Toward Practical Age-of-Information Scheduling in 5G Cellular
Abstract—We consider a 5G cellular network where a gNB schedules time-sensitive uplink transmissions from multiple UEs and forwards received packets to remote destinations. In practical 5G networks, the gNB does not directly observe the destination-side Age of Information (AoI) and must make scheduling decisions under stringent slot-level runtime constraints. In this paper, we develop a low-complexity AoI-aware scheduling policy for 5G cellular under limited observability. We first design a low-complexity estimator that infers UEside packet timestamps and destination-side AoI from gNBvisible observations. Based on these estimates, we propose and implement a Max-Weight policy (MW-LC) in NetSim, a 5G emulator with a standards-compatible protocol stack, to showcase its performance against baseline 5G scheduling policies. Furthermore, we use MATLAB simulations to show that the LC estimator and MW-LC achieve performance close to a richer estimator-based AoI policy from the literature. The estimator may be of independent interest to the community, enabling AoIaware algorithms beyond 5G scheduling. Index Terms—Age of Information, Scheduling, Wireless Networks, Optimization.
I. I NTRODUCTION The Age of Information (AoI) metric has emerged as an important measure of information freshness [1], [2]. Most existing AoI works are theory-oriented and highlight the promise of AoI as a design objective for communication networks [3]. In particular, there is a vast literature on AoI-based scheduling, e.g., [4]–[6], which is especially appealing for 5G cellular networks that support time-sensitive status updates [7]. However, a main challenge in designing practical AoI-based scheduling policies is that the 5G scheduler requires knowledge of two types of timestamps: (i) the generation times of the packets currently stored in the User Equipment (UE), and (ii) the generation times of the freshest packets successfully delivered to the destinations. In general, 5G Radio Access Network (RAN) nodes, referred to as gNBs, as well as other decisionmakers that operate in the middle of the network, do not have direct access to UE- or destination-side timestamps [8]. To overcome this challenge, prior system-oriented works, e.g., [10]–[15], have taken three main directions. A first class of works, e.g., [10]–[12], consider specific network settings in which AoI could be known by the gNB, for example, when UEs generate packets periodically and the AoI is measured with respect to the gNB, as opposed to the final destination. A second class of works, e.g., [13], [14], use Software Defined Radios to implement non-standard-compliant mechanisms to expose timestamp information to the scheduler. A third class of works optimize AoI indirectly, for example, by optimizing a grouped cyclic scheduling structure [15].
MW with No-Feedback Estimation MW with Low-Complexity Estimation Slot Length
102 Runtime per Slot
arXiv:2605.13012v1 [cs.NI] 13 May 2026
Zhuoyi Zhao and Igor Kadota
101
Slot Length = 1 ms
100
10-1 5
10
15
20 25 30 35 Total Number of UEs
40
45
50
Fig. 1: Average per-slot runtime of AoI scheduling with no-feedback estimation [9], and with the proposed low-complexity estimation, versus the total number of users N ∈ {5, 10, . . . , 50}. The dashed red line marks a 1 ms TTI, which serves as a representative runtime budget for 5G scheduling. Runtime measurements were obtained in MATLAB R2025b on a machine with an Apple M2 Pro chip.
We recently proposed an estimator-based method to track the evolution of AoI over time [9]. While that work showed that AoI-aware scheduling can be realized without direct timestamp access, it has two important limitations from a 5G implementation perspective: prohibitively high computational complexity and assuming knowledge of detailed UE packetgeneration statistics. This work shows that AoI-aware scheduling can be realized in a 5G setting when detailed packetgeneration statistics and freshness are not directly observable at the gNB, by combining low-complexity estimation with scheduler design under strict runtime constraints. Figure 1 shows that the per-slot runtime (of estimation and scheduling) of [9] is significantly larger than the 1ḿs transmission time interval (TTI), whereas the proposed solution remains within this runtime budget across the tested system sizes, which motivates the low-complexity design pursued in this paper. Our main contributions can be summarized as follows: We develop a low-complexity (LC) estimator for UE-side packet timestamps and destination-side AoI based on gNBvisible observations. Based on these estimates, we develop an AoI-aware Max-Weight scheduling policy, termed MWLC, that incorporates uplink channel information. The resulting estimator-and-scheduler pipeline has O(N ) perslot complexity under sequential execution and O(1) time under per-UE parallelization. • We implement the proposed estimator and scheduler in NetSim, a standard-compliant 5G emulator. Our emulation results show that MW-LC consistently outperforms conventional baselines such as Proportional Fair (PF) and Round •
denote the generation time of the single packet stored in UE i at the beginning of slot t, and define its system time as ai (t) := t − γiU (t). The system time evolves as ( 0, if gi (t) = 1, (1) ai (t + 1) = ai (t) + 1, otherwise.
