A Semantic-Aware Multiple Access Scheme Leveraging Spatial Redundancy for Uplink-Dominant Network Services

Hamidreza Mazandarani; Masoud Shokrnezhad; Tarik Taleb

Source-linked AI summary

A Semantic-Aware Multiple Access Scheme Leveraging Spatial Redundancy for Uplink-Dominant Network Services

Hamidreza Mazandarani, Masoud Shokrnezhad, Tarik Taleb

arXiv:2609.03559v1Sep 3, 2026cs.NI

TL;DR

The paper addresses the underexplored integration of semantic intelligence into distributed MAC, especially when users share semantic information. It formulates fairness and energy-efficiency problems with variable packet lengths and develops PRISM, a MADRL-based decentralized access scheme. PRISM outperforms semantic-oblivious approaches and approaches optimal performance across many association settings, while the authors identify changing association matrices as a key scope boundary.

Problem
Semantic-aware communication has advanced beyond the physical layer, but semantic intelligence in MAC and distributed access with correlated user data remain underexplored.
Method
The paper formulates variable-packet-length optimization problems for α-fairness and energy efficiency, then develops PRISM using centralized training and decentralized execution with semantic-aware observations and rewards.
Results
PRISM outperforms random and PRISM-lite approaches, approaches optimal performance in many scenarios, and exceeds orthogonal upper bounds when semantic sharing is high.
Takeaways & Limitations
Semantic intelligence in distributed MAC can improve network sustainability by exploiting correlations among users’ source data.
Takeaways & Limitations
The evaluation assumes stationary user-segment associations; changing associations require online policy adaptation, such as through continual learning.

Abstract

from arXiv · show

The transition toward semantic-aware communication offers a paradigm shift for next-generation mobile networks, promising to decouple information significance from raw data transmission. Despite advances in semantic extraction, the integration of semantic intelligence into the Medium Access Control (MAC) layer remains underexplored, particularly in exploiting spatial correlations among users. To address this, we introduce a novel multiple access scheme designed for uplink-dominant network services. This framework optimizes the trade-off between spectrum utilization and sustainability by formulating variable-packet-length access as distinct $α$-fairness and energy efficiency problems. A key innovation of our approach is the quantification of spatial redundancies through novel metrics of self-throughput and assisted-throughput, which account for the semantic correlation of data across user equipment. We analyze these formulations to identify optimal bounds before proposing PRISM (Protocol for Redundancy Identification in Semantic Multiple-access). Grounded in Model-free Multi-Agent Deep Reinforcement Learning (MADRL), PRISM enables devices to autonomously govern spectrum access using only local observations. Extensive evaluations demonstrate that PRISM successfully leverages redundancies to outperform semantic-oblivious schemes, achieving up to ${90\%}$ of the centralized optimal benchmark and improving both objectives by up to ${2\times}$ across diverse user-semantic association matrices. These results validate PRISM as a viable candidate for future distributed mobile network applications, complemented by orthogonal Multiple Access Schemes where signals are multiplexed in the semantic domain.

I. INTRODUCTION

Semantic-aware communication shifts wireless networking from maximizing transmitted bits toward maximizing semantic efficiency and sustainable resource use. This paper applies that shift to distributed MAC by modeling shared semantic information, fairness, and energy efficiency in uplink access.

From bits to semantics: Semantic-aware communication replaces bit-oriented resource expansion and efficiency improvements with semantic efficiency as a basis for sustainable networking.The motivation includes 6G services such as the Metaverse, where efficient and equitable resource use is important.
Spatial redundancy: Overlapping UAV observations can contain shared semantic segments, allowing redundant transmissions to be avoided while preserving cognitive performance.Vehicles in overlapping observations illustrate how semantic redundancy can reduce bandwidth and energy use.
Research gap: MAC-layer semantic awareness remains underexplored, while existing joint multi-user semantic methods can lack adaptability, scalability, or distributed operation for 6G systems.MADRL is identified as a potential way to support decisions among multiple evolving actors.
Semantic interdependence: The framework treats users’ utilities as interdependent because one transmission can benefit other users sharing relevant semantic segments.This interdependence is connected to holographic presence, distributed GenAI, and multi-UAV coverage.
Optimization objectives: The paper formulates variable-packet-length access problems for both α-fairness and energy efficiency, including semantic correlations and encoding-energy costs.The fairness objective uses α-fairness metrics, while energy efficiency accounts for semantic encoding consumption.
Practical solution: PRISM uses MADRL with semantic- and asynchrony-tailored observations and rewards to provide a practical solution when optimal formulations are computationally infeasible.The method uses fully connected approximation layers, history-based context, and objective-specific reward shaping.

