QoS-Aware Task Scheduling using Reinforcement Learning in Long Rage Wide Area Network IoT Application

Objective: The primary objective of this study is to develop a QoS-aware task scheduling algorithm for LoRaWAN IoT applications using a Reinforcement Learning (RL) approach.

Introduction: LoRaWAN is a widely adopted Low Power Wide Area Network (LPWAN) protocol designed for Internet of Things (IoT) applications due to its long-range communication and low power consumption. However, ensuring QoS in LoRaWAN networks remains challenging due to limited bandwidth, high device density, and dynamic traffic patterns. Existing scheduling algorithms often fail to balance competing QoS requirements effectively. Reinforcement Learning (RL) offers a promising solution by enabling intelligent decision-making through interaction with the network environment.

Case representation: The proposed model employs a Deep Q-Network (DQN) to optimize task scheduling in LoRaWAN networks. The RL agent interacts with a simulated LoRaWAN environment built using NS-3, where it learns to make scheduling decisions based on real-time network states. Key parameters, such as delay, PDR, PER, and throughput, are used as inputs to the reward function to guide the learning process. Performance is evaluated against existing models like RT-LoRa, and LoRa+ under varying node densities and traffic scenarios.

Result: The simulation results demonstrate that the proposed RL-based task scheduling algorithm outperforms existing models across multiple Quality of Service (QoS) metrics. It achieves the lowest delay at approximately 40 ms, significantly outperforming RT-LoRa, which has a delay of around 120 ms, and LoRa+, which experiences a delay of about 80ms. In terms of Packet Delivery Ratio (PDR), the model maintains a competitive value of approximately 85%, comparable to LoRa+ at 87%. Additionally, it records the lowest Packet Error Rate (PER) at around 5%, outperforming RT-LoRa and LoRa+, which exhibit PER values of approximately 15% and 10%, respectively. Furthermore, the model achieves the highest throughput of approximately 250 kbps, surpassing RT-LoRa at 150 kbps and LoRa+ at 200 kbps, demonstrating its superior performance in optimizing network efficiency.

Discussion: The proposed model demonstrates significant strengths in reducing delay and PER while maximizing throughput, making it suitable for time-sensitive IoT applications. However, its marginal improvement in PDR compared to existing models highlights an area for further optimization. Additionally, energy efficiency was not explicitly addressed in this study, which is critical for LPWAN applications like LoRaWAN. These limitations suggest potential directions for future research.

Conclusion: This research successfully develops a QoS-aware task scheduling algorithm using reinforcement learning for LoRaWAN IoT applications. By dynamically adapting to network conditions, the proposed model achieves superior performance across multiple QoS metrics compared to state-of-the-art algorithms. Future work will focus on incorporating energy efficiency into the model and extending its applicability to multi-gateway scenarios.

AI: Artificial Intelligence; CF: Carrier Frequency; Data Extraction Rate; Deep Neural Network; DQN: Deep Q-Network; DRL: DRL: Deep Reinforcement Learning; DSR: Design Science Research; ILP: Integer Linear Programming; IoT: Internet of Things; ISM: Industrial Scientific and Medical; LoRa: Long Range (a physical layer technology); LoRaSim: LoRa Network Simulator; LoRaWAN; Long Range Wide Area Network; LPWAN: Low Power Wide Area Network; MAC: Medium Access Control; MCUs: Microcontrollers; MILP: Mixed Integer Linear Programming; NS-3: Network Simulator-3; PDR: Packet Delivery Ratio; PER: Packet Error Rate; PST: Priority Scheduling Technique; QoS: Quality of Service; ReLU; Rectified Linear Unit; RL: Reinforcement learning; RT-LoRa: Real-Time LoRa; SF: Spreading Factor; SINR: Signal-to-Interference-plus-Noise Ratio; TCP/IP: Transmission Control Protocol/Internet Protocol.

The Internet of Things (IoT) encompasses a vast network of interconnected devices that communicate and exchange data over the Internet, impacting various sectors such as smart cities, healthcare, agriculture, and industry. The rapid expansion of IoT applications has created a pressing need for efficient resource allocation and task scheduling mechanisms to optimize resource utilization while meeting Quality of Service (QoS) requirements [1].

LoRaWAN (Long Range Wide Area Network) is highlighted as a significant enabler for IoT, designed to provide long-range communication with low power consumption. This wireless communication protocol is particularly optimized for IoT devices, allowing them to transmit small amounts of data over considerable distances. LoRaWAN's capabilities make it suitable for applications requiring remote monitoring and data acquisition, thus facilitating the expansion of IoT solutions [1,2]. For example, LoRaWAN which is LPWAN technology can connect battery-powered devices at very long distances while consuming minimum power, hence making it affordable [3].

LoRaWAN operates in the unlicensed ISM bands, which vary according to region [4]. It employs chirp spread spectrum modulation techniques to attain long-distance communication with low power [5]. One of the main advantages of LoRaWAN is its remarkable coverage. For this reason, it can transmit data within several kilometers in open settings such as rural areas or large industrial facilities without the need for cellular towers and other infrastructure items. Consequently, LoRaWAN is best suitable for applications that need a wider coverage area, such as smart agriculture, asset tracking, environmental monitoring, and smart city deployments [1]. Hence, LoRaWAN has become an attractive technology for IoT applications due to its unique combination of long-range capability, low power consumption, and cost-effective deployment [6] and LoRaWAN relies on four key components [7].

Figure 1, depicts the overall architecture of a LoRaWAN network, highlighting its key components and their interactions. The architecture consists of end devices (sensors), gateways, a network server, and application servers, illustrating how data flows from the end devices to the application layer. End devices communicate wirelessly with gateways using LoRa technology, which then forwards the data to the network server, where it is processed and routed to the appropriate application server for further analysis or action, showcasing the hierarchical structure and functionality of the LoRaWAN ecosystem.

Figure 1: LoRaWAN network architecture [7].

In LoRaWAN IoT applications, maintaining Quality of Service (QoS) is crucial due to challenges like limited resources, channel congestion, and varying QoS requirements, which can lead to issues such as high latency and packet loss. Reinforcement Learning (RL) is identified as the most suitable machine learning approach for dynamic task scheduling in LoRaWAN networks, as it can adapt to changing conditions and optimize multiple QoS metrics simultaneously. By leveraging RL, nodes can self-optimize scheduling performance, enhancing reliability and efficiency in diverse applications such as smart agriculture, industrial IoT, and smart city management [8,9,6].

