Risk-Budgeted Online Scheduling for Continuous Edge Inference over Evolving Time Horizons ∗
Houyi Qi∗ , Minghui Liwang∗ , Sai Zou† , Wei Ni‡
arXiv:2605.02179v1 [cs.NI] 4 May 2026
Shanghai Research Institute for Intelligent Autonomous Systems, Tongji University, Shanghai, China † College of Big Data and Information Engineering, Guizhou University, Guizhou, China ‡ School of Engineering, Edith Cowan University, Perth, Australia Email: {[email protected], [email protected], [email protected], [email protected]} Abstract—Continuous edge inference necessitates not merely low per-timeslot latency, but sustained timeliness guarantees in the presence of time-varying channels, fluctuating edge workloads, and coupled bandwidth-computing resource constraints. Existing studies predominantly optimize instantaneous delay or per-timeslot utility, while largely overlooking the regulation of cross-time deadline violation dynamics in continuous services. To address this, we propose AEGIS, a prediction-empowered risk-budgeted online scheduling framework for continuous edge inference. AEGIS models deadline-violation tendency as an updatable cross-time control state through dynamic user-level risk budgets, so that online scheduling accounts for both instantaneous efficiency and long-term service stability. To support proactive decision making, AEGIS leverages LSTM-based shortterm state prediction to construct a smooth deadline-violation risk surrogate, and formulates the resulting time-wise resource competition as a potential-aligned game under coupled feasibility constraints. An asynchronous online algorithm is then developed with finite-step convergence. Experiments demonstrate that AEGIS improves the timely inference ratio, reduces average violation risk and violation burst length, and achieves a favorable delay–risk–convergence trade-off over representative baselines. Index Terms—Continuous edge inference, resource scheduling, deadline-sensitive services, risk budget, potential game
I. I NTRODUCTION The rapid innovation of next-generation communication technologies, edge computing, and artificial intelligence (AI) has enabled the proliferation of delay-sensitive intelligent applications, such as continuous video analytics, augmentedreality interaction, real-time speech recognition, and wearable sensing, which increasingly rely on edge-side inference for real-time responses [1], [2]. As such, end devices may continuously generate inference tasks and transfer them to nearby edge platforms, thereby alleviating local computing burden [3], [4]. Unlike conventional one-shot task dispatching, continuous edge inference necessitates both low latency at each decision epoch and long-term timeliness preservation under time-varying wireless channels, dynamic edge workloads, and intertwined communication-computation constraints [3]. However, under persistent channel degradation, edge congestion, and multi-user contention, deadline violations may propagate across consecutive timeslots even if individual scheduling decisions appear feasible. This inter-temporal accumulation of violation risk can cause sustained service degradation for users under unfavorable operating conditions. Therefore, enabling continuous edge inference requires going beyond instantaneous latency minimization and explicitly regulating deadline-violation risk over time.
Existing studies on edge inference/computing generally followed a performance-centric resource optimization paradigm. Representative works improve inference efficiency through joint communication–computing optimization, such as multiuser co-inference with joint model splitting and resource assignment [5], throughput-oriented batching and early exiting [6], and timely joint resource scheduling at the wireless edge [4]. More recent efforts have considered dynamic service management, including service-level-objective (SLO)aware scheduling [7] and deep-reinforcement-learning-based orchestration for edge inference [8]. Most existing efforts remain focused on instantaneous delay, SLO compliance, or per-slot utility under observed network states, without explicitly characterizing the temporal propagation and accumulation of deadline-violation risk in continuous services. Although deadlines, timeliness, and remaining time margins have been used to guide immediate decisions [4], [7], to the best of our knowledge, a unified control framework that explicitly embeds violation propensity as a budget-constrained, dynamically refreshable inter-temporal state variable within joint bandwidth and edge-computing resource assignment remains unexplored. To address the above challenges, we propose AEGIS, a prediction-assisted cross-time risk-budgeted online scheduling framework for joint bandwidth and computing resource scheduling in continuous edge inference. Main contributions are summarized as follows. • To counteract long-term service degradation induced by time-varying wireless channels, fluctuating edge workloads, and multi-user resource contention, we formulate continuous edge inference scheduling as a cross-temporal, risk-budgeted online assignment problem that jointly captures deadline constraints, communication–computation coupling, and dynamic user-level risk budgets. Building upon this, we develop AEGIS, which leverages LSTM-based short-term state prediction to construct forward-looking deadline-violation risk, explicitly regulates violation tendency through adaptive user-level risk budgets, and casts the resulting time-wise resource competition as a potential-aligned game solved by an asynchronous online algorithm with finite-step convergence. • We theoretically prove the key properties of AEGIS, including the exact potential property, pure-strategy equilibrium existence, and finite-step convergence under feasible unilateral deviations. Extensive experiments demonstrate that AEGIS significantly improves timely inference capability, suppresses deadline-violation exposure and violation burstiness, and reduces convergence overhead over representative baselines.
