AoI-aware Resource Allocation

• AoI-aware resource allocation is a framework that prioritizes the freshness of information by directly integrating age metrics into resource scheduling and control decisions.
• It balances conflicting constraints—such as throughput, energy, and computation—by optimizing age metrics like AoI, peak AoI, and AoA across diverse systems including 5G, IoT, and vehicular networks.
• Methodologies range from analytical models (e.g., MDP, convex optimization) to learning-based techniques (PPO, DDPG) that jointly manage various resource dimensions under dynamic network conditions.

AoI-aware resource allocation is the design of scheduling, power control, bandwidth assignment, computation allocation, access control, mobility control, or robot deployment decisions using information freshness as a primary optimization criterion. Across 5G uplink scheduling, C-V2X platooning, MIMO-NOMA IoT, hierarchical NTN/MEC offloading, RIS-aided IoV, smart-factory subnetworks, wireless-powered IoT, and multi-robot sensing, the common objective is to control the evolution of age-type metrics—most often Age of Information (AoI), but also peak AoI, AoI violation probability, Age of Actuation (AoA), and related penalties—under radio, energy, computation, and topology constraints (Wu et al., 2021, Wang et al., 2024, Tadrous, 4 May 2026, Zhao et al., 9 May 2026). In this body of work, freshness is not treated as a by-product of throughput or delay minimization; it is encoded directly in the state, objective, and action space.

1. Freshness metrics and their formal meaning

The canonical metric is AoI, defined as the time elapsed since the generation of the freshest successfully received update. In one 5G uplink formulation, for user equipment $n$,

$$A_n(t) := t - \max_i \{ t_G^n(i) \mid t_G^n(i) \le t \},$$

with slotted-time evolution

$$A_n(t+1)= \begin{cases} A_n(t)+1, & \text{if no fresh update is received},\ t-t_G^n(i), & \text{if a packet generated at } t_G^n(i) \text{ is received}. \end{cases}$$

This reset-or-increment structure recurs in vehicular, IoT, and MEC models, sometimes with the reset value written as one slot and sometimes as the actual generation-to-delivery delay (Wu et al., 2021, Chen et al., 2019, Chen et al., 18 Oct 2025).

A standard misconception is to equate freshness with either throughput or delay. The surveyed works explicitly separate these notions. One mURLLC formulation states that AoI differs from delay because it depends on both delays and update rates, and differs from throughput because it is a timeline metric reflecting whether the receiver’s information is up to date at every instant (Wang et al., 2024). The 5G uplink scheduling example with two nodes, one generating $L$ packets per slot and the other generating one packet per slot, makes the same point operationally: the more throughput-efficient policy can produce a larger AoI for the low-rate node (Wu et al., 2021).

Several variants extend the basic metric. Peak AoI considers the age immediately before a successful update, and one IoT multi-source queueing analysis studies the violation probabilities

$$P_i^{\mathrm{A}}(w_i):=\Pr\{\Delta_i>w_i\}, \qquad P_i^{\mathrm{P}}(p_i):=\Pr\{\Delta_i^{\mathrm{P}}>p_i\},$$

rather than only mean age (Zhang et al., 2022). In finite-blocklength mURLLC, the target is the peak AoI violation probability

$$p_{k,l}^{(\mu,\mathrm{AoI})} \triangleq \Pr\{P_{k,l}^{\mathrm{AoI}(\mu)} > A_{\mathrm{th}}\},$$

which is upper-bounded via stochastic network calculus (Wang et al., 2024). In semantics-aware wireless networked control, AoI is further generalized to task-aware Age of Actuation,

$\Delta_i(t)= t - A_i(t_{n_i(t)}),$

which resets only when a command is successfully executed, not merely received, and is paired with the Cost of Missing Actuation,

$$\mathrm{CMA} = \lim_{T \to \infty} \frac{1}{T} \sum_{t=0}^{T} \sum_{i \in \mathcal{T}} L_i(t) I_i(t),$$

to represent reliability losses under computation constraints (Zhao et al., 9 May 2026).

