Energy-Efficient UAV-assisted LoRa Gateways: A Multi-Agent Optimization Approach

arXiv:2502.03377 · cs.NI, cs.LG · Submitted 2025-02-05 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Computer Science. The day in computer science: software, systems and theory on arXiv.

Tom: We're looking at the paper: "Energy-Efficient UAV-assisted LoRa Gateways".

Paper summary: Tom: So we're looking at this paper today called "Energy-Efficient UAV-assisted LoRa Gateways: A Multi-Agent Optimization Approach." Essentially, they're tackling the problem of managing energy in next generation Internet of Things networks because these things are getting way more numerous and they use up a ton of power >

Jane: Right, so the big idea here is to figure out how to make sure those LoRa networks stay sustainable by optimizing how much power you use across all the UAVs and their connections >

Tom: They frame this as a problem where you've got multiple unmanned aerial vehicles, or UAVs, helping collect data from ground devices, or EDs, and they want to maximize energy efficiency >

Lu: The paper sets up the whole thing as a partially observable stochastic game so they can account for things like changing channel conditions and how mobile the ground devices are >

Meng: So it's not just a static setup; it's dynamic, meaning each UAV only knows what it sees locally at any given time, which makes the decision-making part pretty tricky >

Lalam: The main goal is to find the best way to jointly optimize transmission power, spreading factor, bandwidth, and user association for all those devices and UAVs >

Tom: They propose a two-stage solution for this complex optimization problem because it's NP hard >

Jane: What's the first stage? How do they handle figuring out which ground device connects to which UAV effectively across the network? >

Lu: For that, they use something called a channel-aware matching algorithm to dynamically assign each device to the UAV that has the best channel gain while respecting communication range and capacity limits >

Tom: And then for the second stage, they use this multi-agent reinforcement learning framework called MAPPO with centralized training with decentralized execution, or CTDE >

Meng: CTDE sounds fancy. What does that mean for deployment? Does it mean we train a big central model but the UAVs run their actual decisions independently in the field >

Jane: It means they train the system centrally to learn how to coordinate, but when it's actually flying around, each UAV just uses its own local information to make its move >

Lalam: The reward function they use is what ties it all together: it's a weighted sum of system energy efficiency, some other terms like information rate, and penalties for total power consumption >

Tom: And the simulations show that this approach converges much faster than other state-of-the-art multi-agent reinforcement learning algorithms >

Lu: They also found that during the execution phase of the simulation, their method improved system energy efficiency compared to both on-policy and off-policy reinforcement learning methods >

Meng: That's good, but how does this translate to real world constraints? What about the complexity of running a MAPPO framework at scale on limited UAV hardware >

Jane: It seems like they've put a lot of work into showing the model works well, but we should always keep an eye on whether those computational requirements are practical for actual deployment >

Tom: The paper concludes that this MAPPO approach outperforms other state-of-the-art MARL algorithms in terms of both fast convergence and system energy efficiency >

Lu: Looking ahead, they mention integrating trajectory optimization for multiple UAV mounted LoRa gateways to get adaptive coverage, and exploring hybrid deployments where LoRa uses both the eight hundred sixty-eight MHz and two point four GHz bands >

Jane: So, for someone just listening to this show, the main point is that they've developed a solid framework using reinforcement learning to make sure these low-power communication networks use energy wisely >

Conclusion: Tom: So we've been looking at this paper about "Energy-Efficient UAV-assisted LoRa Gateways: A Multi-Agent Optimization Approach." Basically, they’re talking about using drones to make low power communication networks last longer >

Jane: Right, so the authors are trying to figure out how to balance the energy use of all those drones and the devices they connect back to >

Tom: The core idea is taking four different things—power levels, how wide their signal is, bandwidth—and optimizing them together based on where everything is and what's happening right now >

Lu: They frame it as a partially observable stochastic game so the drones only know what they can see locally at any moment >

Meng: The main result they show is that their method beats other multi-agent reinforcement learning algorithms in terms of speed and actual energy savings when you run the simulation >

Lalam: And they found that their approach consistently gave the best system energy efficiency across all different numbers of devices they tested >

Tom: So what does this mean for us, listeners? It means we can design these networks to use less battery overall without sacrificing the data collection quality or speed >

Jane: It suggests that smart coordination between drones and ground devices is a really powerful way to manage energy in these IoT systems >

Tom: The authors are pointing toward future work involving trajectory optimization for more complex drone setups and using both eight hundred sixty-eight MHz and two point four GHz bands >

Lu: And they’re looking into how this works when you have multiple LoRa gateways flying around trying to cover the area adaptively >

Jane: That’s a big step toward making these networks really robust in real-world environments >

Abdullahi Isa Ahmed, Jamal Bentahar, El Mehdi Amhoud

College of Computing, Mohammed VI Polytechnic University (UM6P) · Department of Computer Science, Khalifa University · Concordia Institute for Information Systems Engineering, Concordia University

cs.NI, cs.LG

Submitted: 2025-02-05

Updated: 2026-03-25

Comments: 6 pages, 5 figures, 2 table

DOI: 10.1109/VTC2026-Spring68026.2026.11757367

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 83/100

The gist: As next-generation Internet of Things (NG-IoT) networks continue to grow, along with their energy demands…

