This manuscript is curently under review.
Value-Based Massive Access through Goal-Oriented Irregular Repetition Slotted ALOHA Pietro Talli∗ , Andrea Munari† , Federico Mason∗ , Federico Chiariotti∗ , and Andrea Zanella∗ ∗ Department of Information Engineering, University of Padova, Padua, Italy † Institute of Communications and Navigation, German Aerospace Center (DLR), Weßling, Germany
arXiv:2609.20569v1 [cs.NI] 17 Sep 2026
Emails: [email protected], [email protected],{federico.mason,federico.chiariotti,andrea.zanella} @unipd.it
Abstract—The goal-oriented communication paradigm is poised to enable novel real-time applications by easing the burden on communication networks while still delivering task-relevant information. However, efforts so far have focused on the encoding problem, while the design of medium access schemes is still in the early stages of development, especially when connectivity is to be provided to a massive number of devices, e.g., for remote monitoring. In this respect, existing goal-oriented approaches are often centralized or based on simplified underlying mechanisms, requiring unrealistic assumptions. In this work, we present the Goal-oriented Irregular Repetition Slotted ALOHA (GO-IRSA) scheme, which combines modern random access techniques with belief-based policies. GO-IRSA does not impose significant computing loads on the sensors or require frequent feedback, and it can reduce the average and worst-case error of the estimate of a distributed Wiener process by over 30% with respect to the optimal centralized solution in a network with thousands of sensors, and is robust to imperfect interference cancellation and inaccurate process knowledge. Index Terms—Goal-oriented communication, massive access, irregular repetition slotted ALOHA
I. I NTRODUCTION S the sixth generation of mobile networks (6G) foresees an ever tighter integration between the physical and digital worlds [1], the Goal-oriented Communication (GoC) paradigm promises to break the strict separation between applications and communication networks [2], allowing network optimization to directly target task performance. This groundbreaking approach has significant potential in diverse scenarios: on the one hand, distributed control systems, such as robot swarms [3], which need to exchange huge volumes of data with strict latency bounds; on the other hand, massive Internet of Things (IoT) deployments need to coordinate thousands of individual low-power nodes to provide timely and accurate monitoring [4]. In typical GoC settings, the physical environment has a digital counterpart in the network, i.e., a Digital Twin (DT) [5],
A
The work of F. Chiariotti and A. Zanella was supported by the Smart Networks and Services Joint Undertaking (SNS JU) under the European Union’s Horizon Europe research and innovation programme under grant agreement no. 101292933 (MAGIC-6G). The work of A. Munari was supported by the Federal Ministry of Research, Technology, and Space (BMFTR) for supporting the xG-RIC project as part of the research program Communication Systems “Souverän. Digital. Vernetzt.” (grant number 16KIS2429K).
which is used to monitor and control it. The network orchestrator can exploit this knowledge to infer which information is more useful for the end applications, determining the Value of Information (VoI) of each potential communication update [6]. In remote monitoring, this allows one to prioritize the transmissions that more strongly improve estimates of the environment state, while in control applications, GoC strategies select the action that leads the receiver to make the most accurate decisions. This interaction between the communication network and the control process allows for a joint optimization that can significantly improve performance over traditional technical transmission [7]. The GoC approach can greatly reduce the amount of data that must be transmitted while preserving task performance [8] in point-to-point communication systems, e.g., by extracting semantic content from video or 3D data and removing the need to transmit the raw information. On the other hand, distributed uplink applications with multiple transmitters, such as vehicular networks and massive IoT systems, introduce coordination challenges that complicate the design of goaloriented schemes. The resulting problem, named Goal-oriented Multiple Access (GoMA), cannot be addressed by straightforward extensions of point-to-point architectures, as interactions between multiple agents lead to game-theoretical challenges and convergence issues [9]. Two complementary approaches to the GoMA problem have emerged in the literature, namely, pull- and push-based access [10]. In a pull-based system, the Base Station (BS) acts as a coordinator and decides the transmission schedule centrally, entirely avoiding interference and packet collisions. The pull-based GoMA problem can be modeled as a single-agent Partially Observable Markov Decision Process (POMDP), in which the agent placed at the BS uses the application performance as the reward function. The Reinforcement Learning (RL) paradigm can then be leveraged to find near-optimal communication policies [11], which can also adapt to changing application and communication channel statistics. On the other hand, push-based GoMA moves transmission decisions from the BS to the individual nodes, which can proactively decide when to access the channel depending on local observations. The nodes know the actual value of their information, and can thus avoid transmitting low-value updates, but distributing the decision-making introduces a risk
of packet collisions. The balance between the higher value of individual updates and the lower channel efficiency is tilted by an additional advantage of push-based systems: as transmission outcomes are the result of distributed decisions based on the individual knowledge of each sensor, they carry implicit information about this knowledge [12]. If a certain sensor did not transmit at a given time, the other nodes can infer that its update was not valuable enough, gaining information that can be used in subsequent decisions. The literature on push-based GoMA is still in its infancy, and even state-of-the-art schemes are based on classical random access policies, such as slotted ALOHA [13]. Overcoming the well-known limitations of these basic schemes is often computationally complex [9] and requires very frequent feedback from the BS [14], limiting the potential applicability of goal-oriented enhancements based on VoI [15]. However, medium access research has recently developed advanced uncoordinated schemes that can mitigate these issues and improve communication performance in large networks, often grouped under the umbrella of unsourced multiple access or modern random access [16]. An interesting example in this direction is Irregular Repetition Slotted ALOHA (IRSA), a frame-based protocol in which transmitting nodes send multiple replicas of their packets, allowing the BS to resolve packet collisions through Successive Interference Cancellation (SIC). Although the scheme is considered a state-of-the-art random access technique for several 6G-oriented scenarios, including massive IoT systems, and its main principles have been recently embraced in 3GPP Release 19 for Narrowband IoT (NB-IoT) [17], its combination with GoC principles is still unexplored in the current literature. In this work, we propose a novel push-based GoMA framework that exploits IRSA to overcome the two main disadvantages of existing schemes: namely, it can limit the amount of feedback needed in a massive access scenario to a realistic value and greatly simplify the decision-making at the sensors, while increasing the performance gains over pullbased approaches. In our scheme, dubbed Goal-oriented IRSA (GO-IRSA), the BS maintains a belief over the accuracy of its estimates, dynamically determining a value threshold and transmitting it to the sensor nodes, which then independently compare it to their own VoI to make transmission decisions. To the best of our knowledge, this work represents the first integration of GoMA principles with massive random access, as well as one of the first push-based works to consider a timing dimension. Our main contributions are the following: • We define GO-IRSA, a novel push-based GoMA model that exploits IRSA as an underlying medium access mechanism, where a group of sensor nodes needs to transmit local observations about a physical environment to a central BS; • We design a belief update mechanism that allows the BS to exploit implicit information to reduce uncertainty over environment estimation; • We optimize transmission decisions of the sensor nodes based on a VoI-based threshold, computed by the BS and
•
then broadcast to all network nodes; We analyze the performance of the proposed GOIRSA scheme, revealing susbtantial gains over pull-based schemes through extensive Monte Carlo simulations, and pinpoint key trade-offs and design principles.