2. Optimization objectives and trade-off structure

AoI-aware formulations are typically multi-objective. A representative wireless scheduler maximizes

$\sum_{n=1}^{N}\bigl[U_n(t)-\beta A_n(t)\bigr],$

where $U_n(t)$ is a sigmoid-like utility based on allocated bandwidth and current arrivals, and $$A_n(t) := t - \max_i \{ t_G^n(i) \mid t_G^n(i) \le t \},$$0 tunes the throughput–freshness trade-off (Wu et al., 2021). In MIMO-NOMA IoT, the objective becomes a weighted sum of average AoI and average energy,

$$A_n(t) := t - \max_i \{ t_G^n(i) \mid t_G^n(i) \le t \},$$1

with per-slot cost $$A_n(t) := t - \max_i \{ t_G^n(i) \mid t_G^n(i) \le t \},$$2 (Zhu et al., 2023). In HAPS-V2X, AoI is minimized jointly with reliability and power through a reward of the form

$$A_n(t) := t - \max_i \{ t_G^n(i) \mid t_G^n(i) \le t \},$$3

so freshness, capacity thresholds, and energy expenditure are optimized simultaneously (Ince et al., 21 Jul 2025).

Other formulations place AoI directly in long-horizon objectives. In multi-BS IIoT offloading,

$$A_n(t) := t - \max_i \{ t_G^n(i) \mid t_G^n(i) \le t \},$$4

is optimized over binary offloading decisions and continuous bandwidth and CPU allocations (Chen et al., 18 Oct 2025). In multi-robot sensing and transport on graphs, the objective is

$$A_n(t) := t - \max_i \{ t_G^n(i) \mid t_G^n(i) \le t \},$$5

and the resulting lower bound decomposes into a sensing term $$A_n(t) := t - \max_i \{ t_G^n(i) \mid t_G^n(i) \le t \},$$6 and a propagation term $$A_n(t) := t - \max_i \{ t_G^n(i) \mid t_G^n(i) \le t \},$$7 (Tadrous, 4 May 2026). In semantics-aware control, the objective is bi-objective: minimize CoMA and the average AoA of regular tasks under an energy constraint (Zhao et al., 9 May 2026).

These formulations all encode a common structural fact: freshness competes with at least one other resource-oriented quantity. Depending on the system, the antagonistic variable is throughput, energy, rate reliability, payload completion, computation occupancy, or transport distance. This suggests that AoI-aware resource allocation is best viewed as constrained sequential control rather than as a single-metric scheduler.

3. Resource dimensions under AoI-aware control

The action space varies widely across applications, but the literature repeatedly treats freshness as coupled to multiple resource dimensions rather than to radio scheduling alone. In 5G uplink scheduling, the controller decides both which UEs are selected, $$A_n(t) := t - \max_i \{ t_G^n(i) \mid t_G^n(i) \le t \},$$8, and how many bandwidth units they receive, $$A_n(t) := t - \max_i \{ t_G^n(i) \mid t_G^n(i) \le t \},$$9, subject to $$A_n(t+1)= \begin{cases} A_n(t)+1, & \text{if no fresh update is received},\ t-t_G^n(i), & \text{if a packet generated at } t_G^n(i) \text{ is received}. \end{cases}$$0 (Wu et al., 2021). Platoon-based C-V2X and HAPS-V2X enlarge this to joint mode selection $$A_n(t+1)= \begin{cases} A_n(t)+1, & \text{if no fresh update is received},\ t-t_G^n(i), & \text{if a packet generated at } t_G^n(i) \text{ is received}. \end{cases}$$1, sub-channel assignment $$A_n(t+1)= \begin{cases} A_n(t)+1, & \text{if no fresh update is received},\ t-t_G^n(i), & \text{if a packet generated at } t_G^n(i) \text{ is received}. \end{cases}$$2, and transmit power $$A_n(t+1)= \begin{cases} A_n(t)+1, & \text{if no fresh update is received},\ t-t_G^n(i), & \text{if a packet generated at } t_G^n(i) \text{ is received}. \end{cases}$$3, because the same radio slot can serve either V2I freshness or V2V payload dissemination (Parvini et al., 2021, Ince et al., 21 Jul 2025).

