Method and system for constructing a lookup table for distributed congestion control, and a distributed congestion control method using such a lookup table

Abstract

The present invention provides a method for configuring a lookup table for distributed congestion control (DCC) in a network comprising a roadside unit (RSU) and a plurality of transmitting user equipments (TX UEs), the TX UEs being capable of communicating with the RSUs and capable of implementing distributed congestion control for sidelink communications of the TX UEs using the lookup table, the method comprising: -Monitoring sidelink traffic conditions in the vicinity of the RSU; training a machine learning model; dynamically configuring a lookup table for a target TX UE in the TX UE based on the monitored sidelink traffic conditions and the trained machine learning model; The present invention relates to a method, including: The present invention improves distributed congestion control.

Description

[Technical Field]

[0001] The present invention relates to distributed congestion control in networks, particularly in networks using sidelink communication techniques. In particular, the present invention relates to a method and system for configuring lookup tables for distributed congestion control. Priority is claimed to European Patent Application No. 23315157.0, filed May 4, 2023, the contents of which are incorporated herein by reference. [Background technology]

[0002] Future intelligent transportation systems are expected to support a wide variety of specialized applications. These applications rely on both vehicle / train-to-infrastructure (V2I / T2I) and vehicle-to-vehicle / train-to-train (V2V / T2T) links to transmit data packets. Currently, several data traffic types exist in LTE V2X, including traditional broadcast safety messages (BSM) and cooperative awareness messages (CAM). However, new V2X application messages are on the rise, such as data traffic for autonomous driving, remote driving, vehicle platooning, and automated valet parking. As these technologies become more powerful, they will require more advanced communication systems to function effectively.

[0003] One promising candidate for both V2I / T2I and V2V / T2T communications is sidelink communications technology. This technology is particularly well suited to distributed communications scenarios and can offer superior performance when used in this manner. Indeed, sidelink communications technology is expected to play a key role in facilitating the development of future intelligent transportation systems.

[0004] Despite the many advantages of sidelink communication technology, several challenges remain that must be overcome before it can be widely adopted. For example, it is necessary to find ways to optimize the performance of sidelink communication systems in complex urban environments with many competing wireless signals. Furthermore, concerns remain about the security of sidelink communication technology, and researchers must work hard to develop secure communication protocols that can prevent unauthorized access to sensitive data.

[0005] Despite these challenges, the potential benefits of sidelink communications technology are too great to ignore. As the development of intelligent transportation systems continues, sidelink technology will play an increasingly important role in facilitating the safe and efficient operation of these systems.

[0006] Particularly in sidelink communications, congestion control is a key mechanism for maintaining proper network functioning. Congestion control regulates the entry of data packets into the network and effectively and fairly manages shared resources to avoid congestion collapse. For example, in 5G NR V2X Mode 2, a distributed congestion control (DCC) scheme can be used, which provides a more decentralized approach to congestion control.

[0007] Instead of defining a specific congestion control algorithm, the DCC scheme defines relevant metrics and possible measures to reduce channel congestion within the standard. Two important metrics measured and calculated by each transmitting user equipment (TX UE) in the DCC scheme are the sidelink channel busy ratio (CBR) and the sidelink channel occupancy ratio (CR). These metrics are important in determining the current state of the network and identifying areas of congestion that need to be addressed.

[0008] To ensure that each TX UE operates within acceptable limits, the maximum channel occupancy (CR) that a TX UE can have for a given CBR range is defined. limitA predetermined lookup table (pre-)configuration signaling is used to indicate the CBR. Based on the measured CBR, each TX UE can obtain the corresponding CR from the lookup table (pre-)configuration signaling. limit can be obtained, and then the measured CR can be expressed as CR limit can be compared to

[0009] The measured CR is CR limit If the CR exceeds 1, implementation-specific DCC action should be taken to reduce the CR. This action may include packet dropping or changing transmission parameters such as TX power, modulation and coding scheme (MCS), or number of transmissions per resource reservation interval. By taking these proactive measures, the DCC scheme can ensure that the network continues to function stably and properly, even during periods of high traffic or congestion.

[0010] DCC schemes in the art that use (pre-)configured look-up tables are discussed in more detail below.

[0011] In 5G NR V2X Mode 2, a distributed congestion control (DCC) scheme is applied to regulate the entry of data packets into the network and manage shared resources in an effective and fair manner to avoid congestion collapses. This scheme is called distributed because each TX UE manages its congestion control based solely on three factors: (1) its own observations of the sidelink channel conditions (sidelink CBR and CR measurements), (2) a (pre-)configured lookup table in this TX UE that indicates whether some action should be taken if the observed measurements belong to certain conditions, and (3) implementation-specific DCC actions that can reduce the congestion condition in this TX UE.

[0012] A (pre)configured lookup table provides multiple CBR ranges and associated CR limitIf a TX UE observes a CBR measurement that falls within the above CBR range, then that CR measurement will be limit The value should not exceed the CR measurement value. limit If CR > CR, the TX UE does not consider the system to be congested and no DCC action is required. limit If , it indicates potential congestion in the system and the TX UE should take DCC action to reduce its sidelink channel utilization. The lookup table can be multi-dimensional to take into account other parameters such as packet priority. In this case, for each priority level, a CBR range and associated CR limit The value is defined.

