Goal-oriented Multiple Access (GoMA)

• GoMA is a multiple-access paradigm that aligns channel access and scheduling with task-specific goals such as AoI, VoI, and control fidelity.
• It integrates cost-aware transmission strategies and advanced decision-making algorithms to optimize remote inference, federated learning, and digital twin performance.
• Implementations span random access, NOMA, and ISAC, revealing practical trade-offs between transmission efficiency, energy consumption, and control performance.

Search arXiv for: Goal-oriented Multiple Access GoMA random access NOMA medium access semantic communications 2024 2025 Goal-oriented Multiple Access (GoMA) is a multiple-access paradigm in which channel-access, scheduling, and sometimes even content-selection decisions are optimized for an end task rather than for content-agnostic communication metrics alone. Across recent formulations, the “goal” is represented by application-level quantities such as a general function of Age of Information (AoI), Value of Information (VoI), control utility, learning accuracy, semantic usefulness, or task-specific loss, so that the medium-access layer is coupled directly to estimation, control, inference, federated learning, or digital-twin maintenance (Chiariotti et al., 26 Aug 2025, Topbas et al., 5 Mar 2025, Liu et al., 18 Mar 2025, Saggese et al., 2 Mar 2026).

1. Conceptual foundations and lineage

In the GoMA view, a shared channel is not primarily a bit pipe. It is a mechanism for deciding which information should be conveyed, by whom, and at what time, so that a remote task performs well. The 2025 theoretical treatment of distributed GoMA frames the problem as multiple intelligent agents coordinating over a collision channel using packet Value of Information, with explicit transmission cost, and shows that medium access itself can be made VoI-aware rather than content-agnostic (Chiariotti et al., 26 Aug 2025). In parallel, work on goal-oriented networking for interplanetary and non-terrestrial networks argues that multi-user scheduling, grant-free access, and flow control should be driven by mission-level performance—remote inference, prediction, control, and safety alerts—under delay processes with memory, rather than by exogenous traffic models alone (Uysal, 2024).

A closely related line begins from AoI. “Goal-Oriented Random Access (GORA)” formulates a multi-user slotted ALOHA setting in which the objective is not throughput or freshness per se, but a task-specific goal function $h_i(\delta)$ of the delivered AoI, with $h_i(\delta)$ explicitly allowed to be non-monotone (Topbas et al., 5 Mar 2025). This is conceptually significant because it breaks the usual identification of communication quality with “smaller AoI,” and replaces it with a task-conditioned objective.

GoMA also has a service-oriented lineage. Multi-service oriented multiple access (MOMA) groups users by QoS profile, uses service-dependent hierarchical spreading, and allocates receiver complexity according to class requirements rather than device type; although it is not formulated in semantic or task-utility terms, it is frequently treated as a precursor of more general goal-oriented access design because its resource partitioning is explicitly driven by service goals (Ksairi et al., 2015, Ksairi et al., 2016, Ksairi et al., 2017).

2. Mathematical objectives and state variables

Recent GoMA formulations differ mainly in the task metric they optimize. In AoI-centric models, the general objective can be written as

$$J = \lim_{T\to\infty} \frac{1}{T} \mathbb E\left[\int_0^T h(\Delta(t))\,dt\right],$$

or, in the symmetric random-access setting of GORA, as

$$L(b,\tau,\Gamma) := \lim_{\ell \rightarrow \infty} \mathbb E_{(b,\tau,\Gamma)}[H(\ell d)],$$

where $b$ is the AoI of the selected buffered sample, $\tau$ is the access probability when active, and $\Gamma$ is the post-success back-off (Topbas et al., 5 Mar 2025). Because $h(\delta)$ is not required to be monotone, the optimum is allowed to prefer moderately aged data over the freshest sample.

In distributed VoI-driven GoMA, each node $n$ observes a VoI random variable $V_n$, chooses a transmission probability $h_i(\delta)$0 conditioned on the realized value $h_i(\delta)$1, and maximizes the expected system reward

$h_i(\delta)$2

where $h_i(\delta)$3 is the probability that all other nodes remain silent and $h_i(\delta)$4 is the transmission cost (Chiariotti et al., 26 Aug 2025). This formulation exposes GoMA as a distributed decision problem in which utility and energy enter at the MAC layer directly.

In control-oriented GoMA, the objective can be a closed-loop cost. The goal-oriented NOMA model introduces the Goal-oriented Tensor (GoT),

$h_i(\delta)$5

which combines context-dependent state cost, transmission cost, and actuation cost, and minimizes the long-term average sum GoT over all loops (Liu et al., 18 Mar 2025). Here, access is inseparable from estimation quality and control decisions.

