PaperlayerView exact revision on arXiv

Source-linked AI summary

WiSDoM: Wireless Sparse Decision Transformer with Mixture-of-Experts for Multi-Task Mobile Network Optimization

Fatih Temiz, Shavbo Salehi, Melike Erol-Kantarci

arXiv:2609.00284v1cs.NIcs.AIcs.LG

TL;DR

Heterogeneous wireless environments make it difficult for a single offline-RL policy to perform consistently across CoMP tasks because shared models face conflicting task requirements. WiSDoM combines a Decision Transformer with sparse Mixture-of-Experts routing, jointly trains across diverse configurations, and reports higher QoE with efficient inference and few-shot adaptation.

  • Problem

    Offline-RL policies struggle to scale consistently across heterogeneous CoMP environments because conventional shared models can suffer conflicting task gradients and limited specialization.

  • Method

    WiSDoM jointly trains a sparse MoE-DT policy across 36 heterogeneous wireless configurations, using routing and trajectory prompts for task-conditioned decisions.

  • Results

    WiSDoM consistently outperforms heuristic, single-task, and conventional multitask baselines, achieving up to 55% higher QoE while activating as few as one-third of dense-model parameters during inference.

  • Takeaways & Limitations

    Sparse expert routing enables one WiSDoM policy to specialize across heterogeneous wireless tasks while supporting adaptation to unseen configurations without retraining or fine-tuning.

Abstract

from arXiv · show

Emerging 6G wireless networks are expected to operate across diverse deployment scenarios, where variations in network topology, user mobility, traffic demand, and radio conditions challenge the scalability of conventional radio resource management (RRM). While offline reinforcement learning (RL) methods have demonstrated strong decision-making capabilities, learning a single policy that performs consistently across heterogeneous wireless environments remains difficult due to conflicting optimization objectives and limited model specialization. These challenges become particularly pronounced in coordinated multipoint (CoMP) transmission, where selecting the optimal serving-cell combination requires sequential decision-making under evolving network conditions. This paper presents the Wireless Sparse Decision Transformer with Mixture of Experts (WiSDoM), a sparse multi-task offline RL framework for adaptive multi-cell selection. WiSDoM combines Decision Transformers (DTs) with a Mixture-of-Experts (MoE) architecture that dynamically activates specialized experts according to task characteristics. This MoE mechanism improves model capacity without proportionally increasing inference cost, mitigates negative transfer, and enables expert specialization across tasks. WiSDoM is trained jointly on diverse network configurations spanning multiple base station and user equipment densities, mobility levels, and scheduler policies. Experimental results show that WiSDoM consistently outperforms heuristic methods, single-task models, and conventional multi-task DTs, improving quality of experience (QoE) by up to 55% while activating approximately one-third of the parameters of its dense counterpart during inference. Furthermore, WiSDoM exhibits strong task generalization and efficiently adapts to unseen wireless scenarios through few-shot prompting without retraining or fine-tuning.

I. INTRODUCTION

WiSDoM addresses the difficulty of scaling offline decision-making across heterogeneous CoMP wireless environments by combining a Decision Transformer with sparsely activated specialized experts. It is jointly trained across diverse configurations and supports few-shot adaptation, with reported QoE gains and reduced active parameters.

  • Motivation: 6G wireless networks require RRM that adapts to diverse devices, mobility patterns, services, and changing channel conditions.AI-native RRM is presented as a data-driven alternative to conventional optimization and rigid rule-based control.
  • Motivation: CoMP coordinates multiple base stations to improve spatial diversity, mitigate inter-cell interference, extend coverage, and enhance reliability.The paper frames adaptive serving-cell selection as a sequential decision problem in dense and heterogeneous deployments.
  • Research gap: Conventional DRL approaches can outperform heuristics in selected settings but scale poorly, while dense PromptDT sharing creates conflicting task gradients and negative transfer.Increasing shared-model size adds capacity without explicitly separating task-specific knowledge and also increases computational cost.
  • Approach: WiSDoM jointly trains a sparse MoE-DT framework across 36 configurations varying UE density, BS deployment, mobility, and scheduling policies.Its router dynamically directs trajectory tokens to specialized experts, increasing capacity and parameter efficiency for multi-discrete actions.
  • Generalization: WiSDoM adapts to previously unseen tasks and network conditions through task-specific trajectory prompts without retraining or fine-tuning.This extends the framework beyond independently trained fixed-configuration policies and single-configuration DT studies.
  • Results: 55% higher QoE is reported over heuristic, single-task, and conventional multitask baselines while inference activates as few as one-third of dense-model parameters.The evaluation is described as extensive and spans heterogeneous CoMP tasks.