The rest of the paper is organized as follows: Sec. II discusses the relevant literature on GoMA and IRSA optimization; Sec. III presents the target scenario and the system model; Sec. V reports our simulation results; finally, Sec. VI presents our conclusions and some avenues for future work. II. R ELATED W ORK The concept of GoC can be traced back to W. Weaver’s preface to the first edition of Shannon’s work on communication theory [18], where the importance of transmitting information is related to its relevance for the final application rather than to the faithful delivery of every bit. An intermediate step towards this vision is represented by semantic communication [19], which focuses the optimization on the encoding: the most common example is represented by Joint Source Channel Coding (JSCC) schemes [20], where the encoder and decoder are jointly optimized to preserve the semantic content of the transmitted information. This approach has demonstrated significant performance gains across several communication tasks, but is primarily designed for point-to-point links [11]. While multi-point coding approaches [21], [22] based on the information bottleneck principle exist, they require a joint training over all transmitters and are generally vulnerable to variations in the channel statistics. The integration of the GoC paradigm with multiple access schemes, namely GoMA, remains a relatively unexplored research direction. The most common and straightforward approach in this respect is to adopt a pull-based architecture, in which a central BS maintains a DT of the entire system and assigns transmission priorities according to its estimate of the VoI associated with each node [23]. Notably, most existing studies on pull-based scheduling optimize communication following proxy metrics such as Age of Information (AoI) or its variants, rather than application-level objectives. In this regard, [24] proposes a query-based sampling framework that models the information freshness associated with multiple remote transmitters. More recent schemes directly target estimation [25] and control [26] performance, learning the optimal query strategy through RL or dynamic programming [27]. As discussed in Sec. I, pull-based schemes have a fundamental limit, as they can only exploit the statistical knowledge at the BS, represented by the DT. On the other hand, push-based schemes are able to use the actual information available to individual sensors, ensuring a higher value for updates, but also typically incurring a significant coordination burden. Event-triggered transmissions are considered in [28], analyzing the effect of basic on-off mechanisms corresponding to value thresholds. A similar model is analyzed in [29], which also includes retransmissions. However, neither work explicitly addresses contention over resource-limited channels.
III. S YSTEM M ODEL We consider a scenario in which a set N of N sensor nodes report data to a BS over a shared wireless channel. Each node n ∈ N measures the state of an independent Wiener process xn [38], which the BS aims to track through the use of a DT. Consequently, if we know the value of the process at time t, its value at time t + τ is normally distributed: xn (t + τ ) ∼ N xn (t), τ σn2 , (1) where σn is a scaling parameter. Time is discretized into slots of duration Ts , which are organized into frames of equal duration Tf , each covering an integer number of slots S = Tf /Ts . We use index j ∈ {1, . . . , S} for slots and k ∈ N
PRB
Frame k
Frame k + 1
ACK
Downlink
Frame information
Frame k − 1 Frequency
A comprehensive push-based GoMA framework is proposed in [30], where multiple sensing agents observe a common process and optimize transmission decisions according to the expected effectiveness of the information. This line of research is further extended in [31], which combines push-based sensing updates with pull-based query mechanisms. Another recent work [9] generalizes the framework considering arbitrary VoI belief distributions, moving towards a more complete theory, but the scalability of the approach is limited by the need for the strategy to be computed by each node at every time step. A simpler belief-based approach was developed in [14], which examines belief updates over slotted ALOHA, obtaining good results in wider networks. However, the analysis in both papers neglects the impact and overhead cost of the downlink feedback channel, which is required at every time step. Recent years have also seen the first steps towards leveraging modern random access schemes for GoC. Initial works evaluated the performance of IRSA in terms of the AoI [32], possibly in combination with an age-based threshold [33], identifying a fundamental trade-off driven by the frame duration, with longer contention intervals allowing for higher throughput while possibly stalling delivery of newly generated packets. Similar effects were observed for other modern random access schemes, such as frameless ALOHA [34], and recent contributions have also started considering asynchronous (unslotted) replica transmissions [35]. In turn, fewer works explored the potential of IRSA towards GoC beyond AoI. An interesting example is provided by [36], where a Threshold IRSA (T-IRSA) scheme based on the Age of Incorrect Information (AoII) was proposed, providing some approximated analysis of the system performance. Notably, the approach does not rely on any distributed knowledge or belief calculation, but rather relies on a fixed and pre-calculated threshold. Another recent work [37] considers an extension of IRSA in which multiple IoT nodes are assigned different priority classes, setting higher transmission probabilities for more urgent users. Although this approach progresses towards the integration of GoC into the medium access layer, the proposed metric is not directly related to the contribution of each packet to the final application objective. Hence, to our knowledge, the extension of IRSA-based methodologies to the GoMA problem remains an open research direction.
Uplink
Time
Time slot
Fig. 1. Frame structure.
for frames, so that the j-th slot of frame k begins at time (j − 1)Ts + kTf . The state of process xn at the beginning of slot j in frame k, which is observed directly and noiselessly by sensor n, is denoted as xn (j, k). A. Application Model We consider the DT for each process to be a belief distribution, updated at each slot and identified by Probability Density Function (PDF) φn;j,k ∈ Ω(R), where Ω(·) denotes the probability simplex over a set. We denote the Maximum A Posteriori (MAP) estimate resulting from the belief distribution φn;j,k by x̂n (j, k) and define the drift error en (j, k) as the squared difference between the real state of the system and the MAP estimate: en (j, k) = |xn (j, k) − x̂n (j, k)|2 .
