1
Sub-Band Full Duplex Resource Allocation: A Predictive Deep Reinforcement Learning Approach
arXiv:2605.14339v1 [cs.NI] 14 May 2026
Abhiram D, Aiswarya Rajan, Arin Shemeem, Vipindev Adat Vasudevan, Abdulla P
Abstract—This paper presents a predictive deep learning framework for dynamic sub-band allocation in Sub-Band Full Duplex (SBFD) systems, addressing the challenge of balancing uplink (UL) and downlink (DL) performance under highly dynamic traffic conditions. The key contribution lies in integrating a hybrid Bidirectional Long Short-Term Memory (Bi-LSTM) model for traffic forecasting with a Double Deep Q-Network (DDQN) for realtime resource allocation. Using both predicted traffic and current queue states, the proposed system enables proactive scheduling based on traffic demand. Evaluation results show that the prediction model achieves high accuracy in capturing bursty traffic patterns, while the DDQN agent effectively adapts UL/DL split ratios according to traffic variations. The framework improves spectrum utilization, reduces queue buildup, and avoids inefficient static configurations. The proposed approach demonstrates that combining predictive intelligence with reinforcement learning significantly enhances the efficiency and adaptability of SBFD systems, making it a strong candidate for autonomous resource management in future 6G networks. Index Terms—Sub-Band Full Duplex, Resource allocation, Traffic Forecasting, Predictive deep reinforcement learning
I. I NTRODUCTION 6G networks are evolving to meet the massive data demands of immersive applications such as augmented reality (AR) [1], virtual reality (VR) [2], and smart cities. It builds on 5G by integrating artificial intelligence (AI) and machine learning across all layers of the network. Ultimately, 6G is expected to achieve ultra-high data rates of up to 1 Tbps, extremely low latency below 0.1 milliseconds, two-to-three times the spectrum efficiency, and up to ten times the energy efficiency of its predecessor [3]. To enable these advanced applications, multiple enhancements in the Radio Access Network (RAN) are considered. In this work, we focus on SBFD as a key 6G enabler. SBFD helps lower networking latency compared to legacy 5G Time Division Duplex (TDD) [4], which typically favours DL transmission and limits the UL. Using the SBFD mode, more UL slots can be allocated dynamically without compromising DL performance. The greatest challenge in deploying SBFD is managing the split allocation to optimally balance the trade-off between UL and DL based on traffic demand [5]. Existing literature and early SBFD evaluations demonstrate that configuring specific Abhiram D*, Aiswarya Rajan, Arin Shemeem, and Abdulla P are with the School of Engineering, Cochin University of Science and Technology, Cochin, Kerala, India and Vipindev Adat Vasudevan is with Adat Nextwave Consulting, Kerala, India and Global Young Researchers’ Academy (GYRA), Oregon, USA. (*[email protected]). This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible.
symbols for SBFD while maintaining others as standard UL or DL can effectively prioritise high-priority traffic [6]. To mitigate inter-subband leakage, current frameworks rely on frequency splits implemented with guard resource blocks (RBs), inserted between the active sub-bands. Native AI is envisioned as a foundational block of 6G architecture [7] to enable smart, autonomous communication systems; its full potential for dynamic, predictive sub-band allocation in SBFD frameworks remains underexplored. In this paper, we address this gap by proposing a Bi-LSTM network inspired by the broader spatio-temporal traffic modeling principles used in [8], paired with a DDQN [9], a reinforcement learning (RL) agent, to dynamically allocate the UL subband size based on predicted traffic demands. By predicting traffic bursts before they occur, our model ensures that the required UL resources are provided with minimal latency, while preventing unnecessary degradation of DL performance. Our main contributions are as follows: 1. Predictive AI-Driven Split Allocation: AI is used to determine the exact split that should be provided for the UL traffic inside a DL sub-band, ensuring that required UL resources are provided without degrading DL performance. 