The study proposes the use of Reinforcement learning (RL) techniques to develop a scheduling algorithm that can adapt to dynamic network conditions, optimize energy consumption, and enhance overall system performance. By leveraging RL, the proposed solution aims to improve latency, reliability, and efficiency in LoRaWAN networks, ultimately contributing to the sustainability and scalability of IoT deployments [9,10].

Motivation of the study

Reinforcement Learning (RL) is uniquely suited for dynamic task scheduling in LoRaWAN due to its ability to learn optimal policies through interaction with the environment without requiring labeled data. Unlike supervised learning, which relies on historical datasets and struggles with real-time adaptability, RL agents make decisions based on feedback (rewards) from the current network state, allowing them to respond to changes in node density, channel interference, and traffic loads dynamically. Moreover, traditional optimization or classification-based ML approaches often lack the capacity to handle sequential decision-making over time, which is essential in scheduling tasks where present actions impact future network states. RL's strength lies in continuously adapting its policy to maximize long-term performance across multiple QoS metrics, making it a natural fit for the resource-constrained, time-sensitive, and stochastic nature of LoRaWAN-based IoT systems.

Contributions of the article

Unlike prior scheduling approaches that rely on static heuristics, clustering, or centralized optimization models, the novelty of the proposed method lies in its integration of a Deep Q-Network (DQN) within a dynamic and adaptive task scheduling framework for LoRaWAN. Specifically, the model introduces a context-aware reinforcement learning agent that continuously observes real-time network states; such as channel conditions, node congestion, task urgency, and signal quality and learns an optimal scheduling policy through interaction with the environment. This real-time adaptability enables the system to prioritize tasks and allocate channels or gateways effectively, even under fluctuating traffic loads and node densities. Furthermore, the design of a multi-metric reward function that simultaneously optimizes delay, PDR, PER, and throughput distinguishes this work from most existing solutions that optimize only one or two QoS metrics. To the best of our knowledge, this is among the first studies to deploy DQN-based scheduling in a simulated LoRaWAN environment using NS-3, validated across multiple network configurations and performance benchmarks.

The remaining parts of this Section are arranged as follows: Section 2, presents a comprehensive review of the existing literature on LoRaWAN technology, scheduling algorithms, and reinforcement learning in IoT. It provides the theoretical foundation for the research and identifies the gaps that this thesis aims to address. Section3 covers the details of the research methodology employed in this study, including the research design, development tools, algorithm design, implementation, complexity analysis, and reward function design. the analysis of the simulation results. It describes the simulation setup and scenarios and discusses the performance of the proposed algorithm in terms of key QoS metrics, will be covered in Section IV. Lastly, the conclusion covered in Section 5.

LPWANs like LoRaWAN have revolutionized the IoT by enabling long-range communication with battery-powered devices. However, IoT applications within the IoT domain demand reliable and expedited data delivery, posing challenges for LoRaWAN due to inherent limitations in range, latency, and energy constraints [11]. This review explores existing research related to Task Scheduling in LoRaWAN.

The paper [12] proposes a dynamic transmission Priority Scheduling Technique (PST) based on an unsupervised learning clustering algorithm for dense LoRaWAN networks. The LoRa gateway classifies nodes into different priority clusters, and the dynamic PST allows the gateway to configure transmission intervals based on cluster priorities. This approach aims to improve transmission delay and decrease energy consumption. Simulation results suggest that the proposed work outperforms conventional LoRaWAN and recent clustering and scheduling schemes, making it potentially well-suited for dense LoRaWAN deployments.

In [13] a Real-Time LoRa (RT-LoRa) communication protocol for industrial Internet of Things applications is introduced. The real-time flow is processed by the RT-LoRa using a medium access strategy. Static and movable nodes are used to build the entire network. The QoS level is regarded as being the same for every static node. Three classes-normal, dependable, and most reliable-are used to categorize the QoS level for flows produced by mobile nodes. The technique distributes SF and CF based on the QoS level. A star topology is used to arrange and connect the mobile and static nodes to the gateway. The following are the important points raised in this paper: For a single gateway network using single-hop communication, the general process is described. Even within 180 meters, this results in a significant transmission delay of up to 28 seconds for the majority of dependable flows. This study has not addressed the need for greater coverage and reduced time delay for industrial data in real time. There are limitations in QoS provisioning because the QoS level is only assigned to mobile nodes and all static node flows are given the same priority level. All nodes need a lot of energy to connect with the central gateway, and nodes farther from the gateway use even more energy.

The paper [14] proposes a method to optimize the performance of LoRaWAN networks by dynamically assigning values for the Spreading Factor and Carrier Frequency radio parameters. This assignment is formulated as a Mixed Integer Linear Programming problem to maximize network metrics like Data Extraction Rate and minimize packet collisions. An approximation algorithm is also developed to solve the problem more efficiently at scale. The results show improved performance for metrics like DER and an average 6-13% fewer packet collisions compared to baseline policies. The performance evaluation of the proposed optimization algorithms is done through simulation using the LoRaSim simulator. The optimization focuses on optimizing just the SF and CF parameters of the LoRa radio configuration. Considering additional parameters could lead to even better performance. The simulations assume a single gateway setup. Therefore, in summary, the key limitations are limited configuration parameters, static network assumptions and evaluation based on few metrics.

In [15], the authors explore the viability of real-time communication within LoRaWAN-based IoT systems. Leveraging an Integer Linear Programming (ILP) model, they assess the feasibility of real-time communication during the network design stage. This model not only determines feasibility but also optimizes the number and placement of gateways necessary to achieve real-time requirements. The paper further validates the model's performance through various scenarios, offering valuable insights into LoRaWAN's scalability and real-time support limitations. However, it is important to note that the model primarily focuses on static network design at deployment. This may not fully capture the dynamic nature of real-world networks, where factors like interference, congestion, and gateway availability can significantly affect real-time QoS performance.

In [16], the authors present a low-overhead synchronization and scheduling concept implemented on top of LoRaWAN Class A. They design and deploy an end-to-end architecture on STM32L0 Microcontrollers (MCUs), where a central entity provides synchronization metrics and allocates transmission slots. By measuring clock drift in devices, the system defines slot lengths within the network. This approach achieves 10-millisecond accuracy and demonstrates significant improvements in packet delivery ratios compared to Aloha-based setups, especially under high network loads. Notably, the paper addresses the gap in the literature regarding experimental approaches to LoRaWAN scheduling and demonstrates the feasibility of the proposed concept. However, the paper does not delve into the energy consumption impact of the implemented scheduling algorithms.

