User Satisfaction-Aware Resource Allocation with Prospect-Theoretic Utility for Multimedia Streaming

TO BE SUBMITTED IN IEEE JOURNAL , VOL. XX, NO. X, XX 2026

1

User Satisfaction-Aware Resource Allocation with Prospect-Theoretic Utility for Multimedia Streaming

arXiv:2608.10625v1 [cs.NI] 11 Aug 2026

Manoj Kumar S∗

Pranjal Varshney∗

Abstract—Traditional resource allocation policies for multimedia streaming have primarily targeted metrics such as throughput, fairness and delay. However, user satisfaction is known to be strongly influenced by variations in throughput. Although some recent works [1]–[3] have considered throughput variations, they overlooked a well-known aspect of human behavior that has reshaped behavioral economics over the past few decades [4]: the same amount of decrease in throughput causes greater annoyance than the satisfaction caused by the increase. Motivated by this insight from the broad field of economics, popularly known as prospect theory, we propose a generic utility (metric) that captures both the average throughput and the asymmetric effect of throughput variations on user satisfaction. Next, we propose a resource allocation policy that reduces abrupt decreases in allocated resources when the system transitions from a resourcerich to resource-scarce state while respecting the priorities of different classes of users. Extensive simulations show that this policy outperforms the existing fair allocation policies. However, the choice of the optimal threshold depends on the network instance, which, in practice, may change frequently. Hence, building on the theoretical analysis of multi-queue systems, we propose fast numerical methods for determining the optimal threshold, which can be used in field deployments. Index Terms—Resource allocation; Prospect theory; Queuing analysis.

I. I NTRODUCTION A characteristic feature of human behavior is loss aversion, as noted in the Nobel-winning work on behavioral economics [4]. In particular, humans are known to experience losses more intensely than the satisfaction derived from equivalent gains. Current quality of experience (QoE) metrics for multimedia streaming largely overlook this inherent asymmetry in human behavior. Specifically, most QoE metrics consider average metrics such as data rate, buffering rate or average delay. Although there is a vast body of literature that studies variations in data rate, the QoE metrics considered therein are symmetric, i.e., the same magnitude of increase and decrease in data rate (delay) is treated identically. In this work, we introduce a general QoE metric that captures this asymmetry by building on prospect theory [4]. We propose a threshold-based resource allocation policy that achieves better QoE than the fair allocation policy, which maximizes the total average throughput. Furthermore, using an analytical approach, we propose computationally efficient methods for obtaining the optimal threshold value, since the system settings change frequently due to user dynamics and The authors are with the Department of Electrical Engineering, Indian Institute of Technology Madras, Chennai, Tamil Nadu 600036, India. E-mails: ee24d020,ee23b180@smail.iitm.ac.in, avhishek@ee.iitm.ac.in. ∗ Student authors are listed alphabetically.

Avhishek Chatterjee

other changes in the network. Our proposed method and the corresponding analysis naturally extends to networks with multiple classes of users. A. Related Literature Resource allocation is fundamental to wireless communication systems, where limited radio resources, such as bandwidth, time slots, power, etc. must be efficiently distributed among users. In traditional resource allocation methods, such as maximum throughput (MT) [5], proportional fair (PF) [6], and modified largest weighted delay first (M-LWDF) [7], the objectives are to optimize throughput or delay, or to enforce fairness. However, user satisfaction in multimedia streaming is governed by the frequency of buffering pauses, multimedia quality and temporal variations of the multimedia quality [8]– [11]. A significant amount of research has been devoted to QoEaware resource allocation, where scheduling decisions are guided by utility functions that map network-level parameters such as throughput, delay, packet loss and bitrate, to userperceived quality. In [12]–[16], cross-layer resource allocation frameworks jointly optimize application-layer video adaptation and network-layer resource allocation to maximize QoE, thereby improving resource utilization compared with the conventional throughput-based schemes. In [17], [18], the authors proposed QoE-aware scheduling frameworks that employ learning-based QoE estimation, QoE-aware prioritization and resource allocation based on channel conditions, buffer status and video requirements to reduce playback interruptions while achieving a balance among QoE, throughput and fairness. In [11], [19], [20], buffer-aware and QoE-aware resource allocation strategies were proposed to improve playback continuity in adaptive video streaming. By incorporating playback buffer status and bitrate requirements into resource allocation decisions, these approaches reduce playback stalls and enhance overall user-perceived quality. In [21]–[23], packet importance, delay sensitivity, video quality metrics and video characteristics are incorporated into resource allocation to improve resource utilization while preserving video quality. In [24], a control-theoretic framework that formulates bitrate adaptation as a stochastic optimization problem is proposed to improve user QoE for HTTP-based adaptive video streaming services, with QoE explicitly defined as a weighted combination of average bitrate, quality variations, rebuffering time and startup delay. There is a line of work [1]–[3] that captured QoE by introducing utility functions that increase with average quality and decrease with variations in quality, and proposes resource allocation algorithms to improve QoE by optimizing these

TO BE SUBMITTED IN IEEE JOURNAL , VOL. XX, NO. X, XX 2026

utility functions or by provisioning constrained rate variability while efficiently utilizing network resources. Furthermore, several multi-class resource allocation algorithms have been proposed to address heterogeneous user requirements by assigning different priorities to different user classes. In [25], [26], a class-based priority scheduling and cross-layer resource allocation schemes assigns priority weights to different user classes and exploit application-layer information, thereby influencing resource allocation along with channel conditions. Similarly, [27] proposed a service-based prioritization approach that preferentially allocates resources to QoE-sensitive video streaming users while distributing the remaining resources among lower-priority services. In [28], [29] QoE fairness-based schemes that prioritize users with poor QoE to ensure a baseline level satisfaction while improving overall system performance. B. Main contribution As discussed above, a well known concept in behavioral economics [4] is that a drop in quality is perceived more intensely than an equivalent improvement in quality. Existing QoE metrics and related utility functions, including those that capture variations in quality [1]–[3], consider the impact of such variations to be symmetric. Furthermore, in most QoEaware algorithms for networks with multiple classes of users, the number of users in each class is assumed to remain static throughout the resource allocation process. To address these limitations, we make the following main contributions. i) We propose a performance metric that captures the asymmetric impact of variations in average quality or data rate. The metric is defined as a linear combination of an average resource allocation metric, which reflects the average amount of resources received by users, and a prospect metric, which captures the asymmetric impact of increase and decrease in the allocated resources. ii) Next, we propose a novel threshold-based resource allocation policy that allocates a stable share of resources when the number of active users is below a threshold, thereby avoiding unnecessary reductions in allocation due to minor changes in the system load. Once the number of users exceeds this threshold, the available resources are distributed equally among all users. Thus, the proposed policy helps reduce abrupt decreases in resource allocation, which are perceived more negatively by users. iii) An M/M/∞ system model was developed to evaluate the performance of the proposed resource allocation algorithm. The simulation results demonstrate that the proposed threshold-based resource allocation policy achieves better performance than the existing policies. iv) Due to the stochastic nature of user arrivals and departures, determining the optimal threshold through simulation requires repeated evaluations over a wide range of threshold values, resulting in high execution time. Therefore, we develop an exact theoretical framework based on numerical analysis, that accurately computes the optimal threshold approximately 100 times faster than the simulation.