2. Hybrid Bi-LSTM Forecasting: To forecast future network states, we propose a hybrid 1D Convolutional Neural Network (1D-CNN) and Bi-LSTM architecture. This enables multistep time series forecasting to faithfully capture the time-varying intensity of network demand and predict rapid load fluctuations. 3. DDQN-Based Dynamic Scheduling: The proposed work addresses dynamic UL scheduling in an SBFD system using an RL framework combined with traffic prediction. The problem is modeled as a Markov Decision Process (MDP) [10], where the DDQN agent makes scheduling decisions on a slot-by-slot basis. II. S UB - BAND F ULL D UPLEX The key objective of SBFD is to enhance UL coverage and reduce UL latency, particularly for power-limited User Equipment (UEs). SBFD enables the simultaneous transmission of DL and reception of UL on the same TDD carrier, within a single configured DL and UL Bandwidth Part (BWP) pair with aligned center frequencies [4]. In this architecture, full duplexing mode is deployed exclusively at the next-generation NodeB (gNB), allowing it to receive UL signals within a transmitting DL sub-band. Meanwhile, the UEs continue to operate in half-duplex mode under the control of the Medium Access Control (MAC) layer scheduler, allowing SBFD-aware and legacy UEs to coexist seamlessly within the same cell. The
2
We model the highly bursty and unpredictable behaviour of real-world 6G traffic using a Markov-Modulated Poisson Process (MMPP) [13]. While this accurately simulates the traffic environment, the scheduling limitations identified in Section II motivate the need for a predictive approach. The Bi-LSTM serves as a forecast mechanism to overcome this limitation, providing the RL agent with advanced knowledge of impending demand shifts. The proposed architecture begins with a time-series prediction module, wherein the 1D-CNN and Bi-LSTM network ingests a sliding window of past traffic intensities and outputs a multi-step forecast of UL and DL demands. This predicted traffic vector is then formulated as the environmental state for the RL agent. For determining the optimal UL/DL resource split ratio, a DDQN agent is used. The DDQN decouples action selection from target Q-value evaluation, ensuring stable convergence under non-stationary traffic conditions. Fig. 1. Slot split in TDD and SBFD
SBFD system partitions resources for DL and UL in both the A. Bi-LSTM time domain and frequency domain, providing flexibility for Our prediction architecture processes the traffic data in adaptation to different types of traffic. In the SBFD pattern, the two specialized stages. First, the 1D-CNN layer utilizes a DL frequency is split, separated by a guard band, and UL is sliding filter with a kernel size of 3 across the time-series sandwiched between two DL resource blocks, such as “DUD”, data. This layer extracts local spatial structures, filtering out within a slot. Based on the latest 5G specification [6], almost noise and detecting sudden, sharp micro-bursts to create a all the bands above 3 GHz are designated as TDD bands for condensed feature map. This map is subsequently fed into the new radio (NR). Fig. 1 demonstrates how the split occurs in an Bi-LSTM layer. Unlike a standard Long Short-Term Memory SBFD pattern in comparison with the TDD slot configuration (LSTM) that processes data in one direction, our Bi-LSTM [11]. evaluates the sequence in both forward and backward directions In every TDD symbol, a frequency split is implemented to simultaneously. This dual-processing maximizes the contextual facilitate SBFD. A single SBFD symbol in a TDD carrier is understanding of the data flow, enabling the model to learn split in the frequency domain into one UL sub-band and one or long-term temporal dependencies and accurately predict abrupt two DL sub-bands [11]. To prevent the leakage of a strong DL load transitions. signal into the UL slot, empty guard RBs are inserted between To execute the predictions, we employ a multi-step time the sub-bands. The base station signals the number of guard series forecasting approach. For simulating continuous realRBs and the duration of the UL sub-band via Radio Resource world network operations, we apply a “Stitched Continuous Control (RRC). Forecasting” method, iteratively advancing 10 slots at a time