Other GoMA variants replace AoI or VoI with semantic effectiveness metrics. The networked intelligent-systems formulation defines a Grade of Effectiveness (GoE) using Effective Discrepancy Error, Effective Resource Consumption, and Effective Utility of Content, and optimizes activation probabilities subject to a minimum required effectiveness $h_i(\delta)$6 (Agheli et al., 2024). In ISAC-enabled digital twins, the objective is the total delivered VoI, but each scheduled transmission must also satisfy both communication-rate and localization requirements; the number of resource elements assigned to a user is therefore

$h_i(\delta)$7

where $h_i(\delta)$8 comes from a spectral-efficiency constraint and $h_i(\delta)$9 from a Position Error Bound requirement (Saggese et al., 2 Mar 2026).

3. Access decisions: who transmits, when, and what is transmitted

A distinctive property of GoMA is that it jointly addresses at least three questions: who should access, when access should occur, and what information should be sent. In GORA, these are made explicit through the tuple $$J = \lim_{T\to\infty} \frac{1}{T} \mathbb E\left[\int_0^T h(\Delta(t))\,dt\right],$$0: $$J = \lim_{T\to\infty} \frac{1}{T} \mathbb E\left[\int_0^T h(\Delta(t))\,dt\right],$$1 chooses the buffer index of the transmitted sample, $$J = \lim_{T\to\infty} \frac{1}{T} \mathbb E\left[\int_0^T h(\Delta(t))\,dt\right],$$2 determines the access probability of active users, and $$J = \lim_{T\to\infty} \frac{1}{T} \mathbb E\left[\int_0^T h(\Delta(t))\,dt\right],$$3 imposes a back-off period after success (Topbas et al., 5 Mar 2025). The resulting policy can intentionally delay access and can intentionally send an aged sample.

The “what to send” dimension appears in several other forms. In federated edge learning, the FM/TMBA design quantizes a model parameter $$J = \lim_{T\to\infty} \frac{1}{T} \mathbb E\left[\int_0^T h(\Delta(t))\,dt\right],$$4 into an index $$J = \lim_{T\to\infty} \frac{1}{T} \mathbb E\left[\int_0^T h(\Delta(t))\,dt\right],$$5, maps that index to an orthogonal MFSK tone, and uses the shared MAC as a function-computing medium so that matched filtering recovers the empirical distribution of parameter values rather than user-wise decoded messages (Martinez-Gost et al., 2024). In blind goal-oriented massive access, the first design target is not channel reconstruction but active-user identification through angular support, so the optimization is constructed to preserve the angle information required to identify the active devices and then recovers preambles and gains only for the active subset (Daei et al., 2022, Daei et al., 2023). In networked intelligent systems, self-decision GoMA compares a content meta-value $$J = \lim_{T\to\infty} \frac{1}{T} \mathbb E\left[\int_0^T h(\Delta(t))\,dt\right],$$6 against a threshold $$J = \lim_{T\to\infty} \frac{1}{T} \mathbb E\left[\int_0^T h(\Delta(t))\,dt\right],$$7, so access is conditioned on semantic usefulness rather than on buffer occupancy alone (Agheli et al., 2024).

The “when to access” question is equally diverse. GoMA proposals include threshold-AoI activation in slotted random access, pull-based controller-triggered uplink updates in remote control, push-based VoI-triggered reservation followed by pull-based scheduled transmission in digital twins, and collision-channel access probabilities derived as game-theoretic best responses (Topbas et al., 5 Mar 2025, Liu et al., 18 Mar 2025, Saggese et al., 2 Mar 2026, Chiariotti et al., 26 Aug 2025). This suggests that GoMA is not tied to a single MAC technology; it is a design criterion that can inhabit random access, scheduled access, NOMA, and waveform-level over-the-air computation.

