Relay load balancing in large-scale UAV networks requires joint decisions on association, bandwidth, power and hovering position, while the dimension grows with the number of UAVs and each UAV sees only its local state. This paper splits the decision into two layers. An anchored potential direction decider emits a closed-form target at every instant and returns its direction only. A gain-learning executor maps a local observation into a positive gain per dimension, composes gain and target, and clips the action against the physical single-step bound; the two layers form PF-Exec. The unit gain is the default start, so the executor reduces to the pure potential policy before training. The load is first-order homogeneous in bandwidth, making the bandwidth target closed form, and the position target is the demand-weighted centroid of the served users. The system uses a two-hop access and millimetre-wave backhaul link with soft association, a 3GPP UMa-AV channel with two shadowing layers, an OU-type social-force crowd and a quadrotor physical layer. Evaluation spans all fleet sizes, two crowd scenarios, independent seeds, the published baselines in their own settings and a degraded-information line. In the highest-pressure burst setting the fixed-step comparison breaks down on part of the seeds, losing fairness and mean load at once, while PF-Exec holds fairness near one. Among the ten comparison methods only the greedy heuristic survives that setting; its load variance stays far higher. Under stale global state and position noise the fairness index stays high at both largest-scale scenarios. Flight energy is nearly identical between the learned and the fixed-step executor, and three constants are fixed once across all settings. The convergence properties of both layers, from the truncated gain iteration to the perturbed tracking bound, are proved in the companion analysis.
Figures & tables
Fig. 1: System model. A host base station sits at the origin and hovers serve ground users over an access hop, drawn solid, while the backhaul hop is drawn dashed. Every UAV keeps a fixed hovering height above a horizontal position inside the field.
Category
Value
field
600×450 m, height 50 m
scale
3, 12, 30, 60, 100, with side ∝n/12 and users, bandwidth ∝n/12
crowd
slow 0.42 m/s, fast 1.39 m/s, burst 2.78 m/s, step 0.1 s
demand
∣N(0.5,0.15)∣ truncated to [0.1,1] , anchor 0.8
access
2 GHz UMa-AV, 10 m grid, shadow 4 dB over 37 m and 6 dB over 50 m, runtime U[0,13.2] dB every hundred steps
backhaul
28 GHz, 100 MHz, 43 dBm, −174 dBm/Hz
TABLE I: Simulation parameters, base setting at n=12 .
Fig. 2: Computation and training flow. The direction decider returns the closed-form targets and the direction package from the local state, the executor, chosen among constant speed, fixed step and learned gain, composes the direction and the residual distance into the five-dimensional action, and the environment advances on the channel, the association, the crowd and the physics before the state is read again. The network of the learned executor takes twenty-two inputs and shares them across all scales.
Fig. 3: Return band over training rounds, with the sliding mean on the vertical axis and the spread across seeds as a shaded band. The return starts from the pure potential starting point of η≡1 , climbs, and enters the flat band that triggers the early stop.
Item
Setting
Observation
22: 12 env, 5 direction, 5 residual
Hidden widths
16, 32, 64, 32, 16
Output
5 positive gains, zero-initialised last layer
Gain map
η=exp(±0.7tanh(⋅))
Gain range
[0.497,2.014]
Relaxation on b,π
0.2
TABLE II: Executor network and training settings, identical at every scale.
n
field (m)
users
total bandwidth
share per UAV
3
300×225
150
25 MHz
8.33 MHz
12
600×450
600
100 MHz
8.33 MHz
30
949×712
1500
250 MHz
8.33 MHz
60
1342×1006
3000
500 MHz
8.33 MHz
100
1732×1299
5000
833 MHz
8.33 MHz
TABLE III: Scale ladder under the density-invariant convention.
Method
Type
Speed or solver
PF-Exec
proposed
learned gains
fixed-step
proposed executor
bandwidth 0.2, price 0.05, rest maximal
constant-speed
proposed executor
always maximal
greedy
untrained
top quarter demand centroid, even bandwidth
fixed-parameter
untrained
fixed position, even bandwidth
round-robin
untrained
99 to 1 rotation
TABLE IV: Comparison methods, all sharing one environment and one evaluation flow.