C. MADRL: An enabler for semantic-aware Multiple Access

The paper motivates MADRL for decentralized semantic-aware multiple access, where shared information, energy sustainability, and variable packet access shape the MAC problem.

C. MADRL: An enabler for semantic-aware Multiple Access: MADRL suits decentralized access because agents adapt to the non-stationarity caused by co-evolving policies.
C. MADRL: An enabler for semantic-aware Multiple Access: Energy efficiency is treated as a critical objective because semantic learning and extraction can add sustainability costs.
C. MADRL: An enabler for semantic-aware Multiple Access: The framework targets shared semantic information rather than independent user utilities, covering applications such as holographic presence, distributed GenAI, and multi-UAV coverage.
C. MADRL: An enabler for semantic-aware Multiple Access: The system models N UEs contending for C slotted uplink channels, where simultaneous transmissions on one channel collide.
C. MADRL: An enabler for semantic-aware Multiple Access: Users are associated with semantic segments through a binary matrix that an orchestration entity generates using service-specific matching rules.

1) Semantic Spatial Correlations:

The framework measures semantic throughput by separating each UE’s direct contribution from assistance provided when other UEs transmit shared segments.

1) Semantic Spatial Correlations:: Semantic throughput is evaluated with a segment transmission indicator that counts whether a segment is delivered during a time slot.
1) Semantic Spatial Correlations:: The normalized throughput of each UE is decomposed into self-throughput and assisted-throughput components.
1) Semantic Spatial Correlations:: Assisted throughput captures other UEs’ contribution when they transmit segments shared with UE i, while duplicate deliveries count only once.
2) Correlated Utilities:: Because users can share semantic segments, allocating resources to one UE can increase another UE’s utility.
2) Correlated Utilities:: The framework applies to contention-based many-to-one uplinks with local feedback, persistent semantic overlap, and duplicate suppression.

3) Fairness-utilization trade-off:

The paper formulates fairness-utilization and energy-efficiency objectives that balance aggregate semantic throughput, user equality, and device energy costs.

3) Fairness-utilization trade-off:: The α-fairness objective spans sum-throughput maximization as α →0 and max-min fairness as α →∞.
3) Fairness-utilization trade-off:: α therefore controls whether access favors aggregate semantic throughput or limits persistent starvation and energy concentration among individual UEs.
4) Energy Awareness:: The energy-efficiency utility subtracts a coefficient-weighted energy cost from throughput, including packet transmission and semantic encoder-inference energy.
4) Energy Awareness:: Semantic energy modeling includes transmission, inference, and encoder-initialization components, whose relative importance affects user utilities and behavior.
4) Energy Awareness:: As ζ increases, energy consumption weighs more heavily and inactive users avoid negative objective values by selecting zero utility.

5) Practical Challenges of Semantic Orchestration:

The paper addresses practical orchestration and centralized-benchmark issues by defining semantic associations, variable-length packet decisions, and channel occupancy over time.

5) Practical Challenges of Semantic Orchestration:: Semantic segment granularity depends on the downstream task, and KB-MANO can support construction of the UE–segment association matrix.
5) Practical Challenges of Semantic Orchestration:: When the base station lacks users’ raw data, UEs can provide semantic metadata such as identifiers, hashes, embeddings, or confidence scores.
5) Practical Challenges of Semantic Orchestration:: The centralized MINLP benchmark uses system-wide information to schedule packet starts, lengths, and channels, excluding collisions by construction.
5) Practical Challenges of Semantic Orchestration:: The packet-start variable records packet size by UE, channel, and time, while auxiliary variables track remaining duration and occupancy.
5) Practical Challenges of Semantic Orchestration:: An occupancy indicator remains one across all slots covered by an ongoing packet and supports constraints preventing simultaneous channel occupation.