B. MoE for Scalable Sequence Modeling

The paper motivates MoE as a way to scale sequence-model capacity through conditional computation, while relating wireless applications to task specialization and efficient inference. WiSDoM is positioned as combining sparse routing with PromptDT-based offline RL across diverse wireless configurations.

  • Scalable sequence modeling: MoE routers activate a small subset of specialized expert networks per input token instead of uniformly activating all dense-transformer parameters.This provides conditional computation as model capacity grows.
  • Wireless applications: Wireless applications of MoE target increased capacity, sparse inference, and dynamic feature learning for time-varying tasks such as modulation classification, interference cancellation, and beamforming.The cited review also discusses MoE integration with generative AI and RL across wireless communication tasks.
  • Related approaches: Prior wireless MoE approaches include independently trained experts combined by gating, including DRL experts adapted to unseen SLA weight vectors.These approaches differ from jointly learning a single internally routed model.
  • Related approaches: An LLM-based approach externally orchestrates independently pretrained RL experts using textual QoS requirements, whereas WiSDoM uses MoE internally within one model.The external approach incurs LLM inference overhead and relies on natural-language task specifications.
  • WiSDoM positioning: WiSDoM is presented as combining sparse MoE routing with PromptDT-based offline RL for wireless communications, jointly learning across diverse configurations and supporting prompt-based adaptation.The paper identifies CoMP as the particularly targeted wireless domain.

III. SYSTEM MODEL AND PROBLEM FORMULATION

The system model represents downlink CoMP as time-varying UE–BS association with adjustable network sizes, mobility, synchronized BS transmission, and alternative scheduling policies. The optimization maximizes bounded long-term QoE subject to SNR and connectivity constraints.

  • System model: The downlink CoMP system contains M UEs and N BSs operating over discrete time steps in a bounded two-dimensional area.Both M and N are adjustable, and UE mobility follows a random waypoint model with constant velocity between random destinations.
  • System model: The propagation model uses carrier frequency and BS and UE antenna heights to compute a mobile-height correction factor and received SNR.The received SNR is computed from path loss together with transmit power and the noise floor.
  • System model: Multiple BSs may simultaneously serve one UE, enabling CoMP transmission under resource-fair or proportional-fair PRB scheduling.Resource-fair scheduling divides PRBs equally, whereas proportional-fair scheduling balances instantaneous rates with historical throughput.
  • System model: C_j(t) denotes the set of BSs serving UE u_j at time t.This serving set determines the UE’s aggregated downlink data rate.
  • Problem formulation: The UE utility maps aggregated downlink rate D_j(t) to a bounded QoE range using configurable scaling coefficients and lower and upper utility bounds.Clipping keeps utility finite and represents diminishing returns in perceived satisfaction as data rate increases.
  • Problem formulation: The association objective selects UE–BS connectivity decisions over time to maximize long-term average QoE across active UEs.The optimization is subject to connectivity and SNR constraints.
  • Problem formulation: A UE–BS connection is permitted only when the received SNR exceeds the minimum threshold.The formulation also limits each UE’s simultaneous connections to the available number of BSs.

B. Problem Formulation

The multi-cell selection problem is modeled as an MDP whose state represents active-UE connectivity, channel quality, and service utility, while actions toggle at most one connection per UE and rewards capture average QoE.

  • The stochastic evolution of user mobility and channel conditions motivates modeling multi-cell selection as a Markov Decision Process.
  • Each system state captures downlink connectivity, channel quality, and service utility for all active UEs.For M(t) active UEs and N base stations, the state is defined over those per-UE quantities.
  • Each UE can modify at most one connection per time step to limit protocol overhead and signaling complexity.
  • An action toggles a UE's connection with one base station when the SNR constraint is satisfied, or performs no operation.The joint action contains the per-UE actions at time t.
  • The reward is the average utility of all active UEs and represents overall system QoE.The formulation encourages maximizing long-term average QoE across users.

IV. WISDOM FOR MULTI-TASK LEARNING IN COMP

WiSDoM addresses the difficulty of generalizing offline sequence-modeling policies across heterogeneous CoMP tasks by combining task prompts with specialized mixture-of-experts modules.