Source
UAVs
users
field
bandwidth, mobility
Cheng [ 36 ]
7
50
1000×1000 m
20 MHz per UAV, static
Jiang [ 34 ]
5
250
500×500 m
180 kHz per user, static
Rizvi [ 37 ]
3
6
1000×1000 m
15 kHz per UAV, linear 10 m/s
Shen [ 26 ]
3
20
2000×2000 m
10 MHz, static
TABLE V: Native settings of the published baselines, every parameter taken from the source.
Block
Content
records
native settings
four settings, proposed and original method, seeds and step horizons
60
main comparison
two scenarios, five scales, eleven methods, seeds and horizons
452
degradation line
two scenarios, five scales, three seeds, three horizons
90
TABLE VI: Composition of the evaluation records, 602 in total.
Fig. 4: Paired panels of the four indicators across the scale ladder, with the proposed method against the fixed-step executor. Seeds are marked separately. Fairness, dead count, coverage and load variance keep their ordering along the whole axis.
Fig. 5: Phase diagram under burst load, with mean load and the logarithm of load variance on the two axes, together with the envelope of the hundred-step peaks after the gathering at step 600. The trajectory slides along the balanced manifold and returns, and the envelope decays within a hundred steps.
Scenario
n
Method
J
uˉ
var u
dead
cov
flow
3
PF-Exec
0.9999–1.000
0.81–0.90
≤9.1e−5
0
≥0.994
fixed-step
0.999–1.000
0.81–0.89
≤7.3e−4
0
≥0.994
12
PF-Exec
0.9997–0.9999
0.80–0.94
≤3.0e−3
0
≥0.988
fixed-step
0.990–0.9998
0.80–0.93
≤8.6e−3
0
≥0.992
30
PF-Exec
0.994–0.9999
0.83–1.03
≤6.4e−3
0
≥0.997
fixed-step
0.9999
0.82
1.0e−4
0
0.999
TABLE VII: Main scenarios, end window at 4000 steps over three seeds. J is the fairness index, uˉ the mean load, var u the load variance, dead the window-mean count of dead UAVs and cov the coverage ratio. Ranges are over seeds.
flow, 12 UAVs
burst, 12 UAVs
burst, 100 UAVs
Method
J
uˉ
var u
J
uˉ
var u
J
uˉ
var u
PF-Exec
0.9999
0.80
4.8e−5
0.9999
0.80
6.7e−5
0.9987
1.21
1.9e−3
fixed-step
0.9998
0.80
1.1e−4
0.9997
0.80
1.7e−4
0.310
1.32
3.88
constant-speed
0.169
176
1.5e5
0.132
225
3.3e5
0.030
1.7e4
9.6e9
greedy
0.924
0.77
4.8e−2
0.967
0.68
1.6e−2
0.976
1.14
3.2e−2
fixed-parameter
0.871
0.79
9.2e−2
0.200
31.9
4.1e3
0.392
5.1e3
4.1e7
TABLE VIII: Comparison system at the end window of 4000 steps, seed 0. Dead counts are window means.
Fig. 6: Average rank over all common configurations, with rank spread as error bars. PF-Exec ranks first, the greedy heuristic and the block-coordinate method follow, and the random and constant-speed policies sit at the bottom.
Fig. 7: Report cards of the four native settings. Each card gives the median fairness of the proposed method and of the original method at the 2000, 3000 and 4000 step horizons, with the median variance on a logarithmic axis.
Fig. 8: Paired differences between the degradation line and the ideal-information line on identical grids. Differences concentrate below 10−3 and the largest, at one hundred UAVs under burst, is 0.002.
Scenario
Method
efly (J)
travel (m)
throughput (Mbps)
flow, 12
PF-Exec
4.623e6
7490
1068
fixed-step
4.623e6
7450
1076
flow, 100
PF-Exec
3.853e7
6.76e4
8728
fixed-step
3.854e7
6.78e4
8693
burst, 12
PF-Exec
4.617e6
5300
1036
fixed-step
4.617e6
5403
1042
TABLE IX: Flight energy, travel and throughput, seed 0 over the full 4000 steps.