In MEC and NTN systems, computation becomes co-equal with radio allocation. Hierarchical HAP–UAV offloading optimizes user powers, subcarriers, UAV–HAP forwarding, UAV CPU shares $$A_n(t+1)= \begin{cases} A_n(t)+1, & \text{if no fresh update is received},\ t-t_G^n(i), & \text{if a packet generated at } t_G^n(i) \text{ is received}. \end{cases}$$4, HAP CPU shares $$A_n(t+1)= \begin{cases} A_n(t)+1, & \text{if no fresh update is received},\ t-t_G^n(i), & \text{if a packet generated at } t_G^n(i) \text{ is received}. \end{cases}$$5, and UAV trajectories, all under an AoI objective tied to task completion rather than packet reception (Ansarifard et al., 2023). The multi-BS IIoT model uses binary offloading variables $$A_n(t+1)= \begin{cases} A_n(t)+1, & \text{if no fresh update is received},\ t-t_G^n(i), & \text{if a packet generated at } t_G^n(i) \text{ is received}. \end{cases}$$6 together with bandwidth $$A_n(t+1)= \begin{cases} A_n(t)+1, & \text{if no fresh update is received},\ t-t_G^n(i), & \text{if a packet generated at } t_G^n(i) \text{ is received}. \end{cases}$$7 and computation $$A_n(t+1)= \begin{cases} A_n(t)+1, & \text{if no fresh update is received},\ t-t_G^n(i), & \text{if a packet generated at } t_G^n(i) \text{ is received}. \end{cases}$$8, and the resource-allocation subproblem reduces AoI by minimizing the sum of transmission and computation delays (Chen et al., 18 Oct 2025).

Some systems add programmable propagation or energy constraints. RIS-aided IoV treats the RIS phase-shift matrix

$$A_n(t+1)= \begin{cases} A_n(t)+1, & \text{if no fresh update is received},\ t-t_G^n(i), & \text{if a packet generated at } t_G^n(i) \text{ is received}. \end{cases}$$9

as a freshness-control variable because it changes both V2I and V2V effective channels (Qi et al., 2024). Wireless-powered IoT chooses among WET, OMA, NOMA, and WET+OMA while tracking battery levels and outage-driven AoI resets (Chen et al., 2024). The smart-factory InF-S model allocates resource blocks and discrete power levels, but does so proactively by minimizing the probability that the predicted next-slot AoI exceeds a threshold $L$0 (Farag et al., 20 Apr 2025).

The same logic extends beyond conventional wireless links. In multi-robot sensing and transport, the resources are sensing robots $L$1 per node and mobile conveyors $L$2 along graph edges; their allocation directly shapes sensing times and hop-by-hop propagation delay (Tadrous, 4 May 2026). In semantics-aware WNCS, differentiated admission probabilities $L$3, transmit powers $L$4, and multi-rate computation units are the relevant resources, because freshness is measured at actuation time rather than reception time (Zhao et al., 9 May 2026).

4. Analytical and learning-based solution methods

A recurrent pattern is to cast the control problem as an MDP and learn a policy from state variables that explicitly include AoI. In the 5G uplink scheduler, the observation includes current AoI, queue status, and recent throughput, and PPO maximizes

$L$5

over long horizons (Wu et al., 2021). In MIMO-NOMA IoT, DDPG controls continuous power allocation after the sampling decision is analytically eliminated through the rule

$L$6

thereby reducing a mixed discrete–continuous AoI problem to continuous control (Zhu et al., 2023).

Multi-agent variants appear when interference and topology are distributed. The platoon-based C-V2X framework uses a global critic for interference-aware cooperation and local task-specific critics for AoI and CAM dissemination, while HAPS-V2X compares centralized-critic DDPG with fully decentralized FD-MADDPG, where each platoon leader learns from local observations only (Parvini et al., 2021, Ince et al., 21 Jul 2025). In RIS-aided IoV, SAC controls V2V power, channel reuse, and RIS phases jointly, leveraging entropy-regularized exploration in a high-dimensional continuous action space (Qi et al., 2024).

Other papers exploit problem structure to avoid purely black-box learning. The multi-robot graph problem yields a separable discrete convex resource-allocation problem,

$L$7

solved optimally by a greedy water-filling algorithm based on marginal benefit

$L$8

followed by an Euler-walk conveyor deployment (Tadrous, 4 May 2026). The IoT multi-source queueing paper proves convexity of the maximal AoI or PAoI violation probability with respect to arrival rates $L$9, and shows that optimal allocations equalize the per-source violation probabilities (Zhang et al., 2022). The IIoT multi-BS paper combines Branching-D3QN for combinatorial offloading with a strictly convex bandwidth/CPU allocation solved by CVX, exploiting the positive definiteness of the Hessian in $$P_i^{\mathrm{A}}(w_i):=\Pr\{\Delta_i>w_i\}, \qquad P_i^{\mathrm{P}}(p_i):=\Pr\{\Delta_i^{\mathrm{P}}>p_i\},$$0 (Chen et al., 18 Oct 2025).