Despite various efforts to improve task scheduling in LoRaWAN, existing methods exhibit notable limitations. Many approaches, such as clustering-based scheduling and MILP-based resource optimization, operate under static assumptions or rely on centralized architectures that limit scalability and adaptability in dynamic environments. Others fail to prioritize tasks effectively or consider a limited subset of QoS parameters. Additionally, most traditional algorithms lack the ability to learn and adapt in real time to fluctuating network conditions, leading to suboptimal performance under high traffic or dense node scenarios. The proposed RL-based scheduling algorithm addresses these gaps by employing a Deep Q-Network (DQN) that continuously learns optimal scheduling policies through interaction with the network environment. Unlike static or rule-based strategies, the RL agent dynamically adapts to changes in node density, traffic load, and channel conditions, optimizing multiple QoS metrics—including delay, Packet Delivery Ratio (PDR), Packet Error Rate (PER), and throughput—simultaneously. This adaptive capability enables the model to outperform existing approaches in both efficiency and scalability, making it more suitable for real-time, large-scale IoT applications (Table 1).

Proposed method

The research methodology focuses on designing and implementing a Reinforcement Learning (RL)--based scheduling algorithm for reliable data delivery in LoRaWAN networks. It adopts a Design Science Research (DSR) approach, which emphasizes systematic development and evaluation of practical solutions to address inefficiencies in existing task scheduling mechanisms. The methodology begins with a detailed description of the research design, which emphasizes the need for a task-scheduling algorithm that can effectively manage resources in dynamic environments. The study identifies the limitations of existing scheduling methods in LoRaWAN networks, particularly their inability to meet the QoS demands of modern IoT applications. To address these challenges, the research proposes a Reinforcement Learning (RL) based algorithm that can adapt to varying network conditions and optimize resource allocation.

Research design

The research employs a mixed-methods approach, combining quantitative research with design science to systematically design, develop, and assess a QoS-aware task-scheduling algorithm. This approach allows for addressing questions related to the effectiveness of the proposed algorithm in improving QoS in dynamic IoT environments.

Algorithm design and implementation

Algorithm design: The design of the RL-based scheduling algorithm focuses on creating an intelligent agent that optimizes task scheduling in a LoRaWAN environment. Key components include defining the state space, action space, and reward function, which guide the agent's learning process to make optimal scheduling decisions based on network conditions.

• State space: The state space encompasses various network parameters, such as node status, channel conditions, and traffic patterns, allowing the agent to assess the current environment effectively.
• Action space: The action space includes possible scheduling actions, such as channel selection, task prioritization, and gateway allocation, enabling the agent to make informed decisions to enhance QoS metrics.
• Reward function: The reward function is designed to provide feedback to the agent based on its actions, encouraging behaviors that lead to improved QoS outcomes, such as reduced delay, increased packet delivery ratio, and minimized packet error rates.
• Policy (π): The policy defines the strategy the agent uses to select actions based on the observed state, enabling it to balance exploration and exploitation during learning.
• Learning algorithm: A suitable reinforcement learning algorithm, such as Deep Q-Networks (DQN), is employed to enable the agent to learn from its experiences and improve its scheduling decisions over time.

Figure 2, illustrates the architecture of a Deep Q-network (DQN), which combines Q-learning with deep neural networks to enable reinforcement learning in complex environments. The architecture typically consists of the following key components:

Figure 2: DQN architecture [17].

• Input layer: This layer receives the state representation of the environment, which can include various features relevant to the task at hand. The input is often a high-dimensional vector that captures the current state of the system.
• Hidden layers: The DQN architecture includes multiple fully connected hidden layers (in this case, two layers) that process the input data. Each hidden layer consists of a specified number of neurons (e.g., 128), which are responsible for extracting features and learning non-linear relationships between the input state and potential actions. ReLU (Rectified Linear Unit) activation functions are commonly used to introduce non-linearity.Output layer: The output layer generates Q-values for each possible action based on the processed input state. These Q-values represent the expected future rewards for taking specific actions in the given state, allowing the agent to make informed decisions about which action to take.
• Experience replay: Although not explicitly shown in the architecture diagram, experience replay is an integral part of the DQN framework. It involves storing past experiences (state, action, reward, next state) in a replay memory, which is sampled during training to improve learning stability and efficiency.

The diagram shows the agent taking action in the environment, receiving a new state and reward, and updating its policy based on the experience. This iterative process allows the agent to learn an optimal policy for maximizing rewards in the environment.

Here's a breakdown of the diagram's elements:

• Agent: This is the decision-making entity. It receives the current state of the environment (s) and uses its policy (π) to select an action (a). The policy is typically implemented as a neural network (DNN) with parameters θ.
• Environment: This is the external world the agent interacts with. It receives the agent's action (a) and provides the agent with a new state (s') and a reward (r).
• State (s): The current situation or observation of the environment.
• Action (a): The decision or move made by the agent.
• Reward (r): A scalar value indicating the outcome of the agent's action. Positive rewards encourage behaviors, while negative rewards discourage them.
• Policy (π): A function that maps states to actions. In DRL, it's often represented as a neural network.

In the context of LoRaWAN, the input layer of the DQN receives a state vector that encapsulates key network parameters representing the current environment status. These include normalized values of channel conditions (e.g., signal-to-interference-plus-noise ratio), gateway congestion levels, task queue lengths, packet retransmission counts, and node-specific information such as remaining transmission energy and task deadlines. By encoding these factors into the input layer, the DQN can interpret the real-time status of the LoRaWAN network and make informed scheduling decisions that adapt to dynamic conditions. This mapping ensures that the agent's learning process is grounded in practical, context-specific observations relevant to QoS optimization in LoRaWAN-based IoT systems.

Training phase of the proposed scheduling algorithm

Figure 3, outlines the training phase of the proposed scheduling algorithm, which utilizes a Deep Q-Network (DQN) approach to optimize task scheduling in a LoRaWAN environment. The training phase consists of several key steps:

Figure 3: Training phase of the proposed scheduling algorithm.