2

v) Furthermore, we introduce an approximate analytical approach derived from the exact theoretical framework by considering a single bystander that remains in the system at all times. This simplifies the analysis while preserving the system’s essential characteristics. Numerical results show that the bystander approach achieves performance metrics and optimal threshold value similar to the exact theoretical analysis, with approximately 30 times faster execution. These computational advantages make the proposed analytical approaches well suited for efficient system design and optimization. We emphasize that our resource allocation operates at the time scale of user dynamics rather than at the time-scale of channel coherence times. For wireless systems with n users and m fading channels (subcarriers), due to concentration of probability, each user 1 ≤ i ≤ n can be allocated ki channels at the highest fading state, the highest permissible moduPi.e., n lation rate, provided that i=1 ki ≤ m, when m ∼ 102 , which is the case for orthogonal frequency-division multiple access (OFDMA) in 4G/5G systems. We refer to [11] for a detailed proof of this result, particularly the proof of Theorem 3 and the accompanying discussions. Hence, we do not consider the effects of fading and adaptive modulation and coding (AMC) in this work, and focus on user dynamics. C. Organization of the paper The rest of the paper is organized as follows. Section II presents the problem formulation including user dynamics, user utility modeling and different user classes. The proposed threshold-based resource allocation policy and weighted threshold-based resource allocation policy for multi-class scenarios, along with their comparison with processor sharing (PS)/ fair allocation (FA) policy is presented in section III. Section IV presents the exact theoretical framework and simulation-based validation for both single-class and doubleclass systems. Section V introduces an approximate analytical framework based on the bystander approach and evaluates its performance for single-class and double-class scenarios. The comparative performance evaluation of the exact theoretical and bystander approaches is provided in section VI. Finally, the paper concludes with section VII. II. P ROBLEM SETTING User Dynamics The system is modeled as an M/M/∞ queue. User arrivals follow a Poisson process with rate λ and occur independently and randomly over time. Each user requires service for a random duration that is exponentially distributed with rate µ. The service times are assumed to be independent across users and independent of the arrival process. The system has an infinite number of servers, ensuring that every arriving user immediately receives service without any waiting. The total available resource in the system is denoted by B > 0. Let ρ = λ/µ denote the traffic intensity of the system. In the steady state, the number of users present in the system is a random variable that follows a Poisson distribution with mean

TO BE SUBMITTED IN IEEE JOURNAL , VOL. XX, NO. X, XX 2026

3

ρ. The stationary probability that the system contains exactly k users is given by π(k) = e−ρ

ρk , k!

k = 0, 1, 2, . . .

User Utility The performance metric used for optimization is the average amount of resources allocated to a user, which is denoted by A(F ), where F is the resource allocation policy. Let T denote the duration for which a user remains in the system. Then, the average amount of resources allocated is given by Z
1 T F q(t) dt (1) A(F ) = T 0 where q(t) denotes the number of users in the system at time t and F (q) denotes the amount of resources allocated to a particular user. When there are q users in the system, the resource allocation is symmetric and each person in the system is allocated F (q) amount of resources. User behavior is also modeled based on prospect theory using two constants, α and β, where α > β. An increase of x units of resources adds βx to the metric, while a decrease of x units of resources subtracts αx. Let si (i = 1, . . . , k1 ) represent all time instants at when a user’s allocated resources increase by an amount of gi (i = 1, . . . , k1 ) and let ti (i = 1, . . . , k2 ) represent all time instants at which the user’s allocated resources decrease by an amount of hi (i = 1, . . . , k2 ). The prospect-based metric P (F ), is given by k2 k1 X X hi (2) gi − α P (F ) = β i=1

heavier overall load on the system than class-1 users. The prospect gain and loss constants for class-1 and class-2 are denoted as βc1 , βc2 , αc1 and αc2 , respectively. The system is modeled under the assumption that the class-1 users are more sensitive to slight degradation in service than class-2 users. Hence, the prospect loss constants satisfy αc1 > αc2 , whereas the prospect gain constants satisfy βc1 < βc2 . The overall performance metric is obtained by aggregating the contributions of both classes. III. R ESOURCE A LLOCATION P OLICY Resource allocation is the process of distributing a fixed amount of resources among all active users in the system to optimize performance, fairness and service quality. In shared systems, the demand for resources fluctuates as users join or leave the system. Therefore, a resource allocation policy is required to determine each user’s share and maximize resource utilization. As discussed in Sec. I-B, thanks to previous work [11] that established concentration of measure for user-to-subcarrier mapping in multichannel wireless systems with fading and AMC, e.g. OFDMA in 4G/5G. Consequently, it is sufficient to study resource allocation at the scale of user dynamics instead of fading. Processor Sharing (PS) / Fair Allocation (FA) Policy [30] This is the simplest resource allocation policy, in which the total available resources are equally shared among all active users in the system. Let q denote the number of active users in the system and B denote the total available resources. Then, the resource allocation based on PS/FA policy is given by

i=1

The overall performance metric M (F ) combines the average resources metric A(f ) and the prospect-based metric P (F ), weighted by a factor γ and is given by M (F ) = A(F ) + γP (F )

(3)

User Class In general, the system may serve users belonging to multiple classes that differ in their traffic intensities and in how they perceive the services. Each class has its own arrival rate, indicating how frequently users of that class enter the system. The service time is modeled as a common parameter, implying that the session durations of users across all classes are statistically identical. Furthermore, the prospect constants are also class dependent. Users belonging to different classes exhibit different degrees of gain and loss sensitivity, represented by separate parameters that quantify the extent to which they perceive resource reductions negatively and value resource increases. This framework allows the system to be modeled with two distinct classes, namely class-1 and class-2, with λc1 and λc2 denoting their respective arrival rates and µ denoting the common service rate. The arrival rate for class-2 is set higher than that of class-1, i.e. λc1 < λc2 . This reflects the assumption that class-2 users have higher arrival intensity and impose a

N (q) =

B q

(4)

This policy is conceptually straightforward and is commonly used in real-world systems because it is easy to implement and automatically adapts to the system load. However, this policy ignores the user’s sensitivity to quality fluctuations. When the number of users in the system is small, a new arrival can cause a large drop in the resources allocated to all the active users, which may be perceived as highly negative under prospecttheory. In other words, this policy can lead to a large and sudden reduction in the service experienced by existing users as the population increases. Threshold-based Resource Allocation Policy To mitigate sudden fluctuations in service, a new thresholdbased resource allocation policy is proposed. When there are q users in the system, the amount of resources allocated to each user under the proposed threshold-based resource allocation is given by B   , q≤t t F (q) =  B , q > t q

(5)

TO BE SUBMITTED IN IEEE JOURNAL , VOL. XX, NO. X, XX 2026

4

where B denotes the total available resource and t is the threshold. This policy intentionally limits the resource allocated under low system load so that the arrival of a new user does not cause a large and sudden drop in the resources allocated to the existing users. This is more beneficial in terms of the prospect metric, under which users are more sensitive to losses than to equivalent gains. The main objective of this policy is to determine the threshold t that maximizes the overall performance metric. Weighted Threshold-based Resource Allocation Policy for multi-class systems As the system can accommodate users belonging to multiple classes, the proposed approach is extended to a weighted threshold-based resource allocation policy. Each class is assigned a positive weight and the system monitors the weighted load rather than the actual the user count. For a system with N classes, the resource allocation to the tagged user belonging to class-i, where i ∈ N is given by    wci B , Si (q) ≤ t  t Gci (qi ) = w B c   i , Si (q) > t  Si (q)