Fig. 2: Network with N UE-destination pairs communicating through a 5G gNB. The gNB schedules K uplink transmissions per slot. UEto-gNB links have reliability pU i (t), and successfully received packets are forwarded to the destinations over links with reliability pD i and transmission delay θi .
Robin (RR) in AoI performance. We further evaluate the LC estimator and the MW-LC policy through MATLAB simulations. Our numerical results show that the LC estimator remains close to richer estimators in estimation accuracy, and that the resulting MW-LC policy achieves AoI performance close to that of richer estimator-based policies. More broadly, the estimator may be of independent interest to the community, enabling AoI-aware algorithms in different domains when the decision-maker is an intermediate entity that does not directly observe source-generation timestamps or destination freshness. II. S YSTEM M ODEL Consider a network with N UE sending time-sensitive uplink (UL) packets to N destinations through a 5G gNB, as illustrated in Fig. 2. Time is slotted, with slot index t ∈ {1, 2, . . . , T }. Each UE i generates packets according to a renewal process. Let gi (t) = 1 indicate that UE i generates a new packet at the beginning of slot t, and let gi (t) = 0 otherwise. Let the inter-generation period Xi denote the number of slots between two consecutive packet generations from UE i, and define the average packet-generation rate as λi := 1/E[Xi ]. We assume1 that the gNB can estimate λi . System Time. Each UE maintains a single-packet queue2 that stores only the freshest generated packet. Whenever a new packet is generated, it replaces the previously stored packet. If no new packet is generated, the stored packet remains available for (potential) repeated transmissions3 . Let γiU (t) 1 5G NR has periodic Scheduling Request (SR) opportunities through which UEs can inform the gNB that UL resources are needed. SR periodicity can be configured to be shorter/longer than a slot, with shorter periods resulting in higher control overhead. Multiple MAC/PHY conditions, including measurement gaps and SR prohibit timers, may prevent SR transmissions even if an SR opportunity exists. Our system model does not rely on real-time, perfect knowledge of UE packet generation. Instead, it relies on a long-term traffic statistic λi that the gNB can estimate from scheduler-visible signaling and historical observations, with SR serving as one possible input. 2 Single-packet queues have been shown to be optimal in terms of AoI in many different scenarios [3]. 3 The scheduler in Sec. III-C will attempt to avoid repeated transmission as they provide no freshness gain to the destination.