2) Temporal Interdependency Constraints:

The constraints coordinate variable-length transmissions over time by preventing overlap, linking packet length to auxiliary variables, and enforcing channel-access limits.

Temporal Interdependency Constraints: Transmission decisions remain inactive during the subsequent k −1 time slots after a packet of length k starts.
Temporal Interdependency Constraints: The packet-length variable r equals z when transmission starts, with linearized forms used to represent this relationship.
Temporal Interdependency Constraints: A single time-slot example makes r, z, m, and d equal binary variables, illustrating the auxiliary-variable relationships.
Temporal Interdependency Constraints: Each UE may use at most one channel per slot, while each channel admits no more than one qualifying simultaneous transmission.

5) Problems:

The paper formulates nonlinear α-fairness and energy-efficiency optimization problems for semantically correlated, variable-length transmissions, then studies their computational limits and optimal solutions.

5) Problems: The α-fairness and energy-efficiency problems remain intractable because their integer and binary variables create exponentially large solution spaces.The stated solution space becomes impractical for large numbers of users, channels, packet sizes, or network lifetime.
C. Optimal Solutions: Increasing shared semantic segments improves both objective functions, although the magnitude depends on the objective and its parameters.
C. Optimal Solutions: A 0.58 correlation links objective-improvement ratios between A4,3 and A4,1 with throughput unfairness.
C. Optimal Solutions: Higher α produces more equal throughputs, while including transcoding costs changes the energy-efficiency optimum toward larger packets.Average packet size is 5 for d1 = 1 and 2.5 for d1 = 0.

IV. PRISM

PRISM converts centralized semantic-aware allocation objectives into decentralized online MAC decisions using one reinforcement-learning agent per UE and variable-duration macro-actions.

IV. PRISM: Macro-actions represent variable-duration sensing or packet transmissions, matching the multi-slot decisions in the MacDec-POMDP formulation.
IV. PRISM: PRISM assigns one DRL agent to each UE and exposes self and assisted transmissions while shaping rewards for fairness-utilization or energy-efficiency objectives.
IV. PRISM: Only UEs that have completed their current macro-actions participate in the next joint action, state, and reward calculations.
1) Action Space: Each active UE can sense a channel with (0, ci) or transmit a packet of length ri > 0 on channel ci.
2) State Space: The state combines normalized delay information, local sensing or transmission outcomes, self and assisted transmissions, recent histories, and longer-term averaged context.

3) Self and Assisted Transmissions:

Self and assisted transmissions quantify how much a UE contributes uniquely and how much of its semantic content is delivered by other UEs.

3) Self and Assisted Transmissions: A UE’s self-transmission value reflects its own involvement in delivering related segments, with examples updating values after successful or unsuccessful transmissions.
3) Self and Assisted Transmissions: Assisted transmission measures the normalized contribution of a UE’s segments transmitted successfully by other UEs.
3) Self and Assisted Transmissions: The metrics account for asynchronous shared-segment delivery by decoding only the largest successful packet containing a segment and using segment-level D2LT.
3) Self and Assisted Transmissions: For two UEs sharing a segment, successful packets of sizes 4 and 3 with segment D2LT 2 yield only 2 uniquely contributing time slots.
3) Self and Assisted Transmissions: The policy estimates semantic contribution from local history and lightweight downlink feedback rather than receiving neighbors’ data, actions, or association rows.

4) Reward:

The cooperative reward averages user rewards while penalizing unsuccessful actions, constraint violations, and strategy-dependent costs. Spatial semantic redundancy provides assisted throughput without retransmission, and explicit retransmission is deferred.