4. Representative architectures and protocol families

A first canonical instantiation is goal-oriented random access. In the symmetric slotted-Aloha model of GORA, successful transmissions form an arithmetic renewal process with inter-renewal time $$J = \lim_{T\to\infty} \frac{1}{T} \mathbb E\left[\int_0^T h(\Delta(t))\,dt\right],$$8, where $$J = \lim_{T\to\infty} \frac{1}{T} \mathbb E\left[\int_0^T h(\Delta(t))\,dt\right],$$9 is geometric with parameter $$L(b,\tau,\Gamma) := \lim_{\ell \rightarrow \infty} \mathbb E_{(b,\tau,\Gamma)}[H(\ell d)],$$0. The long-term average goal cost is then given by a renewal ratio, and the main optimization problem minimizes $$L(b,\tau,\Gamma) := \lim_{\ell \rightarrow \infty} \mathbb E_{(b,\tau,\Gamma)}[H(\ell d)],$$1 over $$L(b,\tau,\Gamma) := \lim_{\ell \rightarrow \infty} \mathbb E_{(b,\tau,\Gamma)}[H(\ell d)],$$2, $$L(b,\tau,\Gamma) := \lim_{\ell \rightarrow \infty} \mathbb E_{(b,\tau,\Gamma)}[H(\ell d)],$$3, and $$L(b,\tau,\Gamma) := \lim_{\ell \rightarrow \infty} \mathbb E_{(b,\tau,\Gamma)}[H(\ell d)],$$4 (Topbas et al., 5 Mar 2025). Numerical results reported for non-monotone goal functions show that GORA yields lower time-average penalty than Threshold ALOHA and Slotted ALOHA, while the advantage vanishes under sufficiently heavy congestion.

A second instantiation is task-oriented multiple access for federated learning. In the FM/TMBA scheme, $$L(b,\tau,\Gamma) := \lim_{\ell \rightarrow \infty} \mathbb E_{(b,\tau,\Gamma)}[H(\ell d)],$$5 devices transmit quantized model parameters over an AWGN MAC using orthogonal tones, the receiver estimates the type vector $$L(b,\tau,\Gamma) := \lim_{\ell \rightarrow \infty} \mathbb E_{(b,\tau,\Gamma)}[H(\ell d)],$$6, and from that estimates the mean model parameter and other statistics. The design is explicitly evaluated in terms of FEEL convergence and test accuracy rather than BER, and it reports 0 dB PAPR for the MFSK/TMBA signal versus 14 dB PAPR for analog DSB AirComp (Martinez-Gost et al., 2024). The same paper reports that MFSK/TMBA maintains model accuracy nearly intact down to SNR $$L(b,\tau,\Gamma) := \lim_{\ell \rightarrow \infty} \mathbb E_{(b,\tau,\Gamma)}[H(\ell d)],$$7 dB on MNIST with a CNN.

A third instantiation is closed-loop GoMA in NOMA. In the goal-oriented NOMA network, $$L(b,\tau,\Gamma) := \lim_{\ell \rightarrow \infty} \mathbb E_{(b,\tau,\Gamma)}[H(\ell d)],$$8 sensor–controller–actuator loops share one uplink resource via superposition coding and SIC, the joint transmission-and-control problem is cast as a POMDP, and a Double-Dueling Deep Q-Network is trained to select both the power vector $$L(b,\tau,\Gamma) := \lim_{\ell \rightarrow \infty} \mathbb E_{(b,\tau,\Gamma)}[H(\ell d)],$$9 and the actuation vector $b$0 (Liu et al., 18 Mar 2025). The reported simulations show a fundamental trade-off between transmission efficiency and control fidelity, and also show lower average GoT for NOMA than OMA in the considered two-loop scenario.

A fourth instantiation is ISAC-enabled digital-twin access. The proposed GOIA mechanism uses a push subframe in which UEs with $b$1 contend through a Framed Slotted ALOHA-like reservation and report their VoI $b$2 and data size $b$3, followed by a pull subframe in which a centralized scheduler solves a VoI-maximization problem under communication and localization constraints (Saggese et al., 2 Mar 2026). This architecture makes explicit that, in some GoMA systems, admission to contention is itself goal-filtered before classical scheduling begins.

A fifth instantiation is self-decision GoMA for networked intelligent systems. There, sensing agents observe a common event, generate updates under uniform, change-aware, or semantics-aware acquisition, and individually decide whether to transmit according to usefulness thresholds derived from optimal activation probabilities. The reported simulations state that the self-decision scheme may attain at least 92% of optimal performance, and that self-decision combined with semantics-aware acquisition can simultaneously improve GoE, reduce channel drop-offs, and lower transmission rate (Agheli et al., 2024).

5. Structural results and solution methods

GoMA has been approached through renewal theory, game theory, variational inference, convex surrogates, and deep reinforcement learning. In GORA, Theorem 1 gives two balance conditions for the optimum $b$4 at fixed $b$5: the penalty at the start of a renewal cycle equals the expected penalty at the end of the cycle, and the average cycle penalty equals the expected end-of-cycle penalty (Topbas et al., 5 Mar 2025). The same work derives corollaries showing that $b$6 need not equal zero for non-monotone $b$7, and that the renewal interval may avoid the global minima of $b$8 when the number of users is large.