Packet Transmission. At the beginning of each slot t, the gNB schedules at most K UEs for UL transmission4 , where K ≤ N . Let ui (t) PN∈ {0, 1} indicate whether UE i is scheduled in slot t, with i=1 ui (t) ≤ K. A scheduled UE transmits its packet over a time-varying UE-to-gNB UL channel. Let cU i (t) ∈ {0, 1} denote the UE-to-gNB link success indicator in slot t, where ui (t)cU i (t) = 1 indicates that UE i’s packet is successfully received by the gNB. We model the UE-to-gNB UL channel using Sounding Reference Signal (SRS)-based measurements available at the gNB. In our model, the gNB obtains SRS-based measurements at the beginning of each frame5 and uses them to estimate the UL SINR. These SRSbased measurements are mapped to a time-dependent UL link U reliability, denoted by pU i (t) ≜ Pr ci (t) = 1 | SRSi (t) . The SRS-based measurements and pU i (t) are kept fixed within each frame and are available to the gNB when making slot-level scheduling decisions, while the realized link success indicator cU i (t) becomes known only after the transmission attempt. Every packet received by the gNB is immediately forwarded to its corresponding destination through a heterogeneous network with long-term delivery reliability pD ∈ (0, 1] and i transmission delay θi ∈ {0, 1, . . .} slots. Let cD i (t) ∈ {0, 1} denote the gNB-to-destination link success indicator in slot D t, where ui (t)cU i (t)ci (t) = 1 indicates that UE i’s packet, transmitted in slot t, is successfully received by destination i in slot t + θi . We assume that {cD over i (t)} is independent D time slots and across UEs, with Pr cD i (t) = 1 = pi . Age of Information (AoI). Let γiD (t) denote the generation time of the freshest packet successfully received by destination i by the beginning of slot t. The corresponding AoI is Ai (t) := t − γiD (t). Its evolution is given by ( ai (t − θi ) + θi + 1, if Λi (t) = 1, Ai (t + 1) = (2) Ai (t) + 1, otherwise, D where Λi (t) := ui (t − θi )cU i (t − θi )ci (t − θi ), ∀t > θi . For simplicity, we assume Ai (1) = 1 and ai (1) = 0 for all i. For a non-anticipative scheduling policy π ∈ Π, the performance metric is the expected weighted sum AoI (EWSAoI): T N 1 XX αi E[Aπi (t)], (3) E[J π ] := lim T →∞ T N t=1 i=1
where αi > 0 is a priority weight for destination i. 4 Our system model can be extended to account for UL transmissions delayed by a constant number of slots. Furthermore, the scheduler in Sec. III-C can be easily modified to accommodate a time-varying K. 5 5G New Radio (NR) has 10 ms frames, 1 ms subframes, and slot duration determined by numerology. In Sec. IV, we assume slot duration of 1 ms.
gNB Observation. The scheduler operates using gNB-side information from successful UL receptions. Upon each successful UL reception, the gNB determines whether the received packet is a repeated copy of the current most recent packet or a newly generated distinct packet6 , and updates the stored packet-transmission times as described below. For each UE i, the gNB keeps track of the two most recently received distinct packets. Let τicur (t) and τ̄icur (t) denote the first and last slots, respectively, in which the current most recent distinct packet was successfully received from UE i by the beginning of slot t. Similarly, let τipre (t) and τ̄ipre (t) denote the first and last slots, respectively, in which the previous distinct packet was successfully received from UE i. Failed UL transmissions are not included in the estimator state, as they do not reveal the identity or generation time of the transmitted packet at the gNB [9]. History from older packets is not retained, as it is captured through the recursive estimator developed in Sec. III. Next, we describe an AoI estimator and scheduler that use knowledge of N, K, pD i , θi , λi , and gNB observation U pre O(t) := {pi (t), τi (t), τ̄ipre (t), τicur (t), τ̄icur (t)}N i=1 . (4) We assume the gNB either knows or can estimate pD i , transmission delays θi , and packet-generation rates λi from network configuration, historical measurements, and schedulervisible control information. In particular, pD i and λi are treated as long-term statistics rather than real-time observables, with pD i ≈ 1 in many practical scenarios. Interestingly, we will see that timestamp estimation and the MW-LC do not depend on θi , which only affects the AoI estimate Âi (t + θi ) in (12). III. L OW-C OMPLEXITY E STIMATION AND S CHEDULING In this section, we develop a Low-Complexity (LC) Estimator for system time and AoI. Relative to [9], two main differences are: (i) assumption of knowledge about λi , as opposed to knowledge of the complete packet generation statistics, namely its probability mass function P(Xi = x), ∀i, x; and (ii) a LC Estimator based on closed-form approximations, as opposed to high-complexity algorithms with nested “for loops”. The resulting estimator trades exactness for tractability by replacing the full inter-generation distribution with a ratebased Bernoulli approximation. Next, we leverage this restricted knowledge to derive approximate closed-form packetgeneration probabilities, develop the LC Estimator and the Max-Weight-LC scheduling policy. A. Packet-Generation Probabilities For each UE i, the two most recent successful UL receptions determine the feasible interval in which the current packet may have been generated. Using a Bernoulli approximation parameterized by λi , we first define λi , (5) ηiD (t) = pre cur 1 − (1 − λi )τi (t)−τ̄i (t) 6 In the NetSim implementation, this distinction is made using the packet ID available in the simulator’s protocol stack. More generally, it requires packet-identity information exposed to the scheduler.