(2)
We also note that, as the BS receives observations of a Wiener process, its belief distribution is symmetrical, and the MAP estimator corresponds to its mode. The error of the system E(k) is measured through the Mean Square Error (MSE), i.e., S
E(k) =
N
1 XX en (j, k). SN j=1 n=1
(3)
B. Communication Model We model the communication system using a common frame structure for push-pull access [39], shown in Fig. 1: each frame has a fixed duration, and follows the Orthogonal Frequency-Division Multiplexing (OFDM) structure, with S time slots and F subcarriers. The duration of a time slot for the application model coincides with the communication time slot. We adopt a Time Division Duplex (TDD) approach, in which all but the first and last symbol of each frame are reserved for uplink communication: the total number of Resource Elements (REs) available to the sensors is thus M = (S −2)F , and each uplink packet transmission can fit into a single RE. At the end of each frame, the BS broadcasts a collective acknowledgment packet to all nodes: this is a simple binary vector with N entries, corresponding to a decoded/not-decoded outcome for each node. On the other hand, at the beginning of the frame, the BS sends a downlink packet, which contains
C. Irregular Repetition Slotted ALOHA (IRSA) When considering push-based schemes, we will focus in particular on IRSA [40]–[42]. In this scheme, a node which decides to send data over a frame transmits multiple identical copies (replicas) of its message over some of the available REs. The number of replicas is drawn independently by each transmitter from a probability distribution, defined in advance and available to all network nodes. Following P a common L ℓ notation, the distribution is indicated as Λ(x) = ℓ=1 Λℓ x , P where ℓ Λℓ = 1, Λℓ denotes the probability for the device to send exactly ℓ replicas, and L is the maximum number of copies that can be transmitted. To distribute its ℓ replicas across the available REs, a device randomly selects ℓ of the S − 2 uplink slots in the frame, choosing one of the F subcarriers for each selected slot. We group the subcarrier and time indexes chosen by node n at frame k in vectors fn (k) and sn (k), with components in {1, . . . , F } and {2, . . . , S − 1}, respectively. Every replica further contains a pointer to the REs where all other copies of the same message are sent, e.g., by means of a short header, or with more advanced solutions [17]. At the receiving end, decoding is performed at the end of the frame, relying on SIC. Specifically, the receiver begins by identifying slots containing a single packet (singleton) and decoding them. Once a packet is decoded, its waveform is reconstructed and subtracted from the overall incoming signal across the other REs where it was transmitted, eliminating its interference contribution. This cancellation may uncover new singleton slots, enabling the decoding of packets that previously suffered collisions. This procedure is iterated until all messages are successfully retrieved or no further decoding is possible. In the latter scenario, some transmitting nodes remain unresolved; in this case, we assume the BS can estimate neither their number nor their identities. The performance of IRSA is fundamentally driven by the degree distribution Λ(x) and by the number M of REs available for transmission. These well-known aspects are exemplified in Fig. 2, where we report the packet loss rate, i.e., the probability for a transmitting user not to be decoded after SIC, as well as the throughput of the scheme against
Λ(1) (x), M = 100 (1)
Λ(2) (x), M = 100 −2 −2 ·0.1 (x), M =8 Λ(2) (x), M = 800 642800 10 0
Throughput [users/RE]
Λ
Packet loss rate
control information for the uplink part. We thus distinguish between pull- and push-based access: in the former, the frame information packet specifies the schedule, i.e., exactly which node should transmit over which RE, and thus requires a payload of M log2 (N ) bits. The sensors then behave as the packet specifies, and access is fully orthogonal. On the other hand, in push-based access, the BS may specify a probability or condition to transmit, and sensors will independently decide whether to do so and over which REs. In the remainder, we assume a collision channel model [13]: whenever two or more packets are present over one slot, none of them can be decoded, whereas a slot containing a single packet (singleton) always leads to retrieval of the information content. The model has been widely adopted in the literature both for GoMA [9], [31] and IRSA [36], [40] in view of its tractability, and captures the fundamental trade-offs of medium contention.
10−1
10−3 0.4
0.6
0.8
Load G [users/RE] (a) Packet loss rate.
0.8
0.6
0.4 0.4
0.6
0.8
Load G [users/RE] (b) Throughput.
IRSA performance using distributions Λ(1) (x) = x4 and Λ(2) (x) =
Fig. 2. 0.86x3 +0.14x8 . Two frame lengths (M = 100 and M = 800) are considered, with F = 10 in both cases. Results were obtained for Poisson traffic of parameter G users/RE over each frame. Throughput is also expressed in terms of served users per RE.
the channel load G. This is defined as the average number of users transmitting over a frame divided by the number of available REs M , and is thus independent of the selected distribution. The results were generated considering two replica distributions: Λ(1) (x) = x4 , i.e., all nodes transmit 4 replicas; and Λ(2) (x) = 0.86x3 + 0.14x8 , originally proposed in [43] as a good solution to operate with small to average frame sizes. In the plots, F = 10 subcarriers are used, comparing the performance of the scheme when S = 12 and S = 82, i.e., M = 100 and M = 800, respectively. Notably, longer frames lead to substantial improvements in terms of reliability and throughput, decreasing the probability to encounter stopping sets that halt the SIC procedures [42], [43]. Operating with a larger S entails a cost in terms of more sporadic transmission decisions, driving the design of effective GoC schemes, as will be discussed in Sec. V. IV. T HE GO-IRSA S CHEME In this section, we present the GO-IRSA solution. The basic design is simple, and requires almost no computational effort from the distributed sensors. At the beginning of frame k, the BS computes a threshold θ(k), then distributes it to the sensors using the control information packet. Upon receiving this, each sensor observes the value of the monitored process, and computes its VoI en (0, k), deciding to participate in the frame if en (0, k) > θ(k). If sensor n decides to send an update, it generates a payload containing the reading obtained in slot inf(sn (k)), i.e., just before its first transmission in the frame, and then sends its replicas containing such value following the IRSA scheme. A. Digital Twin Belief Update At the end of frame k, the BS performs SIC over all uplink REs, attempting to retrieve transmitted packets. As discussed,
We can then obtain an a posteriori belief φ̃m;0,k by applying Bayes’ theorem, truncating the distribution: (φ m;0,k (z) ζm (k) , if z ∈ Xm (x̂m (0, k), θ(k)); φ̃m;0,k = (6) 0, otherwise, where value ζm (k) corresponds to the a priori probability of node m not transmitting: Z ζm (k) = φm;0,k (z)dz. (7) Xm (x̂m (0,k),θ(k))
We then note that the value of the process xm evolved for a full frame,so that we have 2 xm (0, k + 1) ∼ N xm (0, k), Tf σm . (8) In order to obtain φm (0, k+1), we need to take the convolution of (6) and the Gaussian PDF from (8): 2 Z φ̃m;0,k (w) exp − (z−w) 2 2Tf σm p φm;0,k+1 (z) = dw. (9) 2 2πTf σm Xm (x̂m (0,k),θ(k))
On the other hand, if the frame is unresolved, the BS does not know the identity or even the number of unresolved transmissions.2 In this case, we do not use implicit information, and 1 Our formulation can be easily extended to any linear dynamical system, as long as the update equation is known to both the sensor and BS. 2 Estimating the number of transmissions in an unresolved frame is a complex task, as it requires enumerating possible stopping sets for SIC [44], and we consider it to be beyond the scope of our work.
