Graph Attention-based Decentralized Actor-Critic for Dual-Objective Control of Multi-UAV Networks

arXiv:2506.09195 · eess.SP, cs.AI, cs.MA · Submitted 2025-06-10 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: We're looking at the paper: "Graph Attention-based Decentralized Actor-Critic for Dual-Objective Control of Multi-UAV Networks".

Dev: Here's the gist of the paper. Graph Attention-based Decentralized Actor-Critic for Dual-Objective Control of Multi-UAV Swarms focuses on optimizing multi-UAV systems with dual objectives: maximizing service coverage as the primary goal while extending battery lifetime as the secondary objective <ref:2506.09195#pg2>.

Rosa: The proposed approach leverages a graph attention network to process UAVs’ limited local observation and reduce the dimension of the environment states <ref:2506.09195#pg3>.

Dev: Subsequently, an actor-double-critic network is developed to manage dual policies for joint objective optimization <ref:2506.09195#pg3>.

Rosa: The proposed GADC uses a Kullback-Leibler (KL) divergence factor to balance the tradeoff between coverage performance and battery lifetime in the multi-UAV system <ref:2506.09195#pg3>.

Title and authors: Rosa: So we're looking at this paper now, "Graph Attention-based Decentralized Actor-Critic for Dual-Objective Control of Multi-UAV Networks," and it tackles a really specific challenge in multi-UAV systems. It focuses on balancing two things: getting the most service coverage possible while also trying to make those drones last longer on battery.

Dev: That’s right, Rosa. The authors propose something called Graph Attention-based Decentralized Actor-Critic, or GADC, which uses a graph attention network to process what each individual UAV can see locally and then develops an actor-double-critic network to handle those two competing objectives simultaneously.

Taro: When you say local observation for the UAVs, that sounds like they are really dealing with limited information about the whole environment at any given moment.

Rosa: Exactly, Taro. The paper points out that each UAV only has a limited view of the broader area around it, so they use a graph attention network to aggregate data from neighboring UAVs to get a better understanding of what’s going on.

Dev: And that aggregation is key because they build two specialized critic networks for this system: one specifically for coverage effectiveness and another one focused on maximizing the operational duration, which means battery life.

Taro: So, so we have these two separate goals being chased by different critics? How does that help when the drone is flying around a complex area?

Rosa: The way they set up those critics is designed to manage that tradeoff directly because they introduce this Kullback-Leibler divergence factor to control how much the lifetime objective adjusts based on the coverage objective.

Dev: It sounds like they’re using that KL divergence as a constraint, which stops the drone from letting its battery life strategy completely override what’s needed for good service coverage, which is something traditional weighted-sum methods often struggle with when objectives have huge differences in scale.

Taro: That sounds like a way to keep the system stable while still pushing for both goals at the same time? But how do they actually make sure that constraint is enforced during training?

Title and authors: Rosa: They enforce it by using a probability ratio between these two policies to restrict the adjustment of the lifetime-aware updating within a trust region defined by the coverage-effectiveness strategy, which keeps things tightly controlled.

Dev: That level of control over policy adjustment sounds like it would be really important for a real-time system where we need to know what’s happening in terms of latency and safety cases.

Taro: What about the environmental understanding part? The paper mentions that because each UAV is confined to local observations, early studies focused on centralized control or using multi-agent reinforcement learning, but they point out that those things can be expensive in ad hoc scenarios.

Rosa: Right, so this GADC approach specifically leverages graph attention layers within the multi-agent reinforcement learning structure to enhance the ability of the drones to get global information from their neighbors, which improves navigation capabilities beyond what they could manage on their own.

Dev: So it's about using that graph structure—that message passing between UAVs—to give them a richer context when making decisions about where to fly and how much energy to spend.

Taro: So if the environment misbehaves, like a sudden change in service demand or unexpected obstacles appear, how does this system handle it gracefully? Does it just crash, or does it recover well?

Rosa: The GADC architecture is designed so that by having these two distinct critics guiding the dual-objective policy iteratively, they should be able to adjust the UAV policy efficiently without getting stuck in a local optimum for either coverage or lifetime alone.

Dev: But Rosa, what about the practical reality? The paper says they tested this in both ideal settings and NVIDIA Sionna’s realistic ray tracing environment to see how it holds up when things get messy.