which is the conditional generation probability associated with the interval [τ̄ipre (t) + 1, τicur (t)]. Let qiD (φ) denote the conditional probability that the most recently received distinct packet of UE i was generated in slot φ. Under this Bernoulli approximation, it follows that cur
qiD (φ) =
ηiD (t) (1 − ηiD (t))τi (t)−φ . pre cur 1 − (1 − ηiD (t))τi (t)−τ̄i (t)
(6)
Next, for the interval after the most recent successful reception of the current packet, we define ηiU (t) =
λi , cur 1 − (1 − λi )t−τ̄i (t)
(7)
which is the conditional generation probability associated with the interval [τ̄icur (t) + 1, t]. Similarly, let qiU (ϕ) denote the conditional probability that the packet currently stored at UE i was generated in slot ϕ. Under the Bernoulli approximation, it follows that τ cur(t)−ϕ D D Ci (t), τ̄ipre(t)+1 ≤ ϕ ≤ τicur(t), ηi (t) 1−ηi (t) i U qi (ϕ)= ηiU(t) 1−ηiU (t) t−ϕ , τ̄icur(t)+1 ≤ ϕ ≤ t, 0, otherwise. (8) t−τ̄icur (t) τ cur (t)−τ̄ipre (t) where Ci (t):= 1−ηiU (t) / 1− 1−ηiD (t) i . The numerator in (8) corresponds to the event that the current packet is generated in slot φ and no newer packet is generated before slot t, while the denominator normalizes this probability over the feasible generation window. Algorithm 1 LC Estimator for Âi (t + θi ) and âi (t) 1: Input: Observation O(t) and estimate γ̂iD (t + θi − 1) 2: Compute ηiD (t) and ηiU (t) using (5) and (7)
3: Compute qiD (φ) and qiU (φ) using (6) and (8) D,cur
4: Update γ̂iU (t) and γ̂i (t) using (9) 5: If ui (t)cU (t) = 1 then Update γ̂iD (t + θi ) using (10) i D 6: Else Update γ̂i (t + θi ) using (11) 7: Compute Âi (t + θi ) and âi (t) using (12) 8: Output: γ̂iU (t), γ̂iD (t + θi ), Âi (t + θi ), and âi (t)
B. Low-Complexity Estimator Using the packet-generation probabilities in (5)–(8), the gNB constructs approximate MMSE estimates of the UE timestamp and of the generation time of the most recently received distinct packet as t X γ̂iU (t) = φqiU (φ), pre φ=τ̄i (t)+1
τicur (t) X D,cur γ̂i (t) = φqiD (φ). pre φ=τ̄i (t)+1
(9)
To estimate the destination timestamp, the gNB recursively infers whether the most recently received distinct packet is delivered to the destination using the gNB-to-destination reliability pD i . If UE i is successfully received by the gNB in slot t, then D D D,cur γ̂iD (t + θi ) = (1 − pD (t). (10) i )γ̂i (t + θi − 1) + pi γ̂i
(a) Weighted AoI sample path.
Fig. 3: System Diagram of NetSim with 10 UE-destination pairs. TABLE I: Channel Model Setting Setting UE/gNB Height & Antenna Count UE Tx Power CA Type & CA Configuration DL:UL Ratio & Numerology µ Channel BW & Coherence Time MCS Table Pathloss/Shadow Fading Model Outdoor Scenario & LOS/NLOS Fading and Beamforming
Value 1.5 m / 10 m & 1 / 2 23 dBm 2 Bands & n78 1:1 & 0 (15 kHz SCS, 1 ms slot) 40 MHz & 10 ms QAM256 3GPP TR 38.901, Sec. 7.4.1 URBAN MACRO & LOS RAYLEIGH WITH EIGEN
Otherwise, γ̂iD (t + θi ) = γ̂iD (t + θi − 1).