Fig. 1. Framework and procedure of AEGIS, including edge inference task offloading, LSTM-based channel/load prediction, deadline-violation risk construction, dynamic risk-budget update, and potential-aligned game scheduling and bandwidth-computing resource assignment.
II. S YSTEM OVERVIEW AND M ODELING As shown in Fig. 1, we consider a wireless edge inference system where mobile users (MUs) may continuously generate deadline-sensitive inference tasks and request services from an edge service platform (ESP). In each timeslot, only active MUs can participate in scheduling, and for each active one, the ESP jointly handles admission control and bandwidth-computing resource assignment, deciding task acceptance and the allocation levels of resources. The resulting decision determines the end-to-end service latency, including transmission, queuing, and execution delays, and hence task timeliness. To support stable continuous inference, AEGIS first predicts channel and edge-load states, constructs a deadline-violation risk surrogate, and updates dynamic risk budgets. These risk budgets then guide a potential-aligned game scheduler to make time-wise admission and resource scheduling decisions, so as to balance instantaneous timeliness and long-term service stability under dynamic wireless channels, fluctuating edge workloads, and multi-user contention. A. Continuous Edge Inference Task Model Let U = {u1 , . . . , uN } denote the set of MUs. We consider a discrete time horizon T = {1, . . . , T }, where each timeslot has an equal duration of τ . To capture the dynamism of edge inference demand, let χi (t) ∈ {0, 1} denote the task-arrival indicator of MU ui at timeslot t, where χi (t) = 1 indicates that ui generates an active inference task and χi (t) = 0 otherwise. Accordingly, the active-MU set at timeslot t is U A (t) = {ui ∈ U | χi (t) = 1}. For each active MU ui ∈ U A (t), its inference task is characterized by Ji (t) = Li (t), Ci (t), Dimax , wi , where Li (t) is the input data size, Ci (t) denotes the required computing workload, Dimax is the deadline, and wi represents the task weight. Each active task can be transferred to the ESP through an uplink wireless link for edge execution. B. Delay and Timely Service Model For each active MU ui ∈ U A (t), let bi (t) and fi (t) denote the assigned bandwidth and computing resources, respectively. Then, we define the end-to-end (E2E) delay as queue di (t) = dtx (t) + dcomp (t). (1) i (t) + di i Accordingly, the uplink transmission rate is R (t) = i bi (t) log2 1 + SINRi (t) , and the corresponding transmission
delay is thus calculated by dtx i (t) = Li (t)/Ri (t); also, the computing delay is dcomp (t) = Ci (t)/fi (t). To characterize i possible congestion at ESP, let Q(t) denote the background queued workload at timeslot t. Then, the queuing delay experienced by a newly admitted task is approximated by Q(t) dqueue (t) = tot P , (2) i F − uj ∈U A (t) fj (t) + ϵf where F tot is the total computing capacity of ESP and ϵf > 0 describes a small constant introduced for numerical stability. This surrogate captures the fact that the queuing delay increases with the backlog Q(t) and decreases with the residual service capacity available at ESP. The backlog state Q(t) evolves according to the aggregate admitted workload and the available service capacity of ESP; only its realized value and one-step prediction are required by the proposed scheduler. A task is regarded as timely served if its E2E delay does not exceed the deadline.(Accordingly, we have ∀ui ∈ U A (t): 1, if di (t) ≤ Dimax ; (3) 0, otherwise, For inactive MUs with χi (t) = 0, we set bi (t) = fi (t) = 0 and exclude them from the current scheduling process. For notational simplicity, the delay expressions apply to admitted tasks with bi (t) > 0 and fi (t) > 0. If an active MU is not scheduled in timeslot t, i.e., bi (t) = fi (t) = 0, the task is regarded as not admitted, and we set di (t) = +∞ and Ψi (t) = 0. The same convention is applied to the predicted delay and risk surrogate later. Ψi (t) =