• Initialization of DQN parameters: The training process begins with the initialization of essential DQN parameters, including the learning rate, which determines how much the Q-values are updated during training; epsilon, which controls the exploration-exploitation trade-off; and the experience replay buffer, which stores past experiences to enhance training stability.
• Observation of current state: The agent interacts with the OpenAI Gym environment to observe the current state of the network. This state includes various parameters such as network conditions, task queue status, and other relevant metrics that influence scheduling decisions.
• Action selection and execution: Based on the observed state, the agent selects an action using an epsilon-greedy policy, balancing exploration of new actions and exploitation of known rewarding actions. The selected action is then executed within the environment.
• Reward calculation: After executing the action, the agent receives feedback in the form of a reward, which quantifies the effectiveness of the action taken in terms of QoS metrics such as delay, throughput, and packet delivery ratio.
• Experience storage and learning: The agent stores the experience (state, action, reward, next state) in the replay buffer. A mini-batch of experiences is sampled from this buffer to update the Q-values, allowing the agent to learn from past actions and improve its scheduling policy over time.
• Iteration and convergence: The training process continues iteratively, with the agent observing new states, selecting actions, and updating Q-values until a predefined maximum number of training iterations is reached or the performance converges to an acceptable level.

Figure 4, presents a diagram of the trained proposed scheduling algorithm, illustrating the workflow and key components involved in the task scheduling process within a LoRaWAN environment. The diagram outlines the following steps:

Figure 4: The trained proposed scheduling algorithm diagram.

• Receive task request: The process begins with the system receiving a new task-scheduling request, which includes critical parameters such as deadlines and network context. This initiates the scheduling cycle.
• Retrieve network state: The algorithm retrieves the current network state, which encompasses various factors like Signal-to-Interference-plus-Noise Ratio (SINR), existing task queue, and other relevant network conditions that influence scheduling decisions.Generate schedule: Utilizing the learned policy from the training phase, the reinforcement learning (RL) agent generates a schedule by assigning tasks to specific gateways. This assignment is optimized based on Quality of Service (QoS) metrics and the deadlines specified in the task request.
• Evaluate schedule feasibility: The generated schedule is assessed for feasibility, ensuring that it meets all required constraints and QoS criteria. This step is crucial to confirm that tasks can be completed within their deadlines and adhere to the necessary QoS standards.
• Feasibility check: If the schedule is deemed feasible, it is sent to the relevant gateways for execution. If not, the algorithm enters an adjustment phase to refine the schedule.
• Adjust schedule with RL agent: In cases where the initial schedule is infeasible, the RL agent recalibrates the task assignments to meet the QoS requirements, iteratively adjusting the schedule until it becomes feasible or the maximum number of attempts is reached.
• Re-evaluate schedule feasibility: The adjusted schedule undergoes another feasibility evaluation to ensure compliance with the required constraints.
• Final outcome: If a feasible schedule is produced, it is transmitted to the gateways for execution. If a feasible schedule cannot be achieved within the maximum attempts, a failure report is generated, indicating that the task scheduling request could not be fulfilled.

Overall figure 4, effectively illustrates the structured workflow of the trained scheduling algorithm, highlighting the interaction between task requests, network state retrieval, schedule generation, feasibility evaluation, and adjustments made by the RL agent to optimize task scheduling in a LoRaWAN network.

Algorithm implementation

The implementation of the RL-based scheduling algorithm involves translating the designed components into a functional system that operates within the simulated LoRaWAN environment. This process includes several key steps:

• Initialization: The algorithm initializes the RL agent, setting up the state space, action space, and reward structure, along with any necessary parameters for the learning process.
• Training Phase: The agent interacts with the environment through a reinforcement learning loop, where it observes the current state, selects actions based on its policy, receives rewards, and updates its knowledge (Q-values) to improve future decision-making.
• Integration with simulation: The algorithm is integrated with the network simulator (NS-3), allowing for real-time interaction with the simulated LoRaWAN network. This integration enables the agent to adapt its scheduling decisions based on dynamic network conditions and traffic patterns.Evaluation: The performance of the implemented algorithm is evaluated using various QoS metrics, such as delay, packet delivery ratio, and packet error rate, to assess its effectiveness in optimizing task scheduling in the LoRaWAN environment.

Overall, the implementation phase focuses on creating a working model of the algorithm that can learn and adapt to improve network performance in real-time scenarios.

Pseudocode for task scheduling algorithm

The LoRaWAN network improved task scheduling algorithm focuses on channel selection, task priority, and adaptive gateway placement in order to achieve better QoS parameters. The RL agent interacts with the LoRaWAN environment, observes network states, selects actions based on policy, receives rewards, and updates its knowledge to optimize QoS metrics like delay, reliability, throughput, and energy efficiency.

Pseudocode structure

1. Initialization
2. State Observation
3. Action Selection
4. Environment Interaction (OpenAI Gym Integration)
5. Reward Calculation
6. Q-Value Update (Learning)
7. Training Loop
8. Policy Improvement and Execution

Algorithm 1 Initialization

1. Initialize Q-network with random weights
2. Initialize target Q-network with the same weights as Q-network
3. Initialize Replay Memory D with capacity N
4. Set ϵ for ϵ-greedy policy
5. Set learning rate α, discount factor γ, and batch size
6. Define action space A = {channel selection, task prioritization, gateway allocation}
7. Define state space S = {channel status, signal strength, gateway congestion, task deadlines}
8. Define reward function R (s, a) based on QoS metrics
9. Periodically synchronize target Q-network with Q-network weights every K episodes

Algorithm 2 State Observation

1. Function Observe State()
2. Initialize state as an empty list
3. Normalize current channel status, signal strength (SINR), gateway congestion, and task deadlines
4. Append normalized values to state
5. return state

Algorithm 3 Action Selection using ϵ-Greedy Policy

• Function SelectAction(state, ϵ)
• Generate a random number and Î [0, 1]
• if rand < ϵ then
• Choose a random action from action space A
• else
• Compute Q-values for all actions using Q-network
• Choose action argmax(Q-values) // Select the action with the highest Q-value
• end if
• return action

Algorithm 4 Environment Interaction

1. Function PerformAction(action)
2. Initialize the OpenAI Gym environment
3. if action == "channel selection" then
4. Select the channel with the lowest interference and load
5. else if action == "task prioritization" then
6. Prioritize tasks based on deadlines
7. else if action == "gateway allocation" then
8. Assign tasks to gateways with optimal load balancing and signal quality
9. end if
10. Execute the selected action in the LoRaWAN environment via OpenAI Gym
11. Observe the resulting state, reward, and whether the episode is done using GetEnvironmentFeedback() from Gym environment
12. return new state, reward, done

Algorithm 5 Reward Calculation

1. Function CalculateReward(state, action)
2. Initialize reward = 0
3. if QoS metrics are improved then
4. reward += k // Positive reward for improved QoS metrics
5. else
6. reward -= k // Negative reward for decreased QoS metrics
7. end if
8. return reward

Algorithm 6 Q-Value Update (Learning)