φ̃n;0,k (x)
xn φn;0,k+1 (x)
xn φ̃n;0,k (x)
xn
xn φn;0,k+1 (x)
xn
Su
φn;0,k (x)
Silence
Prior (frame k + 1) φn;0,k+1 (x)
φ̃n;j,k (x)
cc e
ss
Posterior (frame k)
d ve ol es
This corresponds to a Gaussian distribution with mean 0 and variance [S − inf(sn (k)]Ts σn2 , following (1). After receiving the acknowledgment packet, the sensor also knows that the BS updated x̂n (inf(sn (k), k), and that the estimate will remain stable, as the Wiener process is a martingale.1 It will then use the estimated value to compute the error en in future frames. We now consider the implicit information in the case in which there are no unresolved collisions: as we know the identity of all nodes that transmitted in frame k, we also know that other nodes did not transmit. Since the decision rule that the nodes use is a threshold policy, the BS can infer that node m, which did not send an update, must have had em (0; k) < θ(k). The set of values Xm (x̂, θ) for which node m chose not to transmit in frame k is h √ i √ (5) Xm (x̂, θ) = x̂ − θ, x̂ + θ .
Prior (frame k)
nr U
there are two possible outcomes: either all packets are successfully retrieved, or there are some unresolved collisions and part of the information is lost. We first consider nodes that successfully transmitted in frame k, whose identity and transmission vectors are recovered from the packets. The BS thus knows the state of the process xn (inf(sn (k), k). However, the process keeps evolving, so that, at the beginning of the next frame, the belief PDF φn;0,k+1 is exp −z 2 (2(S − inf(sn (k))Ts )−1 σn−2 p . (4) φn;0,k+1 (z) = 2π(S − inf(sn (k))Ts σn2
xn
xn
Fig. 3. Diagram showing the outcome of the belief update process for different outcomes of the frame for a given node n.
simply update the belief distribution for any node that was not successfully decoded over frame k, i.e., either remained silent or could not be retrieved due to collisions, as Z ∞ φm;0,k (w) exp − (z−w)22 2Tf σm p φm;0,k+1 (z) = dw. (10) 2 2πT σ −∞ f m Using these equations, we can update the belief distribution for all sensors at the end of each frame, refreshing the DT. An illustrative example of the belief update process is shown in the diagram in Fig. 3: if node n is successful, the posterior belief for the measurement time becomes deterministic, and the prior used for the next frame is a Gaussian variable with a low variance, accounting for changes in the process since the received update. On the other hand, if the node is silent and there are no unresolved collisions in the frame, the DT truncates the belief at the threshold (represented with dashed lines), then adds a Gaussian term to account for changes during the frame. Finally, if the frame is unresolved, the posterior belief is identical to the prior, and the prior for the next frame includes another Gaussian component, as in the silent case. B. Threshold Selection The other component of GO-IRSA is the selection of the threshold θ(k). This choice presents a trade-off: setting a high threshold reduces transmissions and increases the overall MSE, yet aiming for a low threshold increases the risk of unresolved collisions and possibly degrades the overall performance. We thus consider vector φ0,k , including all belief distributions available at the BS at the start of frame k. The threshold leading to an expected load Ḡ over frame k is thus the solution of the integral equation " # Z √θ N X (11) 1 − √ φn;0,k (z)dz = ḠM. n=1
− θ
Target load RL solution
0.25
0.5
0.75
1
Probability
MSE E
2 1.5 1 0.5 0 0.0
6
0.25 0.2 0.15 0.1 0.05 0 -0.15
IRSA GO-IRSA
-0.1
-0.05
Training steps [×10 ]
We can thus simply identify a target load Ḡ and use it as a rule to determine the threshold in each step. As we will see in Sec. V, this simple strategy is quite effective at reducing the MSE of the distributed monitoring. In order to validate this approach, we consider an RL solution as part of our preliminary analysis. In this case, we place the agent at the BS, with a state space corresponding to the possible belief distributions for all nodes, quantized as described in Sec. V. The action of the agent is to set a set of thresholds, one per each possible number of replicas in the IRSA distribution, with an increase in replicas corresponding to lower thresholds. The set of values is distributed in the downlink frame information, and a sensor with error larger than a given threshold contends sending the corresponding number of replicas. We report the MSE performance of the RL agent during its training with Proximal Policy Optimization (PPO) in Fig. 4, comparing it with a system with a fixed target load Ḡ = 0.7, picked to minimize the MSE. The figure clearly shows that the RL agent never outperforms the target load approach, even after convergence. In this respect, it is also interesting to notice that the use of different thresholds, aiming to favor the decoding of nodes with higher errors by having them transmit more replicas, does not play a significant role. Based on these results, the simple target load approach in (11) reduces complexity and training costs and does not degrade performance, and we will consider it as the main solution in the following. V. S IMULATION R ESULTS In this section, we analyze the proposed GO-IRSA scheme by Monte Carlo simulation. We run Nsim = 50 episodes of L = 200 frames for each configuration.3 We consider a scenario with N = 2000 sensing nodes communicating with the BS over an OFDM grid with F = 10 subcarriers and time slots lasting Ts = 10 ms. We vary the number of slots in the frame, with S ∈ {12, 22, 42, 62, 82}, obtaining M ∈ {100, 200, 400, 600, 800} REs for uplink communication when considering the downlink slots for acknowledgments and scheduling information. The Wiener processes that the nodes estimate have the same scaling factor σ = 4 Us−1 , where U is the arbitrary unit measuring the process, e.g., Pascals for a pressure sensor. This assumption will be relaxed in Sec. V-C, studying the case in which each sensor has a different value of σn . The belief distribution φn was also discretized 3 The simulation code is available at https://github.com/pietro-talli/GOIRSA
0.05
0.1
0.15
G − Ḡ Fig. 5. Probability distribution of the difference between the actual and the target load, with M = 400, Ḡ = 0.7, N = 2000.
Throughput [nodes/RE]
Fig. 4. Performance of the RL agent during the training phase, with M = 200.
