1

Human-in-the-Loop Wireless Communications:

Machine Learning and Brain-Aware Resource
Management
Ali Taleb Zadeh Kasgari1 , Walid Saad1, and Mérouane Debbah2
arXiv:1804.00209v1 [cs.IT] 31 Mar 2018

1
Wireless@VT, Electrical and Computer Engineering Department, Virginia Tech, VA, USA,
Emails:{alitk,walids}@vt.edu.
2
Mathematical and Algorithmic Sciences Lab, Huawei France R&D, Paris, France, and CentraleSupelec,
Universite Paris-Saclay, Gif-sur-Yvette, France, Email: [email protected].

Abstract

Human-centric applications such as virtual reality and immersive gaming will be central to the future
wireless networks. Common features of such services include: a) their dependence on the human user’s
behavior and state, and b) their need for more network resources compared to conventional cellular
applications. To successfully deploy such applications over wireless and cellular systems, the network
must be made cognizant of not only the quality-of-service (QoS) needs of the applications, but also of
the perceptions of the human users on this QoS. In this paper, by explicitly modeling the limitations of
the human brain, a concrete measure for the delay perception of human users in a wireless network is
introduced. Then, a novel learning method, called probability distribution identification, is proposed to
find a probabilistic model for this delay perception based on the brain features of a human user. The
proposed learning method uses both supervised and unsupervised learning techniques to build a Gaussian
mixture model of the human brain features. Given a model for the delay perception of the human brain,
a novel brain-aware resource management algorithm based on Lyapunov optimization is proposed for
allocating radio resources to human users while minimizing the transmit power and taking into account
the reliability of both machine type devices and human users. The proposed algorithm is shown to have
a low complexity. Moreover, a closed-form relationship between the reliability measure and wireless
physical layer metrics of the network is derived. Simulation results using real data from actual human
users show that a brain-aware approach can yield savings of up to 78% in power compared to the system

A preliminary version of this work appeared in the proceedings of the 51th Asilomar Conference on Signals, Systems and
Computers, Pacific Grove, CA, USA [1].
This research was supported by the U.S. National Science Foundation under Grants CNS-1460316 and IIS-1633363.
 2

that only considers QoS metrics. The results also show that, compared with QoS-aware, brain-unaware
systems, the brain-aware approach can save substantially more power in low-latency systems.

I. I NTRODUCTION

The next generation of wireless services is expected to be highly human centric. Examples
include virtual reality and interactive/immersive gaming [2], [3]. In order to cope with the quality-
of-service (QoS) needs of such human-centric applications, in terms of data rate and ultra-low
latency, wireless networks will have to allocate and exploit substantially more radio resources by
leveraging heterogeneous spectrum bands across low and high frequencies [4]. However, even
though allocating heterogeneous spectrum resources can potentially increase the raw QoS, given
the human-centric nature of such emerging applications, their users may not be able to perceive
the improved QoS, due to the cognitive limitations of the human brain [5]. Indeed, many empirical
studies (anecdotal and otherwise) have shown that the limitations on the human brain can be
translated into a limitation on how wireless users translate QoS into actual quality-of-experience
(QoE) [6]–[8]. For example, the human brain may not be able to perceive any difference between
videos transmitted with different QoS (e.g., rates or delays) [8], [9]. Hence, in order to deploy
these services over wireless networks, such as 5G cellular systems, there is a need to enable
the system to be strongly cognizant of the human user in the loop. In particular, to deliver
such immersive, human-centric services, the network must tailor the usage and optimization of
wireless resources to the intrinsic features of its human users such as their behavior and brain
processing limitations. By doing so, the network can potentially save resources, accommodate
more users, and provide a more realistic QoE to its users.
Developing resource management mechanisms that can cater to intrinsic needs of wireless
users and their context (e.g., device features or social metrics) has recently been studied in
[4], [10]–[17]. In [10], a context-aware scheduling algorithm for 5G systems is proposed. This
algorithm exploits the context information of user equipments (UEs), such as battery level,
to save energy in the system while satisfying the QoS requirements of users. The authors
in [11], proposed a user-centric resource allocation framework for ultra-dense heterogeneous
networks. Context-aware resource allocation for heterogeneous cellular networks is also studied
in [4], [12], and [13]. In [4], a novel approach to context-aware resource allocation in small
cell networks is introduced. Both wireless physical layer metrics and the social ties of human
users are exploited in [4] to allocate wireless resource blocks. Proactive caching using context
 3

information from social networks is studied in [14]. In this work, it is shown that such a socially-
aware caching technique reduces the peak traffic in 5G networks. Other context-aware resource
allocation algorithms are also studied in [15]–[17]. However, despite this surge in literature on
context-aware networking [4], [10]–[17], this prior art is still reliant on device-level features and
is agnostic to the human users and their features (e.g., brain limitation or behavior) and, hence,
they can potentially waste network resources as they can still allocate more resources to human
users that cannot perceive the associated QoS gains, due to cognitive brain limitations.
A general framework for modeling the intelligence of communication systems which serve
humans is proposed in [18]. The author defines intelligence in terms of predicting and serving
human demands in advance. However, the work in [18] does not account for the cognitive
limitations of a human brain. Moreover, demand prediction, as done in [18], will not be sufficient
to capture the full spectrum of the human user limitations and behavior. By being aware of
brain limitations of each user, the network can provide a unique experience for each user and
optimize its performance. For example, an increase in the delay of a wireless system may have
different effects on the QoE perceived by different human users. In particular, such different delay
perceptions can potentially be exploited by the cellular network to minimize power consumption
and reduce the amount of wasted resources. To our best knowledge, no existing work has studied
the impact of such disparate brain delay perceptions on wireless resource allocation.
The main contribution of the paper is, thus, a novel brain-aware learning and resource
management framework that explicitly factors in the brain state of human users during resource
allocation in a cellular network which has both human and machine type devices. In particular,
we formulate the brain-aware resource allocation problem using a joint learning and optimization
framework. First, we propose a novel learning algorithm to identify the delay perceptions of a
human brain. This learning algorithm employs both supervised and unsupervised learning to
identify the brain limitations and also creates a statistical model for these limitations based on
Gaussian mixture models. Then, using Lyapunov optimization, we address the resource allocation
problem with time varying QoS requirements that captures the learned delay perception. Using
this approach, the network can allocate radio resources to human users while considering the
reliability of both machine type devices and human users. We then identify a closed-form
relationship between system reliability and wireless physical layer metrics and derive a closed-
form expression for the reliability as a function of the human brain’s delay perception. Simulation
results using real data show that the proposed brain-aware approach can substantially save
 4

power in the network while preserving the reliability of the users, particularly in low latency
applications. In particular, the results show that the proposed brain-aware approach can yield
power savings of up to 78% compared to a conventional, brain-unaware system.
The rest of the paper is organized as follows. Section II introduces the system model. Sec-
tions III and IV present the proposed learning algorithm and resource allocation framework,
respectively. Section V presents the simulation results and conclusions are drawn in Section VI.