C. State Prediction and Risk Construction Since both wireless conditions and edge-load states vary over time, resource assignment based only on current observations may underestimate the risk of future deadline violations. To enable proactive scheduling, we first introduce an LSTMbased short-term predictor. For each MU ui , let xch i (t) = gi (t − H + 1), . . . , gi (t) denote the historical channel observation sequence, where gi (t) represents the observed channel quality at timeslot t and H is the observation-window length. Similarly, let xedge (t) = Q(t − H + 1), . . . , Q(t) denote the historical edge-load sequence. Based on these observations, the predictor outputs the one-step-ahead channel forecast ĥi (t + 1) = LSTMch xch (t) and the one-step-ahead edge-load forecast i q̂(t + 1) = LSTMedge xedge (t) . For notational simplicity, the prediction results used for timeslot-t decision-making are denoted by ĥi (t) and q̂(t) in the sequel. Notably, the predictor does not directly output assignment decisions. Instead, it provides predicted operating states for online scheduling. To make the assignment decision explicit, we define the time-wise action of MU ui as ai (t) = bi (t), fi (t) , where bi (t) and fi (t) denote the allocated bandwidth and computing resource, respectively. A positive action with bi (t) > 0 and fi (t) > 0 means that the task of MU ui is admitted and scheduled, while the null action (0, 0) means that the task is not admitted in the current timeslot. Let a(t) = a1 (t), . . . , a|U | (t) denote the joint action profile. Given a(t), the predicted E2E delay of active MU ui is given by ˆqueue (a(t); t) + dˆcomp (a(t); t), (4) dˆi (a(t); t) = dˆtx i (a(t); t) + di i
where dˆtx i (a(t); t) =
Li (t)
, dˆqueue (a(t); t) =
i \ i (t) bi (t) log2 1+SINR q̂(t) comp ˆ P (a(t); t) = F tot − u ∈U A (t) fj (t)+ϵf , and di j
Ci (t) fi (t) .
Here,
s.t. X ui ∈U A (t)
bi (t) ≤ B tot ,
X
fi (t) ≤ F tot ,
∀t,
(8a)
ui ∈U A (t)
\ i (t) depends on the predicted channel state ĥi (t). SINR If an active MU takes the null action (0, 0), the corresponding task is regarded as not admitted, and we directly set dˆi (a(t); t) = +∞, r̂i (a(t); t) = 1, and ŝi (a(t); t) = 0. Based on the predicted delay, we define the deadline margin of MU ui at timeslot t as (5). Apparently, a smaller margin indicates that the task is closer to deadline violation. ∆i (a(t); t) = Dimax − dˆi (a(t); t). (5) To obtain a smooth and bounded measure for online optimization, we define the task-level deadline-violation risk surrogate as follows: 1 , r̂i (a(t); t) = (6) 1 + exp κi ∆i (a(t); t) where κi > 0 is the risk-sensitivity coefficient. Notably, r̂i (a(t); t) is a prediction-driven risk surrogate rather than an exact posterior probability. It provides a smooth approximation of how close the current task is to the deadline-violation boundary under predicted states. Correspondingly, the predicted timely-service surrogate is ŝi (a(t); t) = 1 − r̂i (a(t); t). This quantity will later be used to construct the time-wise scheduling objective, while the binary indicator Ψi (t) is reserved for long-term formulation and performance evaluation.
bi (t) = fi (t) = 0, ∀ui ∈ / U A (t), ∀t, (8c) (7) holds, ∀ui ∈ U, ∀t, (8d) where B tot and F tot denote the total uplink bandwidth and edge computing capacity, respectively. Constraint (8a) limits the aggregate bandwidth and computing-resource assignment in each timeslot. (8b) requires the predicted risk of each admitted active MU to stay within its remaining risk budget, while the null action (0, 0) is always feasible and represents task rejection. (8c) prevents inactive MUs from occupying resources, and (8d) captures the cross-time evolution of risk budgets through risk consumption and recovery. The above long-horizon problem couples binary timely-service decisions, cross-time budget dynamics, and multi-MU resource constraints, making direct optimization intractable. We next replace the original objective with a time-wise surrogate based on ŝi (a(t); t) and reformulate the problem as an online game under coupled feasibility constraints.