[0013] The standardization organization ETSI currently specifies lookup tables for 802.11p and LTE-V2X DCC mechanisms [ETSI TR 101 612, ETSI TS 103 574]. This means that the table structure, how the table is organized, and how the table is used are determined in advance. Lookup tables can also be defined per traffic priority; examples of lookup tables proposed in 802.11p and LTE harmonized specifications at ETSI or 3GPP can be found in [ETSI TS 103 574]. For NR V2X sidelinks, 3GPP has agreed that the lookup table should have a maximum of 16 rows [TS 38.211]; the values ​​within the lookup table have not yet been specified. There has been some discussion within ETSI about extending and validating the LTE V2X lookup table for NR V2X, but no conclusion has yet been reached. The lookup table can be pre-configured, i.e., the values ​​can be written in advance (by the chipset manufacturer or vehicle manufacturer) into the memory of each TX UE, or it can be configured via Uu link signaling when the TX UE is within the coverage of the BS.

[0014] The (pre-)configuration of DCC lookup tables in accordance with the above standards has the following drawbacks, especially when NR V2X sidelink communications are considered:

[0015] First, lookup tables in the art are rigid and can easily become outdated over time. The lookup tables proposed in the harmonized specifications in ETSI and 3GPP® are based on system-level simulations with simple, non-uniform traffic. Due to limited applications and QoS requirements, such assumptions are reasonable for 802.11p and LTE-V2X systems. However, in NR V2X systems, there are various applications with different associated QoS constraints. For example, LTE V2X has a limited set of resource reservation interval (RRI) values: 20 ms, 50 ms, 100 ms, or any multiple of 100 ms up to a maximum value of 1000 ms. However, in NR V2X, any RRI value between 1 and 100 ms is also possible. Due to the potentially high heterogeneity of traffic in NR V2X sidelink communication systems, lookup tables constructed from simple traffic assumptions may not capture the system behavior and therefore may lead to degradation of overall DCC performance.

[0016] Second, the distributed nature of the DCC mechanism leads to local measurement-based congestion control decisions, which cannot take into account the congestion status of other users and cannot make globally optimal decisions. The DCC algorithm attempts to throttle the channel utilization of each TX node below a (pre-)configured threshold, promoting fair channel access among all TXs. Due to its distributed nature, this mechanism is like a gentleman's agreement that should work well when everyone is in a more or less similar situation and makes the same decisions. However, if the congestion status among TXs is unevenly distributed due to local perturbations, such a mechanism may lead to unfair channel utilization and poor overall DCC performance.

[0017] Finally, the DCC mechanism is a closed-loop system, and implementation-specific DCC behavior affects traffic characteristics. Therefore, the optimal thresholds defined by the lookup table (pre-configuration) should be adapted according to shifts in traffic distribution. As mentioned above, all implementation-specific DCC behaviors, such as packet drops and TX parameter changes, including but not limited to TX power changes, modulation coding scheme (MCS) changes, and number of transmissions per RRI change, will in turn change the overall traffic characteristics. Therefore, the optimal lookup table needs to constantly update the thresholds to match the current traffic distribution. This cannot be achieved with a rigid (pre-configuration) based lookup table.

[0018] To better understand DCC using lookup tables in the art, the DCC lookup table configurations in conventional 802.11p and LTE V2X systems will now be discussed.

[0019] In particular, the DCC lookup tables proposed in ETSI or LTE V2X harmonized specifications are based on system-level simulations. For a given deployment scenario (e.g., urban, highway scenario), physical layer simulation parameters and simple and uniform traffic are assumed. To ensure that target TX UEs have fair sidelink channel utilization, the DCC concept in this case is to limit the channel utilization by each UE to a maximum allowed limit of the ratio of total radio resources that can be used in a given geographical area. One way to do this is to divide the maximum allowed proportion of busy radio resources among TX UEs in a given geographical area. In this case, the channel resource utilization limit (CR) is used. limit ) is assumed to be approximated by

number

[0020] However, in an NR V2X system, due to the first drawback of DCC lookup table (pre-)configuration mentioned above, NR V2X traffic can be highly heterogeneous, and a rigid (pre-)configured lookup table built from simple traffic assumptions may not capture the system behavior and therefore may lead to a degradation of overall DCC performance.

[0021] In the art, CBR measurements vs. N neibor It has been proven that it is important to correctly identify the relationship between the CBR and the traffic mix. The problem with the DCC method is that the CBR measurements observed at each TX UE cannot estimate the heterogeneous traffic that leads to such a level of channel busy rate. Therefore, even if the average communication density (average system load) is exactly the same and the TX UEs observe the same level of CBR measurements, the N CBR measurements given the CBR may vary depending on the traffic mix and traffic parameters. neibor As a result, the methods in the art are likely to give incorrect neighbor TX UE estimates, which is due to the N neibor This can lead to incorrect calculation of the DCC frequency and subsequently to poor DCC performance.

[0022] Therefore, the (pre)configured lookup table-based DCC mechanisms in the aforementioned technical fields may not be the best solution for NR V2X DCC mechanisms due to various reasons. One of the main reasons is that the (pre)configured lookup tables may become outdated and inflexible over time. This may lead to a suboptimal decision-making process regarding congestion control. Furthermore, the (pre)configured lookup table-based DCC mechanism only takes into account local measurement-based congestion control decisions, which may not be optimal for making globally beneficial decisions that can take into account the congestion status of other users. Another important factor to consider is that the DCC mechanism operates as a closed-loop system, which means that the specific DCC operation implemented can have a significant impact on traffic characteristics. Taking this into account, it is important to adapt the optimal threshold defined by the (pre)configured lookup table according to shifts in traffic distribution.