Architectural innovation is increasingly prominent. A Transformer-enhanced PPO policy encodes users as tokens, uses self-attention

$$P_i^{\mathrm{A}}(w_i):=\Pr\{\Delta_i>w_i\}, \qquad P_i^{\mathrm{P}}(p_i):=\Pr\{\Delta_i^{\mathrm{P}}>p_i\},$$1

and learns to prioritize users with tight AoI thresholds and high penalty weights in NOMA offloading (Ansarifard et al., 26 Feb 2026). The proactive smart-factory scheme is not RL-based; instead it uses Bayesian Ridge Regression to predict next-slot AoI with mean $$P_i^{\mathrm{A}}(w_i):=\Pr\{\Delta_i>w_i\}, \qquad P_i^{\mathrm{P}}(p_i):=\Pr\{\Delta_i^{\mathrm{P}}>p_i\},$$2 and variance $$P_i^{\mathrm{A}}(w_i):=\Pr\{\Delta_i>w_i\}, \qquad P_i^{\mathrm{P}}(p_i):=\Pr\{\Delta_i^{\mathrm{P}}>p_i\},$$3, then minimizes

$$P_i^{\mathrm{A}}(w_i):=\Pr\{\Delta_i>w_i\}, \qquad P_i^{\mathrm{P}}(p_i):=\Pr\{\Delta_i^{\mathrm{P}}>p_i\},$$4

thereby coupling AoI-risk minimization and exploration (Farag et al., 20 Apr 2025). Earlier vehicular work already anticipated this temporal perspective through LSTM-based DRQN with per-VUE Q-function decomposition under partial observability (Chen et al., 2019).

5. Empirical regimes and reported performance

Several studies report that AoI-aware methods are most valuable under congestion, heterogeneity, or topology variation. In the 5G uplink scheduler, low-load operation produces similar AoI and throughput for round-robin, proportional fair, and PPO, whereas under high load round-robin has the worst AoI and throughput, proportional fair is best, and PPO approaches proportional fair without prior knowledge of traffic statistics (Wu et al., 2021). This suggests that freshness-aware control becomes operationally decisive only when resource scarcity forces nontrivial scheduling.

Topology-sensitive V2X results show similar patterns. In HAPS-V2X, FD-MADDPG yields average AoI of about $$P_i^{\mathrm{A}}(w_i):=\Pr\{\Delta_i>w_i\}, \qquad P_i^{\mathrm{P}}(p_i):=\Pr\{\Delta_i^{\mathrm{P}}>p_i\},$$5 ms versus about $$P_i^{\mathrm{A}}(w_i):=\Pr\{\Delta_i>w_i\}, \qquad P_i^{\mathrm{P}}(p_i):=\Pr\{\Delta_i^{\mathrm{P}}>p_i\},$$6 ms for DDPG at inter-platoon spacing $$P_i^{\mathrm{A}}(w_i):=\Pr\{\Delta_i>w_i\}, \qquad P_i^{\mathrm{P}}(p_i):=\Pr\{\Delta_i^{\mathrm{P}}>p_i\},$$7 m, and when spacing increases to $$P_i^{\mathrm{A}}(w_i):=\Pr\{\Delta_i>w_i\}, \qquad P_i^{\mathrm{P}}(p_i):=\Pr\{\Delta_i^{\mathrm{P}}>p_i\},$$8 m the FD-MADDPG AoI increases by only about $$P_i^{\mathrm{A}}(w_i):=\Pr\{\Delta_i>w_i\}, \qquad P_i^{\mathrm{P}}(p_i):=\Pr\{\Delta_i^{\mathrm{P}}>p_i\},$$9 ms while DDPG increases by almost $$p_{k,l}^{(\mu,\mathrm{AoI})} \triangleq \Pr\{P_{k,l}^{\mathrm{AoI}(\mu)} > A_{\mathrm{th}}\},$$0 ms (Ince et al., 21 Jul 2025). In RIS-aided IoV, increasing the number of RIS elements improves V2I sum rate and reduces V2I AoI, while SAC outperforms PPO, DDPG, TD3, and stochastic baselines in cumulative reward, convergence speed, AoI, and payload transmission probability (Qi et al., 2024).