shared resource limits and per-MU risk budgets: X X max χi (t) wi Ψi (t)
Φ(a(t); t) = X r̂i (a(t); t) wi ŝi (a(t); t) − αi bi (t) − βi fi (t) − γi , (12) Bi (t) + ϵ A
r̂i (a(t); t) ≤ Bi (t), ∀ui ∈ U A (t) with ai (t) ̸= (0, 0), ∀t, (8b)
III. D ESIGN OF AEGIS A. Time-Wise Risk-Budgeted Scheduling Game At each timeslot, active MUs compete for limited wireless and computing resources under shared constraints. Different D. Dynamic Risk Budget and Problem Formulation from conventional myopic assignment, the current scheduling decision is guided not only by instantaneous resource availTo regulate long-term violation exposure across varying ability, but also by the predicted channel state, predicted edgetime, we introduce a dynamic risk budget Bi (t) for each MU load state, and each MU’s remaining risk budget. To enable ui , representing the remaining tolerable risk quota available to low-overhead online scheduling, we replace the long-term ui at timeslot t. Assume Bi (t) ∈ [0, Bimax ], with Bimax ≤ 1, objective with a time-wise surrogate based on the predicted and initialize the budget as Bi (1) = Bimax . Given the timeslot- timely-service measure ŝ (a(t); t) = 1−r̂ (a(t); t), and model i i t scheduling decision, the budget is updated as the resulting interaction as a risk-budgeted scheduling game. n o Bi (t + 1) = min Bimax, max 0, Bi (t) − χi (t)r̂i (a⋆ (t); t) + ρi , Formally, the timeslot-t game is defined as (7) G(t) = U, {Ai (t)}ui ∈U , {Ui (·; t)}ui ∈U , (9) ⋆ where a (t) denotes the timeslot-t scheduling decision, and ρi is the budget recovery rate. The factor χi (t) ensures where Ai (t) is the feasible action set of MU ui and Ui (·; t) is that only active MUs consume risk budget, while inactive its utility function. Following the resource-action definition in ones do not incur risk consumption. This forms a cross- Sec. II, the bandwidth and computing assignment decisions are time, closed-loop optimization: at the beginning of timeslot discretized into finite candidate sets B̃i and F̃i , respectively. t, MUs enter the scheduler with current budget states Bi (t). For active MUs, the discrete action set is given by Ai (t) = {(0, 0)} ∪ {(b, f ) | b ∈ B̃i , f ∈ F̃i }, (10) After the time-wise scheduling is determined, ESP updates the budgets according to the induced prediction-driven risk Inactive MUs are excluded from resource competition and surrogates and obtains the next-timeslot state Bi (t + 1). keep the fixed action (0, 0). Because MUs compete for limited Hence, MUs that repeatedly stay in unfavorable conditions will bandwidth and computation resources, with each active MU gradually deplete their budgets and become more restricted subject to a budget-constrained risk exposure requirement, the in subsequent competition, whereas MUs experiencing benign feasible joint-action set is defined as n o conditions can gradually restore their risk tolerance. In this Afeas (t) = a(t) (8a), (8b) hold, ∀ui ∈ U A (t) . (11) way, deadline-violation tendency is explicitly regulated across timeslots rather than being treated as a passive byproduct of Since inactive MUs can only choose the null action, (8a) and (8b) effectively apply to active MUs. To further characterize instantaneous optimization. Next, we design the long-term scheduling objective to the time-wise scheduling objective, we define the potentialmaximize the weighted timely-service utility while respecting aligned surrogate as
{bi (t),fi (t)}
t∈T ui ∈U
(8)
ui ∈U
(t)
where αi , βi , and γi are the coefficients for bandwidth cost, computing cost, and risk regularization, respectively; and ϵ > 0 is a small positive constant introduced for numerical stability when the remaining risk budget is close to zero. To align time-wise MU incentives with the above surrogate, the utility of MU ui is defined as its marginal contribution: Ui (a(t); t) = Φ(a(t); t) − Φ (0, 0), a−i (t); t , (13) where (0, 0), a−i (t) denotes the joint action profile in which MU ui takes the null action while all other MUs keep their current actions. Because the feasibility of one MU’s deviation depends on the actions of the others, we define the feasible unilateral action setn of MU ui under a−i (t) as o Afeas (a−i (t); t) = ai (t) ∈ Ai (t) i
Input: U , {B̃i , F̃i }, B tot , F tot , historical channel/load observations, current risk budgets {Bi (t)}, parameters {αi , βi , γi , wi }, threshold ε′ , and Kmax . 2 Use LSTM predictors to obtain ĥi (t) and q̂(t). 3 Construct the predicted-delay, risk-surrogate, and timely-service functions dˆi (a(t); t), r̂i (a(t); t), and ŝi (a(t); t), together with Afeas (t) and Φ(a(t); t). (0) (t). 4 Set k ← 0 and generate a feasible joint action a 5 while k < Kmax do (k) 6 For each active MU ui ∈ U A (t), search Afeas (a−i (t); t) and i (k)
obtain a candidate action ãi (t) with potential gain (k) (k) (k) ∆i (t) = Φ(ãi (t), a−i (t); t) − Φ(a(k) (t); t). 7
ai (t), a−i (t) ∈ Afeas (t) .