While the SBFD approach significantly increases UL ca- to generate an unbroken forecast trajectory. pacity and spectrum efficiency, it introduces critical trade-offs The Bi-LSTM framework does not perform the bandwidth and challenges [5]. Because UL and DL share the same fixed allocation itself. It concludes the forecasting stage by providing bandwidth, widening the UL sub-band inherently narrows the the predictive intelligence required for dynamic scheduling. DL sub-band, which can degrade performance during high This predictive foundation overcomes the limitations of purely DL traffic loads. Deploying full-duplex technology at the reactive schedulers that suffer from increased latency when gNB causes severe interference, notably Self-Interference (SI), data queues unexpectedly accumulate. By anticipating incoming where powerful DL signals leak into the UL sub-band. This traffic bursts in advance, the predictions are fed directly into architecture creates Cross-Link Interference (CLI) from the UE the RL framework, enabling the agent to proactively optimize perspective, primarily Co-channel UE-UE CLI and Adjacent- the SBFD frequency split ratio. channel UE-UE CLI [12]. The challenge in an SBFD framework is carefully trading off DL performance loss with improvements in UL performance. B. RL Approach Legacy systems rely on static or semi-static RRC configurations, The proposed work models dynamic SBFD scheduling as an which cannot adapt to rapid, bursty traffic changes. MDP [10], employing a DDQN agent for slot-by-slot decisionmaking. While standard Deep Q-Network (DQN) effectively III. P ROPOSED S YSTEM M ODEL FOR DYNAMIC S UB -BAND handles complex traffic states, it relies on a single network A LLOCATION IN SBFD SLOTS for both action selection and evaluation, which often causes This section details the two-stage architecture from Section overestimation bias. DDQN resolves this by separating the I: a hybrid 1D-CNN and Bi-LSTM for traffic forecasting, processes: an Online Network analyzes the 22-dimensional followed by a DDQN agent for real-time sub-band allocation. state to select the optimal action, while an independent Target
3
Network evaluates that action, ensuring a stable allocation policy. After validation of the Bi-LSTM module, the dynamic resource allocation problem is solved using the DDQN agent. The environment for this DDQN-based scheduler is characterized by three components: the state representation, the action space, and the reward formulation. State Representation: The state vector has 22 dimensions. The first 20 values are the Bi-LSTM output, i.e., predictions for 10 future UL and 10 future DL traffic slots. The remaining 2 values are the current UL and DL queue sizes, normalized by dividing by 10 times the link capacity. This state design gives the agent both future and current traffic status and allows proactive decisions. Action Space: The action space is defined as a discrete set of UL/DL allocation ratios, allowing the agent to adapt the SBFD transmission pattern in response to predicted traffic and instantaneous congestion. Instead of selecting a specific frame pattern, the agent selects a precise resource split ratio between UL and DL, with the strict rule that the DL should never degrade below a minimum value. Reward and Training: After the agent chooses a split, the system calculates a reward based on the actual traffic demand. It assigns higher rewards for efficient data delivery and applies penalties when queues build up. Through training, the agent learns to divide the available bandwidth between UL and DL based on the current traffic conditions. The scheduler then serves the data according to this split. Any unserved data remains in the queue for the next time slot, which mimics how real base stations handle buffered packets rather than simply discarding them. By combining these two models, the system can both understand the traffic behavior and act in real time. The BiLSTM provides the intelligence to read the traffic, and the DDQN provides the ability to make smart allocations based on that reading.