Industrial and factory settings emphasize tail performance. The proactive InF-S method reports a reduction of $$p_{k,l}^{(\mu,\mathrm{AoI})} \triangleq \Pr\{P_{k,l}^{\mathrm{AoI}(\mu)} > A_{\mathrm{th}}\},$$1 in AoI violation probability compared to relevant baseline methods, and shows that moderate exploration improves both average AoI and AoI violation probability, whereas excessive exploration degrades both (Farag et al., 20 Apr 2025). The multi-BS IIoT framework reports at least a $$p_{k,l}^{(\mu,\mathrm{AoI})} \triangleq \Pr\{P_{k,l}^{\mathrm{AoI}(\mu)} > A_{\mathrm{th}}\},$$2 reduction in long-term average AoI and up to a $$p_{k,l}^{(\mu,\mathrm{AoI})} \triangleq \Pr\{P_{k,l}^{\mathrm{AoI}(\mu)} > A_{\mathrm{th}}\},$$3 enhanced convergence speed over comparison methods, with gains becoming larger as the number of devices grows (Chen et al., 18 Oct 2025).

Learning dynamics themselves have been analyzed as part of freshness control. In the Transformer actor-critic model, attention maps evolve from nearly uniform patterns early in training to focused patterns that privilege users with small AoI thresholds and large penalty sensitivities, revealing a learned priority structure aligned with NOMA constraints (Ansarifard et al., 26 Feb 2026). This suggests that, in heterogeneous settings, AoI-aware policies are not merely optimizing scalar ages; they are learning structured prioritization over users, channels, and interference couplings.

6. Misconceptions, limitations, and current directions

A second misconception is that all age-based objectives are interchangeable. The cited literature treats average AoI, peak AoI, AoI violation probability, AoA, and CoMA as distinct objects. Peak AoI and its violation probability emphasize tail behavior (Zhang et al., 2022, Wang et al., 2024); AoA shifts the freshness reference from reception to execution and therefore introduces controller-service availability $$p_{k,l}^{(\mu,\mathrm{AoI})} \triangleq \Pr\{P_{k,l}^{\mathrm{AoI}(\mu)} > A_{\mathrm{th}}\},$$4 and actuation delays into the age metric (Zhao et al., 9 May 2026). These differences matter because an algorithm designed for average AoI need not control age tails or execution freshness.

The surveyed works also rely on strong modeling assumptions. Some 5G and NOMA schedulers assume error-free or idealized successful transmission once resources are allocated (Wu et al., 2021, Ansarifard et al., 26 Feb 2026). Several models are single-cell, single-HAPS, or single-RSU abstractions that omit multi-cell coupling, inter-controller coordination, or mobility of infrastructure (Ince et al., 21 Jul 2025, Qi et al., 2024). The smart-factory predictor uses one-slot AoI forecasting and fixed exploration weights (Farag et al., 20 Apr 2025). The robot-sensing framework assumes a static graph, unit-time edges, and geometric sensing with diminishing returns (Tadrous, 4 May 2026). Semantics-aware AoA analysis assumes an error-free downlink and no buffering at the controller (Zhao et al., 9 May 2026).

Current directions therefore revolve around relaxing these simplifications while preserving freshness-awareness. Explicitly proposed extensions include energy-efficient learning strategies, adaptive reward mechanisms, HAPS mobility, multi-HAPS deployment, federated learning, and more sophisticated MARL in NTN/V2X (Ince et al., 21 Jul 2025); dynamic or directed graphs, non-unit edge times, non-geometric sensing, and energy-constrained conveyors in robot transport (Tadrous, 4 May 2026); multi-slot AoI prediction, multi-agent coordination, and adaptive exploration parameters in InF-S (Farag et al., 20 Apr 2025); service migration and more complex MEC topologies in multi-BS IIoT (Chen et al., 18 Oct 2025). A plausible implication is that future AoI-aware resource allocation will become increasingly hybrid: analytic where convexity or discrete convexity is available, and learning-based where state, action, and interference structures are too rich for closed-form control.