(11)
Accordingly, the AoI and system-time estimates are given by Âi (t + θi ) = t + θi − γ̂iD (t + θi ),
âi (t) = t − γ̂iU (t). (12)
Although Âi (t + θi ) is indexed by the forwarding delay, the LC estimator itself updates the timestamp estimates γ̂iU (t) and γ̂iD (t + θi ) using only O(t), λi , and pD i . The delay θi is only used when converting the destination timestamp estimate into the AoI estimate in (12). Notice that older packet-generation history does not need to be explicitly retained in the online LC update. Once slot t begins, no new information about the generation time of earlier packets becomes available, and their effect on destination-side estimation has already been absorbed into the recursive state γ̂iD (t + θi − 1). Hence, for the purpose of LC estimation, the compressed observation in (4) is sufficient for updating the current UE timestamp and destination AoI estimates under the adopted recursive LC model. The resulting low-complexity estimator is summarized in Algorithm 1. The LC estimator has an average per-slot time complexity of O(N ), with constant computation per UE. Its per-UE updates are separable and can be reduced to O(1) time under fully parallel execution. Therefore, incorporating the LC estimator does not increase the order of computational complexity of the Max-Weight scheduling policy introduced next. C. MW-LC Scheduling Utilizing the LC estimates Âi (t + θi ) and âi (t), we design an AoI-aware Max-Weight scheduling policy motivated by Lyapunov optimization [16]. To account for heterogeneous
(b) Empirical CDF of weighted AoI.
Fig. 4: NetSim 5G NR emulation results with 40 UE-destination pairs and K = 2 scheduled UL transmissions per slot. Bernoulli arrival rates are {0.05, 0.2, 0.5, 1} for UE groups [1 : 10], [11 : 20], [21 : 30], and [31 : 40], respectively.
transmission delays, we introduce the following Lyapunov function: N 1 X L(t) = βi Âi (t + θi ), (13) N i=1 where βi > 0 is a design parameter. Although Âi (t + θi ) refers to a future destination-side AoI estimate, it is deterministically computed from the current gNB observation and the LC estimator recursion. Since scheduling UE i in slot t may affect the destination AoI only at slot t + θi + 1, we consider the following one-slot Lyapunov drift: ∆(O(t)) := E[L(t + 1) − L(t) | O(t)] . Substituting the AoI evolution in (2) into the drift yields N N 1 X 1 X βi − Wi (t) E[ui (t) | O(t)], (14) N i=1 N i=1 D where Wi (t) = βi pU i (t) pi Âi (t + θi ) − âi (t) − θi can be interpreted as the marginal drift-reduction score associated with scheduling UE i under the current estimates and link reliabilities. Notice that Âi (t + θi ) − âi (t) − θi = γ̂iU (t) − γ̂iD (t + θi ). Hence, the explicit delay term cancels in the timestamp-difference form, and MW-LC does not require θi when computing the scheduling weights. To minimize the Lyapunov drift in (14), the scheduler selects in each slot up to K UEs with the largest positive weights Wi (t), with ties broken arbitrarily. We refer to this policy as MW-LC.