(6)

Comparative analysis with PS/FA resource allocation policy A comparative analysis is conducted between the proposed threshold-based resource allocation policy for single-class, the proposed weighted threshold-based resource allocation policy for double-class systems, and the existing PS/FA resource allocation policy. The simulations are carried out using the same experimental setup and performance metrics to ensure a fair comparison. Single-class: Fig. 1 shows the variation of the overall performance metric M [f ] for varying arrival rates λ under the PS/FA and threshold-based resource allocation policies. As the arrival rate increases, the overall performance metric decreases for both policies, indicating that the system performance degrades with increasing traffic load and reduced resource allocation per user. However, the threshold-based allocation policy consistently achieves significantly higher metric values than the PS/FA allocation policy across different arrival rates. Under the PS/FA resource allocation policy, frequent fluctuations in the allocated resources occur whenever the number of users in the system changes. This has a greater impact on the prospect metric, resulting in a lower performance metric than that achieved by the threshold-based resource allocation policy. This demonstrates that the threshold-based resource allocation policy allocates the available resource more efficiently.

where q = [qc1 , qc2 , · · · , qcn ] is the vector representing the number of users in each class and Si denotes the total weighted load, which is given by Si (q) =

X

wcn qcn + wci

n ∈N

Here, wcn is the weight associated with class-n and wci is the weight associated with the class to which the tagged user belongs. Taking into account a system with two classes, namely class-1 and class-2, having weights wc1 and wc2 , respectively, the resource allocation to a tagged user according to the proposed weighted threshold-based resource allocation policy is    wci B ,  t Gci (qc1 , qc2 ) =   wci B ,  Si (qc1 , qc2 )

Si (qc1 , qc2 ) ≤ t (7) Si (qc1 , qc2 ) > t

The total weighted load in this case is given by Si (qc1 , qc2 ) = wc1 qc1 + wc2 qc2 + wci By applying the threshold to the total weighted load and allocating resources proportionally, this policy ensures stable per-user allocation under low system load and fair resource sharing among heterogeneous classes under high system load. The main objective of this policy is to determine the optimal class weights and threshold that jointly maximize the performance metric.

Fig. 1. Comparison between PS/FA and the proposed threshold-based resource allocation policy for varying arrival rates

Double-class: A similar comparison is carried out for the double-class system by comparing the proposed weighted threshold-based allocation policy with PS/FA policy. Under the PS/FA policy, resources are allocated to the two classes in proportion to their arrival rates, λc1 : λc2 and each arriving user can access only the portion assigned to their respective class. Within each class, the allocated resources are shared equally among all users present in the system. This policy therefore serves as a simple baseline without thresholds or weighting mechanisms. Due to the varying arrival rates, the prospect sensitivity is allowed to vary with the relative scarcity of the classes. Let r denote the arrival rate ratio λc r= 2 ≥1 λc1

TO BE SUBMITTED IN IEEE JOURNAL , VOL. XX, NO. X, XX 2026

The class-specific parameters are scaled as αc1 = α0 ra

αc2 = α0 r−a

βc1 = β0 r−b

βc2 = β0 rb

where α0 > β0 > 0 are base parameters and 0 < a < b < 1. The intuition is that when λc2 greatly exceeds λc1 , reductions in allocated resources are perceived more strongly by class-1 users, leading to greater loss sensitivity and weaker response to gains. Conversely, class-2 users becomes less sensitive to losses and more responsive to incremental gains. The proposed scaling therefore captures this behavior while maintaining the structural condition αci > βci for each class. For each parameter configuration, the expected performance metric under the PS/FA policy, E[M (N )] and the expected performance metric under the weighted threshold-based resource allocation policy, E[M (G)] are evaluated. The performance is summarized using the ratio between the two performance metric obtained using the two allocation policies, given by η=

E[M (N )] E[M (G)]

5

Finding the best threshold and weights in real time The above comparisons shows that our threshold-based resource allocation policy achieve significantly better performance. However, to achieve the best performance, the threshold t in the single-class case and the threshold t and the weights wc1 and wc2 in the double-class case, have to be chosen optimally. For the above plots, these parameters were determined through exhaustive search using simulations. In practical field deployments, performing extensive simulations for every network setting is computationally challenging. To address this, analytical expressions are derived using Markov analysis and fast numerical methods for finding the best threshold and weights are developed based on these analytical expressions. The derived analytical expressions are formulated as a system of linear equations and the computation of the performance metrics involves solving this system of linear equations. Detailed analytical derivations are presented in the subsequent sections. IV. E XACT A NALYSIS In the considered system, users enter and depart from the system independently over time, resulting in a dynamically varying set of active users at any given instant. Rather than focusing on an individual user, this approach considered the collective behavior of all the active users in the system. The performance evaluation is done by averaging the resources allocated across all the users, thereby offering an aggregate view of system behavior. Theoretical Analysis

Fig. 2. Heatmap of the ratio of the performance metrics for different combinations of the arrival rates of class-1 and class-2 users

Figure 2 shows that, across the entire parameter grid the ratio remains strictly below 1, indicating that the PS/FA resource allocation policy never outperforms the weighted threshold-based policy in terms of the performance metric E[M ]. Furthermore, the ratio is negative, because the prospect component dominates the allocation term under the PS/FA policy, causing the overall metric E[M (N )] fall below zero. In contrast, the weighted threshold-based policy consistently maintains significantly higher values of E[M (G)] demonstrating a clear advantage. Another notable pattern is observed along the horizontal slices of the heatmap. For a fixed value of λc2 , the ratio exhibits a uni-modal concave profile as λc1 varies. This suggests that the relative performance gap between the two policies varies non-monotonically with the arrival-rate imbalance between the classes.

Using the underlying stochastic model with Markov chain and probabilistic methods, the theoretical analysis aims to derive a closed-form expression for the overall performance metric. Single-Class: In the single-class system, the theoretical analysis focuses on the rewards received by a typical (tagged) user conditioning on the number of other q users present in the system at the time of arrival. By the Poisson arrival see time averages (PASTA) property, the distribution seen by an arriving user is the same as the stationary distribution of the system, i.e. π(q) is the steady state probability of having q other users in the system, which is also the probability that a randomly chosen arriving user observes q other users in the system. A new event corresponds to the arrival of a new user or the departure of another user or the departure of the tagged user from the system. The probabilities of these events are given by P+ (q) =

λ Λ(q)

P− (q) =

qµ Λ(q)

PLeave (q) =

µ Λ(q)

where P+ (q), P− (q) and PLeave (q) are the probability of the arrival of a new user or the departure of another user or the departure of the tagged user from the system respectively, and satisfy P+ (q) + P− (q) + PLeave (q) = 1. The waiting time

TO BE SUBMITTED IN IEEE JOURNAL , VOL. XX, NO. X, XX 2026

6

until the next event is exponentially distributed with rate Λ(q) and is given by Λ(q) = λ + qµ + µ = λ + (q + 1)µ

(8)

One of the main quantities of interest is the expected cumulative resources E[RD (q)] received by the tagged user from the time of arrival until departure, conditioned on there being q other user being in the system. The derivation of E[RD (q)] is given by

