Mobility-First Scheduling Rule

• Mobility-First Scheduling Rule is a set of principles that evaluate mobility-induced availability as the initial filter before final service selection.
• It encompasses various formulations such as predictive location filtering, integrated future service optimization, current-cell priority, and topology-aware control.
• Empirical and theoretical studies show that this approach enhances task completion rates, reduces response times, and improves network throughput in dynamic environments.

Mobility-First Scheduling Rule denotes a family of scheduling principles in which mobility-induced availability, connectivity, or contact opportunity is evaluated before final service selection. In the cited literature, the term does not refer to a single canonical algorithm. Instead, it appears in at least four distinct forms: predictive recruitment based on future location estimates, integrated-service optimization over a scheduling horizon, local service conditioned on current cell membership, and opportunistic priority among users in favorable channel states. A separate line of work shows that some mobility-aware schedulers are not strictly mobility-first at all, because they embed mobility into a broader backlog- and utility-weighted control law rather than assigning unconditional priority to mobility opportunities (Shao et al., 18 Dec 2025, Zhang et al., 2015, Bhattacharjee et al., 2020, Neely, 2010, Ayesta et al., 2011).

1. Taxonomy of the concept

The term is used differently across network models, objective functions, and timescales.

Interpretation	Representative rule	Representative source
Predictive feasibility filter	$$x_{ij} = 0, \quad \text{if } D_j \notin \text{Predicted\_Region}(T_i, d_i)$$	(Shao et al., 18 Dec 2025)
Integrated future service	$S(T_k) = \int_{t_k}^{t_k+T} C(t)\,dt$	(Zhang et al., 2015)
Current-cell local priority	$$i_j^\star(t) \in \arg\max_{i \in \mathcal U_j(t)} h_i(t)$$	(Bhattacharjee et al., 2020)
Topology-aware max-weight control	$\omega(t)=[A(t);S(t)]$ with backlog-weighted control	(Neely, 2010)
Best-state opportunism with myopic tie-breaking	serve users in state $N_k$; tie-break by highest $c_k\mu_{k,N_k}$	(Ayesta et al., 2011)

These formulations share a common structural motif: mobility or state variation constrains which actions are meaningful at a given decision epoch. They differ, however, on whether mobility is a hard feasibility gate, an optimization metric, or merely one component of a broader control objective. This suggests that “mobility-first” is best understood as a comparative label for schedulers that privilege mobility-conditioned serviceability earlier in the decision pipeline than deadline-only, queue-only, or static-utility policies.

2. Predictive mobility as the primary feasibility filter

A direct and explicit formulation appears in AG-MPBS, where task recruitment is organized around behavior-aware mobility prediction rather than deadline-only or static utility ranking (Shao et al., 18 Dec 2025). The framework consists of three modules: a behavior-aware KNN classifier, a time-varying Markov prediction model, and a dynamic priority scheduling mechanism that considers task urgency and base station performance. The rule is explicit: a device is eligible for a task only if mobility prediction indicates that it is likely to appear in the task region during the task window,

$$x_{ij} = 0, \quad \text{if } D_j \notin \text{Predicted\_Region}(T_i, d_i).$$

Mobility prediction is behavior-adaptive. Devices are classified using KNN over visit frequency, location entropy, average displacement, and dwell time into four classes: Regular movers, Semi-regular movers, Localized movers, and Random movers. The reported classification accuracies are Decision Tree: 93.75%, Naive Bayes: 93.25%, and KNN: 95.55%. The class label determines the predictor: localized movers use accumulated dwell-time logic, semi-regular movers use a second-order Markov model, random movers use a third-order Markov model, and regular movers use a time-varying Markov model.

