows: Where ⊙ is the Hadamard product, which means that the elements in the same position of two matrices are multiplied; L(Θ i )yes The number of layers, and The lower layers of The element is 0; the aggregation weight W t,p Same shape as the remaining p higher layers.
8. The vehicle speed prediction method for urban scenes according to claim 1, characterized in that: The multi-head attention-based sequence-to-sequence long short-term memory network model specifically includes: predicting the vehicle speed in the future based on a period of historical input feature data; the model adopts an encoder-decoder architecture; in addition, the model adopts a long short-term memory network and a multi-head attention mechanism; specifically, the model first processes the input sequence through a two-layer long short-term memory network as an encoder to extract a high-dimensional representation of the time series data; then, through a custom multi-head attention layer, the sequence dependencies are captured in multiple subspaces in parallel to enhance the understanding of the complex patterns of the time series; next, the output of the multi-head attention module is sent to a two-layer long short-term memory network as a decoder to gradually build a prediction of the future vehicle speed; finally, the output of the decoder is mapped to the predicted vehicle speed through a fully connected layer.
9. A vehicle speed prediction system for urban scenes, characterized in that: The system adopts the method as claimed in any one of claims 1 to 8.
Citation Information
Patent Citations
Vehicle track prediction method based on dynamic interaction graph convolution
CN114802296A
Personalized federal learning method based on adaptive local aggregation
CN115481755A
Vehicle track prediction method based on intention perception space-time attention network
CN117141518A
Method and system for predicting trajectories of agents
EP4124999A1