II. S YSTEM M ODEL AND P ROBLEM F ORMULATION

Consider the downlink of a cellular network with humans-in-the-loop having a single base
station (BS) serving a set H of N human users with their UEs and a set M of M machine
type devices (MTDs). Each UE or MTD can have a different application with different QoS
requirements such as sending a command to an actuator (for an MTD) or playing a 3D interactive
game (for a UE). We consider a time-slotted system with each slot duration being equal to the
LTE transmission time interval (TTI). We define K as the set of K resource blocks (RBs). In our
model, the packets associated with user i ∈ H ∪ M arrive at the BS according to independent
Poisson processes with rate ai (t). The lengths li , ∀i ∈ H∪M of the packets follow an exponential
distribution. Hence, each user’s buffer at the BS will follow an M/M/1 queuing model. The total
li
queuing and transmission delay of each user i is Di (t) = qi (t) + ri (t)
. The data rate for each
user is given by:
K  
X pij (t)hij (t)
ri (t) = B ρij (t) log2 1 + , (1)
j=1
σ2

where pij (t) is the transmit power between the BS and user i over RB j at time t and hij (t) is
the time-varying Rayleigh fading channel gain. In (1) ρij (t) = 1 if RB j is allocated to user i
at time slot t, and ρij (t) = 0, otherwise. B is the bandwidth of each RB.
The BS seeks to allocate RBs and power to the users according to their delay needs and their
channel state. The delay that MTD i can tolerate is βim , i.e., Dimax (βim ) = βm . βim is known
to the system. The delay perception of the brain of a human user who is using a given UE
is captured by βi (t). βi (t) essentially represents a delay perception threshold for human user i
at time t. If we decrease the delay below the threshold βi (t), the human user will not be able
to discern the difference. This delay perception is determined by the capabilities of the human
brain. By explicitly accounting for the cognitive limitations of the human brain, the BS can better
allocate resources to the users that need it, when they can actually use it. This is in contrast to
 5

conventional brain-agnostic networks [4], [18] in which resources may be wasted, as they are
allocated only based on application QoS without being aware on whether the human user can
indeed process the actual application’s QoS target.
We pose this resource allocation problem as a power minimization problem that is subject to
a brain-aware QoS constraint on the latency:
X h X j X ji
min P̄i + P̄i , (2a)
ρ(t),P (t)
j∈K i∈H i∈M

Pr Di (t) ≥ Dimax βi (t) } ≤ ǫi βi (t) ,




s.t. ∀i ∈ H ∪ M, (2b)

pij (t) ≥ 0, ρij (t) ∈ {0, 1} ∀i ∈ H ∪ M, j ∈ K, (2c)

X
ρij (t) = 1, ∀j ∈ K, (2d)
i∈H∪M

where ρ(t) is an (M + N) × K matrix having each element ρij (t). P (t) is an (M + N) × K

matrix with each element pi,j (t), representing the instantaneous power allocated to user i on RB
j. The term P̄ij = limt→∞ 1t t−1
P
τ =0 ρij pij (τ ) is the time average of the power allocated to user i

on RB j. Dimax βi (t) is the maximum tolerable delay, and 1 − ǫi βi (t) is the reliability of user

i. We define reliability as the proportion of time during which the delay of a given user does not
exceed a threshold. For notational convenience, hereinafter, we use the terms Dimax (βi (t)) and
Dimax interchangeably. The key difference between our problem formulation and conventional
RB allocation problems is seen in the QoS delay requirement in (2b). In (2b), the network
explicitly accounts for the human brain’s (and the MTDs’) delay needs. By taking into account
the features of the brain of the human UEs, the network can avoid wasting resources. This waste
of resources can stem from allocating more power to a UE, solely based on the application QoS,
while ignoring how the brain of the human carrying the UE perceives this QoS. Clearly, ignoring
this human perception can lead to inefficient resource management.
For finding βi (t), we propose a machine learning algorithm to identify the human brain delay
perception. Each human user has p features, (e.g., age, occupation, location) assumed to be
known to the BS. This time-varying feature vector is denoted by xi (t) ∈ Rp . We develop a
learning algorithm to build a model that maps these features to βi (t) for each user. We then
show that being aware of βi (t) can help the resource allocation algorithm to save a significant
amount of resources for low-latency systems. Here, we assume that BS has the access to the
user features xi (t). In practice, the BS can collect such data whenever a given user registers in
the network or by using the sensors of a user’s mobile device.
 6

III. M ACHINE L EARNING FOR P ROBABILITY D ISTRIBUTION I DENTIFICATION

To find the mapping βi (t) = f (xi (t)) between human features xi (t) and the delay percep-
tion of the brain, we use supervised learning [19], [20]. Since reliability is a key factor in a
communication system, we need a supervised learning algorithm that not only predicts βi (t) as
function of xi (t), but also gives a measure of reliability for this prediction. Hence, we cannot
rely on conventional supervised learning methods, such as neural networks [20]. Here, reliability
of predictions is defined as the probability that the prediction of βi (t) lies within a certain range
of the true values for βi (t).
To this end, we propose a novel supervised learning mechanism dubbed as probability distri-
bution identification (PDI) method that can find the mapping between the features of a human
user and the human brain’s perception on delay, as captured by βi (t), while quantifying the
reliability of this mapping. We will use this reliability to determine the overall reliability of
our system. As discussed in [21], the delay perception of a human brain typically follows a
multi-modal distribution. As a result, we design the proposed PDI approach to capture such a
model and find the different modes of a human brain. Then, using the distribution of the brain
delay, the PDI approach can find the effective delay of the human brain. This effective delay
determines relationship between βi (t) and xi (t) along with its reliability.
Consider a dataset {x1 (t), · · · , xn (t)}, where xi (t) ∈ Rp is one sample data vector. The
elements of xi (t) are features which can be both categorical (such as gender) and numerical (such
as age). For each input vector xi (t), we have a corresponding output value of delay perception
βi (t). This data can be collected using experiments or surveys such as those in [22]. Since we
can remove the data’s time dependency of the data using time-series techniques, hereinafter, we
use x instead of x(t). This dataset can be represented by a matrix X ∈ Rn×p , where xTi is row
i of X. Using PDI, we first create an n × (p + 1) dataset matrix W :
   
T T
w x β1 (t)
 1  1
 ..   .. .. 

W = [Xkβ] =  .  =  . . , (3)
   
wTn xTn βn (t)

where w i ∈ Rp+1 is a vector of the delay perception βi (t) and p other correlated features of the
human brain. Then, we fit a Gaussian mixture model (GMM) of m modes to our dataset using
the expectation-maximization (EM) algorithm [23]. In other words, we adopt an EM method for
learning the joint probability distribution using the joint dataset W , i.e., p(x, βi (t)). After finding
 7

the probability distribution p(x, βi (t)), we are able to cluster the data samples and find m modes
in the data. Then, each data vector xi is labeled based on its cluster so that each xi , i = 1, · · · , n
has a label in the cluster set C = {1, · · · , m}. Subsequently, our dataset is labeled using its cluster
number. In the next step, we train a classifier such that it finds a mapping between the input data
xi and its cluster number. These cluster numbers will correspond to the modes of the human brain
that determine its effective delay perception. When this proposed learning approach is deployed
in a wireless network, each user will be classified after connecting to the BS, and its brain mode
will be identified. Then, each user’s mode can be used to derive a probabilistic model of its
delay perception. Before delving into the PDI method for finding the brain’s effective delay, we
describe the Gaussian mixture model that we use for the human brain.
A multi-modal stochastic model is assumed for the brain features w i for user i. The proposed
distribution for w i is given by [19]:
X L
X
p(wi ) = p(z)p(w i |z) = πk f (wi |µk , Σk ), (4)
z k=1

where f (wi |µk , Σk ) is the probability density function for a multivariate normal distribution
with mean vector µk and covariance matrix Σk . Σk and µk represent the covariance matrix
and mean vector for mode k of the human brain, respectively. πk is the mixing weight of mode
z = zk . p(w) is the distribution of data points w i . L is total number of modes in the GMM.
The human brain will be in mode k with probability πk , and its features are generated using a
multivariate normal distribution with mean and covariance µk and Σk , respectively. The posterior
probability, i.e. responsibility, for mode k will be:
πk f (w|µk , Σk )
r(zk ) = PL . (5)
j=1 f (w|µ j , Σ j )
This responsibility can be used for clustering the data as well. After fitting the GMM on the
dataset, we can find the mode with highest responsibility for each data point and assign the data
to this mode. The EM algorithm is used to find these parameters based on a real-time human
brain behavior [23]. The log likelihood function for our dataset can be written as:
X L
X
ln L(Σ, µ, π|w) = ln p(w|Σ, µ, π) = ln πk f (wi |µk , Σk ). (6)
i k=1