additional 30% dropout layer. Finally, a fully connected Dense layer outputs the raw forecasted values, which are reorganized by a Reshape layer into a 10 × 2 matrix. The network is trained using the Adam optimization algorithm [14] alongside the Huber loss function, making the system highly resilient to the sudden traffic outliers and extreme bursts typical in 6G wireless communication. The training process feeds the data to the model in batches of 32 samples and evaluates performance using the Mean Absolute Error (MAE) metric. To maximize efficiency, the training is scheduled for 5 epochs and utilizes two automated monitoring systems. A Model Checkpoint automatically saves the network weights whenever the validation loss reaches a new low. An Early Stopping mechanism with a patience of 4 actively monitors the validation loss; if the model’s accuracy fails to improve for four consecutive steps, the training halts early and automatically restores the weights with the best performance. This ensures the final forecasting model is both highly accurate and well generalised for unseen traffic data. Following the traffic prediction, the DDQN agent observes the 22-dimensional state space to dynamically manage the SBFD network resources. Based on this state, the agent selects an action from five discrete bandwidth split ratios. These choices dictate the exact proportion of network capacity allocated to UL versus DL, ranging sequentially from a 40:60 split down to a strictly DL-focused 0:100 split. Once a bandwidth split is chosen, the environment evaluates the decision by calculating a reward based on queue management efficiency. The core reward is derived from the geometric mean of the UL and DL satisfaction rates, which measure the fraction of successfully served data against the total data waiting in the queues. Specific numerical penalties are subtracted from this base reward. The algorithm penalises the agent for wasting spectrum on empty queues, failing to allocate bandwidth to bursting queues that exceed capacity thresholds, and unnecessarily switching bandwidth ratios between consecutive steps. This targeted penalty system trains the agent to handle sudden traffic bursts without causing IV. M ETHODOLOGY AND E VALUATION disruptive network jitter. A. Methodology The internal DDQN architecture consists of a main network We use a dataset that contains 1,000,000 network traffic and a separate target network, both built using two hidden dense records with two features: UL and DL traffic, both measured layers of 64 neurons each. The agent navigates its environment in bits per slot. Before feeding it to the model, both features using an epsilon-greedy exploration strategy, starting with are normalized to the 0-1 range using min-max scaling. The completely random actions and gradually decaying this randomnormalized data are then split into 3 categories: 80% for training ness to rely on its learned optimal policies. During operation, (800,000 samples), 10% for validation (100,000 samples), and every interaction (the initial state, chosen action, resulting 10% for testing (100,000 samples). The data is restructured reward, and the subsequent state) is saved into an experience using a sliding window technique. In this setup, the algorithm replay memory buffer capable of storing up to 100,000 past takes a continuous history of 30 past time slots as its input to events. To continuously improve its decision-making, the simultaneously forecast the traffic demands for the upcoming agent randomly samples mini-batches from this memory. The decoupled selection-evaluation mechanism described in Section 10 time slots. This structured sequence is processed by the hybrid fore- III-B is optimized using the Mean Squared Error loss function, casting architecture. The input first passes through a 1D-CNN ensuring a reliable resource allocation strategy. The flow chart layer equipped with 64 filters and a kernel size of 3. The in Fig. 2 describes the overall algorithm. output is then refined using a Batch Normalization layer and a MaxPooling layer (with a pool size of 2). A 20% dropout layer B. Evaluation Results is applied immediately after for regularization. Next, a 128-unit To validate the proposed models, we used the same MMPP Bi-LSTM layer extracts temporal dependencies, followed by an dataset [13] discussed earlier. After training, the Bi-LSTM
4
Start
UL/DL Traffic Data
CNN-LSTM Predictor (Traffic Forecasting)
State Construction (22 states | 20 Predicted + 2 Queue Traffic)