Key concepts

POSG
Partially Observable Stochastic Game is a mathematical model used because UAVs cannot see the entire system state. Each UAV makes decisions based only on its limited, imperfect local observations of the environment and other agents' actions.
MAPPO
Multi-Agent Proximal Policy Optimization is a reinforcement learning framework designed to train multiple autonomous agents (the UAVs) to make optimal decisions simultaneously. It uses centralized training with decentralized execution to handle the complexity of coordinating resource allocation.
Energy Efficiency (EE)
Energy Efficiency measures how much useful data collection energy is achieved relative to the total power consumed by all associated end devices and the UAV's hover power. The goal is to maximize this ratio through joint optimization of transmission parameters.
CTDE
Centralized Training with Decentralized Execution is a training paradigm where agents learn collaboratively using global information during training, but once deployed, each agent only uses its local observations to make decisions. This helps agents learn complex coordination strategies.

Terminology

Summary

As next-generation Internet of Things (NG-IoT) networks continue to grow, along with their energy demands, creating challenges for resource management and sustainability <ref:2502. p. 0

Where the paper says it

1–6, 2026. Abstract—As next-generation Internet of Things (NG-IoT) networks continue to grow, the number of connected devices is rapidly increasing, along with their energy demands, creating challenges for resource management and sustainability.

03377#pg2> In this paper, we study Long Range (LoRa) networks supported by multiple unmanned aerial vehicles (UAVs) in an uplink data collection scenario <ref:2502. p. 0
Where the paper says it

Energyefficient communication, particularly for power-limited IoT devices, is therefore a key research focus. In this paper, we study Long Range (LoRa) networks supported by multiple unmanned aerial vehicles (UAVs) in an uplink data collection scenario.

03377#pg2>. Our objective is to maximize system energy efficiency by jointly optimizing transmission power, spreading factor, bandwidth, and user association <ref:2502. p. 0
Where the paper says it

In this paper, we study Long Range (LoRa) networks supported by multiple unmanned aerial vehicles (UAVs) in an uplink data collection scenario. Our objective is to maximize system energy efficiency by jointly optimizing transmission power, spreading factor, bandwidth, and user association.

03377#pg2>.

System Modeling and Problem Formulation

The system is modeled as a partially observable stochastic game (POSG) to account for dynamic channel conditions, end device mobility, and partial observability at each UAV <ref:2502.03377#pg2>. This framework involves multiple decisionmakers interacting within a dynamic environment where each agent must act based solely on its own local, imperfect observations rather than possessing global knowledge of the system state <ref:2502. p. 0

Where the paper says it

To address this problem, we model the system as a partially observable stochastic game (POSG), where multiple decisionmakers interact within a dynamic environment, and each agent must act based solely on its own local, imperfect observations rather than possessing global knowledge of the system state. Based on this framework, we propose a two-stage solution: a channel-aware matching scheme for ED-UAV association and a multi-agent proximal policy optimization (MAPPO) framework for resource allocation.

03377#pg2>. The total energy efficiency (EE) is modeled as a function of the total uplink transmission power for all associated end devices and the UAV's hover power consumption <ref:2502.03377#pg2>.

Proposed Two-Stage Solution

The proposed solution involves a two-stage approach to tackle the NP-hard optimization problem <ref:2502.03377#pg2>.

  1. Channel-aware matching algorithm for end deviceUAV association is used to dynamically assign each ED to the UAV with the strongest channel gain subject to communication range and capacity constraints <ref:2502. p. 3

    Where the paper says it

    To address this challenge, we propose a two-stage approach. First, we employ a channelaware matching algorithm that dynamically assigns each ED to the UAV with the strongest channel gain, subject to communication range and capacity constraints.

    03377#pg2>.

  2. A multi-agent proximal policy optimization (MAPPO) framework based on centralized training with decentralized execution (CTDE) is employed for resource allocation under these associations <ref:2502. p. 0

    Where the paper says it

    To address this challenging problem, we first model it as a partially observable stochastic game (POSG) to account for dynamic channel conditions, end device mobility, and partial observability at each UAV. We then propose a two-stage solution: a channel-aware matching algorithm for end deviceUAV association and a cooperative multi-agent reinforcement learning (MARL) based multi-agent proximal policy optimization (MAPPO) framework for resource allocation under centralized training with decentralized execution (CTDE).

    03377#pg2>.