The likelihood function in (6) has singularities and, hence, it is infeasible to find parameters πk ,
Σk , and µk . The EM algorithm is proposed in [24] to maximize the likelihood function for a
 8

Gaussian mixture model. We first initialize Σk , µk , and πk randomly. Next, we find the respon-
sibility for each mode using (5). Then, we reestimate parameters using current responsibilities.
Finally, the likelihood in (6) is maximized with respect to Σk , µk ,and πk .
In the next step, we use the derived joint distribution of human features and its associated
delay perception to derive a probabilistic model for the human insensitivity of the delay. In other
words, we find the relationship between a certain level of the delay Dimin and the probability
Pr(βi (t) > Dimin ). In the proposed PDI, along with the supervised learning component previously
explained, we also propose to use an unsupervised learning step to measure the reliability of our
predictions. In the unsupervised learning step, the data will be labeled based on the GMM as
follows. For each feature vector w i , the responsibilities r(zk ) = p(zk = 1|wi ) are found using
the EM algorithm. Then, the most probable mode is assigned as the label of this data, i.e.,

c(wi ) = arg max p(zk = 1|wi ). (7)

k

The output of the unsupervised learning step, ci , is used for training the supervised learning
h iT
model. We will form an output vector y which is defined as y = c(w 1 ) · · · c(wn ) . Then,
during the supervised learning step, we train a classifier so that it can find the mode using the
human features x as input. Given the data matrix X and the output vector c(wi ), this supervised
learning builds us a model f such that ci = f (xi ),i.e.,
n  
ξ c(wi ), fˆ(xi ) ,
X
f = arg min (8)
fˆ
i=1

where ξ(.) is a 0-1 loss function. f is approximated using a set of points (xi , ci ) and determines
the relationship between the features of a user and its cluster. After approximating f , given
each human user’s feature vector xi , we find the modes ci using model f . Finally, we bounds
Dimax (βi (t)) based on its features xi . Now that the system can identify the human users’ modes,
we need to find a relationship between a human user’s mode and the probabilistic model of its
delay perception by defining the concept of effective delay.

Definition 1. Given the statistical model for human delay perception, Dimin(ǫ′ ) is the effective
delay for human user i that satisfies:

Pr βi (t) < Dimin(ǫ′ ) < ǫ′ .


(9)

To find the effective delay for human user i, we first find the probability that the delay
perception of human user i is less than a threshold Dimin (ǫ′ ). In other words, we want to find the
 9

70

60

Age of human user (year)

50

40 Dmin
i

30

20
Dmin
i Cluster 1
10 Cluster 2
Predictive coverage 2
Predictive coverage 1
0
0 0.2 0.4 0.6 0.8 1 1.2
Delay perception (ms)

Figure 1: Finding Dimin using a GMM model for two different clusters.

relation between ǫ′ and Dimin(ǫ′ ) in (9). The concept of effective delay is defined using the fact
that delays less than Dimin(ǫ′ ) cannot be sensed by a human with (1 − ǫ′ ) certainty. The relation
between ǫ′ and Dimin(ǫ′ ) in (9) is found in Theorem 1. For notational simplicity, hereinafter, we
use Dimin instead of Dimin(ǫ′ ).

Theorem 1. The delay perception of the identified brain mode k user i is bounded such that
n q o
Pr |βi − µD
K | < Q p+1 (γ)eT
e
p+1 k p+1 > γ,
Σ (10)

where Σk and µk represent, respectively, the covariance matrix and the mean vector of the
identified brain mode k. Also, Qp (γ) is the quantile function of chi-square distribution with p
degrees of freedom, and is defined as
n Z x o
Qp+1 (γ) = inf x ∈ R|γ ≤ χ2p+1 (u)du , (11)
0

and ej is a unit vector in Rp+1 , whose jth element is 1 and all other elements are zero. p is
number of features used for learning. χ2p+1 (x) is the probability density function of a chi-square
random variable with p + 1 degrees of freedom. µD
K is the element p + 1 in vector µk .

Proof: See Appendix A.

As seen from Theorem 1, in addition to the delay perception element µD
K , the only other

parameter that affects the delay is eTp+1 Σk ep+1, which is the (p + 1)th diagonal element of the
covariance matrix Σk . Note that, we did not assume that matrix Σk is diagonal. Fig.1 shows
the relationship between Dimin and GMM. From Fig.1, we can see that, after finding the GMM
for the dataset, one can find the predictive coverage of each Gaussian distribution. Using this
 10

0.5
D D
i
=5 ms, i
=0.1
0.45 D D
i
=5 ms, i
=1
D D
0.4 i
=6 ms, i
=0.1
D D
i
=6 ms, i
=1
0.35

0.3

0.25

0.2

0.15

0.1

0.05

0
1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6
Minimum perceptible delay Dmin
i
( ) (ms)

Figure 2: Relationship between ǫ and Dimin (ǫ) for different values of µD D

K and ΣK .

predictive coverage, we can determine the probability with which βi (t) for a user i will be higher
than a threshold Dimin .
In order to find the effective delay for human user i, we first find the probability with which
the delay perception for human user i will be less than a threshold Dimin. In other words, we
will find the relationship between ǫ and Dimin in (9) using the following corollary that follows
directly from Theorem 1.

Corollary 1. As a direct result of Theorem 1, we can reduce (10) to

n q o 1−γ
D T
Pr βi (t) < µK − Q(γ)ep+1 Σk ep+1 < . (12)
2
Therefore, we find Dimin (ǫ) and ǫ in (9) as
q
Dimin(ǫ) = µD
K − Q(1 − 2ǫ)eTp+1 Σk ep+1. (13)

Since Q(γ) can only be calculated numerically, a closed-form relationship cannot be found
between Dimin(ǫ) and ǫ. However, we can numerically analyze this relationship, as shown in
Fig. 2. From Fig. 2, we can first observe that Dimin(ǫ) is an increasing function. This means
that the probability of the human brain noticing QoS differences for low delays will be much
smaller than for higher delays, which is an intuitive fact. Furthermore, it can be inferred that,
if the delay perception for a group of human users within a cluster is diverse, then the system’s
confidence on the delay perception of this group of humans will decrease, i.e., the estimation
of the delay perception of this group of human users will be less reliable. Next, we determine
constraint (2b) using Dimin (ǫ).
 11