[0023] In summary, a (pre-)configured lookup table-based DCC mechanism can be a useful tool in certain situations, but may not be the best option for an NR V2X DCC mechanism.

[0024] The present invention aims to address these problems in the art in order to improve overall system performance. Summary of the Invention

[0025] In this regard, according to one aspect of the present invention, there is provided a method for configuring a lookup table for distributed congestion control (DCC) in a network comprising a roadside unit (RSU) and a plurality of transmitting user equipments (TX UEs), the TX UEs being capable of communicating with the RSUs and capable of implementing distributed congestion control for sidelink communications of the TX UEs using the lookup table, the method comprising: -Monitoring sidelink traffic conditions in the vicinity of the RSU; training a machine learning model; dynamically configuring a lookup table for a target TX UE in the TX UE based on the monitored sidelink traffic conditions and the trained machine learning model; A method is provided, comprising:

[0026] With such a configuration, it is now possible to dynamically configure lookup tables for distributed congestion control according to actual traffic conditions so as to optimize overall system performance.

[0027] Optionally, the sidelink communication is based on NR V2X sidelink mode 2.

[0028] According to one embodiment of the present invention, the machine learning model is trained to estimate a number of TX UEs performing sidelink transmissions in a neighborhood of the target TX UE, the neighborhood being an area in which the target TX UE is adapted to communicate with other TX UEs via sidelink communication;

[0029] The lookup table for the target TX UE is given by the following equation (Equation 1):

number

[0030] By such configuration based on the channel resource utilization limits for lookup table configuration, distributed congestion control (DCC) schemes based on pre-configured lookup tables can be substantially improved with dynamic characteristics.

[0031] Further, optionally, the method according to the present invention further comprises: classifying the monitored sidelink traffic conditions into multiple traffic clusters, each traffic cluster including multiple traffics having similar traffic characteristic attributes; generating synthetic data for a TX UE performing sidelink communication based on a plurality of traffic clusters; training a machine learning model using the synthetic data as training data; Further includes:

[0032] Clustering sidelink traffic conditions can accelerate the training of machine learning models. However, clustering is not a required step, and it is possible to skip the clustering step and directly train a model based on traffic characteristic attributes, which means that the clustering step is unnecessary.

[0033] Optionally, the multiple traffic clusters are classified by K-means clustering or probabilistic clustering methods, for example, an expectation-maximization clustering algorithm may be used to cluster traffic flows with similar characteristics for sidelink communication.

[0034] Optionally, the multiple traffic clusters are configured with the following traffic characteristic attributes: Resource Reservation Interval (RRI), number of (re)transmissions (N transmission ), and the number of subchannels occupied for the transport block (L subchannel ) are classified according to

[0035] Optionally, the combined data of the TX UEs performing sidelink communication includes a traffic cluster tuple of each TX UE to which its traffic belongs, and each traffic cluster tuple includes a resource reservation interval (RRI), a number of (re)transmissions (N transmission ), the subchannel occupancy of this TX UE for each reservation (L subchannel ), which contains the probability (Pr) of the traffic type to which this TX UE belongs.

[0036] Optionally, the combined data includes a channel busy ratio (CBR) of the target TX UE and the number of TX UEs performing sidelink transmission in the neighborhood of the target TX UE.

[0037] By using this synthetic data as training data, machine learning models can be trained more efficiently and accurately.

[0038] Optionally, according to the invention, these synthetic data are generated by system level simulation, for example a digital twin of a real-world implementation.

[0039] Therefore, according to the present invention, a machine learning model (such as a neural network) trained at the RSU is based on the sidelink traffic conditions (e.g., RRIs for each traffic category of traffic, N transmission , L subchannel , and Pr) and the sidelink resource pool reservation monitored at each TX UE (e.g., CBR measurements for a given observation window) are taken as input training data. The machine learning model classifies traffic into different categories and provides as output the estimated relationship of CBR measurements as a function of the number of actively transmitting neighboring nodes for each traffic category. With such information provided by the RSU, each TX UE can accurately estimate the number of actively transmitting UEs in its neighborhood and build a dynamic lookup table suitable for the current traffic characteristics.

[0040] According to a second aspect of the present invention, there is provided a method for configuring a lookup table for distributed congestion control (DCC) in a network, the network comprising: at least two RSUs; a plurality of TX UEs in the vicinity of each RSU, the TX UEs being capable of communicating with the respective RSUs and capable of implementing distributed congestion control for sidelink communication of the TX UEs using the lookup table; and a parameter server (PS) connecting the RSUs. The method comprises: - monitoring the sidelink traffic situation in the vicinity of each RSU; training a machine learning model for each RSU; aggregating machine learning models by a parameter server; -Updating machine learning models; dynamically configuring a lookup table for a target TX UE in the TX UE based on the monitored sidelink traffic conditions and the updated machine learning model; Includes.

[0041] By using such a collaborative learning scheme according to the present invention, the mismatch of traffic mixtures is taken into account in the collaboration procedure to make the distributed collaboration useful and efficient. In other words, the RSUs in the collaborative learning according to the present invention consider different traffic mixture distributions in other RSUs and construct individual local update models to ensure positive transfer during aggregation and parameter server model update.