For regular movers, the day is divided into four periods: 0:00–7:00 rest period, 7:00–10:00 commute to work, 10:00–17:00 work hours, and 17:00–24:00 commute home. A separate transition matrix $P^{(t)}$ is learned for each period, yielding a time-conditioned location distribution $\pi_d^{(t)}$. Recruitment is then quantified through

$E_i = \sum_{j=1}^{m} P^{(t)}_{j,l_i} \cdot \rho_j,$

where $S(T_k) = \int_{t_k}^{t_k+T} C(t)\,dt$0 is the probability that device $S(T_k) = \int_{t_k}^{t_k+T} C(t)\,dt$1 appears in region $S(T_k) = \int_{t_k}^{t_k+T} C(t)\,dt$2 at time $S(T_k) = \int_{t_k}^{t_k+T} C(t)\,dt$3, and $S(T_k) = \int_{t_k}^{t_k+T} C(t)\,dt$4 is the historical reliability score of device $S(T_k) = \int_{t_k}^{t_k+T} C(t)\,dt$5. The threshold

$S(T_k) = \int_{t_k}^{t_k+T} C(t)\,dt$6

determines whether a task is executable locally.

Only after this mobility-conditioned feasibility stage are tasks ranked. The priority score is

$S(T_k) = \int_{t_k}^{t_k+T} C(t)\,dt$7

where $S(T_k) = \int_{t_k}^{t_k+T} C(t)\,dt$8 is urgency and $S(T_k) = \int_{t_k}^{t_k+T} C(t)\,dt$9 is base-station reliability built from success rate, average response delay, and utilization rate. Tasks are sorted in descending order of $$i_j^\star(t) \in \arg\max_{i \in \mathcal U_j(t)} h_i(t)$$0, and devices are assigned only if they are predicted to appear at the task location, are not already assigned, and the required device count has not yet been reached. In this formulation, mobility is not a secondary feature; it is the gatekeeper for admissible assignments.

3. Mobility as integrated service over the scheduling horizon

A second, equally explicit, interpretation appears in relay-aided vehicular scheduling for high-mobility networks, where mobility-first means optimizing the integrated transmission capability over an interval rather than the instantaneous rate at a single time (Zhang et al., 2015). The central quantity is the mobile service amount,

$$i_j^\star(t) \in \arg\max_{i \in \mathcal U_j(t)} h_i(t)$$1

which replaces instantaneous achievable information rate as the scheduling metric.

The system comprises one roadside base station, $$i_j^\star(t) \in \arg\max_{i \in \mathcal U_j(t)} h_i(t)$$2 moving vehicles, V2I communication via LTE-A, and V2V communication via DSRC / IEEE 802.11p. Vehicles are partitioned into relaying vehicles (RVs), aided vehicles (AVs), and common vehicles (CVs). Decode-and-forward relaying is assumed, so the end-to-end instantaneous rate for AV $$i_j^\star(t) \in \arg\max_{i \in \mathcal U_j(t)} h_i(t)$$3 via relay $$i_j^\star(t) \in \arg\max_{i \in \mathcal U_j(t)} h_i(t)$$4 is

$$i_j^\star(t) \in \arg\max_{i \in \mathcal U_j(t)} h_i(t)$$5

The corresponding mobile service quantities are

$$i_j^\star(t) \in \arg\max_{i \in \mathcal U_j(t)} h_i(t)$$6

$$i_j^\star(t) \in \arg\max_{i \in \mathcal U_j(t)} h_i(t)$$7

$$i_j^\star(t) \in \arg\max_{i \in \mathcal U_j(t)} h_i(t)$$8

The optimization problem is to choose RVs, AVs, CVs, and the one-to-one relay pairing to maximize total mobile service, subject to partition constraints

$$i_j^\star(t) \in \arg\max_{i \in \mathcal U_j(t)} h_i(t)$$9

$\omega(t)=[A(t);S(t)]$0

and

$\omega(t)=[A(t);S(t)]$1

The pairing matrix satisfies

$\omega(t)=[A(t);S(t)]$2

A key structural fact is Lemma 1: if AV $\omega(t)=[A(t);S(t)]$3 is paired with RV $\omega(t)=[A(t);S(t)]$4, then