As stated before, some delays are not perceptible to human users. To capture this feature,
we find Dimax (βi (t)) and ǫ(βi (t)) in problem (2) using Dimin(ǫ). Recall that Dimax (βi (t)) is a
parameter that will be used by the resource allocation system to represent the maximum tolerable
delay for the reliable communication of user i with 1 − ǫ(βi (t)) being the reliability of user i.
There are three possible cases for Dimax based on Dimin of a human user i:
1) Dimax > Dimin: In this case, the system will not be reliable even if we satisfy Pr(D >
Dimax ) < ǫ. The reason is that the human user has a delay perception of less than the maximum
delay Dimax and hence, the system is not reliable.
2) Dimax < Dimin : In this case, if the system is able to satisfy Pr(D > Dimax ) < ǫ, then the
system will be reliable, because user i cannot sense delays less than Dimin and its service delay
will not exceed Dimax .
3) Dimax = Dimin : If this equality holds, the system will be reliable and it will also have
prevented a waste of resources. If any given user cannot perceive delays less than Dimin , then it
is not effective to allocate more resources to this user.
S represents the event where the system delay meets the cognitive perceptions of the human
user, which is the desired result (case 2) and case 3)). In other words, if event S happens, the
system is reliable. Also, we assume that the events E1 and E2 are defined as D < Dimax and
βi (t) > Dimin, respectively. We know that for case 1, event E1 ∩ E2 is a subset of event S, and
in case 2, event S is a subset of event E1 ∩ E2 . Similarly, in case 3, event E1 ∩ E2 is same as
event S. Since the probability of E1 ∩ E2 can be computed, if we set Dimin to Dimax (case 3),
we will be able to find S based on the system parameters and design the resource allocation
system in a way that each user is satisfied. We know that:
 
Pr(E1 ∩ E2 ) = 1 − Pr (D > Dimax ) ∪ (βi (t) < Dimin) (14)
 
= 1 − Pr(D > Dimax ) + Pr(βi (t) < Dimin ) − Pr(D > Dimax )Pr(βi (t) < Dimin) . (15)

(14) follows from De Morgan’s law, and (15) is true since D and βi (t) are two independent
random variables. Therefore, if Dimin = Dimax for user i and ǫǫ′ is small, we can see that
 
1 − Pr(D > Di ) + Pr(βi (t) < Di ) ≥ 1 − (ǫ + ǫ′ ),
max min
(16)

and, hence,
Pr(S) = Pr(E) > 1 − (ǫ + ǫ′ ), (17)
 12

where Pr(S) is the reliability of the system defined in (2). Subsequently, as we design the system,
we consider the reliability as a predetermined target design parameter for the system. Using this
parameter, we can set ǫ and ǫ′ . Given ǫ′ and numerical function Dimin (ǫ′ ) derived in (12), Dimin
can be determined. Now, given ǫ(βi (t)) and Dimax (βi (t)), we can fully characterize problem (2).

IV. B RAIN -AWARE R ESOURCE M ANAGEMENT

To solve problem (2), we propose a novel brain-aware resource management framework that
takes into account the time-varying wireless channel and the time-varying brain-aware delay
constraint (2b). In this section, we transform this constraint into a mathematically tractable form.
First, in the next lemma, the relation between the packet length distribution and the service time
distribution for a packet is shown.

Lemma 1. If a fixed rate ri is allocated to a user and the packet lengths follow an exponential
distribution with parameter χ, then, the distribution of the service time s will also be exponential
with parameter χri .

Proof: The CDF of the exponential distribution is Fψ (l) = Pr(l < ψ) = 1 − e−χψ . Hence,
l
FS (s) = Pr(s < S) = Pr( < S) = Pr(l < ri S) = Fri S (s) = 1 − e−χri S . (18)
ri
This means that the PDF for the service time is fS (s) = e−χri s .
Therefore, without loss of generality, we assume that the service time of each packet is
exponential with parameter ri , which is the same as the rate allocated to the user.
Here we assume that for any given user, the packets arrive with the rate ai (τ ), and the user
data rate is ri (τ ) in slot τ = 1, · · · , t. In the next theorem, We derive the probability with which
the delay of a given user i exceeds a threshold Dimax .

Theorem 2. Assume that user i has a time varying rate riτ at time slot τ . If the duration of each
time slot is long enough for the queue to reach its steady state, i.e.,
1
<< δτ, (19)
ri (τ ) − ai (τ )
then, the probability that the delay exceeds a threshold is
t
1X −


ri (τ )−ai (τ ) Dimax
Pr(D > Dimax ) = lim e , (20)
t→∞ t
τ =1

under the condition that ri (τ ) > ai (τ ) for all τ > 0.

 13

0.9

0.8

0.7

)
0.6

Pr( Di > Dmax

i
0.5

0.4

0.3

0.2

0.1

0
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
Dmax
i
(ms)

Figure 3: Comparison between simulation results and the result of Theorem 2.

Proof: See Appendix B.

Theorem 2 shows that constraint (2b) is satisfied if the network can satisfy the following
condition
t
1X −


ri (τ )−ai (τ ) Dimax
lim e < ǫ. (21)
t→∞ t
τ =1

Fig. 3 shows the relationship between the theoretical result from Theorem 2 and the simulation
results. It shows that the simulation and analytical results are a near-perfect match. The maximum
error between the analytical and simulation results is only 0.0146.

A. Optimal Resource Allocation with Guaranteed Reliability

Constraint (21) is analogous to the drift-plus-penalty method in the Lyapunov optimization

framework [25] and, hence, we will use this framework to solve (2). The problem has a time-
varying nature since the human brain conditions and needs will change from time to time. The
users’ processing state β(t) is also a function of time, and accordingly, the latency needs in
(2b) will be time-varying. Therefore, we need to solve the optimization problem (2) during each
time slot efficiently. Here, we propose an algorithm with a low computational complexity for
solving problem (2). The drift-plus-penalty approach is used to stabilize a queue network while
minimizing time average of a penalty function. In order for constraint (2b) to be satisfied in all
 14

time slots, we need to make sure that (20) will always be smaller than ǫ. For this reason, we
define a virtual queue:


− ri (t+1)−ai (t+1) Dimax
Fi (t + 1) = max{Fi (t) + e − ǫ}. (22)


ri (t+1)−ai (t+1) Dimax
We can see that e− − ǫ < Fi (t + 1) − Fi (t). Consequently, we obtain:
t

− ri (t+1)−ai (t+1) Dimax
X
e − ǫt < Fi (t) − Fi (0). (23)
τ =1

If Fi (0) is bounded, we have:

t
1X − Fi (t)


ri (t+1)−ai (t+1) Dimax
lim e − ǫ < lim . (24)
t→∞ t t→∞ t
τ =1
Fi (t)
If the queue Fi (t) is mean-rate stable, that is, limt→∞ t
= 0, then we have:
t
1X −


ri (t+1)−ai (t+1) Dimax
lim e < ǫ. (25)
t→∞ t
τ =1

The Lyapunov function is defined for all the queues in the base station as
1 X
Y (t) = Fi (t)2 . (26)
2 i∈M∪H

Then, we can find the drift function ∆t = Y (t + 1) − Y (t) as:

1 X 1 X 1 X X
Y (t + 1) = Fi (t + 1)2 ≤ Fi (t)2 + y(t)2 + y(t)Fi (t), (27)
2 i∈M∪H 2 i∈M∪H 2 i∈M∪H i∈M∪H

where

− ri (t+1)−ai (t+1) Dimax
yi (t) = e − ǫ. (28)

Thus,
1 X X
∆t ≤ y(t)2 + y(t)Fi (t). (29)
2 i∈M∪H i∈M∪H
P
We can form the drift-plus-penalty by adding V i,j pij (t) to both sides of inequality (29),
P
where i,j pij is the total power of the BS which we want to minimize, and V is a parameter
that determines how important minimizing the objective function (2a) is in comparison with
satisfying (2b). We can balance the tradeoff between power and delay. The drift-plus-penalty
inequality is
X 1 X X X
∆t + V pij ≤ y(t)2 + V pij (t) + y(t)Fi (t). (30)
i,j
2 i∈M∪H i,j i∈M∪H
 15