  • Multi-task RL learns a shared policy across tasks with potentially different state and action spaces, dynamics, and reward structures.Wireless CoMP heterogeneity includes varying BS and UE counts, mobility, scheduler policies, and network topologies.
  • Offline RL learns policies from fixed trajectory datasets without online interaction, while Decision Transformers predict actions from RTG, state, and action sequences.
  • PromptDT prepends short task-identifying demonstrations to history segments, enabling few-shot generalization to unseen tasks without retraining.
  • A single shared backbone can suffer gradient conflicts as task heterogeneity grows because tasks compete for the same parameters.The proposed MoE augmentation assigns dedicated expert modules to task-specific specialization while retaining shared cross-task knowledge.

A. WiSDoM Framework

WiSDoM uses a four-step framework that trains a shared PromptDT backbone, groups tasks by gradient behavior, specializes experts for those groups, and then combines the resulting components.

  • The WiSDoM framework is organized into four main steps described in Fig. 1 and Algorithms 1 and 2.
  • The offline setting supplies trajectories and demonstrations for training across multiple tasks.Algorithm 1 requires training tasks, task groups, a PromptDT backbone, offline datasets, demonstrations, and training iterations.
  • The shared backbone is trained jointly across tasks with prompted trajectories and masked cross-entropy action prediction.
  • After gradient-conflict analysis groups similar tasks, one task-specific feedforward expert is inserted into each transformer block for every group.
  • The trained backbone and experts form the final specialized model components for inference.Algorithm 1 returns the backbone and the set of experts.

1) Trajectory Collection:

WiSDoM collects heterogeneous CoMP trajectories and trains a shared prompted backbone before independently specializing group-specific experts, while accounting for variable state and action dimensions.

  • 1) Trajectory Collection:: Offline trajectories are generated by expert PPO policies across diverse CoMP environments and may contain successful and suboptimal interactions.
  • 1) Trajectory Collection:: Variable BS and UE counts produce different state and action dimensions across tasks.The tasks span network scale, UE velocity, and ReF/PF scheduler policy variations.
  • 1) Trajectory Collection:: Zero-padding aligns variable dimensions to the maximum training-task sizes, while invalid padded actions are masked during training and inference.
  • 1) Trajectory Collection:: The shared PromptDT backbone is trained jointly by pairing sampled offline trajectories with task-identifying prompts and updating masked cross-entropy loss.
  • 1) Trajectory Collection:: Task groups are formed from inter-task gradient behavior, and each frozen-backbone group receives independently trained block-specific experts.Each expert specializes on its task subset without interference from other groups.

3) Task Grouping:

WiSDoM groups tasks by gradient alignment so its experts specialize in related learning signals rather than only similar network configurations. The six-expert design balances specialization against routing and training overhead.

  • 3) Task Grouping:: Six task groups are formed by clustering gradient-cosine similarities, aligning each expert with tasks that have similar gradients.The grouping partitions T1 through T36 into G1 through G6.
  • 3) Task Grouping:: WiSDoM uses six experts to reduce the learning burden on each parameter subset and improve task scalability.The choice is motivated by prior evidence that more experts can improve scalability through parameter specialization.
  • 3) Task Grouping:: Larger expert counts could improve specialization but would increase routing complexity and training cost.The selected count is presented as a practical trade-off between specialization and computational overhead.

4) Router Training and Inference:

The router is trained after the backbone and experts are frozen, selecting a sparse set of experts for each token and using auxiliary balancing losses. During evaluation, the routed model generates actions from task prompts and interaction history in closed loop.

  • 4) Router Training and Inference:: The router is introduced at each transformer block and trained on multi-task batches while the backbone and experts remain frozen.Only the router parameters are updated during this stage.
  • 4) Router Training and Inference:: For each token, the router computes expert probabilities, selects the top-Kr experts, and combines their weighted outputs with the frozen FFN path.A straight-through estimator supports training through the non-differentiable sparse selection.
  • 4) Router Training and Inference:: The auxiliary load-balancing loss encourages tokens to be distributed across experts rather than concentrated in a small subset.The loss is averaged across transformer blocks and combined with masked cross-entropy.
  • 4) Router Training and Inference:: During evaluation, sparse routing is retained and actions are generated in closed loop from the best-performing prompt, trajectory history, and updated return-to-go.The evaluation procedure updates the trajectory and RTG after each environment interaction.
  • 4) Router Training and Inference:: The experiments use the 36-task configuration suite and simulation and model settings reported in Tables II and III.The supplied passages identify the table titles but do not provide their individual entries.

