Priority-Aware Resource Allocation in NR-V2X Networks Sidelink Using Double Deep Q-Network

The Third Generation Partnership Project (3GPP) standardized NR-V2X resource allocation for V2V communication in Release 16 of the fifth-generation (5G) mobile network specifications. In Mode 2, vehicles autonomously select sidelink transmission resources from configured or preconfigured resource pools. A gNodeB (gNB) can provide resource-pool configuration within network coverage, and vehicles communicate directly through the PC5 interface, including outside network coverage [1]–[3]. Vehicles autonomously sense, select, and reserve transmission resources through sensing-based semi-persistent scheduling (SB-SPS). V2V applications exchange collective perception messages (CPMs), cooperative awareness messages (CAMs), and decentralized environmental notification messages (DENMs). CPMs share information about the surrounding environment obtained from onboard sensors, such as LiDAR, radar, and cameras. CAMs periodically broadcast vehicle information, including position, speed, heading, acceleration, and vehicle dimensions. DENMs are transmitted aperiodically to notify nearby vehicles of unexpected or hazardous road conditions, such as sudden braking, traffic accidents, roadworks, and icy roads [4]–[6]. These three message types have different quality-of-service (QoS) requirements, particularly in terms of latency. In this study, the latency requirements are specified as at least 100 ms for CPMs, 100 ms to 1 s for CAMs, and at most 100 ms for DENMs. When a vehicle needs to transmit a message, it senses sidelink subchannels over the preceding 1 s to identify resources that are in use or reserved by other vehicles. It then uses SB-SPS to select an unoccupied subchannel, determines the RC and RRI according to the amount of data to be transmitted, and reserves the resource. After reservation, the vehicle uses the selected subchannel to communicate with nearby vehicles. In a dense urban environment, multiple vehicles that detect the same accident may attempt to transmit emergency messages simultaneously using SB-SPS. Because their sensing windows overlap, these vehicles may identify and select the same available subchannel, resulting in conflicting resource reservations. Such collisions can cause packet loss and require additional resource sensing, reselection, and reservation for subsequent transmission attempts. Consequently, although vehicles must promptly deliver DENMs to nearby vehicles, the packet delivery ratio (PDR) may decrease and latency may increase, preventing emergency messages from meeting their latency requirements. Previous studies on SB-SPS-based V2V transmission in NR-V2X Mode 2 have considered changes in the amount of data caused by vehicle mobility and variations in the surrounding environment. When the required payload decreases, a vehicle may not use all subchannels previously reserved using its RC and RRI. These unnecessary reservations can increase the channel occupancy ratio (CR) and channel busy ratio (CBR), which are key indicators for distributed congestion control (DCC), and reduce resource utilization. To mitigate this problem, a DDQN-based communication scheme has been proposed to adapt the RC to the surrounding environment and reduce unnecessary sidelink resource reservations when the payload decreases [7]. This paper addresses resource collisions that occur when multiple vehicles in an urban environment detect an accident or another event and simultaneously select and reserve the same sidelink subchannel. We propose a reinforcement learning scheme that incorporates message priority into resource selection. The objective is to reduce collisions, improve PDR, and lower transmission latency in V2V communication. The remainder of this paper is organized as follows. Section II reviews related work. Section III presents the problem formulation and the proposed method. Section IV evaluates the method through simulation. Section V concludes the paper and outlines future research directions.

This section reviews resource allocation methods for NR-V2X V2V communication. The relevant studies include reinforcement learning methods that optimize sidelink resource reservation and allocation, as well as neural network approaches that prioritize emergency message transmissions.

The study in [7] proposes a method for optimizing the RC according to the amount of data when sensing and reserving sidelink resources in NR-V2X V2V Mode 2. In conventional SB-SPS, a vehicle senses subchannels over the sensing window to identify resources that are occupied or reserved. It then selects an unused subchannel and sets the RC and RRI to support the payload and the QoS requirements of the message before reserving the resource and transmitting the data. Vehicle mobility continuously changes the surrounding environment, which can alter both the message type and the amount of data to be transmitted. Conventional SB-SPS establishes reservations using an RC and RRI based on the current transmission requirements. If the payload subsequently decreases, some reserved subchannels may remain unused while other vehicles are unable to reserve them, resulting in wasted resources. To address resource waste caused by decreasing payloads, the prior work applies a DDQN to sidelink reservation. When data arrive for transmission, the reinforcement learning model determines an RC intended to be suitable for the expected changes in the surrounding environment and uses it to reserve subchannels. However, an RC selected from predicted environmental conditions may be difficult to optimize continuously as mobility changes the environment in real time.

The study in [8] proposes a neural network approach to sidelink resource reservation and data transmission in NR-V2X Mode 2. It considers an urban accident scenario in which nearby vehicles detect an event and exchange information through V2V communication. When all detecting vehicles attempt to reserve resources simultaneously using SB-SPS, collisions may occur because multiple vehicles select the same sidelink subchannel. To support emergency communication at an urban intersection, the approach uses a graph neural network (GNN) to determine the transmission order of vehicles. A priority-aware allocation loss (PAAL) accounts for neighboring vehicles within communication range when prioritizing emergency message delivery. The transmission order is determined from vehicle positions and channel conditions. However, the approach does not prioritize transmissions by message type, and sidelink utilization and latency are not explicitly considered.