1. Function UpdateQNetwork()
2. Sample a random minibatch of transitions (state, action, reward, next state) from Replay Memory D
3. for each transition in the minibatch do
4. target = reward
5. if not done then
6. target += γ × max (target Q-network.predict(next state))
7. end if
8. Compute loss as Mean Squared Error (MSE) between target and Q- network.predict(state, action)
9. Perform gradient descent step to minimize loss
10. end for
11. Periodically synchronize target Q-network with Q-network weights

Algorithm 7 Training Loop

1. for episode in range(total_episodes) do
2. state = ObserveState()
3. done = False
4. while not done do
5. action = SelectAction(state, ϵ)
6. new_state, reward, done = PerformAction(action)
7. Store transition (state, action, reward, new_state, done) in Replay Memory D
8. if len(Replay Memory) > batch size then
9. UpdateQNetwork()
10. end if
11. state = new_state
12. end while
13. if ϵ > ϵ_min then
14. ϵ *= epsilon_decay // Decay exploration rate
15. end if
16. if episode % evaluation_interval == 0 then
17. EvaluatePolicyPerformance()
18. end if
19. end for

Algorithm 8 Policy Improvement and Execution

1. Function EvaluatePolicyPerformance()
2. Initialize performance metrics
3. for test episode in range (test episodes) do
4. state = ObserveState()
5. done = False
6. while not done do
7. action = SelectAction(state, ϵ = 0) // Greedy action selection during evaluation
8. new state, reward, done = PerformAction(action)
9. Update performance metrics based on reward and QoS metrics
10. state = new state
11. end while
12. end for
13. Return metrics

Algorithm complexity analysis

The algorithm complexity analysis encompasses three main aspects: time complexity, space complexity, and scalability and feasibility.

• Time Complexity: The training time complexity of the RL-based scheduling algorithm is O (T × (|S| × |A| + L × N² + B log E)), where T is the number of training episodes, |S| is the number of states, |A| is the number of actions, L is the number of layers, N is the number of neurons per layer, B is the mini-batch size, and E is the total experiences stored. This complexity arises from exploring the state-action space, performing neural network computations, and sampling experiences.
• Space Complexity: The space complexity is defined as O (L × N² + E × M), where L × N² accounts for the neural network parameters and E × M represents the memory required for the replay buffer, with M being the memory space per experience tuple. This indicates the memory requirements for both the neural network and the experience replay mechanism.
• Scalability and Feasibility: The DQN-based algorithm is computationally intensive during the training phase due to the complexity of state-action exploration and neural network computations. However, once trained, the decision-making phase is efficient, requiring only a single forward pass through the neural network, making it suitable for real-time scheduling tasks in LoRaWAN networks and enabling scalability to handle large numbers of devices.

Overall, the analysis highlights the algorithm's computational demands and its potential for effective deployment in resource-constrained environments.

Reward function design

The reward function design is a critical component of the proposed scheduling algorithm, as it directly influences the learning process of the Reinforcement Learning (RL) agent and the quality of scheduling decisions. The reward function is structured as a weighted sum of various Quality of Service (QoS) metrics, including delay minimization, reliability maximization, and throughput optimization.

• Balancing Trade-offs: The reward function aims to balance trade-offs among different QoS metrics, ensuring that the RL agent can make informed scheduling decisions that optimize overall network performance while adhering to specific constraints.
• Implementation in Learning: The reward function is integrated into the learning process, guiding the agent's actions based on the observed outcomes and facilitating the continuous improvement of the scheduling policy through experience replay and Q-value updates.

Overall, the reward function design is pivotal in shaping the agent's behavior, promoting effective scheduling strategies that meet the dynamic demands of LoRaWAN networks.

Simulation setup and scenarios

The simulation setup and scenarios section outline the environment and parameters used to evaluate the proposed scheduling algorithm in a LoRaWAN context.

• Simulation Environment: The simulations were conducted using the NS-3 simulator, specifically utilizing the ns-3-lora module to accurately emulate LoRaWAN network characteristics. This environment allows for realistic simulations of long-range, low-power communication in an unlicensed spectrum, with a defined area of 200m x 200m and a maximum distance of 200m to the gateway.
• Parameters: Key simulation parameters include three gateways, 100 IoT devices, one network server, a LoRa Log Normal Shadowing propagation model, a frequency band of 868MHz, and a maximum of five retransmissions. These parameters were selected to create a medium-scale LoRaWAN network that balances complexity, communication reliability, and computational efficiency.
• Tuning Strategies: The performance of the reinforcement learning-based scheduling algorithm is highly dependent on the choice of parameters, such as learning rate, batch size, and discount factor. The section discusses the importance of optimizing these parameters to enhance the convergence rate and overall effectiveness of the algorithm, ensuring it can adapt to varying network conditions and QoS requirements.

Overall, this section emphasizes the careful design of the simulation environment and parameters to facilitate a comprehensive analysis of the proposed scheduling algorithm's performance in realistic scenarios.

Table 2, outlines the key parameters used in the simulation of the LoRaWAN network to evaluate the proposed scheduling algorithm. The parameters include:

• Number of gateways: Set to 3, indicating the infrastructure available for communication within the network.
• Number of IoT devices: A total of 100 devices are simulated, representing the end-user devices that will communicate through the gateways.
• Network server: There is 1 network server managing the communication and data processing for the IoT devices.
• Environment size: The simulation area is defined as 200 m x 200 m, providing a controlled space for the network operations.
• Maximum distance to gateway: The maximum communication distance for devices to the gateway is set at 200m, reflecting the range capabilities of LoRaWAN technology.
• Propagation model: The LoRa Log Normal Shadowing Model is used to simulate realistic signal propagation conditions, accounting for environmental factors.
• Number of retransmissions: A maximum of 5 retransmissions is allowed for packet delivery attempts, enhancing reliability.
• Frequency band: The simulation operates on the 868MHz frequency band, commonly used for LoRaWAN communications.
• Spreading factor: Set to SF7, which determines the data rate and range of communication.

These parameters are carefully chosen to create a realistic medium-scale LoRaWAN environment, enabling the investigation of Quality of Service (QoS) metrics and the effectiveness of the scheduling algorithm.

The above parameters have been chosen in order to perform a realistic LoRaWAN environment and investigate the QoS metrics in IoT applications. The selected parameters aim to simulate a realistic medium-scale LoRaWAN IoT network that offers a good balance between the complexity of the network and communication reliability and computational efficiency for reinforcement learning. They rely on widely adopted real-world LoRaWAN configurations but provide the flexibility needed to effectively test a range of QoS and scheduling algorithms.