Taro: I'm curious about that realism, because testing in a simulation is one thing, but seeing it work when you introduce real physics like ray tracing makes a big difference for deployment.

Title and authors: Rosa: They did extensive testing in both ideal settings and NVIDIA Sionna’s realistic ray tracing environment to show how the GADC performs under those conditions, which suggests its capability extends beyond simple lab tests.

Dev: And the results are pretty strong; they showed that GADC improves convergence speed by fifty percent compared to traditional algorithms in both service coverage and system lifetime metrics.

Taro: Fifty percent improvement in convergence speed is significant, especially when you have to train these decentralized controllers repeatedly for real-world scenarios.

Rosa: It also pointed out that the GADC approach allows the UAVs to achieve a near uniform distribution above the user terminals in urban scenarios, which is different from conventional methods that tend to put drones at the outer edges of their coverage area.

Dev: That means instead of scattering them far apart trying to cover everything, they can actually cluster them smartly around areas where there’s a high density of user terminals.

Taro: That clustering sounds much more efficient for battery life and service quality. So what's the final word on what this paper means for deploying these swarm systems in practice?

Rosa: Essentially, it means we have a framework that lets us optimize dual objectives—coverage and longevity—without having to manually tune a weighting coefficient, which is a big deal because tuning those weights is usually guesswork.

Dev: It moves the optimization from being an ad-hoc balancing act to being handled by this integrated critic structure guided by the KL divergence constraint.

Taro: I just think it shows that combining graph attention for local context with a dual-critic approach is a solid way to tackle complex, competing goals in decentralized systems.

Rosa: Well, that’s what we have on the "Graph Attention-based Decentralized Actor-Critic for Dual-Objective Control of Multi-UAV Networks" paper. It gives us a method to build more robust and smarter multi-UAV swarms by letting them learn how to balance service coverage with battery life automatically through this GADC framework.

Dev: That's all we have time for today, folks. We'll be back next time with another interesting read from arXiv.

The paper's summary: Rosa: So, to wrap up what we just talked about: this GADC system uses graph attention layers so drones can talk to their neighbors and then it has these two separate critic networks—one for coverage and one for battery life—to find a way to do both things at the same time.

Dev: Exactly. It’s basically a dual-objective control method where they use this specific mathematical tool, the Kullback-Leibler divergence factor, to keep those two goals from fighting each other too hard during training.

Taro: What I find really interesting is how they structure that constraint enforcement—it uses a probability ratio between the policies to keep the lifetime objective from completely taking over. That sounds like a way to guarantee that coverage performance stays within an acceptable boundary while still optimizing for energy use.

Rosa: It's about balancing those trade-offs directly, instead of just guessing some number to weigh coverage against battery life. They’ve shown that this structure helps the system find a better solution than those traditional weighted-sum methods do when the objectives are really different in scale.

Dev: And look at the results they showed in both ideal settings and that realistic NVIDIA Sionna ray tracing environment; GADC actually improved convergence speed by fifty percent compared to other algorithms we’ve seen for this problem. That’s a solid number when you're talking about training time.

Taro: Fifty percent faster convergence means the learning process itself is much more efficient, which is crucial if you have to retrain the system quickly when things change in the environment. It also showed that GADC gets drones into a near uniform distribution above user terminals, which is better than conventional methods that just throw them at the edges of where they are needed most.

Rosa: So it means that for future deployments, we might be able to design drone swarms that are smarter about where they fly based on both what’s around them and how much power they have left. It moves the control from just reacting locally to having a more holistic, coordinated strategy across the whole group.

Dev: And it gives us a way to see how these dual objectives interact in practice through this framework instead of just treating them as two separate problems that need separate tuning. But what about when things go wrong? What is the safety case if the drone policy starts behaving strangely because one critic is overriding the other?

Taro: That’s a fair question, Dev. The authors focus on how they manage that during training, trying to keep the updates bounded by that coverage strategy constraint. If it goes completely haywire in a real-world failure mode, we'd need to see if those localized observations from the graph attention network can still give them enough context to pull back gracefully.

Rosa: It’s about making sure that even when things get messy—like sudden changes in service demand or unexpected obstacles—the system doesn't just crash. It needs a mechanism where the information aggregated by those GAT layers helps it recover its intended behavior, whether that’s maintaining coverage or conserving power.