The distributed VoI-driven theory casts GoMA as an exact potential game over a collision channel and proves that the problem is non-convex and may admit multiple Nash equilibria. It then characterizes each node’s best response as a VoI-threshold policy and proposes LIBRA, an iterated best-response algorithm that converges to an $b$9-NE; it also introduces BETA, a distributed semi-bandit learning method that uses limited feedback and unbiased counterfactual reward estimates to learn thresholds without prior knowledge of the VoI distributions (Chiariotti et al., 26 Aug 2025). This formalizes GoMA not merely as heuristic thresholding but as a distributed strategic optimization problem.

Learning-based GoMA also appears in more general heterogeneous networks. Generalizable Multiple Access (GMA) uses context-based meta-reinforcement learning with Mixture of Experts, where a gating network computes expert weights from a context buffer of transitions, expert posteriors are combined into a latent variable $\tau$0, and SAC learns a context-conditioned policy $\tau$1 that can adapt rapidly to previously unseen MAC environments (Liu et al., 2024). Reported results show slightly lower performance than environment-specific baselines on training tasks but faster convergence and higher performance on unseen environments.

At the PHY/MAC boundary, blind goal-oriented massive access uses a reconstruction-free atomic-norm formulation. Rather than estimating every user channel, it solves a goal-oriented SDP whose dual polynomial $\tau$2 peaks at the AoAs of active users, then uses clustering to group angles into users; for mobile devices it adds alternating minimization to recover data and gains for the active subset only (Daei et al., 2022). The asynchronous extension retains the same principle under timing offsets and preamble errors (Daei et al., 2023).

6. Trade-offs, misconceptions, related paradigms, and open problems

A recurring misconception is that the freshest possible information is always optimal. The AoI-based GoMA literature explicitly rejects this: non-monotone goal functions mean that lower AoI is not always better, and GORA provides conditions under which delayed access and transmission of aged samples minimize long-term task cost (Topbas et al., 5 Mar 2025). A second misconception is that throughput- or BER-oriented MAC design is sufficient whenever the underlying application is inference or control. The FEEL, digital-twin, and goal-oriented NOMA results show instead that the relevant objective can be convergence, VoI, localization accuracy, or closed-loop state cost, and that optimizing these can change the access architecture itself (Martinez-Gost et al., 2024, Saggese et al., 2 Mar 2026, Liu et al., 18 Mar 2025).

The main trade-offs are now well delineated. One is between transmission efficiency and control fidelity: increasing the transmission-cost weight $\tau$3 in GoT reduces power usage but worsens control performance (Liu et al., 18 Mar 2025). Another is between delivered VoI and sensing accuracy: in ISAC-enabled digital twins, tighter PEB constraints increase $\tau$4 and thus reduce the number of users that can be served in the pull phase (Saggese et al., 2 Mar 2026). A third is between moderate-load gains from goal-oriented behavior and heavy-congestion collapse to classical policies: when successful transmissions become too rare, the benefit of aged-sample selection in GORA disappears (Topbas et al., 5 Mar 2025).

Related paradigms clarify GoMA’s scope. MOMA demonstrates how class-dependent spreading, overloading, and receiver complexity can be made service-oriented and scalable under massive MIMO, and its asymptotic rate analysis under general channel correlation provides a template for class-differentiated access design even though its objective is QoS rather than semantic utility (Ksairi et al., 2016, Ksairi et al., 2017). Space-network work extends the GoMA viewpoint to DTN, grant-free access, and cross-layer bundle lifetime control, emphasizing that in highly delayed systems traffic generation itself may need to become endogenous and goal-aware (Uysal, 2024).

Open problems identified across the literature include heterogeneous task functions $\tau$5 or $\tau$6, user-specific policies $\tau$7, richer MACs such as TDMA, NOMA, multi-user MIMO, and CSMA, multi-hop and DTN routing integration, explicit energy constraints, learning when the goal function is unknown, and scalable belief-based or meta-RL methods for large systems (Topbas et al., 5 Mar 2025, Uysal, 2024, Liu et al., 2024). A plausible implication is that GoMA is evolving from a collection of special-purpose designs into a general cross-layer principle: channel access is no longer optimized only for successful packet delivery, but for the utility of the resulting state of the remote task.