IF r < ε
Greedy action selection (Random r ∈ [0, 1])
Fig. 3. Bi-LSTM traffic prediction performance for UL and DL over a 10-slot horizon.
else
Exploitation (Choose UL/DL with maximum Queues)
Exploration (Random SBFD action)
Fig. 4. Dynamic Allocation based on UL demand in DDQN Model vs SACDiscrete model in SBFD slots Action Selection UL = {0,10,20,30,40}% DL = {60,70,80,90,100}%
SBFD Resource Allocation
DDQN Training
Transition Memory (State, Action, Reward)
Queue Scheduling
Reward Calculation
DDQN Update
Fig. 5. Dynamic Allocation on DL demand in DDQN model vs SAC-Discrete model in SBFD slots
alongside the agent’s allocation ratios. Fig. 4 shows the UL demand and the corresponding UL allocation by the DDQN agent. During the first 500 slots, UL demand peaks ( 30,000 Fig. 2. Proposed Framework for traffic prediction and dynamic Sub-band bits/slot) and the agent maintains an allocation of 30%. Between allocation. slots 500 and 1380, UL demand drops to near zero; the agent iteratively predicted future slots of UL and DL traffic for a increases UL allocation to 40% to maintain a balanced split. duration of up to 800 slots. The predictions and real traffic From slot 1380 to 1750, a moderate UL surge ( 12,000 bits/slot) almost converged. Fig. 3 visualizes the UL and DL traffic appears, and the agent reduces UL allocation back to 30%. prediction stitched together across time. At time t, the model Fig. 5 shows DL allocation corresponding to DL demand. predicts slots t to t+9, then shifts to t+10 and predicts During slots 0 to 500, DL demand equals 80,000 bits/slot and slots t+10 to t+19, continuing iteratively to produce a single the agent holds DL allocation at 70%, with the remaining 30% unbroken forecast line. reserved for UL. Near slot 1380, UL demand jumps to 12,000 The UL MAE is 969 bits/slot (0.97% of the 100,000 bits/slot bits/slot and DL to 35,000 bits/slot. During the preceding idle capacity), and DL MAE is 849 bits/slot (0.85% of capacity). period, DL was held at 60%; when demand returns, the agent The stitched predictions confirm that the Bi-LSTM maintains immediately raises DL allocation back to 70%. Between slots long-term phase alignment without drifting and recovers from 1400 to 1750, DL demand settles around 35,000 bits/slot and transient errors, keeping the average prediction error below the agent maintains a consistent 70% allocation. 10% over extended periods. Around slot 500, both demands drop to near zero, yet the To evaluate the RL agent, UL and DL demands were plotted agent decreases DL allocation to 60%. This is not a failure; Decay ε (Start New Episode)
5
when no traffic arrives in either direction, any action switch incurs a switching penalty without throughput gain. Since the baseline traffic is roughly a 1:1 ratio, the agent defaults to a balanced split (60/40), minimizing penalties and decreasing wastage of the DL spectrum. This behavior persists through all subsequent dead periods (slots 600 to 1350 and the later idle window). When both demands return near slot 1800, the agent promptly re-adjusts DL to 70%, confirming instant recovery from idle-state behavior. The static baseline, shown in both figures, cannot adapt to demand changes and wastes resources. The DDQN agent avoids rapid switching when both demands are high and holds a stable configuration during idle periods, demonstrating demand-aware behavior suited to bursty, asymmetric SBFD environments. We compared our Bi-LSTM+DDQN scheduler against the SAC-Discrete scheduler proposed in [15]. In [15], the authors use a geometric mean reward of cumulative UL and DL throughput for proportional fairness, selecting between TDD and SBFD frames. We implemented their SAC-Discrete algorithm and trained it on our MMPP dataset. For fair comparison, we restricted the SAC-Discrete agent to the 3 SBFD frames from their work: XXXXX, XXXXU, and DXXXU, and computed the UL/DL capacity of each frame from the resource block specifications in [15]. A key difference is that the SAC-Discrete environment has no queues, and unserved data at each slot is permanently lost. In [15], this limitation does not apply because their evaluation uses a 3GPP-compliant system-level simulator with packet queues, HARQ retransmissions, and buffer management. Over 100,000 