Dev: So this paper shows a structured way to tackle dual objectives using graph structures and KL divergence for control, and the next thing we need to look at is how this framework performs when there are much more complex dynamics involved in the environment itself.

The paper's improvements: Tom: So, we’re moving on to what the paper suggests as improvements for this GADC system: they focus on how to make those drones even better at handling messy, real-world conditions and how to push the boundaries of what it can do.

Rosa: They’re suggesting that incorporating more layers into the graph attention network could allow those drones to look further afield, meaning they could gather information from neighbors much farther away than just their immediate ones. That would help them see patterns across a larger area.

Dev: That makes sense from a control engineering standpoint. If you can get better spatial awareness, your decision-making loop becomes more robust because the input data is richer and less localized. But how does that affect latency? Will adding more graph layers slow down the real-time loop?

Taro: I think the authors are hinting that this could be beneficial for handling sudden environmental changes. If a drone suddenly loses line of sight with one neighbor, having that aggregated information from multiple layers might let it immediately adjust its trajectory without waiting for a full re-observation cycle. That’s about graceful recovery.

Rosa: They also explore using the dual critics to guide exploration differently over time; it’s not just keeping them balanced constantly, but maybe letting one critic take more control when coverage is critical and the other takes more precedence when battery life is getting low. That’s a dynamic strategy change.

Dev: So they are looking at how to make the trade-off between coverage and lifetime adaptive in real-time based on their own internal assessment of which objective is currently more important for the system's survival. That’s a sophisticated level of control logic.

Taro: I wonder if this could lead to swarm behaviors that are much more intelligent than just following simple coverage maps or simple energy-saving paths. Imagine them learning to navigate not just where the users are, but where they *will* be based on the network structure itself.

Rosa: Exactly. The goal seems to be moving beyond a fixed strategy—like "always prioritize coverage" or "always prioritize battery"—to a more nuanced approach that considers the current state of both things simultaneously, which is what this layered graph attention structure enables.

Dev: It shows they are thinking about pushing the limits of decentralized control in scenarios where perfect information isn't available at any single point, but you have local connectivity to leverage. The safety case would revolve around whether that extra data actually helps or introduces too much lag into the decision-making process.

Taro: If they can demonstrate that this enhanced sensing leads to smoother transitions during failure—like successfully navigating around a dynamic obstacle while keeping an eye on a service area—then it’s a big step for autonomous swarms in complex environments.

Rosa: So, the implication here is that future drone swarms won't just be about putting drones where they are needed now; they’ll be about learning how to manage their entire operational life by intelligently sharing and interpreting information across the network.

Conclusion: Rosa: So we’re wrapping up on this Graph Attention-based Decentralized Actor-Critic for Dual-Objective Control of Multi-UAV Networks paper: essentially, they built a system that uses graph attention to let drones talk to each other and employs two separate critics to balance maximizing service coverage against extending battery life.

Dev: Right. It’s about using this specific structure, the KL divergence factor, to control how those two competing goals adjust their policies during training so neither one dominates the other.

Taro: I think the real takeaway is that this isn't just a tweak to an existing algorithm; it’s a new way to structure the learning process that forces coordination between different objectives. It sets a precedent for how we might handle multiple, conflicting goals in decentralized AI systems going forward.

Rosa: It shows that you don't necessarily need one single weighting number to tune performance across two different metrics; you can have an integrated mechanism that manages the trade-off dynamically. This is important because tuning those weights manually is usually a huge guessing game for anyone trying to deploy this kind of swarm in the field.

Dev: And from an engineering standpoint, they proved it works in both simulated ideal settings and that realistic ray tracing environment, showing a fifty percent improvement in convergence speed compared to traditional methods. That’s a solid metric when you’re thinking about deployment latency and training time for these control loops.

Taro: What I really like is how the GAT helps them explore unseen areas by prioritizing information from neighbors based on things like service load and residual energy, which makes the system much more adaptive to changing conditions than just relying on fixed observation ranges.

Rosa: So it means we could build swarms that are much smarter about where they fly, not just where the users are right now, but where they expect them to be based on how the network is structured and how much energy they have left.

Dev: It definitely shifts the focus from just optimizing for one thing—like maximum coverage or minimum energy—to finding that sweet spot in between them through this learned interaction. The safety case here would rely heavily on that KL divergence constraint keeping the policy adjustments within a bounded region.