This paper establishes a layered mathematical char acterisation of PF-Exec, the algorithm of the companion system paper. Demand-proportional bandwidth allocation yields an exact first-order homogeneous decomposition, from which the equal-load bandwidth target and its feasibility bound follow in closed form. The logarithmic knee view preserves strict load order and shares its global optimum with the Jain fairness index, while the demand-weighted centroid minimises the position potential. The damped softmax association block possesses a fixed point on the product of simplices. The truncated one-dimensional gain iteration admits a complete characterisation, with geomet ric convergence below the critical gain, initial-value-dependent bounded oscillation at it, and a family of bounded two-state cycles above it. A global invariant error ball and an explicit tracking recursion complete the execution guarantee. A numerical audit confirms the structural identities on every measured state and quantifies the local amplification of the coupled association and resource dynamics. Together, the results explain how closed form recovery directions and bounded adaptive gains provide stable execution without requiring a global Euclidean contraction assumption.
Hanzhang Jia, Liheng Zeng, Yiran Li +2
University of Sussex · Shanghai Shuyuan Information Technology Co., Ltd. · Zhejiang Gongshang University
Uncrewed aerial vehicle (UAV)-enabled federated learning (FL) can provide flexible, on-demand edge intelligence for large-scale IoT deployments, but operating in shared unlicensed bands makes uplink update delivery interference-coupled and unreliable. In this paper, we develop a packet-level transmission framework that captures buffer overflow, delay violations, and transmission errors, and uses the resulting packet delivery ratio (PDR) to represent partial-update reception through a packetized, Bernoulli-masked FL aggregation process. We then formulate a fairness-consensus bilevel (FCB) optimization that jointly controls (i) transmission thresholds to maximize the average PDR while reaching consensus under partial observability and (ii) transmission powers to improve the worst PDR and enforce fairness across IoT learners. To solve this problem, we propose an alternating FCB optimizer composed of a consensus-based threshold controller (CTC), which drives the IoT learners toward a PDR-efficient consensus on transmission thresholds, and a fairness-based power controller (FPC), which updates transmission powers to improve the worst PDR and ensure fairness under the resulting consensus thresholds. Numerical results on CNN-based FL tasks show that the FCB optimizer improves FL aggregation and training performance by enhancing packet-level update delivery, consistently outperforming baseline transmission policies.
Masoud Ghazikor, Zhou Ni, Morteza Hashemi
Department of Electrical Engineering and Computer Science, University of Kansas
This paper addresses multi-objective conflicts and catastrophic forgetting in uncrewed aerial vehicle (UAV) networks across dynamic spatiotemporal environments. Conventional multi-agent reinforcement learning (MARL) algorithms suffer from severe policy degradation during sequential task transitions. We propose a spatiotemporal continual federated learning (SCFL) framework driven by the group-decoupled multi-agent proximal policy optimization (G-MAPPO) algorithm. SCFL incorporates a three-stage geometric alignment mechanism: it resolves local gradient conflicts via group-decoupled policy optimization (GDPO), mitigates spatial Non-Independent and Identically Distributed (non-IID) client drift through adaptive cosine aggregation, and suppresses inter-task interference via global temporal orthogonal projection without raw experience replay. Evaluations show that SCFL achieves superior robustness over federated baselines, maintaining spatial service reliability above 0.95 and a load balancing index of approximately 0.95 during non-stationary transitions. A longitudinal self-degradation analysis further shows that SCFL preserves historical knowledge with near-zero performance variation in spatial reliability and QoS under moderate loads from 40 to 120 users, while revealing its operating boundary under extreme congestion with 140 users due to hard projection constraints. The framework provides a scalable, communication-efficient approach for autonomous aerial network orchestration.
Chuan-Chi Lai
Department of Communications Engineering, National Chung Cheng University, Minxiong Township, Chiayi County 621301, Taiwan · Advanced Institute of Manufacturing with High-tech Innovations (AIM-HI), National Chung Cheng University, Minxiong Township, Chiayi County 621301, Taiwan