∆(O(t)) =
IV. E MULATION AND S IMULATION R ESULTS This section evaluates the proposed framework from two complementary perspectives. We first implement the scheduler in a 5G NR emulator using the LC estimator-based policy and only scheduler-visible information at the gNB. We then implement the MW-LC scheduler in MATLAB, compare it
Weighted Sum AoI
300 MW with No-Feedback Estimation MW with Low-Complexity Estimation MW with Low-Complexity Estimation in NetSim
250
200
150
5
10
15
20
Average Inter-Generation Period
Fig. 5: Comparison of NetSim LC estimator and MATLAB estimators under mixed periodic and Bernoulli arrivals, with N = 40, K = 2, and common expected inter-generation period E[Xi ]. 180 Bernoulli + No-Feedback Bernoulli + Low-Complexity Periodic + No-Feedback Periodic + Low-Complexity Uniform + No-Feedback Uniform + Low-Complexity
Weighted Sum AoI
160 140 120 100 80 60 40
3
6
9
12
15
18
21
24
27
30
Average Inter-Generation Period
Fig. 6: EWSAoI versus E[Xi ] for N = 10, K = 2, α = [1, 2, 3, 4, 5, 2, 4, 6, 8, 10], NetSim-distributed uplink reliabilities D pU i (t), pi = 0.8, and θi = 5, under uniform, Bernoulli, and periodic packet generation.
with the NetSim results, and compare it with the highercomplexity estimator-based policy7 from [9]. A. NetSim Emulation We implement the scheduler in NetSim [18], a discreteevent simulator that models the 5G NR protocol stack, including the application, MAC, and PHY layers. Packet generation occurs at the application layer, UL transmissions are scheduled at the MAC layer, and decoding outcomes together with SRSbased UL channel measurements are produced at the PHY layer. In line with our model assumptions, the NetSim gNB can only observe network configuration parameters including D N, K, pU i (t), pi , λi , transmission outcomes, and SRS-based UL channel-state information, from which it executes the MW-LC scheduler without direct access to UE/destination timestamps. In implementation, the mapping from SRS-based UL measurements to the reliability pU i (t) is estimated offline via Monte Carlo simulation under the same channel model. Figure 3 illustrates the network implemented in NetSim. The wireless channel and PHY-layer parameters follow the 3GPP-compliant configuration provided by the simulator [18] and are summarized in Table I. NetSim Results. We compare against the original estimatorbased Max-Weight policy (MW-O) in [9], as well as RoundRobin (RR) and Proportional Fairness (PF) [19]. MW-O assumes an i.i.d. UE-to-gNB channel and uses long-term UL 7 As shown in Fig. 1, the high-complexity estimator incurs substantial per-slot runtime and is therefore too costly for direct implementation in the NetSim 5G NR protocol stack. We thus evaluate this baseline only in MATLAB. The MATLAB code is available at [17].
reliabilities. All policies are evaluated under identical channel realizations and traffic processes. We do not include [10]–[12] as direct numerical baselines because they are formulated for different scheduling objectives and gNB observability models, which may conflate the comparison. In particular, our focus is the setting in which the gNB does not directly observe UEside timestamps or destination freshness and must estimate them from gNB-visible observations. We emulate a network with N = 40 UEs and schedule K = 2 UL transmissions per slot. The UEs are divided into four traffic classes with Bernoulli arrival rates 0.05, 0.2, 0.5, and 1, respectively, and heterogeneous priority weights αi across users. Channel reliabilities and transmission delays are jointly determined by the user locations, the channel configuration in Table I, and the protocol settings of the emulator. Fig. 4 shows the weighted Sum AoI sample path and CDF obtained from the NetSim emulation. The gNB executes the LC estimator using only protocol-level observations, and the true AoI at the destinations is never revealed to the scheduler. This shows that the proposed observation model and scheduling policy can be realized within the NetSim protocol stack, under the observability assumptions described in Sec. II, without revealing explicit AoI values to the scheduler. Among all implemented policies, MW-LC achieves the lowest AoI in the tested NetSim scenarios. Conventional channel-driven schedulers such as PF can repeatedly favor users in good channel states and may ignore stale users for extended periods, leading to large AoI build-ups. Moreover, because MW-O relies on long-term channel reliabilities, it may