 λ  − α F (q + 1) − F (q + 2) Λ(q)
 qµ  + β F (q) − F (q + 1) Λ(q)  λ  + E[RP (q + 1)] Λ(q)  qµ  + E[RP (q − 1)] (11) Λ(q)

E[RP (q)] =

Using the PASTA property, the expected time-averaged resource metric and the expected prospect metric are given by

E[RD (q)] = P+ (q) E[RD (q) | q + 1] + P− (q) E[RD (q) | q − 1]

E[A(F )] =

+ P Leave (q) E[RD (q) | q Leave ]

∞ X

π(q) E[RA (q)]

(12)

π(q) E[RP (q)]

(13)

q=0

Applying the event probabilities and conditional expectations,

E[P (F )] =

∞ X q=0

  λ F (q + 1) + E[RD (q + 1)] E[RD (q)] = Λ(q) Λ(q)   qµ F (q + 1) + + E[RD (q − 1)] Λ(q) Λ(q)   F (q + 1) µ + Λ(q) Λ(q)

Finally, the theoretical expression for the overall performance metric E[M (F )] under the threshold-based resource allocation policy is given by E[M (F )] = E[A(F )] + γ E[P (F )]

This can be further simplified as F (q + 1) Λ(q) λ + E[RD (q + 1)] Λ(q) qµ + E[RD (q − 1)] Λ(q)

E[RD (q)] =

(9)

By the Renewal Reward Theorem (RRT), the long-run average reward rate is given by the ratio of the expected cumulative resource received during a renewal cycle to the expected cycle length. Therefore, the expected time-averaged resource E[RA (q)] conditioned on the tagged user observing q other users upon arrival arrival is given by

E[RA (q)] =

1 E[RD (q)] µ

Double-class: Theoretical analysis is carried out by considering the arrival of tagged user from either of the two classes with varying number of users in each class class. Let qc1 and qc2 denote the number of class-1 and class-2 users respectively. By the PASTA property, the joint distribution of (qc1 , qc2 ) observed at the arrival instant coincides with the stationary distribution of the underlying Markov chain. Since the users of both classes are modeled as independent Poisson random variables, the probability that the tagged user finds qc1 users of class-1 and qc2 users of class-2 in the system is Π(qc1 , qc2 ) = π(q1 ) π(qc2 ) The total rate of (qc1 , qc2 ) with respect to class-1 arrival, class-2 arrival, departure of other class-1 user, departure of other class-2 user and departure of the tagged user is given by Λ(qc1 , qc2 ) = λc1 + λc2 + (qc1 + qc2 + 1)µ

(10)

The other main quantity is the expected cumulative prospect contribution E[RP (q)] that captures the gains and losses in the resources allocated to the tagged user resulting from the arrival or departure of another user in the system. The quantity E[RP (q)] is given by

The event probabilities are given by λc1 Λ(qc1 , qc2 ) qc1 µ P−c1 (qc1 , qc2 ) = Λ(qc1 , qc2 ) µ PLeave (qc1 , qc2 ) = Λ(qc1 , qc2 ) P+c1 (qc1 , qc2 ) =

λc2 Λ(qc1 , qc2 ) qc2 µ P−c2 (qc1 , qc2 ) = Λ(qc1 , qc2 ) P+c2 (qc1 , qc2 ) =

(c )

E[RP (q)] = P+ (q) E[RP (q) | q + 1] + P− (q) E[RP (q) | q − 1]

(14)

For a tagged user of class i ∈ {1, 2}, E[RDi (qc1 , qc2 )] denotes the expected cumulative resources acquired by the tagged user from arrival until departure, conditioned on the

TO BE SUBMITTED IN IEEE JOURNAL , VOL. XX, NO. X, XX 2026

7

(c )

initial state (qc1 , qc2 ). The recursion for E[RDi (qc1 , qc2 )] is derived as

  λc1 (c ) E[RP i (qc1 + 1, qc2 )] Λ(qc1 , qc2 )   λc1 − αci ∆(qc1 + 1, qc2 ) + Λ(qc1 , qc2 )   λc2 (c ) + E[RP i (qc1 , qc2 + 1)] Λ(qc1 , qc2 )   λc2 + − αci ∆(qc1 , qc2 + 1) Λ(qc1 , qc2 )  qc1 µ  (c ) E[RP i (qc1 − 1, qc2 )] + Λ(qc1 , qc2 )  qc1 µ  + βci ∆(qc1 − 1, qc2 ) Λ(qc1 , qc2 )  qc2 µ  (c ) + E[RP i (qc1 , qc2 − 1)] Λ(qc1 , qc2 )  qc2 µ  + βci ∆(qc1 , qc2 − 1) (17) Λ(qc1 , qc2 )

(c )

E[RP i (qc1 , qc2 )] = (c )

(c )

E[RDi (qc1 , qc2 )] = P+c1 (qc1 , qc2 ) E[RDi (qc1 , qc2 ) | qc1 + 1] (c )

+ P+c2 (qc1 , qc2 ) E[RDi (qc1 , qc2 ) | qc2 + 1] (c )

+ P−c1 (qc1 , qc2 ) E[RDi (qc1 , qc2 ) | qc1 − 1] (c )

+ P−c2 (qc1 , qc2 ) E[RDi (qc1 , qc2 ) | qc2 − 1] (c )

(c )

i + P Leave (qc1 , qc2 ) E[RDi (qc1 , qc2 ) | q Leave ]

µ Gci (qc1 , qc2 ) (c ) E[RDi (qc1 , qc2 )] = Λ(qc1 , qc2 ) Λ(qc1 , qc2 )   Gci (qc1 , qc2 ) λc1 (ci ) E[RD (qc1 + 1, qc2 )] + + Λ(qc1 , qc2 ) Λ(qc1 , qc2 )   λc2 Gci (qc1 , qc2 ) (ci ) + E[RD (qc1 , qc2 + 1)] + Λ(qc1 , qc2 ) Λ(qc1 , qc2 )   Gc (qc , qc ) qc1 µ (c ) E[RDi (qc1 − 1, qc2 )] + i 1 2 + Λ(qc1 , qc2 ) Λ(qc1 , qc2 )   qc2 µ G c (qc , qc ) (c ) + E[RDi (qc1 , qc2 − 1)] + i 1 2 Λ(qc1 , qc2 ) Λ(qc1 , qc2 )

where ∆(qc1 + 1, qc2 ) and ∆(qc1 , qc2 + 1) are the decrease in the allocated resources when a class-1 user and a class-2 user joins the system, respectively. Similarly, ∆(qc1 − 1, qc2 ) and ∆(qc1 , qc2 − 1) denote the increase in the resource when a class-1 user and a class-2 user departs from the system, respectively.