Given that we assumed ri (t) > ai (t) for all t, we know that |yi (t)| < 1 ∀t, i ∈ M ∪ H, and
hence, we can rewrite (31) as
X X X
∆t + V pij ≤ B + V pij (t) + y(t)Fi (t), (31)
i,j i,j i∈M∪H

1 |H|+|M|
y(t)2, and is equal to
P
where B is the upper bound of 2 i∈M∪H 2
. |A| is the cardinality of
set A. Using the drift-plus-penalty algorithm [26], we know that, by minimizing the right hand
side of equation (31), queue Fi (t) will be mean-rate stable, and hence, the condition yi (t) < 0
will be satisfied. As a result, constraint (2b) will also be satisfied. Furthermore, we know that
by minimizing the right hand side of (31), cost function (2a) is also minimized, owing to the
fact that (2a) is defined as a penalty function. By minimizing the right hand side of (31), our
optimization problem can be converted to the following time-varying problem:
X X
min V pij (t) + yi (t)Fi (t), (32a)
ρ(t),P (t)
i,j i∈M∪H

s.t. ri (t) > ai (t), ∀i ∈ H ∪ M (32b)

pij (t) ≥ 0, ρij (t) ∈ {0, 1}, ∀i ∈ H ∪ M, j ∈ K, (32c)

X
ρij (t) = 1, ∀j ∈ K. (32d)
i∈H∪M

As discussed earlier, the cost function in (32a) is equivalent to (2a) and (2b) in the original
optimization problem. Learning the effective delay of each human user using our proposed PDI
method determines the parameters yi (t) and Fi (t) in the problem (32a). However, in order to
satisfy (2b), we need to also satisfy (32b). The reason for adding (32b) is that if this constraint
is not satisfied in any time slot, the queue length will approach infinity. Constraints (32c) and
(32d) are feasibility conditions and remain the same as (2). Hence, by solving (32) in each time
slot, the original problem (2) will be solved.
Nonetheless, problem (2a) is not a convex optimization problem, due to the fact that it is
a mixed integer problem and its complexity increases exponentially with the number of users.
Since (2a) needs to be solved at each time slot, this exponential order of complexity makes the
implementation infeasible. Consequently, we should use a dual decomposition method to break
down optimization problem (32) to smaller subproblems, and find the optimal solution to (32)
using a low complexity method.
It is rather challenging to solve (32) using a dual decomposition method, as the structure of
yi (t) makes it infeasible to decompose the objective function for each resource block. In order
 16

to overcome this challenge, we convert (32) to a decomposable form. Then, we will show that
this converted problem is equivalent to (32).
For this purpose, the Lagrangian for problem (32) is written as
X X X

V pij (t) + yi (t)Fi (t) + λi ai (t) − ri (t) , (33)
i,j i∈M∪H i∈M∪H


− ri (t+1)−ai (t+1) Dimax
where λi is the Lagrange multiplier. As we know, yi (t) = e − ǫ. Therefore,
the only decision variables are allocation of resource blocks to the users and allocating power
to each resource block. Although Fi (t) is a function of yi (t), it is not a decision variable and is
treated as a constant. Hence, (33) can be rewritten as
X X − r (t+1)−a (t+1)
Dmax X

V pij (t) + e i i i Fi (t) + λi ai (t) − ri (t) . (34)
i,j i∈M∪H i∈M∪H

The main optimization problem consists of two components. First, minimizing the total power
of the BS with weight V , and second, minimizing the summation i∈M∪H e−ri (t+1) which has
P
max
a weight Fi (t)e−ai (t+1)Di for each user i.
As we can see, (34) is not decomposable for each resource block. Here we will have an
approximation of (32) and then propose an algorithm to solve

this approximation efficiently. In
max
this C-additive approximation, i∈M∪H e− ri (t+1)−ai (t+1) Di Fi (t) in (34) is substituted with
P

its linear approximation of exponential term e−x at x = 0.

X

− ri (t + 1) − ai (t + 1) Fi (t). (35)
i∈M∪H

In the original problem, if yi (t) starts to become greater than
zero for user i, then Fi (t) will
max
increase and it will give more weight to the term e− ri (t+1)−ai (t+1) Di . As a
result, the algorithm
max
allocates more resources to user i such that it minimizes
e− ri (t+1)−ai (t+1) Di for user i, and
max
accordingly, yi (t) decreases. Hence, Fi (t)e− ri (t+1)−ai (t+1) Di plays the role of feedback in the
system. As we can see from (35), this approximation will not change this feedback mechanism
and plays the same role in the system. Therefore, we can write
 X X
− ri (t + 1) − ai (t + 1) Dimax Fi (t)


min V pij (t) +
P ,ρ
i,j i∈M∪H


ri (t+1)−ai (t+1) Dimax
 X X −
< C + min V pij (t) + e Fi (t) . (36)
P ,ρ
i,j i∈M∪H
 17

Using this C-additive approximation, it can be easily proved that all terms are mean-rate stable.
Hence, (2b) in the original problem is satisfied [25]. Finally, problem (2) can be presented as:
X X
ri (t + 1) − ai (t + 1) Dimax Fi (t),


min V pij (t) −
ρ(t),P (t)
i,j i∈M∪H

s.t. ri (t) > ai (t), (37a)

Pin (t) ≥ 0, ∀i ∈ H ∪ M, n ∈ N , (37b)

ρni (t) ∈ {0, 1}, ∀i ∈ H ∪ M, n ∈ N , (37c)

X
ρij (t) = 1, ∀j ∈ K. (37d)
i∈H∪M

In order to solve this problem, we can decompose it into K subproblems. Since these subproblems
are coupled through constraint (37d), we use the dual decomposition method for solving (37)
[27]. First, the Lagrangian is written for problem (37), and in the second step, it is decomposed
for each resource block. After that, the resource block allocation and the power of each RB are
found in terms of the Lagrange multiplier λ. Finally, λ is calculated using an ellipsoid method.
The Lagrangian for problem (37) is
X X
− ri (t) − ai (t) Dimax Fi (t) − λi ri (t) − ai (t)



L(P , ρ, λ) = V pi,j (t) +
i,j i∈M∪H
X X
λi + Dimax Fi (t) ri (t) − ai (t) .



=V pi,j (t) − (38)
i,j i∈M∪H

One major difference between our problem and conventional power minimization problems is
that there is an additional term Dimax Fi (t) added to the Lagrange multiplier (the shadow price).
In this problem, Dimax Fi (t) plays the role of a bias term. Therefore, a new hypothetical Lagrange
multiplier λ′i is assumed and defined as

λ′i = λi + Dimax Fi (t). (39)

This means that adding constraint (2b) to the problem instead of constraint (37a) increases the
shadow price by a factor of Dimax Fi (t). Increasing the shadow price for a constraint makes it
looser. As a result, in many time slots, constraint (37a) will not be a tight constraint and the
+
Lagrange multiplier will be set to λi = λ′i − Dimax Fi (t) . the Lagrange dual function is


g(λ) = min L(P , ρ, λ). (40)

ρ(t),P (t)
 18

The minimization problem (40) can be decomposed to K subproblems. gj′ (λ) can be written as
X X hi,j pi,j

gj′ (λ) = min V λi + Dimax Fi (t) W log2 (1 + K


pi,j − ) , (41)
P (t)
i i∈M∪H
σ2
where D is a set of feasible pij s in which for RB j, there is only one i that pij 6= 0. Hence,
g(λ) is
X X
gj′ (λ) + λi + Dimax Fi (t) ai (t) .