0
IRSA (M = 200) IRSA (M = 400) IRSA (M = 600)
GO-IRSA (M = 200) GO-IRSA (M = 400) GO-IRSA (M = 600)
0.8 0.7 0.6 0.5 0.4 0.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
Actual load G Fig. 6. Throughput achieved by IRSA and GO-IRSA.
to V = 1000 possible values for computational reasons. The main simulation parameters are listed in Table I. We compare the performance of GO-IRSA against the optimal pull-based solution, which was run with S = 82, as it was the setting under which it performed best. In the case where σ is the same for all nodes, this corresponds to a simple round robin policy, but when we introduce different values of σn throughout the network, the pull-based schedule ranks nodes by their expected VoI, following an indexing approach similar to [45]. We note that, since the pull-based approach is collision-free, its throughput is 1 packet per RE, giving it a significant advantage over random access schemes. A value-unaware IRSA approach was also considered as a potential benchmark, but its MSE was at least twice as high as for the pull-based solution in all scenarios, so it is not shown in the plots. As part of our robustness analysis, we will also compare it to a T-IRSA solution with a fixed threshold, similar to the solution in [36]. We used the replica distribution Λ(2) (x) = 0.86x3 + 0.14x8 for all IRSA-based schemes. A. Throughput Analysis We first consider the throughput that GO-IRSA can obtain. Analytical results on IRSA such as the one in Fig. 2 often TABLE I S IMULATION PARAMETERS . Parameter
Value
Parameter
Value
Nrun F S σ
50 episodes 10 {12, 22, 42, 62, 82} slots 4 U/s
L Ts N V
200 frames 10 ms 2000 nodes 1000 values
Pull-based GO-IRSA (M = 400)
2 (M = 100) GO-IRSA 1.5 1−2(M = 600) 0.5 GO-IRSA 0 ·10 Quantile Q95 (e)
0.2
MSE Ē
0.18 0.16 0.14 0.12 0.1 0.5
0.55
0.6
0.65
0.7
0.75
0.8
GO-IRSA (M = 200) GO-IRSA (M = 800)
0.8 0.7 0.6 0.5 0.4 0.5
0.55
0.6
Target load Ḡ
0.65
0.7
0.75
0.8
Target load Ḡ
(a) MSE.
(b) 95th percentile of the drift error.
assume Poisson arrivals, while we have a limited population of nodes. Assuming a standard IRSA scheme, the actual load G in a given frame then follows a binomial distribution, which may overshoot the target Ḡ. On the other hand, the load offered by GO-IRSA depends on the belief distributions, and may have an even wider variance. Fig. 5 shows a simulation of the actual load offered by IRSA and GO-IRSA with the considered N = 2000 nodes and a target load Ḡ = 0.7. We note that, while the load for IRSA has a small variance, the load offered by GO-IRSA has a larger variance, with a long tail and a negative expected value E G − Ḡ = −0.013. The randomness of the load can affect the performance of the scheme: as Fig. 2a shows, small changes in the load can increase the packet loss rate, especially when operating close to the peak-throughput point, leading to more unresolved frames. This has a negative effect on throughput: Fig. 6 compares the throughput of IRSA and GO-IRSA for different values of M , showing that GO-IRSA cannot support higher target loads without degrading throughput. As we will discuss in the following, this disadvantage is more than compensated by the goal-oriented nature of GO-IRSA, which can ensure that the delivered packets are highly informative and improve the estimation quality even with a slightly reduced throughput. We remark again that operating IRSA or GO-IRSA over longer frames always improves throughput performance due to the more efficient operation of SIC [42], [43]. B. Performance Analysis Although throughput represents a major performance indicator for traditional communication systems, the key principle of GoC is that directly targeting application performance will provide significant benefits. We thus consider the estimation error of the DT, which we measure through the MSE Ē, computed as the average over the whole episode of the framelevel MSE from (3), and the worst-case performance Q95 (e), corresponding to the 95th percentile of the error in each frame. Fig. 7a reports the MSE of GO-IRSA as a function of the target load Ḡ, comparing it to the MSE obtained by the pullbased scheme. Considering the optimal working point Ḡ∗ for each configuration, GO-IRSA can reduce the MSE by up to 30%, in spite of the collision-free operation of the pull-based
MSE E
Fig. 7. Performance of GO-IRSA in the symmetric scenario (σn = σ ∀n).
0.2 0.18 0.16 0.14 0.12 0.1 0.5
Pull-based T-IRSA
0.55
GO-IRSA No implicit inf.
0.6
0.65
0.7
0.75
0.8
Actual load G Fig. 8. Performance of GO-IRSA against T-IRSA an adaptive threshold with no implicit information, with M = 200.
approach. We can also note that the schemes are robust to different settings of Ḡ, maintaining a significant gain even if the target load is misconfigured. Similar trends outcome are shown in Fig. 7b for the worst-case performance. It is interesting to note that, while targeting throughput would lead to longer frames to reduce the risk of unresolved frames, the performance of GO-IRSA is maximized when setting relatively short frames (M = 200): shorter frames allow nodes to promptly report updates, reducing the risk of high values going unnoticed for a long time and increasing the frequency with which the DT gets implicit information about nodes that do not transmit in fully resolved frames. Naturally, there is a trade-off between throughput efficiency and responsiveness, and setting M = 100 degrades performance due to the high packet loss rate. On the other hand, higher levels of M suffer because of their low responsiveness. We then compare the proposed GO-IRSA scheme against two other IRSA-based goal-oriented schemes: the first is the T-IRSA scheme using a fixed threshold inspired by [36], while the second is a modified version of GO-IRSA that does not use implicit information. In this scheme, the posterior belief distribution is only updated for nodes who successfully transmit, while all other nodes are updated as in unresolved frames. In other words, the “silence” branch in the diagram from Fig. 3 is removed, and all nodes that did not have a successful transmission go through the “unresolved” branch. This comparison is shown in Fig. 8. We note that, as the concept of target load cannot be applied to T-IRSA, we plotted the results against the actual load G, i.e., the average number of sensors
G(k)
Threshold θ(k)
Load G(k)
1 0.8 0.6 0.4
0
50
100
150
200
−2 θ(k) 8642 ·0.1 10 0 −2 E(k)
Unresolved frame
0.2
1
MSE E(k)
Ḡ
0.75 0.5 0.25
0
50
Frame k
100
150
0.15 0.1 0.05
200
0
50
Frame k
(a) Offered IRSA load.
100
150
200
Frame k (c) Process MSE.
(b) Transmission threshold.