$\omega(t)=[A(t);S(t)]$5

Hence vehicles with smaller V2I mobile service are natural AV candidates, whereas vehicles with larger V2I mobile service remain direct users or relays. The proposed MSRS algorithm computes $\omega(t)=[A(t);S(t)]$6, sorts vehicles by V2I mobile service, uses the Hungarian algorithm to solve the assignment problem for a fixed $\omega(t)=[A(t);S(t)]$7, and then uses binary search or golden-section search over $\omega(t)=[A(t);S(t)]$8. Its total complexity is

$\omega(t)=[A(t);S(t)]$9

Here mobility-first means that the scheduler optimizes what the links can deliver over the whole interval under mobility, not merely what they can deliver at the instant when the decision is made.

4. Mobility-aware control without strict mobility priority

A different perspective is provided by universal scheduling for networks with arbitrary traffic, channels, and mobility. That framework is mobility-aware but explicitly not mobility-first in the sense of prioritizing mobility over backlog or utility (Neely, 2010). The network state is represented as an arbitrary sample path

$N_k$0

and in the dynamic network specialization

$N_k$1

where $N_k$2 denotes exogenous arrivals and $N_k$3 the topology state on slot $N_k$4. Mobility is therefore modeled indirectly through time-varying topology, link availability, and capacities, not through an explicit physical motion model.

The controller observes $N_k$5 only at the beginning of slot $N_k$6, with no probabilistic model required. The algorithm chooses decisions slot-by-slot using only the present state and queue backlogs. Its control-action step minimizes

$N_k$7

In the network model, resource allocation is chosen to maximize

$N_k$8

with differential backlog weights

$N_k$9

This formulation is important because it delineates a conceptual boundary. Mobility changes the feasible action set $c_k\mu_{k,N_k}$0 and the rate opportunities $c_k\mu_{k,N_k}$1, but the priority metric remains backpressure. The paper is explicit that the scheduler exploits mobility-induced links opportunistically when they align with backlog and utility weights; it does not prefer mobility opportunities unconditionally. A common misconception is therefore that any mobility-aware max-weight rule is mobility-first. In this framework, the more precise description is mobility-aware, topology-aware, and backpressure-driven.

5. Current-state mobility and opportunistic tie-breaking

In adversarial Age-of-Information scheduling, mobility-first appears in a fully local and causal form (Bhattacharjee et al., 2020). The system has $c_k\mu_{k,N_k}$2 mobile users, $c_k\mu_{k,N_k}$3 base stations, disjoint cell coverage regions, no CSIT, and channel and mobility processes that may be dictated by an adversary. Each BS can serve at most one user in its current coverage area per slot. The performance metric is the maximum AoI,

$c_k\mu_{k,N_k}$4

with cumulative cost

$c_k\mu_{k,N_k}$5

The Cellular Max-Age rule is

$c_k\mu_{k,N_k}$6

where $c_k\mu_{k,N_k}$7 is the set of users currently in cell $c_k\mu_{k,N_k}$8. The mobility-first aspect is that service decisions are conditioned on current mobility-induced cell membership before any AoI comparison is made within that cell.

A related but distinct formulation appears in flow-level scheduling in a random environment (Ayesta et al., 2011). There, the central structural result is that maximum stable policies are Best Rate (BR) policies: whenever there are users present that are currently in their best possible channel condition, i.e., state $c_k\mu_{k,N_k}$9, such a user is served. The maximum stability condition is

$$x_{ij} = 0, \quad \text{if } D_j \notin \text{Predicted\_Region}(T_i, d_i).$$0

Among users already in their best state, the paper defines a myopic tie-breaking rule that selects the user with the largest value of $$x_{ij} = 0, \quad \text{if } D_j \notin \text{Predicted\_Region}(T_i, d_i).$$1. Best Rate Priority (BRP) policies are BR policies with this myopic tie-breaking rule. The distinction is crucial: BR guarantees maximum stability, while BRP is the family shown to be asymptotically fluid optimal.