(14) A joint action a (t) ∈ A (t) is called a pure-strategy equilibrium under feasible unilateral deviations if, for every MU ui ∈ U, we have Ui a⋆ (t); t ≥ Ui ai , a⋆−i (t); t , ∀ai ∈ Afeas (a⋆−i (t); t). (15) i At such a point, no MU can improve its utility through any unilateral deviation that preserves joint feasibility under the current predicted state and budget profile. ⋆
Algorithm 1: Proposed AEGIS 1
feas
8 9 10 11 12
Construct the improving-MU set (k) I (k) (t) = {i | ui ∈ U A (t), ∆i (t) > ε′ }. if I (k) (t) = ∅ then break Select one index i⋆ ∈ I (k) (t). (k+1) (k) (k+1) (k) Set ai⋆ (t) = ãi⋆ (t) and keep aj (t) = aj (t) for all other MUs uj . Set k ← k + 1.
Set a⋆ (t) = a(k) (t). Update each risk budget Bi (t + 1) according to (7). ⋆ 15 Output: a (t) and {Bi (t + 1)}ui ∈U . 13 14
B. Potential Property and Equilibrium Existence The resulting time-wise scheduling formulation exhibits a potential-aligned structure, rendering the problem amenable to a tractable asynchronous online solution. This structural property is advantageous in environments characterized by shared-resource competition, coupled feasibility constraints, and dynamically evolving risk-budget limits. Theorem 1 (Potential property under feasible unilateral deviations). Given the feasible joint-action set Afeas (t), the timeslot-t scheduling game G(t) admits the exact potential function: P (a(t); t) = Φ(a(t); t). (16) Consequently, the game admits at least one pure-strategy equilibrium under feasible unilateral deviations and satisfies the finite improvement property (FIP) [9], [10]. Proof. Consider two feasible joint actions a(t) = (ai (t), a−i (t)) and a′ (t) = (a′i (t), a−i (t)) that differ only in MU ui ’s action and both satisfy joint feasibility. From (13), we have Ui (a′ (t); t) − Ui (a(t); t) = Φ(a′ (t); t) − Φ(a(t); t) = P (a′ (t); t) − P (a(t); t).
(17)
Hence, every feasible unilateral utility improvement corresponds exactly to an increment of the potential function. Since each MU has a finite action set, the resulting feasible jointaction space is also finite. At least one pure-strategy equilibrium exists under feasible unilateral deviations. Moreover, each strict feasible-improvement step strictly increases the potential function, implying that any feasible-improvement path must terminate in a finite number of updates. The above analysis establishes that the proposed formulation admits an exact-potential structure. The time-wise scheduling game is guaranteed to possess at least one pure-strategy equilibrium and admit finite-step asynchronous improvement
dynamics, providing a tractable foundation for online implementation. C. Solution Design: Structure of AEGIS For a given a−i (t), the feasible best response of MU ui is defined as abr (a−i (t);t) Ui (ai , a−i (t); t). i (t) ∈ arg maxai ∈Afeas i Since the action set is discrete, the feasible response can be obtained by finite enumeration or lightweight local search. Based on this potential property, we develop AEGIS, a predictionassisted risk-budgeted asynchronous online scheduling procedure, as summarized in Algorithm 1. Algorithm 1 describes the closed-loop operation of AEGIS within a timeslot. First, the LSTM predictors generate the short-term channel and edge-load estimates, which are used to construct the predicted delay, deadline-violation risk surrogate, feasible action set, and potential function. Then, given the current risk budgets, AEGIS solves the resulting timewise scheduling game through asynchronous feasible improvements. In each iteration, MUs generate feasible unilateral candidate actions, and only candidates with positive potential gains larger than ε′ are retained. If no improving MU exists, the algorithm terminates; otherwise, one improving MU updates its action while all other actions remain unchanged. This asynchronous update avoids conflicts caused by simultaneous changes and preserves joint feasibility. According to Theorem 1, each accepted update strictly increases the potential function. Since the feasible joint-action set is finite, the asynchronous improvement process terminates in finite steps when no feasible improving action exists. When ε′ = 0, the converged output corresponds to a pure-strategy equilibrium under feasible unilateral deviations. The parameter Kmax is used only to limit online computation; if this limit is reached before convergence, AEGIS returns the last feasible joint action. Finally, after obtaining a⋆ (t), AEGIS updates