[0042] Optionally, the step of updating the machine learning model is repeated until a predetermined condition is met. For example, the predetermined condition is: - maximum number of iterations, -Model convergence speed, - maximum execution time, or -Includes model accuracy.

[0043] According to a third aspect of the present invention, there is provided a system for configuring a lookup table for distributed congestion control (DCC) in a network, the system comprising: a network including at least two RSUs; a plurality of TX UEs in the vicinity of each RSU, the TX UEs being capable of communicating with each RSU and capable of implementing the distributed congestion control using the lookup table; and a parameter server (PS) connecting the RSUs; RSUs are - monitoring the sidelink traffic situation in the vicinity of each RSU; configured to train a machine learning model for each RSU; The parameter server is -Aggregate machine learning models, - configured to update machine learning models, RSUs are A system is provided, further configured to dynamically configure a lookup table for the target TX UE in the TX UE based on the monitored sidelink traffic conditions and the updated machine learning model.

[0044] Such a system can dynamically configure lookup tables for distributed congestion control optimization.

[0045] According to a fourth aspect of the present invention, there is provided a distributed congestion control method for a network comprising at least one RSU and a plurality of TX UEs capable of communicating with each other via sidelink communication, the method comprising: A method is provided that includes implementing decentralized congestion control using a lookup table constructed by a method as described above.

[0046] According to a fifth aspect of the present invention there is provided a computer program comprising instructions which, when executed by a computer, cause the computer to perform a method as set out above.

[0047] Thus, the present invention overcomes the shortcomings of state-of-the-art (pre-)configured lookup table based DCC mechanisms by providing a more flexible and adaptive approach that takes into account the congestion status of other users and shifts in traffic distribution, allowing for more effective management of congestion and optimizing network performance over time. [Brief explanation of the drawings]

[0048] Other features and advantages of the present invention will become apparent from the following description, taken in conjunction with the accompanying drawings. [Figure 1] 1 shows an Intelligent Transport System (ITS) implementing the method according to the invention; [Figure 2a] 1 is a flowchart of an exemplary embodiment illustrating machine learning model training for dynamically configuring DCC lookup tables at individual RSUs according to the present invention. [Figure 2b] 1 is a flowchart of an exemplary embodiment according to the present invention illustrating collaborative learning for updating trained machine learning models of different RSUs. [Figure 3] RSU-assisted for cooperative DCC: We show a block diagram of machine learning for non-uniform traffic flow clustering and learning of CBR measurements versus N_neighborhood relationships per traffic flow cluster. [Figure 4] FIG. 10 illustrates a resource pool at a TX UE for evaluation of its CBR measurements versus N_neighbor relationship. [Figure 5] FIG. 1 is a diagram illustrating an example of collaborative learning according to the present invention. [Figure 6] FIG. 10 illustrates local training of N_neighborhood estimation model parameters per cluster for RSU i in federated learning. DETAILED DESCRIPTION OF THE INVENTION

[0049] As discussed above, the present invention proposes a method for dynamically configuring the lookup table used for DCC for each transmitting user equipment (TX UE) in the vicinity (coverage) of an RSU with the help of an RSU that constantly monitors the traffic situation within the deployment area. The optimal dynamic lookup table is obtained by machine learning to parameterize a neural network at each RSU. Multiple RSUs in the system are also connected to a centralized parameter server, which can aggregate the parameterized models of the neural networks trained at each RSU and provide the updated models to each RSU in a collaborative learning manner. In this way, the overall system performance in terms of packet reception rate is improved for all TX UEs that implement the DCC mechanism.

[0050] 1 illustrates an intelligent transportation system (ITS) scenario for implementing a method according to an embodiment of the present invention. The intelligent transportation system (ITS) uses information and communication technologies to improve the safety, efficiency, and overall performance of transportation networks. The intelligent transportation system includes RSUs, TX UEs (e.g., vehicles), and TX UEs (e.g., vehicles) that use various communication modes, such as vehicle-to-vehicle (V2V) communication and vehicle-to-infrastructure (V2I) communication, based on, for example, NR V2X sidelink mode 2.

[0051] In such scenarios, RSUs are fixed infrastructure components that act as communication hubs within the ITS. RSUs are typically installed along roads, at intersections, or in other strategic locations to facilitate vehicle-to-infrastructure (V2I) communication. RSUs collect and process data from various sources, such as traffic sensors, cameras, or connected vehicles, and can use this information to make decisions or issue commands that help manage traffic, optimize signal timing, or respond to accidents.

[0052] TX UEs are mobile devices or vehicles equipped with communication modules that enable them to participate in ITS. These TX UEs can exchange information with each other and with RSUs using vehicle-to-vehicle (V2V) and V2I communications, respectively. TX UEs can share data such as their location, speed, direction, or other relevant information, enabling collaborative decision-making and improving situational awareness for all connected road users. V2V communication allows vehicles to communicate directly with each other, exchange safety-critical information, and enable coordinated actions such as collision avoidance, platooning, or lane change assistance. This direct communication can help prevent accidents, improve traffic flow, and increase overall road safety. V2I communication connects vehicles to the transportation infrastructure, allowing them to receive real-time traffic updates, road condition information, or other relevant data from RSUs. This communication can help vehicles make informed routing decisions, optimize fuel consumption, and contribute to a more efficient transportation system.