∆(qc1 + 1, qc2 ) = G(ci ) (qc1 , qc2 ) − G(ci ) (qc1 + 1, qc2 ) (c )

Gci (qc1 , qc2 ) Λ(qc1 , qc2 ) λc1 (c ) + E[RDi (qc1 + 1, qc2 )] Λ(qc1 , qc2 ) λc2 (c ) + E[RDi (qc1 , qc2 + 1)] Λ(qc1 , qc2 ) qc1 µ (c ) E[RDi (qc1 − 1, qc2 )] + Λ(qc1 , qc2 ) qc2 µ (c ) + E[RDi (qc1 , qc2 − 1)] (15) Λ(qc1 , qc2 )

E[RDi (qc1 , qc2 )] =

∆(qc1 , qc2 + 1) = G(ci ) (qc1 , qc2 ) − G(ci ) (qc1 , qc2 + 1) ∆(qc1 − 1, qc2 ) = G(ci ) (qc1 − 1, qc2 ) − G(ci ) (qc1 , qc2 ) ∆(qc1 , qc2 − 1) = G(ci ) (qc1 , qc2 − 1) − G(ci ) (qc1 , qc2 ) The expected time-averaged resource metric and the expected prospect metric are given by

E[A(ci ) (G)] =

X

(c )

(18)

(c )

(19)

Π(qc1 , qc2 ) E[RA i (qc1 , qc2 )]

qc1 ,qc2

The expected time-averaged resource for class i is given by 1 (c ) (c ) E[RA i (qc1 , qc2 )] = E[RDi (qc1 , qc2 )] µ

(16)

Similarly, for a tagged user of class i ∈ {1, 2}, (c ) E[RP i (qc1 , qc2 )] denotes the expected cumulative prospect contribution received by the tagged user from arrival until departure conditioned on the initial state (qc1 , qc2 ). The recursion (c ) for E[RP i (qc1 , qc2 )] is derived as

(c )

(c )

E[P (ci ) (G)] =

X

Π(qc1 , qc2 ) E[RP i (qc1 , qc2 )]

qc1 ,qc2

The class-wise overall metric is E[M (ci ) (G)] = E[A(ci ) (G)] + γ E[P (ci ) (G)]

(20)

Since a randomly chosen arrival belongs to class-1 with probability λc1 /(λc1 + λc2 ) and to class-2 with probability λc2 /(λc1 + λc2 ), the combined overall performance metric is given by

E[RP i (qc1 , qc2 )] = P+c1 (qc1 , qc2 ) E[RP i (qc1 , qc2 ) | qc1 + 1] (c )

+ P+c2 (qc1 , qc2 ) E[RP i (qc1 , qc2 ) | qc2 + 1] (c ) + P−c1 (qc1 , qc2 ) E[RP i (qc1 , qc2 ) | qc1 − 1] (c ) + P−c2 (qc1 , qc2 ) E[RP i (qc1 , qc2 ) | qc2 − 1]

λc1 E[M (c1 ) (G)] λc1 + λc2 λc2 E[M (c2 ) (G)] + λc1 + λc2

E[M (G)] =

(21)

TO BE SUBMITTED IN IEEE JOURNAL , VOL. XX, NO. X, XX 2026

Validation through Simulation Analysis Simulation-based evaluation involves the numerical simulation of the M/M/∞ system by randomly generating arrival and departure events and analyzing the user experience throughout the user’s stay in the system. The simulation provides empirical estimates of the expected average resource, expected prospect contribution and overall performance metric for a range of threshold values. The simulation is organized as a sequence of discrete events. At each event, an arrival or a departure occurs randomly, with the time between successive events drawn from an exponential distribution. Between consecutive events, all the users in the system receive a constant resource allocation determined by threshold-based policy, which is accumulated into each user’s cumulative resource. When arrival or departure occurs, the number of users in the system increases or decreases respectively and the corresponding prospect loss or gain is added to all the active users in the system. The time-averaged resource and the prospect contribution of the departing users are used to determine the overall performance metric. To eliminate transient effects, an initial set of departing users is discarded and the simulation is continued until either a sufficient number of users have departed or the maximum number of events has occurred. Single-class: For single-class simulation analysis, the M/M/∞ system is configured with an arrival rate λ = 4, a service rate µ = 1 and a total available resources of B = 100. The Prospect loss sensitivity and gain sensitivity constants are set to α = 2 and β = 1, respectively. The scalar weight γ = 1 is used to combine the average resource and the prospect contribution. To obtain a reliable steady-state estimates, the first 500 departing users are discarded. The maximum number of departing users and events considered in the simulation are 20,000 and 50,000,000 respectively. The overall performance metric is evaluated over a threshold range of 1 to 16. The comparison of the overall performance metric obtained from the simulation and the theoretical analysis for varying threshold value is shown in Fig. 3.

8

threshold illustrates the trade-off introduced by the thresholdbased resource allocation policy. When the threshold is very low, the available resource is divided more aggressively. As a result, frequent reductions in the allocated resources occur whenever a new user arrives. As the threshold increases, the system becomes more tolerant to the arrival of new users. This enables a more balanced allocation of the resource, resulting in a steady increase in the overall performance metric. The metric reaches its maximum at the threshold t = 6, indicating that this threshold provides an efficient and balanced allocation of resources with minimal reductions caused by fluctuations in the allocated resources. Beyond this point, the metric gradually decreases as the system allocates resources more conservatively when the number of users exceeds the threshold. Consequently, the per-user resource allocation decreases significantly, thereby reducing the overall performance metric. In addition, the close agreement between the theoretical and simulation curves over entire threshold values validates the correctness of the analytical model and confirms that the theoretical formulation accurately captures the system behavior. Double-class: The system is modeled with two classes, class-1 and class-2 with arrival rate λc1 = 2 and λc2 = 6 respectively, common service parameter µ = 1 and total available resource of B = 100. The prospect loss constants are set to αc1 = 4 and αc2 = 3, while the prospect gain constants are set to βc1 = 1 and βc2 = 2. The scalar weight used to compute the overall metric is set to γ = 1. Similar to the single-class simulation, the first 500 departing users are discarded and the maximum number of departing users and events considered are 20,000 and 50,000,000 respectively. The weight of class-1 is varied as wc1 ∈ {0.5, 1, 2, 3, 4, 5}, while the weight of class-2 is kept constant at wc2 = 1. The simulation is carried out over a threshold values range of 1 to 30 and the overall performance metric is evaluated.

Fig. 4. Heatmap of performance metric across threshold and class-1 weight values under exact theoretical analysis Fig. 3. Overall performance metric comparison obtained through simulation and exact theoretical analysis for single-class scenario

The variation of the overall performance metric with the

The heatmap in Fig. 4 and Fig. 5 depicts how the overall performance metric for users from two different classes varies with different class weights and threshold values under

TO BE SUBMITTED IN IEEE JOURNAL , VOL. XX, NO. X, XX 2026

9

Fig. 6. Comparison of optimal threshold for different class-1 weights under theoretical and simulation analysis Fig. 5. Heatmap of performance metric across threshold and class-1 weight values under simulation analysis

theoretical and simulation analyses, respectively. From the figures, it can be observed that, for each value of wc1 , the metric initially increases as threshold increases and reaches a maximum. At very small threshold values, the metric value is low, as indicated by the darker shades in the heatmaps. This suggests that the resource allocation is highly sensitive to the instantaneous number of users in the system, resulting in less efficient resource distribution. As the threshold increases, the metric gradually increases, as indicated by the transition from darker to brighter shades in the heatmaps. This increase in the overall metric indicates reduced fluctuations in the resources allocated among users. As a result, users in both the class receive more consistent service. Another important observation from the heatmaps is that the highest metric value shifts towards larger threshold values as the weight increases. As wc1 increases, the weighted demand contributed by the class-1 users also increases, effectively increasing the weighted load in the system. To maintain a balanced resource allocation, a larger threshold is required. This explains the diagonal pattern of the high metric region observed in the heatmaps. Furthermore, beyond the optimal threshold, increasing the threshold value allocates the resources more uniformly without providing additional benefits. Consequently, the overall metric decreases significantly. The heatmaps clearly demonstrate the importance of properly tuning the threshold values for different weights to achieve optimal system performance. Figure 6 shows the comparison of the optimal threshold for different class-1 weights obtained under the theoretical and simulation analyses. From the figure, it is evident that the optimal threshold increases as the weight increases, indicating the influence of the class weight on the optimal threshold. As the weight wc1 increases, the class-1 users contribute more heavily to the total weighted load of the system. In this case, low optimal threshold value results in unstable resource allocation. Therefore, a larger threshold is required to maintain a stable allocation process across all the users. The theoretical and

simulation curves overlap exactly, indicating that the optimal threshold obtained from the simulation matches accurately with that predicted by the theoretical analysis.