repeatedly select a user whose instantaneous channel remains poor within a transmission frame, whereas the SRS-aware MW-LC policy mitigates this effect by incorporating framelevel UL reliability into the scheduling weight. In the scenario of Fig. 4, this corresponds to average AoI reductions of 63%, 44%, and 21% relative to PF, RR, and MW-O, respectively. B. MATLAB Simulation We next use MATLAB simulations to compare MW-LC against the Max-Weight with no-feedback Estimation (MWEnF) [9], whose richer estimator exploits the probability mass function of the inter-generation period P(Xi = x), ∀i, x. The performance of both policies is averaged over 10 independent runs with horizon T = 105 slots. Fig. 5 compares the LC estimator implemented in the NetSim 5G NR emulator with the MATLAB-based estimators under mixed periodic and Bernoulli arrivals. The gap between MW-LC and the richer estimators remains small across the tested inter-generation periods, suggesting that the Bernoulli approximation preserves much of the information most relevant for scheduling. Figs. 6 evaluates the policies under uniform, Bernoulli, and periodic packet-generation processes, respectively. Across all these generation processes MW-EnF and MW-LC exhibit very similar performance. Figs. 8 further compare the true AoI and the LC-estimated AoI for a representative stream. When pD i = 1, the estimated AoI closely matches the true AoI
70
50 Estimated AoI (stream 1) True AoI (stream 1)
45
60
pD = 0.8, MW with No-Feedback Estimation pD = 0.8, MW with Low-Complexity Estimation pD = 1.0, MW with No-Feedback Estimation pD = 1.0, MW with Low-Complexity Estimation
Estimated AoI (stream 1) True AoI (stream 1)
10
40
0
50 35 40
NMSE
AoI
AoI
30 25
30
10
-1
10
-2
10
-3
20 20 15 10
0 200
10
250
300
350
400
450
500
Slot index
5 200
250
300
350
400
450
500
Slot index
4
6
8
10
12
14
Average Inter-Generation Time
D (a) AoI sample path of Stream 1 with pD i = (b) AoI sample path of Stream 1 with pi = 1, ∀i. Fig. 9: Estimator NMSE versus the expected 0.8, ∀i. inter-generation period E[Xi ] for N = 10, K=
Fig. 8: True and estimated AoI sample paths for a representative stream under the LC estimator. Parameters: N = 10, K = 2, α = [1, 2, 3, 4, 5, 2, 4, 6, 8, 10], uplink reliabilities pU i (t) distributed as in the first 10 UEs of the NetSim configuration, θi = 5, and Xi ∼ D U [4, 8]. The destination-link reliability is pD i = 0.8 in (a) and pi = 1 in (b).
because every forwarded packet is delivered. When pD i = 0.8, a mismatch appears because the gNB cannot directly observe destination receptions and must instead infer them recursively. Nevertheless, the two trajectories remain well aligned in their drop instants and relative magnitudes, indicating that the LC estimator preserves the urgency information needed for scheduling. This observation is further supported by Fig. 9, where the estimator NMSE remains moderate while the AoI performance of MW-LC stays close to that of MW-EnF. V. F INAL R EMARKS This paper studied AoI-aware scheduling for 5G cellular under limited freshness observability and slot-level runtime constraints. We developed a low-complexity estimator (based on gNB-visible observations) and a MW-LC scheduling policy that leverages SRS-based UL CSI. We implemented the estimator and scheduler in NetSim, a 5G NR emulator with a standards-compatible protocol stack, and showed that MWLC outperforms 5G baselines. We used MATLAB simulations to evaluate both the estimator and scheduler under a broader set of traffic and network conditions. Our current design assumes access to long-term traffic parameters, packetidentity information needed for distinct-packet tracking, and a Bernoulli approximation of packet generation. Interesting directions for future work include implementing MW-LC on Software Defined Radios and within the OpenAirInterface software stack, as well as applying our low-complexity estimator framework to other domains in which the decision-maker is an intermediate entity (e.g., a network cache, federated learning server, network controller, or digital twin) that does not directly observe source-generation timestamps or destination freshness. R EFERENCES [1] S. Kaul, M. Gruteser, V. Rai, and J. Kenney, “Minimizing age of information in vehicular networks,” in Proc. IEEE SECON, 2011. [2] Y. Sun, E. Uysal-Biyikoglu, R. D. Yates, C. E. Koksal, and N. B. Shroff, “Update or wait: How to keep your data fresh,” IEEE Trans. Inf. Theory, vol. 63, no. 11, pp. 7492–7508, 2017.