Multi-Agent Reinforcement Learning Framework

The problem is formulated as a POSG defined by the tuple (U, S, Au, Ou, T, Zu, Ru, γ, µ0) where each UAV acts as a decentralized agent under partial observability <ref:2502.03377#pg2>. The agents' action space includes the spreading factor ψ[t], the transmission power p[t], and the bandwidth w[t] <ref:2502.03377#pg2>. The shared reward function is defined as r[t] = ω1 · EEsys[t] + ω2 · Ξ[t] + ω3 · β[t] − ω4 · Ptotal[t], which ensures all UAVs collaboratively optimize the global objective <ref:2502. p. 3

Where the paper says it

(12), Ξ[t] ∈ [0, 1] is the communication success rate, β[t] is a shaped term that rewards positive average SNR margin and heavily penalizes negative margins which address constraint (13h), Ptotal[t] is the instantaneous total power consumption of the system, and ω1, ω2, ω3, ω4 are weighting coefficients. This shared reward ensures all UAVs collaboratively optimize the global objective despite relying only on their partial local observations ou[t].

03377#pg2>.

Simulation and Performance Results

Simulation results show that the proposed approach significantly outperforms conventional off-policy and on-policy MARL algorithms <ref:2502.03377#pg2>. Specifically, the MAPPO approach converged faster than other state-of-the-art MARL benchmarks <ref:2502. p. 4

Where the paper says it

3(a), we plot the cumulative rewards over environment timesteps for all compared algorithms with 2 UAVs and 10 active EDs. As training progresses, our proposed MAPPO approach converges faster and obtains higher rewards as compared to other stateof-the-art MARL benchmarks.

03377#pg2>. Furthermore, during execution phase, our approach improved the system EE compared to both on-policy and off-policy RL <ref:2502. p. 1
Where the paper says it

• Our simulations result show that our approach requires fewer environment steps to achieve convergence compared to other state-of-the-art MARL algorithms. Additionally, during execution phase, our approach improved the system EE compared to both on-policy and off-policy RL.

03377#pg2>. The approach consistently achieved the highest performance across all numbers of active EDs when compared against QMIX and VDN <ref:2502. p. 4
Where the paper says it

As expected, the figure presents the final average EE over three seeds, and it shows that our proposed MAPPO approach consistently achieves the highest performance across all numbers of active EDs. Compared to the second-best algorithm, VDN, our approach improves EE by 36.

03377#pg2>.

Conclusion

The paper concludes that the proposed MAPPO approach significantly outperforms state-of-the-art MARL algorithms in terms of fast convergence and system energy efficiency <ref:2502.03377#pg2>. Future directions include integrating trajectory optimization for multiple UAVmounted LoRa gateways to enable adaptive coverage and exploring hybrid deployments where LoRa operates in both the 868 MHz and 2. p. 5

Where the paper says it

Simulation results show that our proposed MAPPO approach significantly outperforms state-of-the-art MARL algorithms in terms of fast convergence and system energy efficiency. Future directions include integrating trajectory optimization for multiple UAVmounted LoRa gateways to enable adaptive coverage and exploring hybrid deployments where LoRa operates in both the 868 MHz and 2.

4 GHz bands <ref:2502. p. 5
Where the paper says it

Future directions include integrating trajectory optimization for multiple UAVmounted LoRa gateways to enable adaptive coverage and exploring hybrid deployments where LoRa operates in both the 868 MHz and 2. 4 GHz bands.

03377#pg2>.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Channel-aware matching algorithm for ED-UAV association: This scheme dynamically assigns each ED to UAV with maximum gain subject to communication range and capacity constraints, as described by av[t] = arg max u∈U n Gu,v[t] du,v[t] ≤ Rcomm.

  2. MAPPO-based resource allocation under CTDE: The system can jointly optimize spreading factor, transmission power, bandwidth, and user association by using a Multi-Agent Proximal Policy Optimization framework with centralized training and decentralized execution.

  3. Reward function design: The reward function is explicitly defined as r[t] = ω1 · EEsys[t] + ω2 · Ξ[t] + ω3 · β[t] − ω4 · Ptotal[t], which rewards system energy efficiency, communication success rate, and SNR margin while penalizing total power consumption.

  4. Performance benchmarking: The proposed approach can be compared against state-of-the-art MARL algorithms like COMA, VDN, and QMIX to demonstrate superior convergence speed and higher system energy efficiency in uplink data collection scenarios.

Abstract

As next-generation Internet of Things (NG-IoT) networks continue to grow, the number of connected devices is rapidly increasing, along with their energy demands, creating challenges for resource management and sustainability. Energy-efficient communication, particularly for power-limited IoT devices, is therefore a key research focus. In this paper, we study Long Range (LoRa) networks supported by multiple unmanned aerial vehicles (UAVs) in an uplink data collection scenario. Our objective is to maximize system energy efficiency by jointly optimizing transmission power, spreading factor, bandwidth, and user association. To address this challenging problem, we first model it as a partially observable stochastic game (POSG) to account for dynamic channel conditions, end device mobility, and partial observability at each UAV. We then propose a two-stage solution: a channel-aware matching algorithm for end device-UAV association and a cooperative multi-agent reinforcement learning (MARL) based multi-agent proximal policy optimization (MAPPO) framework for resource allocation under centralized training with decentralized execution (CTDE). Simulation results show that our proposed approach significantly outperforms conventional off-policy and on-policy MARL algorithms.

Sources

Related papers