Fig. 9. Performance of GO-IRSA during a single episode with target load Ḡ = 0.7. G(k)
−2 First 8 10 6 4 2 ·0.1 0 −2unresolved frame
0.6 0.4 0.2 0
0
50 100 150 200
4
MSE E
Load G(k)
Load G(k)
1 0.8
2 0
0
50 100 150 200
Frame k
Frame k
(a) Good episode.
(b) Bad episode.
0.2 0.18 0.16 0.14 0.12 0.1 0.5
Pull-based GO-IRSA (M = 200) GO-IRSA (M = 600)
GO-IRSA (M = 100) GO-IRSA (M = 400) GO-IRSA (M = 800)
0.55
0.7
0.6
0.65
0.75
0.8
Target load Ḡ
Fig. 10. Actual load G during a single episode with fixed threshold. Fig. 11. Performance of GO-IRSA with imperfect SIC (η = 0.95).
transmitting per RE as measured in the simulations. We also set M = 200 for all IRSA-based policies: in the T-IRSA case, this corresponds to the optimal frame length, while the version with no implicit information would benefit from longer frames, although its performance is always worse than the pull-based scheme. This is an indication of the importance of implicit information: without these updates, GO-IRSA would not have a performance advantage, as its belief distribution φn would be overly pessimistic and misleading. On the other hand, T-IRSA performs rather well when considering lower values of the load G, but there is a critical point over which decreasing the threshold by even a small amount leads to significant performance degradation and increased load: setting θ = 0.6585, we get a load G ∼ 0.6, while setting θ = 0.6582, the load is over 0.8, with a significant increase in the MSE. As we will discuss below, this is caused by its fundamental instability, which appears as soon as the packet loss rate becomes non-negligible. We can better understand the advantages of GO-IRSA by examining Fig. 9, which illustrates the performance of the proposed approach over a single episode. As shown by Fig. 9a, GO-IRSA succeeds in maintaining the system load G close to the target value Ḡ = 0.7: after an unresolved frame, i.e., a frame in which the BS is not able to successfully decode all packets, G tends to drop significantly, quickly returning to the target value after a couple of frames. Fig. 9b provides an explanation for this: GO-IRSA immediately increases the threshold θ after each unresolved frame, as not being able to use implicit information increases its uncertainty. This reduced
load allows collided nodes, which presumably have highvalue information, to successfully send their updates, and the system quickly returns to the baseline. The MSE performance, shown in Fig. 9c, reflects the threshold changes, with some spikes corresponding to the unresolved frames, from which the performance quickly recovers, remaining stable for long periods in which all frames are fully resolved. As this recovery is enabled by dynamically adapting the threshold, it is obviously not possible for T-IRSA: under a fixed threshold, any unresolved frame leads to increased load in the next frame, as more nodes get higher values and attempt to transmit, which, in turn, increases the probability that the next frame will be unresolved as well. This positive feedback loop causes a snowball effect that is clear from Fig. 10, reporting the load generated by T-IRSA in two episodes. In the first one, shown in Fig. 10a, the evolution of the Wiener process allows for recovery, but in the second, shown in Fig. 10b, the snowball effect quickly leads to an unsustainable load, as more and more nodes try to transmit. This effect is the cause of the instability we discussed above, and makes its performance highly sensitive to system parameters. C. Robustness Analysis In the following, we evaluate the robustness of GO-IRSA when some of the assumptions in the above are violated. We first consider the possibility that the SIC process does not guarantee perfect decoding due to the wireless channel. Specifically, after the waveform for each decoded packet is subtracted from slots with other replicas, we consider the
Pull-based GO-IRSA (M = 400)
2 (M = 100) GO-IRSA 1.5 1−2(M = 600) 0.5 GO-IRSA 0 ·10 Quantile Q95 (e)
MSE E
0.3 0.25 0.2 0.15 0.1 0.5
0.55
0.6
0.65
0.7
0.75
0.8
GO-IRSA (M = 200) GO-IRSA (M = 800)
1 0.8 0.6 0.4 0.5
0.55
Target load Ḡ
Quantile Q95 (e)
MSE E
0.25 0.2 0.15 0.6
0.65
0.7
0.7
0.75
0.8
(b) Full information scenario, 95th percentile error.
0.3
0.55
0.65
Target load Ḡ
(a) Full information scenario, MSE.
0.1 0.5
0.6
0.75
0.8
Target load Ḡ
1 0.8 0.6 0.4 0.5
0.55
0.6
0.65
0.7
0.75
0.8
Target load Ḡ
(d) Partial information scenario, 95th percentile error. Fig. 12. Performance of GO-IRSA with variable σn ∈ U σ2 , 3σ . 2
(c) Partial information scenario, MSE.
possibility that the received signal may still be too degraded to retrieve additional packets. Each round of SIC then succeeds with probability η = 0.95. This model is often used to easily represent realistic SIC in IRSA [46] and in other nonorthogonal communication schemes. Fig. 11 reports the performance of GO-IRSA in this scenario. The large number of unresolved frames due to imperfect SIC reduces the robustness of the scheme, limiting its applicability to lower target loads. However, we still obtain a 25% gain over the pull-based scheme when setting Ḡ∗ = 0.6. Finally, we consider an asymmetric case in which nodes do not have the same σ. We draw the scaling parameter for each node from a uniform distribution in [σ/2, 3σ/2], i.e., σn ∼ U(σ/2, 3σ/2), and we consider two cases. In the full information scenario, the exact value of σn for each node is known to the BS, while in the partial information scenario, the BS only knows σ, and thus has an imperfect knowledge of the statistics of the Wiener processes. The results for this scenario are shown in Fig. 12. We first consider performance in the full information scenario, shown in Fig. 12a-b: even when using the optimal indexing policy, the pull-based benchmark’s performance degrades, as the MSE increases from E ≃ 0.17 to E ≃ 0.22, while the worst-case error increases from Q95 (e) ≃ 0.74 in the symmetric case to Q95 (e) ≃ 0.84. On the other hand, GO-IRSA’s performance is similar to the symmetric case: if we consider the system with M = 200 and Ḡ∗ = 0.7 as a target load, the MSE is Ē ≃ 0.12 in both cases, while the error percentile goes from Q95 (e) = 0.42 in the symmetric case to Q95 (e) = 0.44. The robustness to misconfigurations of Ḡ and M is also preserved,