g(λ) = (42)
j i∈M∪H

If λ is fixed, gj′ (λ) is a convex function of P . Therefore, P is found by taking a derivate with
respect to pij and setting it to zero. This results in
h λ + D max F (t)
W σ 2 i+
i i i
pij = − . (43)
V log2 Khij
The optimal RB allocation for RB j is k(j), and can be written as
X X hi,j pi,j

λi + Dimax Fi (t) W log2 (1 + K


k(j) = argminV pi,j − ) , (44)
i i i∈M∪H
σ2
X X hi,j pi,j

gj′ (λ) = minV λi + Dimax Fi (t) W log2 (1 + K


pi,j − 2
) . (45)
i
i i∈M∪H
σ
Thus, ρ∗ij and p∗ij will be given by:
 
1, i = k(j),
 pij ,

i = k(j),
∗
ρij = p∗ij = (46)
0, otherwise.
 0,

otherwise.
hi,j p∗i,j
Hence, the optimal rate becomes ri∗ =
P
j W log2 (1 + K σ2
). The only parameter that
affects this joint RB and power allocation is λ. As the number of RBs increases, the duality gap
in this problem approaches zero [27]. We know that the optimal value is found by maximization
of g(λ) with respect to λ. In order to find λ, we use the ellipsoid method [28], and to do so,
we have to find the sub-gradient for the dual objective g(λ). The following theorem will show
that the subgradient for (38) is a vector with elements di = ai − ri .

Theorem 3. The subgradient of the dual optimization problem with dual objective defined in
(42), is the vector d whose elements di , ∀i ∈ H ∪ M are given by:

ai − r ∗ , ai ≥ r ∗ ,

i i
di = (47)
0, ai < ri∗ .


Proof: Since
g(λ) = min L(P , ρ, λ) = L(P ∗ , ρ∗ , λ), (48)
P ,ρ
 19

we have:

g(δ) ≤L(P ∗ , ρ∗ , δ)
X X
p∗i,j (t) − δi + Dimax Fi (t) ri∗ (t) − ai (t)



=V
i,j i∈M∪H
X X
p∗i,j (t) − λi + Dimax Fi (t) ri∗ (t) − ai (t)



=V
i,j i∈M∪H

+ (λi − δi ) ri∗ (t) − ai (t) = g(λ) + (λ − δ)T d′ ,

(49)

where
h iT
d′ = r1∗ − a1 · · · rN
∗
+M − aN +M . (50)

However, because of the term Dimax Fi (t), when λ = 0 and ai < ri∗ , then the direction of
d′ is infeasible. Using the projected subgradient method [29], we can transform this infeasible
direction to a feasible one. The update rule for projected subgradient is:

λ(k+1) = Π(λ(k) − αk d′k ) (51)

where αk is the step size and Π is the Euclidan projection on the feasible set. Since the feasible
set is λi > 0, we can see that

Π(λ(k) − αk d′k ) = λ(k) − αk dk , (52)

where:  
 d′ , d′ ≥ 0
 ai − r ∗ ,

ai ≥ ri∗ ,
i i i
di = = (53)
0, d′i < 0 0, ai < ri∗ .
 

B. Complexity Analysis

Next, we find the complexity of our algorithm which needs to be run in each iteration. There
are K RBs in our problem, for each of which (44) needs to be evaluated for M + N users.


It takes O (M + N)K times to solve a primal problem. Subsequently, the dual problem will
be solved, which gives us the optimal value of λ in an M + N dimensional space and has a
complexity of O (M + N)2 . Therefore, the overall complexity should be O (M + N)3 K .

However, as mentioned before, adding Dimax Fi (t) to the Lagrange multiplier sets a major part


of it to zero, and as a result, the order of complexity will decrease to O (M + N)K .
 20

300

250

200

Number of users 150

100

50

0
20 40 60 80 100 120 140 160
i

Figure 4: Distribution of βi for the 1000 users in dataset.

V. S IMULATION R ESULTS AND A NALYSIS

For our simulations, we consider the dataset in [22] to model the delay perception of a human
user. In [22], the authors conducted human subject studies using 30 human users, where each
subject is asked to rate the quality of 5 movies while the delay and packet loss in the system is
being increased. We used the average score of each human user to estimate their delay perception.
We also used a variation of the bootstrap method [30] to increase the number of data points to
1000. We can see the histogram of the delay perception for these 1000 data points in Fig. 4.
Also, since the dataset has no features for each user, we attributed three random continuous
features to each user. Hence, each user is associated with a vector w ∈ R4 .
We consider a network with a bandwidth of 10 MHz, ai (t) = 1 Mbps, σ 2 = −173.9 dBm, and
ǫ = 0.05. We use a circular cell with the cell radius of 1.5 km. We set the path loss exponent
to 3 (urban area) and the carrier frequency to 900 MHz. The packet length is an exponential
random variable with an average size of 10 kbits. We use 5 MTD and 5 UE in the system and
we set Dimax to 20 ms for them, unless otherwise mentioned. For the brain aware users, we
arbitrarily select 5 UE in the system out of all data points.
Fig. 5 shows the within cluster point scatter for the EM algorithm in our dataset. This within
cluster point scatter for a clustering C is defined as [30]:
n
1X X X
W (C) = d(xi , xi′ ), (54)
2
k=1 c(i)=k c(i )=k
′
 21

108
2

1.8

1.6

Within cluster point scatter

1.4

1.2

0.8

0.6

0.4

0.2

0
1 2 3 4 5 6 7 8 9 10
Number of clusters

Figure 5: Within point scatter for the EM clustering method on the datasest.

where d is an arbitrary distance metric. In essence, the within cluster point scatter is a loss
function that allows the determination of hyper-parameters in the clustering algorithm. The hyper-
parameter that we seek to find here is the number of clusters in the dataset. As we can see from
Fig. 5, after the number of clusters reaches 5, increasing the number of clusters does not decrease
the within cluster point scatter substantially. Hence, the optimal number of clusters with is 5.
Fig. 6 shows the total BS power resulting from the proposed brain-aware case and from a
brain-unaware case in which UEs have a fixed constraint (2b) with Dimax between 10 ms to
60 ms. Here, the total power is the objective of main optimization problem (2). Fig. 6 shows
that, as the latency increases, the total power decreases, because it is easier to satisfy constraint
(2b) at higher latencies. Also, at higher delays, being brain-aware will no longer yield substantial
gains, since βi (t) and Dimax become close to each other and learning βi (t) cannot save resources
for the system. In contrast, in Fig. 6, we can see that for stringent low-latency requirements, the
proposed brain-aware approach yields significant gains in terms of saving power. In particular,
for 10 ms delay in (2b), Fig. 6 shows that the BS in brain-unaware approach uses 44 % more
power compared to the brain-aware case. These results stem from the fact that a brain-aware
approach can minimize waste of resources and provide service to the users more precisely based
on their real brain processing power. Fig. 7 shows average BS power for different number of
MTDs. As we can see from Fig. 7, the brain-aware approach will always outperform the brain-
unaware approach as the number of MTD increases. For the case of 30 MTD user, the BS in
brain-unaware approach uses 16 % more power compared to the brain-aware case. This is due
 22

0.4
3.5

0.35 3
Average power (W)

2.5
0.3

Average power (W)

2

0.25
1.5

0.2 1

0.5
0.15
10 15 20 25 30 35 40 45 50 55 60
0
Maximum tolerable delay Di max (ms) 0 5 10 15 20 25 30