SLO-Edge 1
DeadlineFirst
A. Experimental Setup, Benchmarks, and Evaluation Metrics We consider a dynamic edge inference environment where MUs may become active over time, generate time-sensitive inference tasks, and compete for limited wireless bandwidth and edge computing resources. Unless otherwise specified, the total number of MUs varies across experiments, while the total bandwidth and computing capacity are fixed as B tot = 50 MHz and F tot = 155 GHz. Each simulation episode consists of 180 timeslots. For each active MU ui ∈ U A (t), the task tuple Ji (t) = Li (t), Ci (t), Dimax , wi is generated as follows: Li (t) = [0.12, 0.90] MB, Ci (t) = [0.08, 0.95] × 109 CPU cycles, and Dimax = [0.28, 0.82] s [6], [11], [12]. To capture heterogeneous MU activity levels, we calibrate the user activation probabilities using real-world mobility traces from the Chicago taxi trips dataset [13]. Following [14], we select the 77th community area as the target region and extract the traces of 200 taxis in January 2013, where each taxi is mapped to one MU. The activity probability of MU ui is denoted by pi 1 and obtained by counting the number of days on which the corresponding taxi appears in the target region and dividing by 31; at each timeslot t, the activation indicator χi (t) is generated according to a Bernoulli distribution with parameter pi . The wireless channel process and edge-load process evolve over time according to the simulation environment, where the edge-load state Q(t) is updated based on the aggregate admitted workload and the available computing capacity of ESP. To support prediction-driven online scheduling, we adopt an LSTM-based predictor with observation-window length H = 8 for one-step-ahead estimation of short-term channel and edge-load states, i.e., ĥi (t + 1) and q̂(t + 1). The predictor is trained using rolling historical samples generated from the simulated channel and load traces; its outputs are used to construct the deadline-violation risk surrogate r̂i (a(t); t) for time-wise resource competition. We compare AEGIS with the following representative benchmark methods: (i) SLO-Edge, an SLO-aware priority benchmark inspired by [7] that prioritizes active MUs with smaller remaining delay margins; (ii) DeadlineFirst, a deadline-priority policy inspired by [4] that schedules resources to active MUs with smaller task deadlines first; (iii) BCLF, a channel-aware latency-oriented benchmark inspired by [6] that prioritizes MUs with favorable wireless conditions to improve instantaneous transmission efficiency; (iv) EqualShare, a fairness-oriented comparator inspired by [8] that equally splits the available bandwidth and computing resources among all active MUs; and (v) AEGISNoBudget, an ablation version of AEGIS that preserves the LSTM predictor and 1 Note that the Chicago taxi traces are used only to emulate heterogeneous MU activity patterns, rather than to model the semantic content or computational characteristics of inference tasks.
0.6 0.4 0.2 10
EqualShare
AEGISNoBudget 1.8
0.4
20
30
40
0 10
1.4 1.2
0.2
Number of MUs (a)
AEGIS
1.6
0.6
AVR
We conduct extensive experiments to evaluate the effectiveness of AEGIS for continuous edge inference. All experiments are implemented in Python 3.10 with a 12th Gen Intel Core i9-12900H processor.
TIR
0.8
IV. E VALUATION
BCLF 0.8
DVBL
the risk budgets according to (7), connecting the current scheduling outcome with the next-timeslot risk state.
20
30
Number of MUs (b)
40
1 10
20
30
40
Number of MUs (c)
Fig. 2. Service reliability and long-term stability versus the number of MUs: (a) TIR, (b) AVR, and (c) DVBL.