[0053] In the present invention, the RSU is adapted to monitor what is happening in its vicinity via a sidelink channel using sidelink communication such as NR V2X sidelink mode 2 (the RSU decodes the first-stage SCI). For example, for RSU-to-vehicle signaling, any type of communication (Uu link, sidelink PC5, Wi-Fi, etc.) can be implemented to configure the lookup table. In addition, all RSUs are connected to a parameter server used for collaborative learning model parameter aggregation and update diffusion. The link between the RSU and the parameter server can be either a wired backhaul connection (optical fiber, copper wire, etc.) or a wireless backhaul connection.

[0054] More specifically, the ITS as shown in FIG. 1 is a system for configuring a lookup table for distributed congestion control (DCC) of a network, the network including at least two RSUs (such as the three RSUs in FIG. 1 ), a plurality of TX UEs (vehicles) in the vicinity of each RSU, the TX UEs (vehicles) being capable of communicating with each RSU and using the lookup table to implement distributed congestion control, and a parameter server (PS) connecting the RSUs; RSUs are - monitoring the sidelink traffic situation in the vicinity of each RSU; configured to train a machine learning model for each RSU; The parameter server is -Aggregate machine learning models, - configured to update machine learning models, RSUs are The system is further configured to dynamically configure a lookup table for the target TX UE in the TX UE based on the monitored sidelink traffic conditions and the updated machine learning model.

[0055] Once the lookup table is constructed, various techniques in the art can be used to implement a distributed congestion control (DCC) based on the constructed lookup table.

[0056] Further details of the method for dynamically configuring lookup tables for distributed congestion control (DCC) in accordance with the present invention are discussed below.

[0057] 2a and 2b are flowcharts of an exemplary embodiment according to the present invention illustrating machine learning model training for dynamically configuring DCC lookup tables in two local RSUs and federated learning for updating the trained machine learning model by a parameter server.

[0058] 2a illustrates a lookup table for distributed congestion control (DCC) in a network with an RSU and multiple TX UEs, where the multiple TX UEs can communicate with the RSU, and the lookup table can be used to implement distributed congestion control for sidelink communication of the TX UEs. (S10) Monitoring sidelink traffic conditions in the vicinity of the RSU; (S11) Training a machine learning model; (S12) dynamically configuring a lookup table for a target TX UE in the TX UE based on the monitored sidelink traffic conditions and the trained machine learning model; Includes.

[0059] As an example, the machine learning model is trained to estimate the number of TX UEs performing sidelink transmissions in the neighborhood of the target TX UE, where the neighborhood is an area in which the target TX UE is adapted to communicate with other TX UEs via sidelink communication, and such an area is determined by factors such as communication range, device density, network topology, application requirements, and real-time network conditions.

[0060] Furthermore, the lookup table for the target TX UE is expressed as follows (Equation 1):

number

[0061] 2b further illustrates a method for configuring a lookup table for distributed congestion control (DCC) in a network, the network comprising at least two RSUs, a plurality of TX UEs in the vicinity of each RSU, the TX UEs being capable of communicating with the respective RSUs and capable of implementing distributed congestion control for sidelink communication of the TX UEs using the lookup table, and a parameter server (PS) connecting the RSUs. The method is based on a collaborative learning procedure and includes the following steps: (S20) monitoring the sidelink traffic situation in the vicinity of each RSU; (S21) Training a machine learning model for each RSU; (S22) aggregating the machine learning models; (S23) updating the machine learning model, where if a predetermined iteration condition is met, the next step is continued; if not, returning to step S20 and repeating the above steps S21 to S23, the predetermined condition may include, for example, the maximum number of iterations, the model convergence speed, the maximum execution time, or the model accuracy; (S24) dynamically configuring a lookup table of the target TX UE in the TX UE based on the monitored sidelink traffic situation and the updated machine learning model; Includes.

[0062] Referring now to Figure 3, we consider machine learning, specifically for non-uniform traffic clustering, and the CBR measure vs. N neibor 1 illustrates one embodiment of training a machine learning model according to the present invention for learning types of relationships per traffic cluster.

[0063] In this embodiment, as shown in Fig. 3, the present invention first proposes a machine learning offline training algorithm for RSU-assisted cooperative DCC. For simplicity, we assume that all messages transmitted by TX UEs are in broadcast mode. The RSU acquires data of sidelink resource reservations in the deployment continuously in time. This is achieved by the RSU detecting the sidelink resource pool by decoding the first-stage sidelink control information (SCI) in the physical sidelink control channel (PSCCH). In the first-stage SCI, the RSU can acquire the following information about the traffic associated with the reservation: the RRI of the traffic (if the traffic is periodic and semi-persistent scheduling is applied), the frequency resource such as the number of subchannels in the PSSCH carrying the current (re)transmission of the transport block (TB), and the number of (re)transmissions reserved in the RRI.

[0064] In the first step of machine learning, while constantly monitoring the sidelink traffic reservation in time, the RSU learns the following attributes: RRI, number of (re)transmissions N transmission , and the number of subchannels occupied for TB, L subchannel Traffic flows can be first classified into different clusters according to

[0000] . If there is prior information on the proposed total number of traffic clusters, traffic flows with similar characteristics can be clustered for sidelink communication using traditional machine learning algorithms such as k-means clustering or probabilistic clustering methods such as expectation-maximization clustering algorithms. If there is no prior information on the total number of clusters, k=1, 2, 3, ..., K clusters can be assumed, and then the suitability can be evaluated using holdout data or cross-validation methods. When traffic flow clustering is performed, in addition to the traffic clusters, the existence probability of each cluster of traffic across the entire traffic mixture is also calculated.