2, α = [1, 2, 3,4,5,2,4,6,8,10], pU i (t) distributed as in the first 10 UEs of the NetSim configuration, and θi = 1. Each UE follows Xi ∼ U [0.8 E[Xi ], 1.2 E[Xi ]]. The destination-link D reliabilities are pD i =0.8 and pi =1.
[3] R. D. Yates, Y. Sun, D. R. Brown, S. K. Kaul, E. Modiano, and S. Ulukus, “Age of information: An introduction and survey,” IEEE J. Sel. Areas Commun., vol. 39, no. 5, pp. 1183–1210, 2021. [4] I. Kadota, A. Sinha, and E. Modiano, “Scheduling algorithms for optimizing age of information in wireless networks with throughput constraints,” IEEE/ACM Trans. Netw., vol. 27, no. 4, pp. 1359–1372, 2019. [5] A. R. Gorle, S. Bhattacharya, and J. M. Cioffi, “AoI-QPS: Age-ofinformation aware efficient queue-proportional scheduling,” in Proc. IEEE GLOBECOM, 2025. [6] Z. Zhao, V. Tripathi, and I. Kadota, “Optimizing age of information in networks with large and small updates,” in Proc. IEEE WiOpt, 2025. [7] H. Zhou, Y. Deng, X. Liu, N. Pappas, and A. Nallanathan, “Goaloriented semantic communications for 6G networks,” IEEE Internet Things Mag., vol. 7, no. 5, pp. 104–110, 2024. [8] 3GPP TS 38.211 V16.2.0: NR; Physical Channels and Modulation, 3rd Generation Partnership Project (3GPP) Technical Specification 3GPP TS 38.211, July 2020, release 16. [9] Z. Zhao and I. Kadota, “Optimizing age of information without knowing the age of information,” in Proc. IEEE INFOCOM, 2025. [10] V. Tripathi, I. Kadota, E. Tal, M. S. Rahman, A. Warren, S. Karaman, and E. Modiano, “WiSwarm: Age-of-information-based wireless networking for collaborative teams of UAVs,” in Proc. IEEE INFOCOM, 2023. [11] C. Li, Y. Huang, S. Li, Y. Chen, B. A. Jalaian, Y. T. Hou, W. Lou, J. H. Reed, and S. Kompella, “Minimizing AoI in a 5G-based IoT network under varying channel conditions,” IEEE Internet Things J., vol. 8, no. 19, pp. 14 543–14 558, 2021. [12] C. Li, Q. Liu, Y. T. Hou, W. Lou, and S. Kompella, “Aequitas: A 5G scheduler for minimizing outdated information in IoT networks,” IEEE Internet Things J., 2024. [13] I. Kadota, M. Rahman, and E. Modiano, “WiFresh: Age-of-information from theory to implementation,” in Proc. IEEE ICCCN, 2021. [14] O. Ayan, H. Y. Özkan, and W. Kellerer, “An experimental framework for age of information and networked control via software-defined radios,” in Proc. IEEE ICC, 2021. [15] B. Han, Y. Zhu, Z. Jiang, M. Sun, and H. D. Schotten, “Fairness for freshness: Optimal age of information based OFDMA scheduling with minimal knowledge,” IEEE Trans. Wireless Commun., vol. 20, no. 12, pp. 7903–7919, 2021. [16] M. Neely, Stochastic Network Optimization With Application to Communication and Queueing Systems. Springer Nature, 2022. [17] Z. Zhao, “MATLAB code for Low-Complexity AoI estimation,” https://github.com/zhuoyijoeyzhao/AoI LCEstimator, 2026. [18] NetSim Network Simulator, Tetcos, Bangalore, India, 2024, version 14.2. [19] F. P. Kelly, A. K. Maulloo, and D. K. H. Tan, “Rate control for communication networks: Shadow prices, proportional fairness and stability,” J. Oper. Res. Soc., vol. 49, no. 3, pp. 237–252, 1998.