online game-based resource competition, but removes dynamic risk-budget regularization during the scheduling process. We evaluate all methods using the following metrics: (i) Timely inference ratio (TIR), indicating the ratio of tasks completed before their deadlines among all active tasks; (ii) Average violation risk (AVR), measuring the average predictiondriven deadline-violation risk surrogate; (iii) Deadline violation burst length (DVBL), capturing the average length of consecutive deadline failures; (iv) Average E2E delay (AED), denoting the average delay of admitted tasks; (v) Average system utility (ASU), calculating the average time-wise scheduling utility balancing timely-service gain, resource cost, and risk regularization; and (vi) Convergence rounds (CR), describing the number of online iterations required to converge. B. Performance Evaluation We first evaluate service reliability and long-term stability under different MU populations in Fig. 2. As the number of MUs increases, the TIR in Fig. 2(a) decreases for all methods, since a larger MU population intensifies competition for limited communication and computing resources. Nevertheless, AEGIS and AEGISNoBudget consistently remain among the best-performing group over the entire range and both outperform the conventional baseline schemes. This indicates that our prediction-driven online scheduling can preserve timely service capability even under increasingly heavy load. Fig. 2(b) reports AVR under different MU scales. As expected, AVR increases with a growing number of MUs for all methods due to stronger contention and tighter service competition. However, AEGIS maintains one of the lowest risk levels and remains consistently below most benchmarks. This verifies that the proposed risk-budget mechanism can suppress deadline-violation exposure in dynamic edge inference systems. Fig. 2(c) plots the performance on DVBL. AEGIS achieves shorter burst lengths than the baselines, including SLO-Edge, DeadlineFirst, and BCLF, indicating that it can better prevent persistent consecutive deadline failures and thus provide stable continuous inference services. This is critical in continuous edge intelligence settings, where MUs are sensitive not only to isolated deadline misses but also to sustained service degradation over time. We next evaluate efficiency, utility, and online overhead in Fig. 3. In Fig. 3(a), AEGIS and AEGISNoBudget achieve consistently lower AED than the benchmarks across different MU scales. This indicates that the proposed gametheoretic joint scheduling framework can effectively reduce E2E inference latency under dynamic multi-MU contention. In contrast, the benchmarks suffer from rapidly increasing delay
SLO-Edge 1
DeadlineFirst
EqualShare
AEGISNoBudget 8
CR
0.6
0
0.4 0.2 10
AEGIS
6
5
ASU
AED
0.8
BCLF 10
4 2
-5 20
30
40
Number of MUs (a)
10
20
30
40
Number of MUs (b)
0 10
20
30
40
Number of MUs (c)
Fig. 3. Efficiency, utility, and online overhead versus the number of MUs: (a) AED, (b) ASU, and (c) CR.
as the MU population grows, primarily because they rely on myopic priority rules or static resource splitting and thus have limited capability to coordinate coupled bandwidth-computing resources. Fig. 3(b) compares the average system utility. AEGIS maintains strong utility performance and markedly outperforms the benchmarks, especially under medium- and high-load regimes. This verifies that the proposed utility design can jointly balance timely-service reward, resource consumption cost, and risk-aware regulation. Note that AEGISNoBudget may obtain slightly higher ASU in some cases, because removing budget-aware regulation allows the scheduler to pursue more aggressive instantaneous utility. However, such short-term utility gain is accompanied by higher online search overhead. As shown in Fig. 3(c), AEGIS requires fewer convergence rounds than AEGISNoBudget, demonstrating that the dynamic risk budget not only regulates cross-time violation exposure but also helps restrict the feasible search space and improve online tractability. Compared with AEGISNoBudget, AEGIS achieves a more balanced trade-off among delay efficiency, utility, risk control, and convergence overhead. To isolate the impact of state prediction, we further compare AEGIS with AEGISNoPred, which keeps the same risk-budgeted scheduling game but replaces the predicted channel and edge-load states with the latest observed states. As shown in Table I, AEGIS achieves slightly higher TIR, lower AVR, higher ASU, and fewer convergence rounds than AEGISNoPred at the fixed scale of 20 MUs, while their AED and DVBL remain close. This indicates that short-term state prediction provides useful forward-looking information