Fig. 7. Comparison of overall metric value for different class-1 weights under theoretical and simulation analysis

Figure 7 compares the theoretical and simulated values of the overall performance metric for different class-1 weight values. From the figure, it can be observed that the performance metric gradually decreases as the weight wc1 increases. This behavior can be explained by the fact that increasing wc1 biases the resource allocation mechanism more toward class-1 users. As a result, a larger portion of the resource is allocated to the class-1 users, leaving relatively fewer resources for the class-2 users. Since the overall performance metric is computed as the weighted average across both classes, excessive prioritization of one class can lead to reduction in the aggregate performance measure. Furthermore, across the entire range of weight values, the simulated performance metric closely matches the theoretical values, with only negligible deviation. This confirms the accuracy of the mathematical formulation and numerical implementation.

TO BE SUBMITTED IN IEEE JOURNAL , VOL. XX, NO. X, XX 2026

10

Comparison of computational Time: The theoretical analysis significantly reduces the computational time required to compute the overall performance metrics for different threshold values and determine the optimal threshold for a given system, compared with the simulation-based analysis. It is observed that, in single-class case, the theoretical analysis is 110 times faster than the corresponding simulation analysis. In double-class case, the analysis is performed for different class1 weight values and threshold values, where the theoretical analysis is 30 times faster than the corresponding simulation analysis. V. B YSTANDER A PPROACH The bystander approach simplifies the analysis by introducing a hypothetical user who remains in the system at all times and evaluating the performance metrics through the timeaveraged resource allocation and prospect-based outcomes experienced by this bystander. This approach is an approximation of the exact theoretical framework because the bystander is always present, hence all state transitions are conditioned on this fixed user. In contrast, an actual arriving user has a finite sojourn time and the distribution of the system observed upon arrival corresponds to the stationary distribution of other users. The performance metric is calculated based on the resource allocated to the bystander over the duration of bystander’s stay in the system. Single-class: Let q denote the total number of users in the system including the bystander. The probability that there are q − 1 users in the M/M/∞ queue is given by the distribution π(q − 1) distribution. The expected value of the time-averaged allocation of resources E[A(F )] is given by E[A(F )] =

∞ X

π(q − 1) F (q)

(22)

q=1

Using the stationarity of the M/M/∞ queue, the time average reduces to a stationary expectation. The expected prospect contribution, E[P (F )] is determined by calculating the expected number of times the transitions q → q ± 1 occur during a time interval of T . Based on properties of M/M/∞ queues, the average number of times q → q + 1 transition takes place in time T is λ × π(q − 1), whereas the average number of times a q → q − 1 transition takes place in time T is µ × (q − 1) × π(q − 1). Hence, E[P (F )] is given by E[P (F )] = β −α

∞ X q=2 ∞ X

! π(q − 1) × (q − 1) × µ ∆(q − 1) !

(23)

π(q − 1) × λ × ∆(q + 1)

To evaluate the performance of this approach, the system is modeled using an M/M/∞ queue with fixed arrival and service rates and a total available resource of 100 units is considered for allocation to the bystander. The parameters associated with the prospect term are kept identical to the exact analytical approach, so that the impact of the prospect term can be analyzed consistently. The overall metric is evaluated over a threshold ranged of 1 to 16 and the result is shown in Fig. 8.

Fig. 8. Overall performance metric for single-class using bystander approach

Similar to the exact theoretical framework, the overall performance metric is low at smaller threshold values due to the fluctuations in the allocated resources. As threshold increases, the performance metric increases and reaches its maximum at the threshold t = 6. This corresponds to the optimal threshold, where the system allocates the resources more efficiently. Beyond this point, the performance metric gradually decreases as the threshold continues to increase. Double-class: The bystander approach is extended to the double-class setting, where users belong to either class-1 or class-2. Let (qc1 , qc2 ) denote the total numbers of class-1 and class-2 users in the system, including the bystander. Let πc1 (·) and πc2 (·) denote the stationary distributions of the number of class-1 and class-2 users excluding the bystander in the corresponding M/M/∞ queues. When the bystander belongs to class-1, the system state (qc1 , qc2 ) corresponds to (qc1 − 1, qc2 ) other users. Similarly, when the bystander belongs to class-2, the system state corresponds to (qc1 , qc2 − 1) other users. Let Gc1 (qc1 , qc2 ) and Gc2 (qc1 , qc2 ) denote the resource allocation to a class-1 and class-2 bystander respectively. The expected time-averaged resource allocated is given by

q=1

Where ∆(q + 1) = F (q) − F (q + 1) and ∆(q − 1) = F (q − 1) − F (q) denote the change in the allocated resources when a user arrives at or departs from the system, respectively. The overall performance metric E[M (F )] is obtained by combining the expected time-average resource allocation and the expected prospect contribution. E[M (F )] = E[A(F )] + γ E[P (F )]

(24)

(c )

E[A(G)] = pb 1

∞ X

πc1 (qc1 − 1) πc2 (qc2 ) Gc1 (qc1 , qc2 )

qc1 =1 qc2 =0 (c )

+ pb 2

∞ X

πc1 (qc1 ) πc2 (qc2 − 1) Gc2 (qc1 , qc2 )

qc1 =0 qc2 =1

(25)

TO BE SUBMITTED IN IEEE JOURNAL , VOL. XX, NO. X, XX 2026

(q

)

(q

11

)

where pb c1 and pb c2 denote the probabilities that the bystander belongs to class-1 and class-2 respectively. Using the stationarity of the M/M/∞ queues, the time average again reduces to a stationary expectation. Analogous to the single-class case, the expected allocation change rates due to arrivals and departures of other users are accounted for determining E[P (G)]. For a class-1 bystander in state (qc1 , qc1 ), the arrivals of class-1 and class-2 users occur at rates λc1 and λc2 respectively, each contributing a loss proportional to the decrease in resource allocation. The departures of other class-1 and class-2 users occur at rates (qc1 − 1)µ and qc2 µ, producing gains due to increase in resource allocation. Combining these contributions yields the expected prospect contribution for a class-1 bystander. An analogous expression holds for a class-2 bystander, with departures occurring at rates qc1 µ and (qc2 − 1)µ. Hence, the expected prospect term is given by (c )

E[P (G)] = pb 1

∞ X

The bystander approach for double-class case is evaluated by considering the same system settings as those used in the exact theoretical framework to ensure consistency in evaluation. The system consists of two classes with arrival rates λc1 = 2 and λc2 = 6 and a common service parameter µ = 1. The same prospect metric parameters are also used. The overall performance metric is evaluated for varying class1 weight values and threshold values. The heat map for the overall performance metric over varying class-1 weight values and threshold values is shown in Fig. 9.