Taro: I just think we need to keep pushing on those real-world failure modes, because while the training is good, seeing how it handles genuinely unexpected environmental chaos in a live trial is what will really tell us if this structure holds up when things go completely sideways.

Rosa: That’s right. The Graph Attention-based Decentralized Actor-Critic for Dual-Objective Control of Multi-UAV Networks offers a solid framework for making those complex trade-offs manageable through coordinated learning and constraint enforcement.

Dev: It’s a big step toward building more robust, multi-objective control systems in decentralized settings, and it gives us a lot to think about regarding how we structure the learning objective itself.

The Chinese University of Hong Kong

eess.SP, cs.AI, cs.MA

Submitted: 2025-06-10

Updated: 2026-11-09

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 95/100

The gist: Graph Attention-based Decentralized Actor-Critic for Dual-Objective Control of Multi-UAV Swarms focuses on optimizing multi-UAV systems with dual objectives…

Key concepts

Graph Attention Network (GAT)
A GAT is a neural network structure that allows the system to intelligently aggregate information from neighboring UAVs. It learns which neighbors are most important for a specific UAV's decision-making, helping the system understand the environment better by focusing on relevant local data.
Dual-Objective Control
This means controlling the UAV swarm to achieve two goals simultaneously: first, maximizing how many ground users (UTs) they can serve (coverage), and second, making sure their batteries last as long as possible (lifetime). The system must find a balance between these two competing needs.
Kullback-Leibler (KL) Divergence Factor
This mathematical factor acts like a constraint or a balancing mechanism. It restricts how much the battery life objective can change relative to the coverage objective. This prevents one goal from completely dominating the optimization, ensuring both objectives are considered fairly.
Actor-Double-Critic Network
This is the core AI structure used to manage the control policies. The 'actor' decides what action (like where to fly) to take, while the 'double critic' networks provide separate evaluations for coverage and battery life. This separation allows for fine-tuned optimization of both objectives.

Terminology

Summary

Graph Attention-based Decentralized Actor-Critic for Dual-Objective Control of Multi-UAV Swarms focuses on optimizing multi-UAV systems with dual objectives: maximizing service coverage as the primary goal while extending battery lifetime as the secondary objective <ref:2506. p. 009195#pg2>. The proposed approach leverages a graph attention network to process UAVs’ limited local observation and reduce the dimension of the environment states <ref:2506. p. 009195#pg3>. Subsequently, an actor-double-critic network is developed to manage dual policies for joint objective optimization <ref:2506. p. 009195#pg3>. The proposed GADC uses a Kullback-Leibler (KL) divergence factor to balance the tradeoff between coverage performance and battery lifetime in the multi-UAV system <ref:2506. p. 009195#pg3>.

System Modeling and Objectives

The system model considers that UAVs function as flying BSs to serve ground UTs, maintaining two types of communications: UAV-UAV links for information exchange between neighboring UAVs and UAV-UT links for service provisioning <ref:2506.09195#pg4>. The number of UTs served by the nth UAV during time slot t is defined by a summation formula involving service coverage indicator <ref:2506. p. 209195#pg4>. UAV energy consumption consists of four components, including navigation energy, ground services energy, inter-UAV communication energy, and hovering energy <ref:2506. p. 309195#pg5>. The system aims to maximize the discounted total reward J = E "∑t=0 γt r(t) for each UAV <ref:2506.09195#pg5>.

GADC Architecture and Information Processing

The GADC framework employs a Graph Attention Network (GAT) with multihead convolutional kernels to aggregate information from neighboring UAVs using a graph attention network (GAT) <ref:2506.09195#pg3>. The architecture utilizes two specialized critic networks: a coverage-effectiveness critic for coverage optimization and a lifetime-aware critic for maximizing operational duration <ref:2506. p. 109195#pg3>. The coverage-effectiveness critic primarily focuses on maximizing the coverage of UTs by the multi-UAV system, while the lifetime-aware critic serves as a secondary critic, aiming to extend the UAV system’s battery life without compromising the coverage performance beyond a controllable threshold <ref:2506. p. 109195#pg3>.

Dual-Objective Optimization Mechanism

The coverage-effectiveness critic and the lifetime-aware critic iteratively guide the training process of the dual-objective policy <ref:2506.09195#pg5>. Specifically, a stringent constraint on the Kullback-Leibler (KL) divergence is imposed to restrict the adjustment of the lifetime-aware updating within a trust region bounded by the coverage-effectiveness strategy, using the probability ratio between these policies to enforce this constraint <ref:2506. p. 109195#pg3>. This approach prevents the larger-valued objective from dominating the solution space, unlike traditional weighted-sum methods <ref:2506. p. 109195#pg3>.

