5g Madrl Sumrate Eval

5g-madrl-sumrate-eval

Dual-Attention Deep Reinforcement Learning for Multi-MAP 3D Trajectory Optimization in Dynamic 5G Networks — Catté et al. (2023) (arXiv:2303.05233, 2023)

What this evaluates

Evaluates the ability of a multi-agent deep reinforcement learning framework to optimize the 3D placement and trajectory of mobile access points in dynamic 5G networks, balancing sum-rate maximization against user mobility and interference.

Datasets

• Custom 5G Network Simulation — total ?; splits: train (-1), test (-1)

Metrics

• sum-rate (primary) — range: Mbps
  • Total achievable data rate across all connected users and access points at each time step, aggregated over the simulation horizon. Computed using Nakagami fast-fading channel models with specified path loss and bandwidth parameters.

Input / output format

Input: State vectors for each MAP and UE (position, mobility, traffic demand), processed through dual-attention message encoders and MLPs to form agent observations.

Output: 3D position/trajectory updates for each of the M=3 MAPs per time step.

Scoring recipe

# Evaluate average instantaneous sum-rate over test horizon T
total_rate = 0.0
for t in range(T):
    for ue in UEs:
        sinr = compute_sinr(ue, maps, interference)
        rate = ue.bandwidth * log2(1 + sinr)
        total_rate += rate
return total_rate / T  # Average over time slots
# Results averaged over 50 random deployments

Common pitfalls

• The centralized benchmark's performance heavily depends on the clustering update period τ_c; comparing it without fixing τ_c leads to unfair baselines.
• Testing uses random waypoint mobility at 0.8 m/s, not the static mobility used during training, which tests generalization to unseen dynamics.
• Results are averaged over 50 random deployments, not single runs, to account for stochastic UE placement and fading.

Evidence (verbatim from paper)

We perform the training of our proposed MADRL algorithm for 10000 Monte-Carlo runs. For each run, we deploy M=3 MAPs moving with a step size Δℓ=5 m; we also deploy K=25 UEs in M centroids of radius 25 m, randomly sampled in a 200 m by 200 m area. During the training phase, the UEs are static for Te=300 iterations while they follow a random way-point centroid mobility at 0.8m/s during the testing phase. ... Figure 4 shows the instantaneous sum-rate comparison between our solution and the benchmark for static and mobile UEs over time. Results are averaged over 50 random deployments.

Citation

@misc{catte2023dualattention,
  title={Dual-Attention Deep Reinforcement Learning for Multi-MAP 3D Trajectory Optimization in Dynamic 5G Networks},
  author={Catté et al. (2023)},
  year={2023},
  note={arXiv:2303.05233}
}

• arXiv: 2303.05233