WiSDoM: Wireless Sparse Decision Transformer with Mixture-of-Experts for Multi-Task Mobile Network Optimization
Fatih Temiz, Shavbo Salehi, Melike Erol-Kantarci
cs.NI · cs.AI · cs.LG
Quick verdict
MoE-boosted Decision Transformer for multi-task CoMP cell selection; solid sim gains (QoE 0.73 vs PPO 0.63) but no error bars, no code, no interference.
Summary
This paper proposes WiSDoM, an offline multi-task reinforcement learning framework for CoMP multi-cell selection. It augments a prompt-based Decision Transformer with token-level sparse Mixture-of-Experts routing: a PromptDT-15M backbone is trained across 36 simulated CoMP tasks (3-5 base stations, 3-8 UEs, three mobility levels, PF/ReF schedulers; trajectories from PPO experts in mobile-env); tasks are clustered into six groups by gradient-agreement similarity, one expert per transformer block per group is fine-tuned on the frozen backbone, and a top-K router with load-balancing loss is trained last. WiSDoM-75M with Top-1 routing (26.9M of 75.2M parameters active) achieves an average episode return of 7.25 versus 7.04 for dense PromptDT-75M, average QoE of 0.73 versus 0.63 for per-task PPO (+15.9%) and up to 55% above SINR-threshold heuristics (+9% versus the best-tuned one), and 23.4-36.0% lower rate-outage probability than PPO. On eight unseen configurations it beats task-specific PPO on five. Main limitations: no confidence intervals, seed counts or variance reporting; an interference-free simulation that sidesteps CoMP's core challenge; the 55% headline targets the weakest heuristic; no ablation against the MoE-DT and grouping methods it builds on; and no code release.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A single sparse MoE-augmented PromptDT trained jointly on 36 CoMP tasks outperforms dense multi-task DTs while activating about one-third of its parameters (return 7.25 vs 7.04; 26.9M of 75.2M active).
- WiSDoM improves average QoE to 0.73 versus 0.63 for per-task PPO (+15.87%) and up to 55% over SINR-threshold heuristics.
- It reduces rate-outage probability by 8.6-14.3% versus dense PromptDT-75M and 23.4-36.0% versus PPO, winning 32 of 36 task-wise comparisons.
- Gradient-agreement task grouping yields experts mixing specialization (experts 0 and 5) and generalist behavior (experts 1, 3, 4) without expert collapse.
- It adapts to eight unseen configurations via trajectory prompts without retraining, beating task-specific PPO on five of eight.
Strengths
Genuinely multi-axis task suite
36 tasks (Table II) vary BS count (3-5), UE count (3-8), velocity (0.5/1.5/3 m/s), and scheduler (PF/ReF) simultaneously, and the 8 unseen tasks (Fig. 7) extrapolate along new UE counts and velocities (1.0-2.5 m/s) — far stronger than the single-configuration evals typical of wireless RL papers.
Well-specified four-stage pipeline
Algs. 1-2 and Table III give enough detail to reimplement: 2000 PPO-expert episodes/task, gradient-agreement clustering into six groups (HarmonyDT-style, by optimization behavior rather than config labels), STE for Top-K routing, load-balancing loss (Eq. 11, lambda_aux=0.01), invalid-action masking over zero-padded spaces, and single-RTX-5090 training scale.
Evaluation links RL return to network KPIs
Efficiency accounting is internally consistent
Candid at the edges
Weaknesses
No uncertainty quantification anywhere
Interference-free simulation for a CoMP paper
Selective headline framing
Missing baselines the method is built from
Unseen-task protocol has a test-time selection hole
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
WiSDoM — Structured Preliminary Review
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is WiSDoM in wireless networks?
WiSDoM (arXiv 2609.00284, 2026) is a method that combines a prompt-based Decision Transformer — an offline reinforcement-learning sequence model — with sparse Mixture-of-Experts routing, so one policy handles CoMP multi-cell selection across 36 simulated network configurations while activating only about a third of its parameters per inference. New configurations are handled by feeding short trajectory demonstrations (prompts) instead of retraining.
How much does WiSDoM improve QoE compared to baselines?
In the paper's mobile-env simulations, WiSDoM reaches an average QoE of 0.73 versus 0.63 for per-task PPO agents, a 15.9% gain. The advertised 'up to 55%' is measured against the weakest SINR-threshold heuristic (QoE 0.47); against the best-tuned heuristic (QoE 0.67) the improvement is about 9%. It also cuts rate-outage probability by 23.4-36.0% versus PPO.
Is WiSDoM's code or data available?
No code or data release is mentioned in the paper. The underlying environment (mobile-env) is open source and hyperparameters are tabulated (Table III), but no random seeds, confidence intervals, or variance figures are reported, so exact numbers would be hard to reproduce bit-for-bit.
Does WiSDoM work on real 6G networks?
Not demonstrated. All results are simulation-only, in an environment that neglects inter-cell interference — the main problem CoMP transmission exists to solve — and inference was measured on a consumer GPU (8-20 ms per decision). The authors list O-RAN integration and robustness beyond trained dimensions as future work.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Numerical series values of Figs. 2-9 appear as flattened number streams in the extracted text; value-to-model attributions (e.g., Fig. 3's 0.47 to the weakest heuristic) were relied on only where the prose explicitly states them.
- The abstract in the supplied metadata cuts off mid-sentence, but the complete abstract appears at the head of the extracted full text, so no content was lost.
- The body runs complete through the conclusion and 49 references; no end-of-text truncation was detected.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.