Reward: The cooperative system reward is defined as the average of user rewards based on each action and resulting system state.Successful actions receive normalized reward, while collisions or unsuccessful actions receive a negative normalized reward and inactive actions receive none.
Reward: Collision penalties trigger a fresh action decision rather than fixed retransmission, allowing a UE to sense, switch channels, or shorten its packet.The retry timing, channel, and packet length are learned from the policy rather than imposed by an ARQ timer or backoff rule.
Reward: Spatial semantic redundancy acts as an implicit repetition code across UEs because a segment is delivered whenever any associated UE succeeds.This mechanism underlies assisted throughput and differs from retransmitting the same UE’s packet across time.
Reward: Explicit retransmission with soft combining would add temporal redundancy but is deferred to future work.

5) Training Process:

PRISM uses centralized training with decentralized execution, evaluating semantic-aware and semantic-oblivious policies against centralized, random, and orthogonal references. Across fairness experiments, PRISM approaches optimal performance and benefits from higher semantic sharing.

Training Process: PRISM follows centralized training and decentralized execution, using global state at the SBS during learning and local information for UE decisions.Its methodology lies between VDN and QPLEX, with centralized value estimation supporting decentralized policies.
Training Process: The D3QL backbone decomposes each UE’s Q-value into state-value and action-advantage components to improve learning stability and performance.The decomposition is expressed as Q_i = V_i + adv_i.
Training Process: The experiments use synthetic stress-test association matrices rather than community benchmarks, with fixed association statistics during each run.Within-run adaptation to changing associations is left for future work.
Training Process: The evaluation compares PRISM with PRISM-lite, a semantic-oblivious CTDE policy, plus CB, RND, and ORTH reference schemes.CB is a model-specific centralized benchmark, while ORTH is an idealized semantic-orthogonality reference rather than a measured implementation or formal upper bound.
Training Process: PRISM outperforms random and PRISM-lite approaches, approaching optimal performance across α values and association matrices.The aggregate comparison is reported for experiments A and B, with objective values averaged over the final 1000 time slots and five simulation rounds.
Training Process: Higher semantic sharing improves both self and assisted throughputs, and PRISM can exceed the idealized orthogonal reference when shared semantics are prevalent.ORTH is more competitive at low sharing because its non-overlap condition is then easier to satisfy.

B. Energy Efficiency Maximization

PRISM is evaluated for energy efficiency under semantic transcoding costs and for sensitivity to network size. It approaches optimal energy-efficiency performance, remains the strongest decentralized method as users or channels vary, and is limited by stationary association assumptions and untested cross-objective evaluation.

B. Energy Efficiency Maximization: PRISM outperforms random and PRISM-lite approaches and approaches optimal energy-efficiency performance with and without semantic transcoding costs.The energy-efficiency experiment sets d1 to 0 to exclude transcoding costs and 1 to include them.
B. Energy Efficiency Maximization: Higher semantic sharing improves energy efficiency, while PRISM outperforms orthogonal approaches when semantic sharing rates are high.The results support complementary use of semantic-similarity and semantic-orthogonality approaches.
B. Energy Efficiency Maximization: The study optimizes the fairness and energy-efficiency objectives separately, so optimizing one does not guarantee the other.Cross-evaluation and scalarized fairness–energy Pareto analysis are left for future work.
C. Sensitivity to the Number of UEs and Channels: As the number of UEs increases, PRISM’s performance advantage over PRISM-lite widens across energy efficiency and α-fairness.The larger UE pool provides more potentially related semantic segments, while ORTH improves more slowly because it depends on semantic non-overlap.
C. Sensitivity to the Number of UEs and Channels: Increasing the number of channels reduces collisions and improves PRISM-lite and RND, but PRISM remains the strongest decentralized method through combined collision reduction and semantic assistance.With d1 = 1, additional channels can also incur transcoding costs, so energy efficiency is not strictly monotonic at every point.
C. Sensitivity to the Number of UEs and Channels: The evaluation assumes stationary association statistics, leaving adaptation to within-run association drift for future work.The planned extension targets dynamic or imperfectly characterized association matrices and online policy adaptation.