Figure 6: Average power usage of the system

Figure 7: Average power usage of the system
as function of different latency requirements
for different number of MTDs and 5 UEs.
for the users.
36

34

32

30
Average power (dBm)

28

26

24

22

20

18

16
0 5 10 15 20 25 30

Figure 8: Average power usage of the system for different number of UEs.

to fact that brain-aware approach can allocate resources more efficiently in case of a shortage
in resources.
In Fig. 8, we show the average power usage of the system when the number of UEs increases
from 2 to 30 with Dimax set to 20 ms. As the number of users increases, the average power
consumption of the system will also increase. This is due to the fact that increasing the number
of users will decrease the bandwidth per user. Since the delay and rate requirements of each
 23

0.03 3

0.025 2.5

0.02 2

User rate (Mbps)

Total power (W)

0.015 1.5

0.01 1

0.005 0.5

0 0
20 40 60 80 100 120 140 160 180 200 20 40 60 80 100 120 140 160 180 200
Time slot Time slot

Figure 9: Transmit power for 4 different users. Figure 10: Transmission rate for four different
The delay perception of two of the users is users. The delay perception of two of the users
learned. is learned.

user are still unchanged, the system needs to use more power to compensate for the bandwidth
deficiency. From Fig. 8, we can see that, in the case of 30 users, the brain-aware system is able
to save 6.7 dB (78%) on average in the BS power. The brain-aware system can allocate resources
based on each user’s actual requirement instead of the predefined metrics and this leads to this
significant saving in the power consumption of the BS.
In Fig. 9, Fig. 10, and Fig. 11, we consider the case of 7 UEs and 5 MTDs. Two UEs are
chosen as brain-aware users and their delay perception is learned by the PDI method. One of
the brain-aware UEs has a delay perception of βi = 133.73 ms, and the other one has βi equal
to 26.8 ms. The system does not learn the delay perception of the 5 remaining UEs and, hence,
it allocates resources to them by using a predefined delay requirement (brain-unaware users).
As we can see in Fig. 9, the power consumption of the first two brain-aware users will be less
than that of the brain-unaware users. Moreover, the power consumption for a user with higher
delay perception will be less than that of a user with lower delay perception. This shows that
the system can successfully allocate resources according to the delay perception of the users.
Furthermore, the power consumption related to each user with predetermined delay requirements
is different, due to their different channel gains. However, as we will see later, the system is
robust to such differences and can guarantee the reliability and rate requirements for users having
different channel gains.
 24

1 0.4

0.9
0.35
0.8
V=1
0.3 V=1.3
0.7 V=1.6
V=1.9

Total power (W)

0.25 V=2.2
0.6
Reliability

0.5 0.2

0.4
0.15
0.3
0.1
0.2
0.05
0.1

0 0
10 20 30 40 50 60 70 80 90 100 5 10 15 20 25 30 35 40 45 50
Time slot Time slot

Figure 11: Reliability for four different users. Figure 12: Effect of parameter V on the
The delay perception of two of the users is convergence of the resource allocation
learned. algorithm.

In Fig. 10, we show the transmission rate for four different users. We can see that the rate for
brain-unaware users with predetermined delay will converge to 2.5 Mbps. This rate will ensure
the reliability for these users. However, the rate of the users with learned delay perception
will converge to a smaller rate. This is due to the fact that these users’ actual requirements
are known to the system, and the system uses this knowledge to avoid unnecessarily wasting
resources. However, as we will see next, this rate reduction does not change the reliability for
these users.
Fig. 11 shows the reliability for the four aforementioned users. As we can see, the reliability
of all the users will converge to 95 %, which is the target reliability value for the users. We can
see that the system is able to ensure reliability for the users with identified delay perceptions as
well as the users with predefined delay requirements. However, the system used 45% less power
for those users for which the delay perception is learned.
Finally, Fig. 12 investigates the effect of parameter V for the system with 5 MTDs and 5 UEs.
We can see that, as V increases from 1 to 1.9, the convergence time decreases from 40 iterations
to 15 iterations. Nevertheless, increasing V will make the algorithm unstable, and as we can see,
increasing it to 2.2 will create an overshoot which is 11% higher than the final value. Hence,
parameter V , if adjusted correctly, can create a balance between stability and convergence rate
of our algorithm.
 25

VI. C ONCLUSION

In this paper, we have introduced and formulated the notion of delay perception of a human
brain, in wireless networks with humans-in-the-loop. Using this notion, we have defined the
concept of effective delay of human brain. To quantify this effective delay, we have proposed a
novel learning method, named PDI, which consists of an unsupervised and supervised learning
part. We have then shown that PDI can predict the effective delay for the human users and
find the reliability of this prediction. Then, we have derived a closed-form relationship between
the reliability measure and wireless physical layer metrics. Next, using this relationship and the
PDI method, we have proposed a novel approach based on Lyapunov optimization for allocating
radio resources to human users while considering the reliability of both machine type devices
and human users. Our results have shown that the proposed brain-aware approach can save a
significant amount of power in the system, particularly for low-latency applications and congested
networks. To our best knowledge, this is the first study on the effect of human brain limitations
in wireless network design.

A PPENDIX A
P ROOF OF T HEOREM 1

We assume that a single brain mode is dominant for each user at each time. We index this
single mode as k. For each user i with this dominant mode, wi = [w1 , · · · , wn ] has the following
probability density function:
1
h 1 i
f (w i ) = |2πΣk |− 2 exp − (wi − µk )T Σ−1
k (w i − µ )
k . (55)
2
We want to find the smallest region D in Rp+1 , in which the delay perception lies with probability
γ, i.e., Z Z
··· f (w1 , w2 , . . . , wn ) dw1· · · dwn = γ. (56)
D

D is not a unique region. However, the objective is to find the smallest region. To this end, we
need to find the region where f (w1 , w2 , . . . , wn ) has the greatest value, i.e., if
Z Z Z Z
··· f (w1 , w2 , . . . , wn ) dw1· · · dwn = · · · f (y1, y2 , . . . , yn ) dy1· · · dyn , (57)
D1 D2

and also
f (y1, y2 , . . . , yn ) ≤ f (w1 , w2 , . . . , wn ) ∀y ∈ D2 , ∀w i ∈ D1 , (58)
 26

then Z Z Z Z
··· dw1 · · · dwn ≤ ··· dy1 · · · dyn , (59)
D1 D2

which implies that the volume of the region D1 is smaller than the volume of D2 . Hence, if we
find the region D for which (56) holds, and, using (58), show that all other regions for which
(56) holds have greater volumes, then, we would have found the smallest region D, in which
the human behavior will stay with the probability γ.
Since w i is distributed according to a multivariate Gaussian, we can find the region where it

has the highest probability density, i.e., wi |f (wi ) > C1 . This region can be written as:
  h 1 i 
1
− −1
 T
w i  |2πΣk | 2 exp − (wi − µk ) Σk (wi − µk ) > C1 , (60)
2
which is equivalent to
n  o
T −1
D = w i (w i − µk ) Σk (w i − µk ) < C2 , (61)