slots, the SAC-Discrete agent selected Action 1 (XXXXU: 36.1% UL / 60.9% DL) in 98,424 slots (98.4%), Action 0 (XXXXX: 20.1% UL / 76.2% DL) in 1,576 slots (1.6%), and Action 2 (DXXXU: 32.1% UL / 65.7% DL) zero times. This extreme convergence indicates a fixed policy that fails to dynamically adjust the subband split. In Fig. 5, the SAC-Discrete agent holds constant DL allocation at 60.9%, while our DDQN agent varies between 60% and 70%. During peak DL demand (80,000 bits/slot, slots 0 to 500), the SAC-Discrete’s 60.9% falls short of both the static baseline’s 80% and the DDQN’s 70%. During dead periods, the DDQN adjusts DL to 60%, while the SAC-Discrete remains at 60.9%. The reason the SAC-Discrete agent locks onto a single frame is its geometric mean reward, which is sensitive to the smaller component. Since UL demand averages roughly three times less than DL, the agent identifies UL starvation as the greatest risk. XXXXU offers the highest UL capacity (36.1%), making it the safest choice. Without queues, unserved data is lost each slot, so the agent is never penalized for over-allocating to an idle direction and has no incentive to switch frames during idle periods. Action 2 (DXXXU: 32.1% UL) offers slightly less UL than XXXXU with only marginal DL gain, making it suboptimal under the geometric mean objective. Table I summarizes the architectural and behavioral differences between our proposed model and the SAC-Discrete baseline [15]. The inability of the SAC-Discrete agent to perform dynamic allocation stems from its lack of traffic forecasting and the absence of stateful queues in its environment. Without
TABLE I C OMPARISON OF P ROPOSED B I -LSTM + DDQN AND SAC-D ISCRETE A PPROACHES Feature
Bi-LSTM + DDQN (Proposed)
SAC-Discrete [15]
RL Algorithm
Double Deep Q-Network (DDQN)
Soft Actor-Critic Discrete (SAC-D)
Traffic Predictor
Bi-LSTM (predicts next 10 slots)
None (uses instantaneous demand)
Dataset
MMPP
MMPP
Frame Type
SBFD only
SBFD only ( [15] uses TDD + SBFD)
Action Space
5 (UL: 40%, 30%, 20%, 10%, 0%)
3 (XXXXX: 20.2% UL, XXXXU: 36.1% UL, DXXXU: 32.1% UL)
States
Bi-LSTM predictions (20) + queue sizes (2) = 22 dimensions
Instantaneous UL/DL demand + cumulative throughput
Reward Function
Geometric mean of UL/DL satisfaction − conditional queue & switching penalties
Incremental geometric mean of cumulative UL & DL
Queue
Yes (unserved traffic is buffered)
No (unserved traffic is lost)
Allocation Behavior
Dynamic: UL varies 10% to 40%, DL varies 60% to 90% based on real-time traffic demand
Fixed: converges to a single frame (XXXXU, 36.1% UL/ 60.9% DL) in 98.4% of slots
queue accumulation to penalize poor resource allocation during bursts or idle periods, the SAC-Discrete agent safely defaults to a single, statically optimal frame configuration. In contrast, the queue-based feedback and proactive 10-slot lookahead force our DDQN agent to develop a genuinely dynamic, demand-aware policy. The performance gains of the DDQN agent over both baselines are as follows: On the UL side, the static baseline fixes UL at 20% (20,000 bits/slot). During peak UL demand of 30,000 bits/slot, this leaves a shortfall of 10,000 bits/slot per slot and queues grow continuously. The DDQN agent, using the BiLSTM forecast, raises UL allocation to 30% (30,000 bits/slot), which exactly matches the demand and keeps the queue empty. This is a 100% reduction in UL queue buildup at peak load, and more than 90% overall across all active traffic periods compared to the static baseline. On the DL side, during peak demand of 80,000 bits/slot (slots 0–500), the DDQN allocates 70% of the 100,000 bits/slot link capacity, giving 70,000 bits/slot for DL. The SAC-Discrete stays fixed at 60.9%, giving only 60,900 bits/slot. This is a difference of 9,100 bits/slot, which is a DL throughput gain of about 14.9% for the DDQN over the SAC-Discrete. The static baseline gives 80% to DL but only 20% to UL. Since peak UL demand is 30,000 bits/slot and the static provides only 20,000 bits/slot, peak UL utilization under the static is just 66.7%, and queue buildup persists. The DDQN avoids this by matching its UL allocation to the actual demand, achieving 100% peak UL utilization while also providing 9.1 percentage points more DL capacity than the SAC-Discrete in every active period.