Performance and Validation

Extensive testing in both ideal settings and NVIDIA Sionna’s realistic ray tracing environment demonstrates GADC’s superior performance <ref:2506.09195#pg3>. Simulations show that GADC improves the convergence speed by 50% and outperforms the traditional algorithm in both service coverage and system lifetime <ref:2506. p. 109195#pg4>. The proposed GADC achieves nearuniform UAV distribution above the UTs, while conventional MADDPG positions UAVs at the outer edges of the coverage area, farther from the central UT concentration <ref:2506. p. 909195#pg14>. The performance evaluation in Sionna shows that GADC achieves convergence at 500 episodes, while the GAT-based MADDPG and MADDPG require more than 700 episodes to converge. p. 9 In the digital twin environment, GADC significantly outperforms the benchmark algorithms in convergence speed. p. 1 The proposed GADC exhibits faster convergence without the need of careful selection of the weighting coefficient <ref:2506. p. 909195#pg13>. The KL divergence factor constrains the variation of the UAV policy within a specified region of the effectiveness objective while optimizing the UAV policy using the lifetime-aware objective at each iteration. p. 9

Conclusion

This study focused on optimizing dual objectives in multiUAV systems: maximizing user terminal (UT) coverage while extending UAV battery longevity <ref:2506.09195#pg2>. We developed the Graph Attention-based Decentralized Actor-Critic (GADC) method, which leverages Graph Attention Network layers to aggregate information from neighboring UAVs to efficiently explore unobserved areas <ref:2506. p. 009195#pg7>. By implementing dual critics, the GADC achieves rapid convergence toward optimal solutions for both objectives independently <ref:2506. p. 109195#pg3>. We tested the proposed scheme in NVidia Sionna, where the UAV-UT channels are generated by ray tracing <ref:2506. p. 009195#pg14>. Numerical results demonstrate that GADC significantly improves both system lifetime and service coverage compared to existing approaches <ref:2506. p. 109195#pg7>.

REFERENCES

[1] Z. Ye, et al., “Multi-UAV navigation for partially observable communication coverage by graph reinforcement learning,” IEEE Trans. p. 0 Mob. Comput., vol. 22, no. 7, pp. 4056–4069, Jan. 2023

[2] Z. Yang, et al., “Online trajectory and resource optimization for stochastic UAV-enabled MEC systems,” IEEE Trans. p. 1 Wireless Commun. p. 2 p. 2 p. 2 p. 2 p. 2, vol. 21, no. 7, pp. 5629–5643, Jul. 2022

[3] Y. Liu, et al., “Fair and energy-efficient coverage optimization for UAV placement problem in the cellular network,” IEEE Trans. p. 2 Commun., vol. 70, no. 6, pp. 4222–4235, Jun. p. 0 2022

[4] R. Madan, et al., “Distributed algorithms for maximum lifetime routing in wireless sensor networks,” IEEE Trans. Wireless Commun., vol. 5, no. 8, pp. 2185–2193, Aug. 2006

[5] X. Zhang, et al. p. 0, “Cooperative trajectory design of multiple UAV base stations with heterogeneous graph neural networks,” IEEE Trans. p. 1 Wireless Commun., vol. 22, no. 3, pp. 1495–1509, Mar. 2023

[6] J. Schulman, et al., “Trust region policy optimization,” in Proc. p. 1 the 32nd Int. Conf. Mach. Learn., vol. 37, Lille, France, Jul. 2015, pp. 1889–1897

[7] ——, “Proximal policy optimization algorithms,” arXiv preprint arXiv:1707.06347, 2017

[8] H. Peng, et al. p. 0, “Energy harvesting reconfigurable intelligent surface for UAV based on robust deep reinforcement learning,” IEEE Trans. p. 3 Wireless Commun., vol. 22, no. 10, pp. 6826–6838, Oct. 2023

[9] L. T. Hoang, et al., “Deep reinforcement learning-based online resource management for UAV-assisted edge computing with dual connectivity,” IEEE/ACM Trans. p. 1 Netw., vol. 31, no. 6, pp. 2761–2776, Dec. 2023