as the scheme is insensitive to small errors in its settings. We can note a similar trend even in the partial information case, shown in Fig. 12c-d: in this case, the pull-based scheme corresponds to a round robin policy, increasing its MSE from Ē ≃ 0.221 in the full information case to Ē ≃ 0.226 and its worst-case error from Q95 (e) ≃ 0.84 to Q95 (e) ≃ 0.97. On the other hand, GO-IRSA is still extremely robust, as the best MSE goes from Ē ≃ 0.118 in the full information case to Ē ≃ 0.126, while the worst-case error goes from Q95 (e) ≃ 0.44 to Q95 (e) ≃ 0.47. However, we note an interesting effect: the optimal target load Ḡ∗ = 0.7 for the full information case, but Ḡ∗ = 0.65 under partial information, and performance degrades for higher target loads. This is particularly evident for M = 100, whose performance is strongly degraded, and which is even worse than the pull-based schemes if Ḡ > 0.75. As the DT has an imperfect model of the Wiener process, the belief distributions for nodes become less accurate, and while the threshold is generally correct for lower loads, this inaccuracy can lead to more unresolved frames for higher target loads. As shorter frames are more vulnerable to variations in the target loads and have higher packet loss rates, this phenomenon is particularly evident for M = 100, but it is also noticeable for M = 200. However, the GO-IRSA scheme still preserves its significant performance gain over pull-based solutions, as well as its robustness to small changes in the parameter settings. VI. C ONCLUSIONS In this work, we introduced GO-IRSA, a new GoC protocol for push-based massive access. GO-IRSA operates over realistic frame structures, without requiring either extremely
frequent feedback or complex calculations on the sensor side. This framework integrates IRSA with a goal-oriented approach, which allows the BS to maintain a belief over the potential VoI for each node, setting and broadcasting a threshold for nodes to transmit and exploiting implicit information from silent nodes. Our simulation results show that GO-IRSA consistently outperform conventional pull-based strategies, reducing the average and worst-case error by about 30% and providing robustness to various system imperfections. To the best of our knowledge, this work represents one of the first attempts to extend push-based GoMA protocols to massive access scenarios. In future work, we will focus on validating the proposed framework in more realistic communication environments and protocols such as NB-IoT. U SE OF AI D ISCLOSURE No AI system was used in any aspect of the present work that affects the scientific conclusions of the paper. R EFERENCES [1] H. Pennanen, T. Hänninen, O. Tervo et al., “6G: The intelligent network of everything,” IEEE Access, vol. 13, pp. 1319–1421, Jan. 2025. [2] D. Gündüz, Z. Qin, I. E. Aguerri, H. S. Dhillon, Z. Yang, A. Yener, K. K. Wong, and C.-B. Chae, “Beyond transmitting bits: Context, semantics, and task-oriented communications,” IEEE J. Sel. Areas Commun., vol. 41, pp. 5–41, Jan. 2023. [3] X. Fang, C. Lei, W. Feng et al., “Sensing-communication-computingcontrol closed-loop optimization for 6G digital twin-empowered robotic systems,” IEEE J Sel. Areas Commun., vol. 43, no. 10, pp. 3330–3346, Oct. 2025. [4] A. E. Kalør, G. Durisi, S. Coleri et al., “Wireless 6G connectivity for massive number of devices and critical services,” Proc. IEEE, vol. 113, no. 9, pp. 826–848, Sep. 2025. [5] Z. Yang, M. Chen, Y. Liu, and Z. Zhang, “A joint communication and computation framework for digital twin over wireless networks,” IEEE J. Sel. Topics Signal Proc., vol. 18, no. 1, pp. 6–17, Jan. 2024. [6] C. Feng, J. Pei, Z. Qin et al., “Goal-oriented communications for future cyber-physical systems,” Nature Rev. Electrical Engin., Jul. 2026. [7] J. Luo and N. Pappas, “Semantic-aware remote estimation of multiple Markov sources under constraints,” IEEE Trans. Commun., vol. 73, no. 11, pp. 11 093–11 105, Nov. 2025. [8] R. Cheng, Y. Sun, A. Taha et al., “A survey on semantic communication for vision: Categories, frameworks, enabling techniques, and applications,” IEEE Trans. Netw. Sci. & Engin., vol. 13, pp. 8080–8103, Mar. 2026. [9] F. Chiariotti and A. Zanella, “A theory of goal-oriented medium access: Protocol design and distributed bandit learning,” in Proc. Conf. Comput. Commun. (INFOCOM). IEEE, May 2026. [10] P. Talli, E. D. Santi, F. Chiariotti et al., “Pragmatic communication for remote control of finite-state Markov processes,” IEEE J. Sel. Areas Commun., vol. 43, no. 7, pp. 2589–2603, Jul. 2025. [11] T.-Y. Tung, S. Kobus, J. P. Roig, and D. Gündüz, “Effective communications: A joint learning and communication framework for multiagent reinforcement learning over noisy channels,” IEEE J. Sel. Areas Commun., vol. 39, no. 8, pp. 2590–2603, Aug. 2021. [12] Y. Xiao, Z. Sun, G. Shi, and D. Niyato, “Imitation learning-based implicit semantic-aware communication networks: Multi-layer representation and collaborative reasoning,” IEEE J. Sel. Areas Commun., vol. 41, no. 3, pp. 639–658, Mar. 2022. [13] N. Abramson, “The throughput of packet broadcasting channels,” IEEE Trans. Commun., vol. 25, no. 1, pp. 117–128, Jan. 1977. [14] F. Chiariotti, A. Munari, L. Badia, and P. Popovski, “Goal-oriented medium access with distributed belief processing,” IEEE Trans. Netw., vol. 34, pp. 1643–1658, Jan. 2026. [15] A. Topbas, C. Ari, O. Kaya, and E. Uysal, “Goal-oriented random access (GORA),” IEEE Wireless Commun. Lett., vol. 14, no. 8, pp. 2292–2296, Aug. 2025.