These two formulations illuminate different senses of “first.” In CMA, mobility first means local reachability is the first filter. In BRP, the first filter is best-state opportunism, and the second-stage tie-breaker prioritizes the user with the highest departure probability. The latter is only mobility-first in an extended sense tied to random environmental state rather than explicit geometric mobility.

6. Performance guarantees, empirical behavior, and conceptual limits

The empirical and theoretical guarantees associated with mobility-first formulations vary with the model class. In AG-MPBS, experiments use the GeoLife GPS trajectory dataset with 182 users, 17,621 trajectories, over 24 million GPS points, Beijing region, 300 simulated devices, 1000 tasks per run, 96 time slots per day, and 10 repetitions. The reported performance measures are Task Completion Rate (TCR), Average Response Time (ART), and Device Utilization (DU). MPBS is reported to outperform Greedy, HSF, and PPO across node scales, to achieve higher TCR, lower ART, and higher DU, and to maintain task completion rate above 97% in resource-limited settings; other comparisons include EDF and LSF (Shao et al., 18 Dec 2025).

In vehicular relaying, MSRS is reported to achieve the optimal results with an optimal approximation ratio larger than 96.5%. The reported gains are an increment of 3.63% compared with instantaneous achievable information rate based scheduling and 15% compared with traditional non-cooperation scheduling. The paper also states that MSRS degrades more slowly than IRRS as vehicle speed increases, and that the advantage is most pronounced in medium-to-high mobility regimes (Zhang et al., 2015).

In universal mobility-aware scheduling, the main guarantees are benchmark-relative rather than empirical. For arbitrary sample paths, queue growth satisfies

$$x_{ij} = 0, \quad \text{if } D_j \notin \text{Predicted\_Region}(T_i, d_i).$$2

so queue magnitude grows at most like $$x_{ij} = 0, \quad \text{if } D_j \notin \text{Predicted\_Region}(T_i, d_i).$$3 in the arbitrary-sample-path setting. Against an ideal $$x_{ij} = 0, \quad \text{if } D_j \notin \text{Predicted\_Region}(T_i, d_i).$$4-slot lookahead policy, the paper emphasizes a utility-gap form

$$x_{ij} = 0, \quad \text{if } D_j \notin \text{Predicted\_Region}(T_i, d_i).$$5

In the ergodic special case, the same policy approaches the optimal infinite-lookahead or stationary randomized optimum $$x_{ij} = 0, \quad \text{if } D_j \notin \text{Predicted\_Region}(T_i, d_i).$$6 up to the stated $$x_{ij} = 0, \quad \text{if } D_j \notin \text{Predicted\_Region}(T_i, d_i).$$7-dependent terms (Neely, 2010).

In adversarial AoI scheduling, the central guarantee for CMA is

$$x_{ij} = 0, \quad \text{if } D_j \notin \text{Predicted\_Region}(T_i, d_i).$$8

while the converse result states that for any online policy $$x_{ij} = 0, \quad \text{if } D_j \notin \text{Predicted\_Region}(T_i, d_i).$$9,

$P^{(t)}$0

Hence CMA is competitively optimal up to an $P^{(t)}$1 factor (Bhattacharjee et al., 2020). In the random-environment flow model, BR policies are maximum stable under the condition in (17), and Theorem VI.3 states that any BRP policy is asymptotically fluid optimal (Ayesta et al., 2011).

Taken together, these results delimit the scope of the concept. A mobility-first scheduling rule may be predictive or non-predictive; centralized or distributed; throughput-oriented, utility-oriented, or freshness-oriented; and may or may not be queue-aware. What unifies the category is not a shared optimization objective but a common ordering principle: mobility-conditioned serviceability enters before final prioritization. The principal conceptual limit is that this ordering is not universal. Some of the strongest mobility-aware results arise precisely in frameworks where mobility is only one state component inside a larger drift-plus-penalty or backpressure criterion, rather than the dominant priority variable (Neely, 2010).