[10] C. Li, et al., “Deep reinforcement learning-based mining task offloading scheme for intelligent connected vehicles in UAV-aided MEC,” ACM Trans. p. 0 Des. Autom. Electron. Syst., vol. 29, no. 3, pp. 1–29, May 2024

[11] Z. Yang, et al., “Dynamic offloading and trajectory control for UAVenabled mobile edge computing system with energy harvesting devices,” IEEE Trans. p. 1 Wireless Commun., vol. 21, no. 12, pp. 10 515–10 528, Dec. 2022

[12] R. Liu, et al., “Resource allocation for NOMA-enabled cognitive satellite–UAV–terrestrial networks with imperfect CSI,” IEEE Trans. Cog. Commun. Netw., vol. 9, no. 4, pp. 963–976, Aug. 2023

[13] Y. Liu, et al., “Space-air-ground integrated networks: Spherical stochastic geometry-based uplink connectivity analysis,” IEEE J. Sel. Areas Commun. p. 1 p. 1, vol. 42, no. 5, pp. 1387–1402, May 2024

[14] Z. Yin, et al., “UAV-assisted secure uplink communications in satellitesupported IoT: Secrecy fairness approach,” IEEE Internet Things J. p. 0, vol. 11, no. 4, pp. 6904–6915, Feb. 2024

[15] L. Zhou, et al., “Intelligent UAV swarm cooperation for multiple targets tracking,” IEEE Internet Things J. p. 2, vol. 9, no. 1, pp. 743–754, Jan. 2022

[16] G. Raja, et al., “Efficient and secured swarm pattern multi-UAV communication,” IEEE Trans. p. 2 Veh. Technol., vol. 70, no. 7, pp. 7050–7058, Jul. 2021

[17] C. H. Liu, et al., “Energy-efficient UAV control for effective and fair communication coverage: A deep reinforcement learning approach,” IEEE J. p. 2 Sel. Areas Commun., vol. 36, no. 9, pp. 2059–2070, Sep. 2018

[18] Z. Feng, et al., “Multi-agent reinforcement learning with policy clipping and average evaluation for UAV-assisted communication markov game,” IEEE Trans. p. 0 Intell. Transp. Syst., vol. 24, no. 12, pp. 14 281–14 293, Dec. 2023

[19] Y. Hou, et al., “UAV swarm cooperative target search: A multi-agent reinforcement learning approach,” IEEE Trans. p. 0 Intell. Veh., vol. 9, no. 1, pp. 568–578, Jan. 2024

[20] S. Kakade, et al., “Approximately optimal approximate reinforcement learning,” in Proc. p. 0 Int. Conf. Mach. Learn., San Francisco, CA, Jul. 2002, pp.

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. Graph Attention based Decentralized Actor Critic (GADC) system can optimize dual objectives by separately optimizing coverage-effectiveness and lifetime-aware objectives using two specialized critic networks. The GADC uses a Kullback-Leibler KL divergence factor to balance the tradeoff between coverage performance and battery lifetime, preventing the larger-valued objective from dominating the solution space.

  2. Multi layer Graph Attention Network for Enhanced Environmental Understanding allows UAVs to aggregate information from neighboring UAVs (nodes) and investigate unobserved areas using multiple GAT layers. This enables comprehensive environmental data assimilation beyond immediate observational coverage by prioritizing information based on service load, position, and residual energy.

  3. The system can achieve superior convergence speed in both ideal settings and realistic ray tracing environments like NVIDIA Sionna's digital twin by leveraging the GADC architecture. Simulations show that GADC improves the convergence speed by 50% compared to traditional algorithms in both service coverage and system lifetime.

  4. The AI system can be deployed with a policy that prioritizes maximizing service coverage as the primary objective while simultaneously seeking to extend battery life as a secondary objective. This is achieved through the proposed GADC framework which prevents the larger-valued objective from overshadowing the solution space, unlike weighted-sum methods.

  5. The system can maintain optimal UAV deployment by achieving near uniform distribution above UTs in urban scenarios by leveraging GAT's ability to aggregate observations from neighboring UAVs to identify high UT density regions. This is demonstrated by GADC controlling UAVs to maintain a clustered distribution within the observation range of their nearest neighbors, unlike conventional MADDPG which positions them at the outer edges.

Sources

Related papers