V. SIMULATION SETTINGS AND NUMERICAL RESULTS

The evaluation covers 36 heterogeneous CoMP tasks, held-out unseen tasks, and multiple baselines across performance, generalization, and complexity axes. WiSDoM improves episode return and QoE while using sparse inference and a single shared model.

  • V. SIMULATION SETTINGS AND NUMERICAL RESULTS: The study constructs 36 tasks varying base-station count, user-equipment count, UE velocity, and scheduler policy using the mobile-env simulator.All models were trained on a single NVIDIA RTX 5090 GPU.
  • V. SIMULATION SETTINGS AND NUMERICAL RESULTS: All baselines are evaluated on the same 36-task suite and held-out unseen tasks using identical closed-loop rollouts of 50 episodes per task.The unseen set includes, for example, eight additional tasks.
  • V. SIMULATION SETTINGS AND NUMERICAL RESULTS: WiSDoM uses Top-1 or Top-2 token-level routing over task-group experts, with 75M total parameters and 27M or 37M active parameters.The two variants are named WiSDoM-75M-A27M and WiSDoM-75M-A37M.
  • V. SIMULATION SETTINGS AND NUMERICAL RESULTS: The evaluation compares methods across RL performance, network performance, task-wise consistency, unseen-task generalization, and inference complexity.Complexity includes parameter count, active parameters per forward pass, and inference latency.
  • V. SIMULATION SETTINGS AND NUMERICAL RESULTS: 7.25 average episode return is achieved by WiSDoM-75M-A27M, exceeding dense PromptDT-75M at 7.04 while using up to 3x fewer active parameters.WiSDoM-75M-A37M achieves 7.11 under Top-2 routing.
  • V. SIMULATION SETTINGS AND NUMERICAL RESULTS: 15.87% QoE improvement is reported over PPO, increasing users’ QoE from 0.63 to 0.73; WiSDoM also exceeds heuristic schedulers by up to 55%.The comparison spans the evaluated task suite and includes heuristic, single-task, and multi-task baselines.

2) Network Performance:

WiSDoM improves QoE and rate reliability across diverse network configurations, while few-shot prompting enables transfer to unseen tasks without target-task fine-tuning.

  • Performance differences among learning-based methods are not explained by serving more UEs or establishing substantially more CoMP links.Across configurations, connected UEs and established links are relatively similar among learning-based approaches; association and resource utilization better explain QoE differences.
  • WiSDoM achieves a QoE of 0.60 in the 3BS-5UE configuration, a 27.7% improvement over PPO’s 0.47.
  • WiSDoM-75M-A27M reduces outage probability by 8.6–14.3% versus PromptDT-75M and by 23.4–36.0% versus PPO across four thresholds.
  • WiSDoM achieves lower outage probability than PPO in 32 of 36 tasks, including all 18 tasks using PF scheduling.At 3 m/s with PF scheduling, outage falls from 6.22% to 0.16% in the 3BS-5UE configuration.
  • Using demonstration trajectories without parameter updates or fine-tuning, WiSDoM outperforms task-specific PPO on five of eight unseen tasks.On 5BS-7UE, WiSDoM achieves a return of 3.4 versus 0.2 for PPO and 1.3 for PromptDT-75M.

C. Complexity Analysis and Ablation Study

WiSDoM uses sparse expert activation to reduce inference computation while retaining performance, with expert selection showing task-group specialization and shared knowledge.

  • Complexity Analysis: WiSDoM-75M-A27M activates 26.9M parameters and WiSDoM-75M-A37M activates 36.6M, compared with 76.3M for dense PromptDT-75M.
  • Complexity Analysis: WiSDoM-75M-A27M requires 4.044 GFLOPs per inference step, or 35% of the dense model’s 11.47 GFLOPs.WiSDoM-75M-A37M requires 5.490 GFLOPs, or 48% of the dense model.
  • Complexity Analysis: Sparse activation reduces arithmetic computation while maintaining and often improving task performance, supporting a performance–efficiency tradeoff for resource-constrained near-real-time RIC deployment.
  • Ablation Study: Experts 0 and 5 exhibit task-dependent specialization, while the expert-selection analysis also observes shared knowledge across task groups.Mean selection distributions are computed across all six transformer blocks, with each task-group row summing to 1.
  • Ablation Study: Across 36 CoMP tasks, WiSDoM consistently outperforms heuristic, single-task, and conventional multi-task baselines in QoE and rate outage probability, with up to 55% higher QoE.The framework also generalizes to unseen configurations through few-shot adaptation without retraining or fine-tuning.
Loading 2609.00284v1…