In NR-V2X sidelink Mode 2, vehicles autonomously reserve and select resources from a sidelink resource pool to communicate with nearby vehicles using SB-SPS. When several vehicles detect an accident at an urban intersection, they may need to sense and reserve sidelink resources at the same time. Simultaneous selection of the same idle subchannel can therefore produce duplicate reservations and resource collisions. Fig. 3 illustrates the V2V transmission scenario following event detection at an intersection. CAMs, CPMs, and DENMs have different transmission requirements, particularly in terms of latency. After detecting an accident or another event, a vehicle must transmit a DENM to nearby vehicles. In the conventional transmission procedure considered in this study, sidelink resources are reserved according to the order in which data require transmission, without explicitly incorporating application-level message-type priority into resource selection. When many vehicles need to deliver DENMs simultaneously, this procedure does not adequately account for message urgency or sidelink utilization.

After multiple vehicles detect an accident at an urban intersection, they use SB-SPS to sense and reserve sidelink resources for emergency message delivery in Mode 2. To mitigate collisions caused by simultaneous selection of the same subchannel, we employ DDQN for resource selection. We first define the priority of each message type. CPMs convey perception information obtained from LiDAR, radar, and cameras; CAMs periodically share vehicle status; and DENMs report unexpected or hazardous road conditions aperiodically. A weight is assigned to each message type according to its relative urgency. The message priority weight is defined as The QoS weights are set as follows: These weights are researcher-defined QoS policy values that reflect relative urgency; they are not fixed message-type weights specified by ETSI or 3GPP. A larger research weight represents greater urgency, whereas a smaller standardized sl-Priority value represents higher priority. The two quantities should therefore not be equated directly. Based on these message weights, the main parameters of the resource selection problem and the DDQN model are defined below.

The state includes message priority, the density of nearby vehicles, the queue state associated with pending data, and future resource reservations inferred from sidelink control information (SCI) obtained through sensing: Here, w_i is the message priority weight, V_i is the density of nearby vehicles, Q_i is the queue state, and F_i represents future reservation information.

Let the set of candidate subchannels available to vehicle v_i be The action is defined as Here, a_i (t) is the index of the subchannel selected by vehicle i, and K_i is the number of candidate subchannels.

The environment is represented by Here, M(t) denotes the set of messages to be transmitted, and R(t) denotes the set of subchannels available at time t.

The reward is determined by transmission success in the environment: The message weight modulates the reward associated with successful transmission.

We evaluate PDR and latency at an urban intersection using DDQN with message-dependent weights under different traffic conditions, including changes in vehicle count and mobility. Table I summarizes the simulation settings. The number of vehicles is varied over 20, 40, 60, 80, and 100. The evaluation compares DQN, Priority-Aware DQN, and the proposed Priority-Aware DDQN in terms of PDR, average delivery latency, and resource collision rate as the vehicle count increases. As shown in Fig. 4, increasing the number of vehicles from 20 to 100 reduces the PDR of DQN from 98% to 80%, that of Priority-Aware DQN from 98% to 84%, and that of the proposed method from 99% to 88%. With 100 vehicles, the proposed method achieves a PDR that is 8 percentage points higher than DQN and 4 percentage points higher than Priority-Aware DQN. The relative PDR improvement over DQN is 10.0%. Fig. 5 shows that the average latency of DQN increases from 17 ms with 20 vehicles to 59 ms with 100 vehicles. The latency of Priority-Aware DQN increases from 16 to 52 ms, whereas that of the proposed method increases from 14 to 44 ms. With 100 vehicles, the proposed method reduces average latency by 15 ms relative to DQN and by 8 ms relative to Priority-Aware DQN. The relative latency reduction compared with Priority-Aware DQN is approximately 15.4%. 2.5% to 15.0%, that of Priority-Aware DQN from 2.2% to 12.5%, and that of the proposed method from 1.8% to 9.5%. With 100 vehicles, the collision rate of the proposed method is 5.5 percentage points lower than DQN and 3.0 percentage points lower than Priority-Aware DQN. The relative reduction compared with Priority-Aware DQN is 24.0%.

This paper addressed duplicate candidate resource selection in NR-V2X Mode 2 when multiple vehicles simultaneously generate emergency messages at an urban intersection. We proposed a Priority-Aware DDQN method that selects candidate resources using message urgency and local resource conditions while retaining the existing sensing and candidate generation procedures. Message weights are incorporated into the learning reward, and online action selection is separated from target value evaluation to learn the resource selection policy. With 100 vehicles, the proposed method improves PDR by 8 percentage points, reduces average latency by approximately 25.4%, and reduces the resource collision rate by approximately 36.7% relative to DQN. Relative to Priority-Aware DQN, the corresponding latency and collision-rate reductions are approximately 15.4% and 24.0%, respectively. These results indicate that the proposed method achieves better performance than the two DQN baselines under the evaluated conditions. Future work will evaluate simultaneous DENM bursts immediately after an accident in a standards-based NR-V2X physical and medium access control (PHY/MAC) environment at an urban intersection. The sidelink configuration will be fixed, and the evaluation will account for vehicle mobility, packet size, retransmissions, and message generation and transmission deadlines. In addition to message-specific PDR and latency distributions, we will measure the probability of delivery within the deadline to directly assess the protection of emergency messages.

This work was supported by the 2026 National University Development Project funded by the Ministry of Education, Republic of Korea. This researchwas supported by 2026 KyungpookNational University BK21 FOURGraduate InnovationProject (International Joint ResearchProject for GraduateStudents).