for online scheduling, helping the scheduler better anticipate time-varying wireless and edge-load conditions. Meanwhile, the relatively small gap is reasonable since the channel and edge-load processes are temporally correlated and the latest observation serves as a strong no-prediction baseline. The above results verify the effectiveness of AEGIS from both service and algorithmic perspectives. Compared with the benchmarks, AEGIS achieves higher timely-service capability, lower E2E delay, and better long-term service stability. The comparison with AEGISNoBudget confirms that the dynamic risk budget is not merely an auxiliary regularization term, but a key mechanism for stabilizing continuous edge inference by jointly regulating cross-time violation exposure and scheduling TABLE I A BLATION STUDY ON STATE PREDICTION AT A FIXED SCALE OF 20 MU S . Method AEGISNoPred AEGIS
TIR ↑ 0.7711 0.7722
AVR ↓ 0.3152 0.3143
AED ↓ 0.3274 0.3297
ASU ↑ 7.7633 7.7791
DVBL ↓ 1.1304 1.1311
CR ↓ 3.1105 3.0640
tractability. V. C ONCLUSION We investigated prediction-assisted online scheduling for continuous deadline-sensitive edge inference and proposed AEGIS, a risk-budgeted framework that regulates cross-time deadline-violation tendency. By leveraging LSTM-based shortterm state prediction, AEGIS constructs a smooth deadlineviolation risk surrogate and introduces dynamic user-level risk budgets to incorporate long-term violation exposure into joint wireless–computing resource competition. Based on this design, the time-wise scheduling problem is cast as a potentialaligned game under coupled feasibility constraints and solved via the proposed asynchronous online algorithm with finitestep convergence. Simulation results demonstrated that AEGIS improves timely inference, suppresses long-term violation risk and burstiness, and achieves favorable delay–risk–convergence trade-offs over benchmarks. Future work will extend AEGIS to more general multi-platform edge inference scenarios and adaptive/intelligent risk-control mechanisms under richer network dynamics. R EFERENCES [1] R. F. El-Khatib, N. Zorba, and H. S. Hassanein, “Collaborative inference in the extreme edge: A proactive task allocation scheme,” in IEEE GLOBECOM 2025, pp. 980–985, IEEE, 2025. [2] J. Fang, Y. He, F. R. Yu, and J. Du, “Resource allocation for video diffusion task offloading in cloud-edge networks: A deep active inference approach,” in IEEE GLOBECOM 2024, pp. 2021–2026, IEEE, 2024. [3] Y. Mao, X. Yu, K. Huang, Y.-J. A. Zhang, and J. Zhang, “Green edge ai: A contemporary survey,” Proc. IEEE, vol. 112, no. 7, pp. 880–911, 2024. [4] M. K. C. Shisher, A. Piaseczny, Y. Sun, and C. G. Brinton, “Computation and communication co-scheduling for timely multi-task inference at the wireless edge,” in IEEE INFOCOM 2025, pp. 1–10, IEEE, 2025. [5] X. Li and S. Bi, “Optimal ai model splitting and resource allocation for device-edge co-inference in multi-user wireless sensing systems,” IEEE Trans. Wireless Commun., vol. 23, no. 9, pp. 11094–11108, 2024. [6] Z. Liu, Q. Lan, and K. Huang, “Resource allocation for multiuser edge inference with batching and early exiting,” IEEE J. Sel. Areas Commun., vol. 41, no. 4, pp. 1186–1200, 2023. [7] X. Zhang and D. Kim, “Enabling slo-aware 5g multi-access edge computing with smec,” arXiv preprint arXiv:2601.19162, 2026. [8] W. Fan, Y. Yu, C. Bao, and Y. Liu, “Vehicular edge intelligence: Drl-based resource orchestration for task inference in vehicle-rsu-edge collaborative networks,” IEEE Trans. Mobile Comput., vol. 24, no. 10, pp. 10927–10944, 2025. [9] D. Monderer and L. S. Shapley, “Potential games,” Games and economic behavior, vol. 14, no. 1, pp. 124–143, 1996. [10] Y. Ding, K. Li, C. Liu, and K. Li, “A potential game theoretic approach to computation offloading strategy optimization in end-edgecloud computing,” IEEE Trans. Parallel Distrib. Syst., vol. 33, no. 6, pp. 1503–1519, 2021. [11] H. Qi, M. Liwang, S. Hosseinalipour, L. Fu, S. Zou, and W. Ni, “Future resource bank for isac: Achieving fast and stable win-win matching for both individuals and coalitions,” IEEE J. Sel. Areas Commun., vol. 44, pp. 513–530, 2026. [12] 3GPP, “5G; Study on Scenarios and Requirements for Next Generation Access Technologies,” Technical Report TR 38.913 V17.0.0, 3GPP, May 2022. Release 17. [13] City of Chicago, “Taxi trips 2013.” [Online]. Available: https://data.cityofchicago.org/Transportation/Taxi-Trips-2013/6h2xdrp2, 2013. [14] H. Qi, M. Liwang, X. Wang, L. Fu, Y. Hong, L. Li, and Z. Cheng, “Accelerating stable matching between workers and spatial-temporal tasks for dynamic mcs: a stagewise service trading approach,” IEEE Trans. Mobile Comput., vol. 25, no. 2, pp. 2878–2894, 2026.