[0065] Once traffic flow behavior clustering is performed and the probability of existence of each cluster of traffic across the traffic mixture is calculated, system-level simulation

number

[0066] A system-level simulation digital twin is a software implementation of its real-world counterpart (physical twin). The deployment scenario of the RSUs and the geometry of the environment in the system-level simulation are the same as in the real-world implementation. The sidelink communication channel model in the system-level simulation is calibrated to have the same statistical channel behavior as in the real-world implementation. The system-level simulation allows for the correct simulation of the sidelink transmission behavior of multiple simultaneous TX UE transmissions. Estimating the number of neighboring UEs, which is very difficult in a real-world implementation, can also be easily achieved in the system-level simulation.

[0067] The second step of machine learning is to estimate the N sigma-based CBR of any TX UE given its CBR measurements. neibor To find the best function approximation of k , N transmission,k , L subchannel,k , Pr k) tuple to train the respective machine learning sub-models for each traffic cluster. More precisely, based on system-level simulation, for each TX UE, window CBR measurements and true N neibor The measurements can be calculated as follows: sc Subchannel and T window / T slot Assume that the time slots are 10 ms (see FIG. 4). Through system-level simulation, the CBR measurement value CBR k is calculated for the above TX UE k based on the following equation, Eq. k can be simulated.

number

number

number

number

[0068] In real-world implementations, the number of neighbors is a hidden parameter for each TX UE that can only be roughly estimated. Estimation methods include (1) geographic distance-based algorithms, such as counting the number of entries in a network layer location table, or (2) sidelink received signal strength indicator (RSSI)-based algorithms, such as having the TX UE measure the RSSI on each subchannel to identify whether it is occupied and estimate the number of active neighbor TX UEs accordingly. However, because this is signal strength, the TX UE cannot know who the source of this signal is. In addition, this signal can be from multiple transmitters transmitting simultaneously, and the received power is additive, so the TX UE cannot separate the sources. In addition to the above difficulties, there is another fact that makes estimating the number of neighbors difficult: heterogeneous traffic conditions in the system. Due to the fact that sidelink traffic in a 5G NR V2X system can be highly heterogeneous and that DCC is not a mechanism for optimizing the performance of a single user but rather a mechanism for optimizing the performance of a system with many users, neighbor number estimation heavily depends on the sidelink traffic conditions in the vicinity of the target TX UE.

[0069] In a real-world system, it is not possible to calculate the true number of actively transmitting TX UEs as in (Equation 2). This is because if there are multiple TX UEs transmitting to TX UE k in the same resource grid (I,j), TX UE k will receive the received power P m,k and the RSSI as defined in (Equation 1) i,j Therefore, the true CBR measurement vs. N neibor The relationship is that TX UE k is assumed to be a "genie" and that the received power P m,k are hidden functions that can only be evaluated by system-level simulation, which can be simulated separately.

[0070] The digital twin can not only simulate the received sidelink signal strength for each pair of communication devices, but also correctly identify the TX UE source occupying each subchannel, allowing for accurate estimation of the number of neighbors according to CBR measurements and sidelink traffic conditions. Therefore, RSU monitoring and traffic clustering in the physical twin can be used to abstract the traffic conditions with their associated distributions, which can then be switched to the digital twin to generate synthetic data. The synthetic data can be used to train a neighbor number estimation model. Because the system-level simulator preserves the statistical properties of real-world systems, which are important for DCC lookup table design, training this digital twin synthetic database can reduce the complexity of model training and simultaneously improve model accuracy.

[0071] After the second step, we include sub-models of different clusters, e.g., according to which cluster the traffic flow of this TX UE belongs to, we estimate N of the target TX UE based on the similarity of the traffic characteristics of this TX UE with the cluster (sub-model). neibor We derived a trained machine learning model (trained neural network) that can provide the neural network and traffic flow cluster tuples (RRI). k , N transmission,k , L subchannel,k , Pr k ), and the RSU can dynamically configure the lookup table for each TX UE according to the flow cluster it belongs to. An example of a lookup table for a TX UE with traffic flows in cluster k is:

number

number

number

[0072] An embodiment of a system that applies the above offline training-based RSU assistance for cooperative DCC is described below. Suppose an RSU monitors sidelink resource allocation in a highway environment, and traffic flow distributions are quite different during peak and off-peak hours. In addition, at night, there may be more truck platoons, and the sidelink communication traffic flow distribution is different from that during the day. Therefore, it is possible to extract RSU monitoring data for different times of the day and use the above offline trained machine learning model (including submodels for each cluster) to calculate an appropriate DCC lookup table for each flow cluster for each time of the day. The RSU can broadcast this lookup table to TX UEs within its coverage and update the DCC lookup table according to the current time.

[0073] It should be noted that in the above embodiment, clustering is to aggregate similar traffic configurations and accelerate model training (i.e., training of sub-models for each cluster), and can also help understand which traffic characteristic attributes are more relevant, so that better results can be obtained when using them to cluster similar traffic and train a neural network accordingly. However, clustering is not a required step, and it is possible to skip the clustering step and directly train a model based on traffic characteristic attributes, which means that the clustering step is unnecessary.

[0074] The above scheme is already superior to the 3GPP (pre-)configured DCC lookup table, since it takes into account the traffic flow distribution and the evolution of the distribution over time.