[16] G. Liva and Y. Polyanskiy, “Unsourced multiple access: A coding paradigm for massive random access,” Proc. IEEE, vol. 112, no. 9, pp. 1214–1229, Sep. 2024. [17] A. Munari, S. Cioni, H. Krishnamurthy et al., “Contention-based early data transmission for 3GPP IoT over non-terrestrial networks,” IEEE Commun. Standards Mag., vol. 10, no. 1, pp. 116–122, Mar. 2026. [18] C. E. Shannon and W. Weaver, The Mathematical Theory of Communication. University of Illinois Press, Dec. 1949. [19] W. Yang, H. Du, Z. Q. Liew et al., “Semantic communications for future internet: Fundamentals, applications, and challenges,” IEEE Commun. Surveys & Tut., vol. 25, no. 1, pp. 213–250, Mar. 2023. [20] D. Gündüz, M. A. Wigger, T.-Y. Tung et al., “Joint source-channel coding: Fundamentals and recent progress in practical designs,” Proc. IEEE, vol. 113, no. 9, pp. 888–919, Sep. 2025. [21] Q. Yuan, Q. Guo, Z. Wang, and W. Yang, “Design of nonorthogonal multiple access systems: An information bottleneck perspective,” IEEE Commu. Lett., vol. 26, no. 8, pp. 1750–1754, Aug. 2022. [22] S. Hassanpour, A. Danaee, D. Wübben, and A. Dekorsy, “Multi-source distributed data compression based on information bottleneck principle,” IEEE Open J. Commun. Soc., vol. 5, pp. 4171–4185, Jul. 2024. [23] S. Chen, E. Spyrakos-Papastavridis, Y. Jin, and Y. Deng, “Goal-oriented semantic communication for robot arm reconstruction in digital twin: Feature and temporal selections,” IEEE J. Sel. Areas Commun., vol. 43, no. 9, pp. 3072–3087, Sep. 2025. [24] N. Akar and S. Ulukus, “Query-based sampling of heterogeneous CTMCs: Modeling and optimization with binary freshness,” IEEE Trans. Commun., vol. 72, no. 12, pp. 7705–7714, Dec. 2024. [25] J. Holm, F. Chiariotti, A. E. Kalør et al., “Goal-oriented scheduling in sensor networks with application timing awareness,” IEEE Trans. Commun., vol. 71, no. 8, pp. 4513–4527, Aug. 2023. [26] O. Ayan, S. Hirche, A. Ephremides, and W. Kellerer, “Optimal finite horizon scheduling of wireless networked control systems,” IEEE/ACM Trans, Netw., vol. 32, no. 2, pp. 927–942, Apr. 2024. [27] P. Agheli, N. Pappas, and M. Kountouris, “Pull-based query scheduling for goal-oriented semantic communication,” IEEE Trans. Commun., vol. 74, pp. 3845–3857, Jan. 2026. [28] S. Wang, Q. Liu, P. U. Abara et al., “Value of information in networked control systems subject to delay,” in Proc. 60th Conf. Decision & Control (CDC). IEEE, Dec. 2021, pp. 1275–1280. [29] W. Wu, Y. Yang, Y. Deng, and A. Hamid Aghvami, “Goal-oriented semantic communications for robotic waypoint transmission: The value and age of information approach,” IEEE Trans. Wireless Commun., vol. 23, no. 12, pp. 18 903–18 915, Dec. 2024. [30] P. Agheli, N. Pappas, and M. Kountouris, “Goal-oriented multiple access connectivity for networked intelligent systems,” IEEE Commun. Lett., vol. 28, no. 8, pp. 1795–1799, Aug. 2024. [31] P. Agheli, N. Pappas, P. Popovski, and M. Kountouris, “Integrated pushand-pull update model for goal-oriented effective communication,” IEEE Trans. Commun., vol. 73, no. 11, pp. 10 914–10 928, Nov. 2025. [32] A. Munari, “Modern random access: An age of information perspective on irregular repetition slotted ALOHA,” IEEE Trans. Commun., vol. 69, no. 6, pp. 3572–3585, Jun. 2021. [33] H. Asgari, A. Munari, and G. Liva, “On the performance of irregular repetition slotted ALOHA with an age of information threshold,” in Proc. 3rd Wkshp. Next Gen. Netw. & Appl. (NGNA), Nov. 2022. [34] A. Munari, F. Lázaro, G. Durisi, and G. Liva, “The dynamic behavior of frameless ALOHA: Drift analysis, throughput, and age of information,” IEEE Trans. Commun., vol. 71, no. 12, pp. 6914–6927, Dec. 2023. [35] S. Chakraborty, A. Li, Y. Ince et al., “ASTRA: Asynchronous ageaware satellite random access via mean-field control,” arXiv Preprint 2605.18282, May 2026. [36] M. Mubarak and B. S. Vineeth, “On threshold policies for minimizing average age of incorrect information in IRSA for remote estimation,” IEEE Commun. Lett., vol. 30, pp. 2053–2057, Apr. 2026. [37] B. Zhang, Y. Zhang, Y. Wang, and Z. Yang, “Value-optimal priorityaware irregular repetition slotted ALOHA in satellite-integrated internet of things via noncooperative game,” IEEE Internet Things J., vol. 11, no. 7, pp. 12 495–12 509, Apr. 2024. [38] N. Wiener, “The average of an analytic functional and the Brownian movement,” Proc. Nat. Acad. Sci., vol. 7, no. 10, pp. 294–298, Oct. 1921. [39] S. R. Pandey, F. Saggese, J. Shiraishi et al., “Medium access for pushpull data transmission in 6G wireless systems,” IEEE Commun. Mag., vol. 64, no. 4, pp. 172–178, Apr. 2026.
[40] E. Paolini, G. Liva, and M. Chiani, “Coded slotted ALOHA: A graphbased method for uncoordinated multiple access,” IEEE Trans. Information Theory, vol. 61, no. 12, pp. 6815–6832, Dec. 2015. [41] E. Casini, R. D. Gaudenzi, and O. del Rio Herrero, “Contention resolution diversity slotted ALOHA (CRDSA): An enhanced random access scheme for satellite access packet networks,” IEEE Trans. Wireless Commun., vol. 6, no. 4, pp. 1408–1419, Apr. 2007. [42] G. Liva, “Graph-based analysis and optimization of contention resolution diversity slotted ALOHA,” IEEE Trans. Commun., vol. 59, no. 2, pp. 477–487, Feb. 2011. [43] M. Ivanov, F. Brännström, A. Graell i Amat, and P. Popovski, “Broadcast coded slotted ALOHA: A finite frame length analysis,” IEEE Trans. Commun., vol. 65, no. 2, pp. 651–662, Feb. 2017. [44] M. C. Moroğlu, H. M. Gürsu, F. Clazzer, and W. Kellerer, “Short frame length approximation for IRSA,” IEEE Wireless Commun. Lett., vol. 9, no. 11, pp. 1933–1936, Nov. 2020. [45] T. Z. Ornee and Y. Sun, “Remote estimation of Gauss-Markov processes over multiple channels: A Whittle index policy,” IEEE Trans. Netw., vol. 34, pp. 5818–5830, Jun. 2026. [46] C. Dumas, L. Salaün, I. Hmedoush et al., “Design of coded slotted ALOHA with interference cancellation errors,” IEEE Trans. Veh. Tech., vol. 70, no. 12, pp. 12 742–12 757, Dec. 2021.