The Spreading Factors (SF7 to SF12) listed in table 2 represent the range of modulation configurations available in LoRaWAN, each offering a trade-off between data rate and transmission range. In the simulation, these SFs are dynamically assigned by the reinforcement learning agent based on the current network state. Nodes farther from the gateway or in poor signal conditions are assigned higher SFs (e.g., SF11 or SF12) to ensure reliable communication, albeit with lower data rates. Conversely, nodes with stronger signal quality or closer proximity to gateways use lower SFs (e.g., SF7 or SF8) to achieve faster data transmission and reduce channel occupancy. This adaptive SF allocation is a key aspect of the scheduling strategy, contributing to optimized QoS across diverse network scenarios.

Parameters and tuning strategies

The selection of the algorithm's parameters generally affects a sizable portion of the outcomes of the RL-based scheduling method. The ensuing sections outline the recommended practices for modifying the primary parameters as well as how the modifications affect algorithms.

The parameters and tuning strategies for the algorithm are crucial for optimizing performance:

• Learning rate (α): Set at 0.001, it determines how much new information influences existing knowledge, balancing convergence speed and stability.
• Exploration-exploitation balance (ε in ε-greedy strategy): The exploration rate starts at 1 and decays to 0.1, allowing the agent to explore initially while gradually favoring known actions.
• Discount factor (γ): Optimized at 0.95, it affects the importance of future rewards, promoting a balance between long-term and short-term rewards.
• Batch size for training: An optimal batch size of 128 is used to achieve faster convergence and effective generalization, avoiding overfitting or underfitting issues associated with larger or smaller sizes.

Table 3, presents the key parameters utilized in the reinforcement learning-based scheduling algorithm, which are crucial for its performance and effectiveness. The parameters include:

• Number of hidden layers: Set to 2, indicating the depth of the neural network used in the scheduling algorithm.
• Number of neurons per layer: Each hidden layer contains 128 neurons, which influences the network's capacity to learn complex patterns and relationships in the data.
• Learning rate (α): Fixed at 0.001, this parameter controls the magnitude of updates to the network weights during training, impacting convergence speed and stability.
• Discount factor (Gamma): Set to 0.95, this factor balances the importance of immediate rewards versus future rewards, guiding the agent's long-term decision-making.
• Exploration rate (Epsilon): Initialized at 1.0, this rate determines the likelihood of the agent exploring new actions versus exploiting known actions, promoting exploration in the early training stages.
• Exploration decay rate: Set at 0.995, this parameter gradually reduces the exploration rate over time, allowing the agent to focus more on exploitation as it learns.
• Minimum exploration rate: Fixed at 0.01, this ensures that the agent retains a small chance of exploring new actions even after extensive training.
• Replay buffer size: Set to 30,000, this parameter defines the capacity of the experience replay buffer, which stores past experiences for training stability.
• Batch size: Fixed at 64, this parameter determines the number of experiences sampled for each training iteration, balancing convergence speed and generalization.
• Target network update frequency: Set to every 500 steps, this parameter specifies how often the target network's weights are synchronized with the main Q-network, aiding in stable learning.

These algorithm parameters are essential for tuning the performance of the scheduling algorithm, ensuring effective learning and adaptation to the dynamic conditions of the LoRaWAN network.

Performance metrics analysis

The Performance Metrics Analysis evaluates the effectiveness of the proposed algorithm using key indicators such as delay, reliability, and throughput. The analysis demonstrates significant improvements in these metrics compared to baseline policies, highlighting the algorithm's ability to optimize QoS in LoRaWAN networks. Overall, the results indicate that the RL-based scheduling approach enhances network performance, particularly in managing overlapping QoS requirements.

Network delay: Figure 5, delay performance in different LoRaWAN node configurations. The results show that the delay for the RL-Based Algorithm (DQN) is considerably lower than LoRa+ and RT- LoRa. It is because the RL-Based Algorithm adopts adaptive decisions, optimally scheduling tasks and allocating resources. Illustrates the relationship between network delay and the number of nodes in a LoRaWAN environment.

Figure 5: Delay vs., number of nodes.

• Trend analysis: The graph typically shows that as the number of nodes increases, the delay experienced in the network also increases. This trend is indicative of the growing contention for communication resources, leading to longer wait times for packet transmission.
• Comparison of algorithms: The figure likely compares the delay performance of different scheduling algorithms, such as the proposed RL-based algorithm versus traditional methods like LoRa+ and RT-LoRa. The RL-based algorithm is expected to demonstrate significantly lower delays, showcasing its effectiveness in optimizing resource allocation and scheduling tasks.
• Implications for QoS: The results presented in this figure highlight the importance of efficient scheduling in maintaining low latency, especially in scenarios with a high density of nodes. This is crucial for applications requiring real-time data transmission, emphasizing the need for advanced algorithms to manage network performance effectively.

Overall, figure 5, indicates, the proposed RL-based model achieves the lowest delay, approximately 40ms, significantly outperforming all other models. In comparison, RT-LoRa exhibits the highest delay at around 120 ms, while LoRa+ perform better than RT-LoRa but still experience delays exceeding 80ms. Adaptive Spatial Scheduling demonstrates moderate performance with a delay of approximately 90ms. The strength of the RL-based model lies in its ability to optimize task scheduling effectively, minimizing delays and making it highly suitable for time-sensitive IoT applications.

• Packet Delivery Ratio (PDR): Figure 6: represents the PDR versus the number of nodes in LoRaWAN. It is remarked that the proposed RL-Based Algorithm DQN reaches the highest value of the PDR compared to the other two algorithms in this network, namely RT-LoRa and LoRa+. Also depicts the relationship between the Packet Delivery Ratio (PDR) and the number of nodes in a LoRaWAN network.PDR Trends: The graph typically shows that as the number of nodes increases, the PDR may initially rise but eventually plateaus or decline. This behavior indicates that while more nodes can enhance network coverage, increased contention and potential collisions can negatively impact the successful delivery of packets.
• Algorithm Comparison: The figure highlights the performance of the proposed RL-based algorithm (DQN) in achieving the highest PDR compared to other algorithms like RT-LoRa and LoRa+. This superiority suggests that the RL-based approach effectively manages scheduling and resource allocation, minimizing packet losses.
• Significance for Network Performance: The PDR is a critical metric for assessing the reliability of communication in IoT networks. A higher PDR indicates better performance and reliability, which is essential for applications that require consistent data transmission, reinforcing the importance of advanced scheduling techniques in optimizing network performance.

Figure 6: Packet Delivery Ratio (PDR) vs., number of nodes.