[0075] Referring now to FIG. 5, there is shown one embodiment of joint learning-based cooperative DCC among three RSUs according to the present invention.

[0076] As discussed above, after training the machine learning model used to dynamically configure the DCC lookup table for TX UEs within the coverage of one RSU, the present invention further improves the scheme using federated learning-based cooperative DCC with the assistance of multiple RSUs, which is especially useful in scenarios with more than two RSUs. The previous training procedure uses a dataset generated via system-level simulation and simulates the dynamic configuration of the DCC lookup table for a given traffic mixture distribution tuple (RRI). k , N transmission,k , L subchannel,k , Pr k ), we can proceed with the same step 1 traffic flow clustering for each RSU as discussed above, and for another RSU with a different traffic mixture composition,

number

[0077] Such individual training on different RSUs is inefficient. First, if both RSUs observe the same traffic mix, neibor The estimated model should be identical for the two RSUs, thus improving the model accuracy and increasing the number of N per cluster. neiborIt is clear that it is possible to use joint learning to accelerate the convergence speed for training a neural network that performs estimation. Even if the traffic mixture distributions are different at two RSUs, it is still possible to exploit the similarity of the traffic mixture distributions observed at different RSUs and apply RSU-centric joint learning between different RSUs. In other words, the present invention proposes to apply RSU-centric joint learning to make the traffic mixture distributions cooperate while taking into account their discrepancies in traffic mixture distributions, and in this way, potential negative transfer caused by differences in traffic mixture distributions can be avoided.

[0078] Therefore, this invention proposes a federated learning-based RSU-assisted cooperative DCC scheme. The architecture is based on the traditional federated learning architecture shown in Fig. 5, for example, when there are three RSUs participating in federated learning.

[0079] According to an embodiment as shown in FIG. 5, the federated learning may include the following steps. 1. Initialization of model parameters θ and distribution by the aggregation server (parameter server) 2. The local node performs training based on the local dataset: θ i ←Client update (θ,i) 3. The local node updates the model for the aggregation. 4. The AS aggregates the updates and returns the new model to the local node:

number

[0080] Each is a local distribution

number

number

number

number

number

number

[0081] The local training step performed by each RSU is the same as those two steps discussed above. Let the flow mixture distribution obtained at RSUi after the clustering in the first step be denoted as

number

number

[0082] N per cluster of RSUi neibor Local training of estimated model parameters is further illustrated in FIG. 6, where a training approach similar to that of FIG. 3 is implemented.

[0083] Unlike conventional federated learning algorithms where the distributed nodes have a data set drawn from the same distribution, in the present case, the flow mixture distribution

number

number

number

[0084] w i,j An example of this is:

number

number

[0085] w i,j Another example of

number

number

[0086] w i,j Another example of

number

number

[0087] The blending coefficients are calculated by the PS during a special round before the joint training. During this round, the PS calculates

number

number

number

number

number

number

[0088] In the multiple RSU model aggregation and update phase, each RSU updates its local model according to the following set of user-centric aggregation steps at the aggregation server (parameter server).

number

number

number

number

[0089] In consideration of the above, the present invention proposes a method for dynamically configuring the lookup table used for DCC for each TX UE with the help of an RSU that constantly monitors the traffic situation within the deployment area. The optimal dynamic lookup table is obtained by a machine learning algorithm for parameterizing a neural network at each RSU. Furthermore, multiple RSUs in the system are connected to a centralized parameter server, which can aggregate the parameterized models of the neural networks trained at each RSU and provide the updated models to each RSU in a collaborative learning manner. In this way, the overall system performance in terms of packet reception rate is improved for all TX UEs that implement the DCC mechanism.

[0090] More specifically, a local neural network of the machine learning model trained at each RSU takes as input training data the sidelink resource pool reservations monitored at the RSU for each given observation window. The neural network classifies traffic into different categories and provides as output an estimated relationship of CBR measurements as a function of the number of actively transmitting neighboring nodes for each traffic category. With such information provided by the RSU, each TX UE can accurately estimate the number of actively transmitting UEs in its neighborhood and build a dynamic lookup table suited to the current traffic characteristics.

[0091] In addition, the present invention also differs from conventional collaborative learning procedures because the sidelink resource pool monitoring at each RSU may observe different, heterogeneous traffic mixes due to different vehicle densities and different V2X applications at different locations. When the RSUs interact with the parameter server to update their local learning models, such traffic mix inconsistencies are taken into account in the collaborative procedure to make distributed collaboration useful and efficient. In other words, RSUs in collaborative learning according to the present invention consider different traffic mix distributions at other RSUs and construct individual local update models to ensure positive transfer during aggregation and parameter server model updates.

[0092] As will be known to those skilled in the art, the foregoing examples described above can be implemented in many ways in accordance with the present invention, such as program instructions for execution by a processor, software modules, microcode, a computer program product on a computer-readable medium, logic circuits, application specific integrated circuits, firmware, etc. Embodiments of the present invention can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment containing both hardware and software elements. In a preferred embodiment, the present invention is implemented in software, which includes but is not limited to firmware, resident software, microcode, etc.