6
V. C ONCLUSION In this paper, we introduced a dynamic sub-band allocation method for SBFD systems using a combined Bi-LSTM and DDQN approach. The Bi-LSTM acts as a predictor that forecasts upcoming UL and DL traffic, providing the DDQN agent with the exact information it needs to adjust resource split ratios dynamically. Legacy static allocation methods and models from existing literature like the SAC-Discrete scheduler [15] waste the available spectrum because they keep the resource split almost fixed, ignoring the actual traffic demand. Instead of depending on these static or semi-static policies, our model constantly adapts based on real-time network conditions. The simulation results confirm that our approach tracks rapid traffic changes much better than the static allocation and the SACDiscrete model, which struggles to adjust to dynamic loads. During peak traffic bursts, the proposed method reduces the buildup of the UL queue by more than 90% compared to the static baseline. The SAC-Discrete scheduler sacrifices approximately 9.1 percentage points of DL capacity by locking its allocation at 60.9% regardless of demand, while our DDQN dynamically raises DL to 70%, delivering 9,100 more bits/slot to DL in every active period. This adaptation yields 100% peak UL spectrum utilization and improves DL throughput by approximately 15% (14.9%) relative to the SAC-Discrete baseline. By maintaining stable split ratios during idle periods, the model also minimizes switching overhead. The system maximizes spectrum efficiency by efficiently allocating total bandwidth based on traffic demand. Future work will focus on testing the model using actual network traffic, with multiple cells and base stations, and deploying reinforcement learning with multiple agents. R EFERENCES [1] S. I. Loutfi, U. Tureli, and I. Shayea, “Augmented reality with mobility awareness in mobile edge computing over 6g network: A survey,” in 2023 10th International Conference on Wireless Networks and Mobile Communications (WINCOM), 2023, pp. 1–6. [2] H. H. J. Mahdi, L. Fouad, F. H. T. Hussain, A. J. Kadhim, M. A. Mohammed, and N. A. Othman, “Augmented and virtual reality services supported by 6g for improving smart cities,” in 2024 IEEE 9th International Conference on Engineering Technologies and Applied Sciences (ICETAS), 2024, pp. 1–6. [3] W. Jiang, B. Han, M. A. Habibi, and H. D. Schotten, “The road towards 6g: A comprehensive survey,” IEEE Open Journal of the Communications Society, vol. 2, pp. 334–366, 2021. [4] X. Wei, J. Li, C. Liang, and R. Liu, “Performance analysis of subband full duplex for 5g-advanced and 6g networks through simulations and field tests,” IEEE Open Journal of the Communications Society, vol. 4, pp. 2572–2585, 2023. [5] T. Chen, S. Garimapati, I. Kadota, T. Dinc, S. L. Garimella, M. Kohli, A. S. Levin, G. Zussman, and H. Krishnaswamy, “Subband full-duplex large-scale deployed network designs and tradeoffs,” Proceedings of the IEEE, vol. 112, no. 8, pp. 1054–1084, 2024. [6] H. Li, C. Sun, S. Wang, T. Cui, X. Wang, Y. Gong, and W. Zhang, “5g subband full duplex: 3gpp standardization progress and performance analysis,” in 2024 IEEE/CIC International Conference on Communications in China (ICCC), 2024, pp. 1–6. [7] M. Uusitalo et al., “6g vision, value, use cases and technologies from european 6g flagship project hexa-x,” IEEE Access, vol. 11, pp. 26 004– 26 020, 2023. [8] S. Siami-Namini, N. Tavakoli, and A. S. Namin, “The performance of lstm and bilstm in forecasting time series,” in 2019 IEEE International Conference on Big Data (Big Data), 2019, pp. 3285–3292. [9] A. Iqbal, M.-L. Tham, and Y. C. Chang, “Double deep q-network-based energy-efficient resource allocation in cloud radio access network,” IEEE Access, vol. 9, pp. 20 440–20 449, 2021.
[10] R. Bellman, “A markovian decision process,” Indiana University Mathematics Journal, vol. 6, no. 4, pp. 679–684, 1957. [11] S. Wu, S. Zhang, Z. Xu, and Z. Pan, “Flexible resource allocation scheme for non-overlapping subband full duplex systems,” in 2023 International Conference on Wireless Communications and Signal Processing (WCSP), 2023, pp. 1067–1072. [12] X. Han, R. Liu, X. Liu, C. Liang, X. Wei, Y. Hao, Z. Zhang, and S. Jin, “Interference mitigation for non-overlapping sub-band full duplex for 5g-advanced wireless networks,” IEEE Access, vol. 11, pp. 1894–1910, 2022. [13] A. H. El Fawal, A. Mansour, and A. Nasser, “Markov-modulated poisson process modeling for machine-to-machine heterogeneous traffic,” Applied Sciences, vol. 14, no. 18, p. 8561, 2024. [14] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” in 3rd International Conference on Learning Representations (ICLR), San Diego, CA, USA, 2015. [15] M. Mokhtari, P. Kela, G. Pocovi, R. Maldonado, and K. I. Pedersen, “Deep reinforcement learning based dynamic sub-band full duplex for 5gadvanced and 6g,” in 2025 IEEE 101st Vehicular Technology Conference (VTC2025-Spring), 2025, pp. 1–6.