πc1 (qc1 − 1) πc2 (qc2 )

qc1 =1 qc2 =0



 − αc1 ∆c1 ,+1 (qc1 , qc2 ) + βc1 ∆c1 ,−1 (qc1 , qc2 ) (c )

+ pb 2

∞ X

πc1 (qc1 ) πc2 (qc2 − 1)

qc1 =0 qc2 =1



 − αc2 ∆c2 ,+1 (qc1 , qc2 ) + βc2 ∆c2 ,−1 (qc1 , qc2 ) (26) where ∆c1 ,+1 (qc1 , qc2 ) and ∆c2 ,+1 (qc1 , qc2 ) denote the changes in the allocated resources when a class-1 and class2 user enters the system respectively. Similarly, the changes in the allocated resources when a class-1 and class-2 user departs from the system are denoted by ∆c1 ,−1 (qc1 , qc2 ) and ∆c2 ,−1 (qc1 , qc2 ) respectively.   ∆c1 ,+1 (qc1 , qc2 ) = λc1 Gc1 (qc1 , qc2 ) − Gc1 (qc1 + 1, qc2 )   + λc2 Gc1 (qc1 , qc2 ) − Gc1 (qc1 , qc2 + 1)   ∆c1 ,−1 (qc1 , qc2 ) = qc2 µ Gc1 (qc1 , qc2 − 1) − Gc1 (qc1 , qc2 )   + (qc1 − 1)µ Gc1 (qc1 − 1, qc2 ) − Gc1 (qc1 , qc2 )   ∆c2 ,+1 (qc1 , qc2 ) = λc1 Gc2 (qc1 , qc2 ) − Gc2 (qc1 + 1, qc2 )   + λc2 Gc2 (qc1 , qc2 ) − Gc2 (qc1 , qc2 + 1)   ∆c2 ,−1 (qc1 , qc2 ) = qc1 µ Gc2 (qc1 − 1, qc2 ) − Gc2 (qc1 , qc2 )   + (qc2 − 1)µ Gc2 (qc1 , qc2 − 1) − Gc2 (qc1 , qc2 ) The overall performance metric is given by E[M (G)] = E[A(G)] + γ E[P (G)].

(27)

This expression depends only on the threshold and the relative weights.

Fig. 9. Heatmap of performance metric across threshold and class-1 weight values under bystander approach

Similarly to the heatmaps obtained using the exact theoretical analysis, the darker region in the heatmap indicates that the overall performance metric is very low at lower threshold values. This is due to the frequent fluctuations in the allocated resources that occur when the number of users exceeds the threshold. Then, for all weight values, the performance metric gradually increases as the threshold increases and reaches a peak. This corresponds to the optimal threshold, where the system allocates the resources in a balanced and efficient way. Beyond the optimal threshold, the performance metric gradually decreases. This is due to the conservative allocation of the resources at high threshold values. Furthermore, as the class-1 weight increases, the peak shifts to the right, indicating that a higher threshold is required to maximize the performance metric when class-1 users are assigned greater importance. From the heatmap, the optimal threshold values can be identified for achieving better resource allocation at different weight values. VI. C OMPARATIVE P ERFORMANCE E VALUATION BETWEEN TWO APPROACHES

Single-class Figure 10 presents the performance comparison between the bystander approach and the exact theoretical analysis for determining the optimal threshold that maximizes the expected performance metric E[M (F )]. In both approaches, the optimal threshold is obtained using a uni-modal binary

TO BE SUBMITTED IN IEEE JOURNAL , VOL. XX, NO. X, XX 2026

12

search over the same parameter range. For the system with parameters λ = 4, µ = 1, B = 100, α = 2, β = 1 and γ = 1, both approaches produce the same optimal threshold, t = 6 with the an expected overall performance metric value E[M (F )] = 13.085506.

Fig. 11. Performance comparison between the two approaches

both approaches produce similar overall performance metric values with only slight deviations. Fig. 10. Performance comparison between the two approaches

The agreement between the results confirms that both formulations compute the same performance metric despite their different analytical structures. Although both methods produce the same metric and the same optimal threshold, their execution time differ noticeably. The bystander approach is approximately 35 times faster than the exact theoretical analysis, which is approximately 110 times faster than the corresponding simulation-based analysis. Therefore, the bystander approach offers a substantial computational advantage, despite both approaches yielding identical optimal results. Double-class The computational performance comparison is carried out for the double-class setting, where the optimal threshold t is determined using uni-modal binary search for the system with parameters λc1 = 2, λc2 = 6, µ = 1, B = 100, αc1 = 4, αc2 = 3, βc1 = 1, βc2 = 2, γ = 1 and weights wc1 = 3 and wc2 = 1. Both approaches yield the same optimal threshold t = 21. Similar to the single-class case, this confirms the consistency of the two approaches despite their different analytical structures. Furthermore, the bystander approach is approximately 10 times faster than exact theoretical analysis, which is approximately 30 times faster than the corresponding simulation-based analysis in double-class scenario. Furthermore, the two approaches are compared for different class-1 weight values and the relative error between the performance metrics obtained using the two approaches is determined. The heatmap of the relative error is shown in Fig. 11. From the heatmap, the relative error is slightly higher at smaller threshold values, particularly when the class-1 weight is low, as indicated by brighter regions. As the threshold increases, the relative error decreases significantly, as reflected by the darker regions. Overall, the heatmap demonstrates that

VII. C ONCLUSION Motivated by decades of highly influential work in behavioral economics, we introduce a general utility metric for user satisfaction during multimedia streaming, which captures the asymmetric humans response to the same amount of increase and decrease in service quality. We propose a threshold-based resource allocation policy that reduces both the frequency and the amount of drops in allocated resources and, consequently, in service quality, while respecting the priorities of different classes of users. Through extensive simulations, we observe that, for a well chosen threshold, which may depend on the network configuration, this policy outperforms the existing fair allocation policy. Building on the detailed queuing theoretic analysis of the system, we propose a fast method for determining the exact optimal threshold and further propose an even faster method to find the threshold approximately. These methods are of practical importance since the network scenarios change dynamically in practice, thereby requiring the optimal threshold to be recalculated frequently. R EFERENCES [1] V. Joseph, G. de Veciana, and A. Arapostathis, “Resource allocation: Realizing mean-variability-fairness tradeoffs,” IEEE Transactions on Automatic Control, vol. 60, no. 1, pp. 19–33, Jan. 2015. [2] F. Mehmeti, A. Papa, W. Kellerer, and T. F. L. Porta, “Minimizing rate variability with effective resource utilization in cellular networks,” IEEE Transactions on Mobile Computing, vol. 23, no. 12, pp. 11 061–11 079, 2024. [3] F. Mehmeti, T. F. L. Porta, and W. Kellerer, “Efficient resource allocation with provisioning constrained rate variability in cellular networks,” IEEE Transactions on Mobile Computing, vol. 23, no. 5, pp. 5130–5146, 2024. [4] D. Kahneman and A. Tversky, “Prospect theory: An analysis of decision under risk,” Econometrica, vol. 47, no. 2, pp. 263–292, 1979. [5] E. A. Jorswieck, A. Sezgin, and X. Zhang, “Framework for analysis of opportunistic schedulers: Average sum rate vs. average fairness,” in IEEE International Conference on Communications (ICC), 2008. [6] P. Viswanath, D. N. C. Tse, and R. Laroia, “Opportunistic beamforming using dumb antennas,” IEEE Transactions on Information Theory, vol. 48, no. 6, pp. 1277–1294, 2002.