[0093] Furthermore, embodiments of the present invention may take the form of a computer program product accessible from a computer-usable or computer-readable medium providing program code for use by or in connection with a computer, processing device, or any instruction execution system. For purposes of this description, a computer-usable or computer-readable medium may be any apparatus that can contain, store, communicate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device. The medium may be an electronic, magnetic, optical, or semiconductor system (or apparatus or device). Examples of computer-readable media include, but are not limited to, semiconductor or solid-state memory, magnetic tape, removable computer diskettes, RAM, read-only memory (ROM), rigid magnetic disks, optical disks, and the like. Current examples of optical disks include compact disc-read-only memory (CD-ROM), compact disc-read / write (CD-R / W), and DVD.

[0094] The above-described embodiments are illustrative of the present invention, and various modifications can be made thereto without departing from the scope of the invention which arises from the appended claims.

Claims

1. 1. A method for configuring a lookup table for distributed congestion control (DCC) in a network comprising a roadside unit (RSU) and a plurality of transmitting user equipments (TX UEs), the TX UEs being capable of communicating with the RSUs and using the lookup table to implement distributed congestion control for sidelink communications of the TX UEs, the method comprising: - monitoring sidelink traffic conditions in the vicinity of the RSU; - training a machine learning model; - dynamically configuring the lookup table for a target TX UE in the TX UE based on the monitored sidelink traffic conditions and the trained machine learning model; A method comprising:

2. 2. The method of claim 1 , wherein the sidelink communication is based on NR V2X Sidelink Mode 2.

3. the machine learning model is trained to estimate a number of TX UEs performing sidelink transmissions in a neighborhood of the target TX UE, the neighborhood being an area in which the target TX UE is adapted to communicate with other TX UEs via sidelink communication; The lookup table for the target TX UE is expressed as follows (Equation 1): [Equation 1] The channel resource utilization limit (CR ｌｉｍｉｔ ) is constructed based on Here, CBR ｍｅａｓｕｒｅｄ is the measured sidelink channel busy rate of the target TX UE, and N ｎｅｉｂｏｒ 3. The method of claim 1, wherein Λ is the number of TX UEs performing sidelink transmission in the vicinity of the target TX UE.

4. The method comprises: - classifying the monitored sidelink traffic conditions into a plurality of traffic clusters, each traffic cluster including a plurality of traffics having similar traffic characteristic attributes; generating composite data for TX UEs performing sidelink communication based on the plurality of traffic clusters; - training the machine learning model using the synthetic data as training data; The method of claim 3 further comprising:

5. The method of claim 4 , wherein the plurality of traffic clusters are classified by K-means clustering or probabilistic clustering methods.

6. The plurality of traffic clusters are classified into the following traffic characteristic attributes: Resource Reservation Interval (RRI), number of (re)transmissions (N ｔｒａｎｓｍｉｓｓｉｏｎ ), and the number of subchannels occupied for the transport block (L ｓｕｂｃｈａｎｎｅｌ 5. The method of claim 4, wherein the particles are classified according to the following:

7. The combined data of the TX UEs performing sidelink communication includes a traffic cluster tuple for each TX UE to which the traffic belongs, and each traffic cluster tuple includes a resource reservation interval (RRI), a number of (re)transmissions (N ｔｒａｎｓｍｉｓｓｉｏｎ信 ), the subchannel occupancy of this TX UE for each reservation (L ｓｕｂｃｈａｎｎｅｌ ), the probability (Pr) of the traffic type to which this TX UE belongs.

8. 5. The method of claim 4, wherein the combined data includes a channel busy ratio (CBR) of the target TX UE and a number of TX UEs performing sidelink transmissions in the vicinity of the target TX UE.

9. The method of claim 4 , wherein the synthetic data is generated by a system-level simulation that is a digital twin of a real-world implementation.

10. 1. A method for configuring a lookup table for distributed congestion control (DCC) in a network, the network including at least two RSUs, a plurality of TX UEs in the vicinity of each RSU, the TX UEs being capable of communicating with each RSU and using the lookup table to implement distributed congestion control for sidelink communications of the TX UEs, and a parameter server (PS) connecting the RSUs, the method comprising: - monitoring the sidelink traffic conditions in the vicinity of each RSU; - training a machine learning model for each RSU; aggregating said machine learning models; - updating the machine learning model; - dynamically configuring a lookup table for a target TX UE in the TX UE based on the monitored sidelink traffic conditions and the updated machine learning model; A method comprising:

11. The method of claim 10 , wherein the step of updating the machine learning model is repeated until a predetermined condition is met.

12. The predetermined condition is - maximum number of iterations, - model convergence speed, - maximum execution time, or The method of claim 11, including model accuracy.

13. 1. A system for configuring a lookup table for distributed congestion control (DCC) of a network, the network including at least two RSUs, a plurality of TX UEs in the vicinity of each RSU, the TX UEs being capable of communicating with each RSU and using the lookup table to implement distributed congestion control, and a parameter server (PS) connecting the RSUs; The RSU is - monitoring the sidelink traffic conditions in the vicinity of each RSU; - configured to train a machine learning model for each RSU; The parameter server: - aggregating said machine learning models; - configured to update the machine learning model; The RSU is and dynamically configuring a lookup table for a target TX UE in the TX UE based on the monitored sidelink traffic conditions and the updated machine learning model.

14. 1. A distributed congestion control method for a network comprising at least one RSU and a plurality of TX UEs capable of communicating with each other via sidelink communication, comprising: - A method comprising using said look-up table configured according to the method of claims 1 to 12 to implement decentralized congestion control.

15. A computer program comprising instructions that, when said program is executed by a computer, cause said computer to perform the method of claim 1.

Images