Overall figure 6, indicates, the proposed RL-based model achieves a Packet Delivery Ratio (PDR) of approximately 85%, which is comparable to other models. RT-LoRa and LoRa also attain PDRs in the range of 80% to 85%. The strength of the RL-based model lies in its ability to maintain high reliability in data delivery while optimizing for other performance metrics. However, its weakness is the marginal improvement over existing models, indicating limited novelty in this aspect.

Packet Error Rate (PER)

It can be easily observed that the minimum PER is given by the RL-Based Algorithm, while the RT-LoRa and LoRa+ give higher values. The low PER of the RL-Based Algorithm is mainly contributed by its dynamic optimization of scheduling and resource allocation, which decreases packet collisions, interference, and transmission errors by a great amount. Figure 7, illustrates the relationship between Packet Error Rate (PER) and the number of nodes in a LoRaWAN network.

Figure 7: Packet error rate (per) vs., number of nodes.

• PER trends: The graph typically shows that as the number of nodes increases, the PER tends to rise, indicating a higher percentage of packets experiencing errors during transmission. This trend reflects the increased likelihood of packet collisions and interference in a congested network environment.
• Algorithm performance: The figure highlights that the RL-based algorithm exhibits the lowest PER compared to other algorithms like RT-LoRa and LoRa+. This lower PER is attributed to the dynamic optimization of scheduling and resource allocation performed by the RL-based approach, which effectively reduces packet collisions and transmission errors.
• Implications for network reliability: A lower PER is crucial for ensuring reliable communication in IoT applications, as it directly impacts the overall performance and efficiency of the network. The results presented in this figure underscore the importance of employing advanced scheduling algorithms to enhance network reliability and minimize transmission errors, especially in scenarios with a high number of nodes.

Overall, figure 7, emphasizes the correlation between node density and packet error rates, showcasing the effectiveness of the proposed RL-based algorithm in maintaining low error rates in a congested network. The proposed RL-based model exhibits the lowest Packet Error Rate (PER) at approximately 5%, outperforming all other models. In comparison, RT-LoRa has the highest PER at around 15%, while LoRa+ show moderate values of approximately 10%. The strength of the RL-based model lies in its ability to minimize errors, demonstrating its robustness in maintaining data integrity. No weaknesses were observed in this metric.

Throughput: Figure 8, illustrates the relationship between throughput and the number of nodes in a LoRaWAN network. It shows that as the number of nodes increases, the throughput achieved by the RL-based algorithm (DQN) remains significantly higher compared to other algorithms like RT-LoRa and LoRa+. This superior performance is attributed to the RL-based algorithm's dynamic optimization of scheduling decisions, which effectively balances network load and minimizes collisions, resulting in enhanced data transmission rates even as node density increases.

Figure 8: Throughput vs., number of nodes.

The proposed RL-based model achieves the highest throughput, approximately 250 kbps, significantly outperforming all other models. In contrast, RT-LoRa has the lowest throughput at around 150 kbps, while LoRa+ throughput of approximately 200 kbps. The strength of the RL-based model lies in its ability to maximize network efficiency by optimizing resource allocation.

Statistical analysis for the Study

The study compares the performance of the proposed RL-based task scheduling algorithm with other existing algorithms across several QoS metrics (delay, PDR, PER, throughput), the following statistical tests are recommended:

1. Validate the sample size: Ensure that the number of simulation runs is statistically sufficient to draw reliable conclusions.
2. Assess the efficiency of the proposed feature selection method: Evaluate whether the selected features (QoS metrics) significantly impact the performance of the proposed model.
3. Analyze differences in performance among classifiers: Compare the performance of the proposed RL-based model against other models (RT-LoRa, LoRa+, DDQN-PER) using statistical tests.

Validate the Sample Size

• Approach: Power Analysis
• Objective: Ensure that the number of simulation runs is sufficient to detect significant differences in performance metrics (e.g., delay, PDR, PER, throughput).
• Method: Use a power analysis to calculate the minimum required sample size for detecting a meaningful effect size.
• Parameters:
• Significance level (αα): 0.05
• Power (1−β1−β): 0.8
• Effect size (dd): Medium (0.5), based on Cohen's guidelines.
• Perform power analysis using Python (stats models’ library) or R (per package).
• Interpretation:  If the number of simulations runs per scenario in the study meets or exceeds this calculated sample size, the results are statistically reliable.

If not, additional simulation runs are necessary.

Assess the efficiency of feature selection

• Objective: Determine whether the selected QoS metrics (delay, PDR, PER, throughput) significantly contribute to the model's performance.
• Approach: Compare model performance with and without feature selection.

Train two versions of the RL-based model:

1. Using all available features (e.g., raw network parameters and QoS metrics).
2. Using only selected QoS metrics as features.

• Use a paired t-test or Wilcoxon signed-rank test to compare performance metrics (delay, PDR, PER, throughput) between these two models.
• Interpretation:
• If p < 0.05 p < 0.05, feature selection significantly improves performance.
• If p ≥ 0.05 p≥ 0.05, feature selection does not have a significant impact.

Analyze differences in performance among classifiers

• Objective:  Compare the performance of the proposed RL-based model with RT-LoRa, LoRa+, and DDQN-PER across all QoS metrics.
• Approach:  Use one-way ANOVA to analyze differences among multiple groups (models).
• Post-hoc tests (e.g., Tukey's HSD) identify which pairs of models differ significantly.
• If data are non-normal or variances are unequal, use Kruskal-Wallis test with Dunn's post-hoc test.
• Interpretation: If p < 0.05 p < 0.05, there are significant differences among models for a given metric.
• Post-hoc tests identify which specific models differ significantly (Table 4).

By applying these statistical tests to your study's data and simulation results (Figures 5,8), you can validate your findings rigorously and provide evidence for your conclusions regarding sample size adequacy, feature selection efficiency, and classifier performance differences.

Performance comparison of the proposed RL-based model with RT-LoRa, and LoRa+.

The performance of the proposed RL-based model is compared with other related work models (RT-LoRa, and LoRa+) across all QoS metrics (delay, PDR, PER, and throughput). The analysis is based on the results presented in figures 5,8 of the thesis (Table 5).