TO BE SUBMITTED IN IEEE JOURNAL , VOL. XX, NO. X, XX 2026

[7] M. Andrews, K. Kumaran, K. Ramanan, A. L. Stolyar, P. Whiting, and R. Vijayakumar, “Providing quality of service over a shared wireless link,” IEEE Communications Magazine, vol. 39, no. 2, pp. 150–154, 2001. [8] F. Dobrian, V. Sekar, A. Awan, I. Stoica, D. Joseph, A. Ganjam, J. Zhan, and H. Zhang, “Understanding the impact of video quality on user engagement,” in Proceedings of the ACM SIGCOMM 2011 Conference. ACM, 2011, pp. 362–373. [9] N. Barman and M. G. Martini, “Qoe modeling for http adaptive video streaming—a survey and open challenges,” IEEE Access, vol. 7, pp. 30 831–30 859, 2019. [10] S. Cheng, H. Hu, X. Zhang, and Z. Guo, “Rebuffering but not suffering: Exploring continuous-time quantitative qoe by user’s exiting behaviors,” in IEEE INFOCOM 2023 - IEEE Conference on Computer Communications. IEEE, 2023, pp. 1–10. [11] A. Bhimaraju, A. A. Zacharias, and A. Chatterjee, “Multichannel resource allocation for smooth streaming: Non-convexity and bandits,” IEEE Transactions on Communications, vol. 70, no. 8, pp. 5085–5097, Aug. 2022. [12] M. Shehada, S. Thakolsri, Z. Despotovic, and W. Kellerer, “Qoe-based cross-layer optimization for video delivery in long term evolution mobile networks,” in Proceedings of the IEEE International Conference on Communications (ICC). IEEE, 2011, pp. 1–6. [13] S. Thakolsri, S. U. Khan, E. Steinbach, and W. Kellerer, “Qoe-driven cross-layer optimization for high speed downlink packet access,” Journal of Communications, vol. 4, no. 9, pp. 669–680, 2009. [14] D. S. Rao and V. B. Hency, “A novel qoe based cross-layer scheduling scheme for video applications in 802.11 wireless lans,” SN Applied Sciences, vol. 3, p. 344, 2021. [15] A. Bentaleb, A. C. Begen, R. Zimmermann, and S. Harous, “Sdnhas: An sdn-enabled architecture to optimize qoe in http adaptive streaming,” IEEE Transactions on Multimedia, vol. 19, no. 10, pp. 2136–2151, 2017. [16] J. Yang, B. Yang, S. Chen, Y. Zhang, Y. Zhang, and L. Hanzo, “Dynamic resource allocation for streaming scalable videos in sdn-aided dense small-cell networks,” IEEE Transactions on Communications, vol. 67, no. 3, pp. 2114–2131, March 2019. [17] K. Piamrat, K. S. Singh, A. Ksentini, C. Viho, and J.-M. Bonnin, “Qoeaware scheduling for video-streaming in high speed downlink packet access,” in Proceedings of the IEEE Wireless Communications and Networking Conference (WCNC). IEEE, 2010, pp. 1–6. [18] Y. Yuan, W. Wang, Y. Wang, S. S. Adhatarao, B. Ren, K. Zheng, and X. Fu, “Joint optimization of qoe and fairness for adaptive video streaming in heterogeneous mobile environments,” IEEE/ACM Transactions on Networking, vol. 32, no. 2, pp. 1618–1632, 2024. [19] T. Wirth, Y. Sánchez, B. Holfeld, and T. Schierl, “Advanced downlink lte radio resource management for http streaming,” in Proceedings of the 20th ACM International Conference on Multimedia (ACM MM). ACM, 2012, pp. 219–228. [20] M. Seufert, N. Wehner, and P. Casas, “A fair share for all: Tcp-inspired adaptation logic for qoe fairness among heterogeneous http adaptive video streaming clients,” IEEE Transactions on Network and Service Management, vol. 16, no. 2, pp. 475–488, 2019. [21] M. Li, P. H. J. Tan, S. Sun, and Y. H. Chew, “Qoe-aware scheduling for video streaming in 802.11n/ac-based high user density networks,” in Proceedings of the IEEE Vehicular Technology Conference (VTC Spring). IEEE, 2016, pp. 1–5. [22] W. u. Rahman and E.-N. Huh, “Content-aware qoe optimization in mecassisted mobile video streaming,” Multimedia Tools and Applications, vol. 82, pp. 42 053–42 085, 2023. [23] Y. Xu, R. Q. Hu, Y. Qian, and T. Znati, “Video quality-based spectral and energy efficient mobile association in heterogeneous wireless networks,” IEEE Transactions on Communications, vol. 64, no. 2, pp. 805–820, Feb. 2016. [24] X. Yin, A. Jindal, V. Sekar, and B. Sinopoli, “A control-theoretic approach for dynamic adaptive video streaming over http,” in Proceedings of the ACM SIGCOMM 2015 Conference. ACM, 2015, pp. 325–338. [25] C.-H. Hsieh and J. C. Hou, “Heavy-traffic analysis of qoe optimality for on-demand video streams over fading channels,” IEEE/ACM Transactions on Networking, vol. 26, no. 5, pp. 2216–2230, 2018. [26] D. Wang, L. Toni, P. C. Cosman, and L. B. Milstein, “Uplink resource management for multiuser ofdm video transmission systems: Analysis and algorithm design,” IEEE Transactions on Communications, vol. 61, no. 5, pp. 2060–2074, May 2013. [27] R. El-Azouzi, Y. Xu, E. Altman, and A. El Fawal, “Qoe-driven resource allocation for massive video distribution,” Ad Hoc Networks, vol. 86, pp. 56–68, 2019.

13

[28] M. Rugelj, M. Volk, and A. Kos, “Novel cross-layer qoe-aware radio resource allocation algorithms in multiuser ofdma systems,” IEEE Transactions on Communications, vol. 62, no. 9, pp. 3196–3207, 2014. [29] Y. Hori and T. Ohtsuki, “Qoe and throughput aware radio resource allocation algorithm in lte network with users using different applications,” in Proceedings of the IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC). IEEE, 2016, pp. 1–6. [30] R. Srikant and L. Ying, Communication Networks: An Optimization, Control, and Stochastic Networks Perspective. Cambridge, United Kingdom: Cambridge University Press, 2014.

Related documents

• Operational Considerations for Streaming Media#502678
• Mend the Measurement Gap: Latent User Preference Modeling for Short-Form Video Recommendation#1117170
• Network Access to Multimedia Information#131857
• All BANG, little buck: Need-related experiences are weakly linked with behavior in the video game domain#976936
• Finite-Capacity Prospect Theory: State-Dependent Risk Preferences under Resource, Attention, and Boundary-Risk Constraints#203517