Conclusion

This study presents a reinforcement learning (RL)-based task scheduling algorithm designed to enhance Quality of Service (QoS) in LoRaWAN-based IoT networks. By leveraging a Deep Q-Network (DQN), the proposed model dynamically learns and adapts to changing network conditions, effectively optimizing key performance metrics such as delay, throughput, packet delivery ratio (PDR), and packet error rate (PER). Unlike traditional scheduling techniques, which often rely on static heuristics or centralized logic, the RL agent autonomously adjusts its actions through continuous interaction with the network environment, demonstrating superior responsiveness and scalability. The proposed solution marks a significant step forward in intelligent IoT network management by introducing a context-aware and self-optimizing scheduler. It highlights the potential of RL to serve as a flexible and powerful alternative to conventional methods, especially in resource-constrained and dynamic LPWAN environments. This adaptability is critical for supporting the next generation of IoT applications, including smart cities, agriculture, industrial monitoring, and remote healthcare.

Limitations

Despite its promising performance, this work has several limitations. The evaluation was conducted exclusively in a simulated environment using NS-3, which, while realistic, cannot fully capture the unpredictability of real-world deployments, including environmental interference, hardware variability, and packet collisions from non-LoRaWAN devices. The model assumes a stable network topology and complete observability, which may not be feasible in large-scale or mobile settings. Additionally, the algorithm does not explicitly incorporate energy efficiency into its decision-making process—an essential factor in battery-powered LPWAN applications. Addressing these limitations through real-world validation, energy-aware policy design, and support for partially observable or mobile scenarios remains an important area for future work.

Further work

To further enhance the effectiveness and applicability of the proposed RL-based task scheduling algorithm for LoRaWAN, several directions are recommended. Future research should explore advanced reinforcement learning techniques such as Double DQN, Dueling DQN, and Proximal Policy Optimization (PPO) to improve learning stability and decision accuracy in dynamic environments. Additionally, incorporating energy consumption into the reward function is essential to extend the operational lifetime of battery-powered IoT nodes. Expanding the model to support multi-gateway and mobile node scenarios through multi-agent reinforcement learning (MARL) can improve scalability and adaptability in real-world deployments. Addressing partial observability using RNNs or POMDP-based methods will also enhance robustness in large-scale networks with incomplete information. Finally, implementing and evaluating the algorithm in real-world testbeds like The Things Network (TTN) or ChirpStack will provide valuable insights into practical challenges such as interference, hardware limitations, and protocol integration.

1. Mahmood NH, Marchenko N, Gidlund M, Popovski P. Wireless networks and industrial IoT: Applications, challenges and enablers. 2020. doi: 10.1007/978-3-030-51473-0.
2. de Oliveira LR, de Moraes P, Neto LPS, da Conceição AF. Review of LoRaWAN applications. 2020. doi: 10.48550/arXiv.2004.05871.
3. Marais JM, Malekian R, Abu-Mahfouz AM. LoRa and LoRaWAN testbeds: A review. IEEE AFRICON Sci. Technol Innov Africa AFRICON. 2017;1496-1501. doi: 10.1109/AFRCON.2017.8095703.
4. K. Mekki, E. Bajic, F. Chaxel, F. Meyer. Overview of cellular LPWAN technologies for IoT deployment: Sigfox, LoRaWAN, and NB-IoT,” 2018 IEEE Int. Conf. Pervasive Comput. Commun. Work. PerCom Work. 2018;197-202. doi: 10.1109/PERCOMW.2018.8480255.
5. Bouguera T, Diouris JF, Chaillout JJ, Jaouadi R, Andrieux G. Energy consumption model for sensor nodes based on LoRa and LoRaWAN. Sensors (Switzerland). 2018;18(7):1-23. doi: 10.3390/s18072104.
6. Augustin A, Yi J, Clausen T, Townsley WM. A Study of LoRa: Long Range & Low Power Networks for the Internet of Things. Sensors (Basel). 2016 Sep 9;16(9):1466. doi: 10.3390/s16091466. PMID: 27618064; PMCID: PMC5038744.
7. Ragnoli M, Barile G, Leoni A, Ferri G, Stornelli V. An autonomous low-power lora-based flood-monitoring system. J Low Power Electron Appl. 2020;10:2. doi: 10.3390/jlpea10020015.
8. Haxhibeqiri J, Moerman I, Hoebeke J. Low overhead scheduling of LoRa transmissions for improved scalability. IEEE Internet Things J. 2019;6(2)3:097-3109. doi: 10.1109/JIOT.2018.2878942.
9. Sutton RS, Barto AG. Reinforcement learning_Peter. Learning. 2012;3:9.
10. Petäjäjärvi J, Mikhaylov  K, Pettissalo M, Janhunen J, Iinatti J. Performance of a low-power wide-area network based on lora technology: Doppler robustness, scalability, and coverage. Int J Distrib Sens Networks. 2017;13:3. doi: 10.1177/1550147717699412.
11. Polonelli T, Brunelli D, Marzocchi A, Benini L. Slotted ALOHA on LoRaWAN-Design, Analysis, and Deployment. Sensors (Basel). 2019 Feb 18;19(4):838. doi: 10.3390/s19040838. PMID: 30781662; PMCID: PMC6413417.
12. Alenezi M, Chai KK, Alam AS, Chen Y, Jimaa S. Unsupervised learning clustering and dynamic transmission scheduling for efficient dense LoRaWAN networks. IEEE Access. 2020;8:191495-191509. doi: 10.1109/ACCESS.2020.3031974.
13. Leonardi L, Battaglia F, Lo Bello L. RT-LoRa: A medium access strategy to support real-time flows over LoRa-based networks for industrial IoT applications. IEEE Internet Things J. 2019;6(6):10812-10823. doi: 10.1109/JIOT.2019.2942776.
14. Sallum E, Pereira N, Alves M, Santos M. Improving quality-of-service in LOra low-power wide-area networks through optimized radio resource management. J Sens Actuator Networks. 2020;9(1)1-26. doi: 10.3390/jsan9010010.
15. Micheletto M, Zabala P, Ochoa SF, Meseguer R, Santos R. Determining Real-Time Communication Feasibility in IoT Systems Supported by LoRaWAN. Sensors (Basel). 2023 Apr 26;23(9):4281. doi: 10.3390/s23094281. PMID: 37177485; PMCID: PMC10181743.
16. Garrido-Hidalgo C, Haxhibeqiri J, Moons B, Hoebeke J, Olivares T, Ramirez FJ. LoRaWAN scheduling: From concept to implementation. IEEE Internet Things J. 2021;8(16):12919-12933. doi: 10.1109/JIOT.2021.3064430.
17. Siddiqi UF, Sait SM, Uysal M. Deep reinforcement based power allocation for the max-min optimization in non-orthogonal multiple access. IEEE Access. 2020;8:211235-211247. doi: 10.1109/ACCESS.2020.3038923.