1
where C2 is a positive constant and equals − ln |2πΣk | 2 C1 . Since Σk is a positive definite matrix,
(61) is the inner volume of an ellipsoid in a p dimensional space.
We now conjecture that this ellipsoid D is the smallest region, in which the delay perception
lies with probability γ, i.e., the probability of w i being in this region is γ. We use a proof by
contradiction to show this. Consider that there exists any other space E which is smaller than D,
and the probability of w i being in this region is γ. We can partition E into two parts A = E ∩ D
and E2 = E ∩ D ′, where D ′ is the complement of the set D. We also define D2 = D ∩ E ′ . We
know that
Z Z Z
f (wi )dwi = f (w i )dwi + f (wi )dwi (62)
D A D2
Z Z Z
= f (wi )dwi = f (wi )dwi + f (wi )dwi (63)
E A E2

=γ. (64)
R R
Hence, D2
f (wi )dwi = E2
f (wi )dw i . Since

f (wi ) < C1 ≤ f (y) ∀w i ∈ E2 , y ∈ D2 , (65)

using (57) and (58) we have Z Z

dwi < dwi . (66)
E2 D2
 27

This means that the set E has a bigger volume than D, which is a contradiction to our first
assumption. This proves that region D is the smallest region in Rp+1 that has the probability γ.
R
Next, we find the relation between C2 and γ. γ can be defined as D f (wi )dwi and can be
calculated using chi-square distribution [31]. The region D can be written as

D = wi |(w i − µk )T Σ−1

k (w i − µk ) ≤ Qp (γ) , (67)

where Qp (γ) is the quantile function of the chi-square distribution with p degrees of freedom.
It is defined as
n Z x o
Qp (γ) = inf x ∈ R|P c ≤ χ2p (u)du . (68)
0

Having defined the confidence region based on γ, we now must find the edges of this ellipsoid.
We know that the center of this ellipsoid is µk . We need to solve the following optimization
problem:
min or max eTi w i , s.t. w i ∈ D, (69)
wi wi

where ei is a unit vector in Rp , having 1 in its ith element and zero otherwise. Using KKT
conditions for solving the above problem, we have:

ei + λ Σ−1
k (w i − µk ) = 0, (70a)

(w i − µk )T Σ−1
k (w i − µk ) ≤ Qp (γ), (70b)
 
λ (wi − µk )T Σ−1k (w i − µ k ) − Qp (γ) = 0, λ ≥ 0. (70c)

The inequality in (70b) is tight. With some algebraic manipulation, we have w i −µk (i) = λ1 Σei ,
and so, e Σk Σ−1
1 T
λ2 i k Σk ei= Q(γ). Therefore,
s
Q(γ)
wi = ± T
Σk ei + µk , (71)
ei Σk ei
s
eTi Σk ei
λ=± , (72)
Q(γ)
q
T
ei w i = ± Q(γ)eTi Σk ei + µk (i). (73)
p
If λ is positive, we can find the maximum which is + Q(γ)eTi Σk ei + µk (i), and if λ is
p
negative, we can find the minimum which is − Q(γ)eTi Σk ei + µk (i). Here, µk (i) is the ith
element of µk . If we set i = p + 1, then the delay perception of user i is in the following range:
q q
D
− Q(γ)ep+1 Σk ep+1 < βi − µk < Q(γ)eTp+1 Σk ep+1 ,
T
(74)
 28

at least with probability γ. Hence, Theorem 1 is proved.

A PPENDIX B
P ROOF OF T HEOREM 2

Since the queuing delay is much smaller than the duration of each time slot, we can assume
that each packet arriving at a specific time slot will be served at the same time slot. For analyzing
the packet delay, we consider a packet that just arrives in the system in time slot τk , and find
Pr(D > Dimax ) for this packet. When this packet arrives, there are m packets in the system.
From lemma 1, we know that the serving time will be an exponential random variable. Since
the exponential distribution is memoryless, there is no distinction between a packet already in
service and the other packets. Therefore, the waiting time for the packet that has just arrived is
the summation of m exponential distributions. Also, the transmission delay for this packet will
be another exponential random variable. Hence, the delay of a packet which arrives at time slot
τk while there are m packets in the system can be written as:

d(τk , m) = ts + t1 (τk ) + t2 (τk ) + · · · + tm−1 (τk ) + tc (τk ), (75)

where ti (τk ) is the service time for packet i in the queue, and tc (τk ) is the service time for
packet already in service. Also, ts is the service time for the packet that has just arrived. we
seek to find Pr(d(τk , m) > Dimax ) which can be written as

X
Pr d(τk , m) > Dimax = Pr(D > Dimax |m, τk )Pr(m, τk )
m,k
X
= Pr(D > Dimax |m, τk )Pr(m|τk )Pr(τk ). (76)
m,k

The probability that there are m users in an M/M/1 queue at time slot τk , i.e. Pr(m|τk ), can be
written as (see [32]):  m  
ai (τk ) ai (τk )
Pr(m|τk ) = 1− . (77)
ri (τk ) ri (τk )
Since we assumed the time slots have equal lengths, the packets arrive at each time slot with
equal probability of Pr(τk ) = 1t , where t is the total number of time slots.
1
The sum of m + 1 identically independent exponential random variables with the mean ri (τk )

is a gamma random variable. Consequently, if the users arrive at time slot τk while there are m
users in the system at the time of arrival, the distribution of delay is
ri (τk )m+1 m −ri (τk )φ
fD (φ|m, τk ) = φ e . (78)
Γ(m + 1)
 29

As a result, we can write the probability of delay exceeding a threshold Dimax as

Z ∞ X
max
Pr(D > Di ) = fD (φ|m, τk )Pr(m|τk )Pr(τk )dφ (79)
Dimax m,k
∞
1 X ri (τk )m+1 m −ri (τk )φ ai (τk ) m ai (τk )
Z
= φ e ( ) (1 − )dφ (80)
Dimax t m,k m! ri (τk ) ri (τk )
t Z ∞
m
1X ∞ −ri (τk )φ
X φai (τk )
= (ri (τk ) − ai (τk ))e dφ (81)
t k=1 Dimax m=0
m!
t ∞
1X
Z

− ri (τk )−ai (τk ) φ
= (ri (τk ) − ai (τk ))e dφ (82)
t k=1 Dimax
t
1X −


ri (τk )−ai (τk ) Dimax
= e , (83)
t k=1
which proves the theorem.

R EFERENCES
[1] A. T. Z. Kasgari, W. Saad, and M. Debbah, “Brain-aware wireless networks: Learning and resource management,” in Proc.
51th Asilomar Conference on Signals, Systems and Computers, Pacific Grove, CA, USA, Nov 2017.
[2] E. Gobbetti and R. Scateni, “Virtual reality: Past, present, and future,” Virtual environments in clinical psychology and
neuroscience: Methods and techniques in advanced patient-therapist interaction, Nov 1998.
[3] M. Chen, W. Saad, and C. Yin, “Virtual reality over wireless networks: Quality-of-service model and learning-based
resource management,” arXiv preprint arXiv:1703.04209, Mar 2017.
[4] O. Semiari, W. Saad, S. Valentin, M. Bennis, and H. V. Poor, “Context-aware small cell networks: How social metrics
improve wireless resource allocation,” IEEE Transactions on Wireless Communications, vol. 14, no. 11, pp. 5927–5940,
Nov 2015.
[5] H. Intraub, “Rapid conceptual identification of sequentially presented pictures,” Journal of Experimental Psychology:
Human Perception and Performance, vol. 7, no. 3, pp. 604–610, 1981.
[6] K. ur Rehman Laghari, R. Gupta, S. Arndt, J. N. Antons, R. Schleicher, S. Möller, and T. H. Falk, “Neurophysiological
experimental facility for quality of experience (QoE) assessment,” in Proc. of IFIP/IEEE International Symposium on
Integrated Network Management (IM 2013), Ghent, Belgium, May 2013, pp. 1300–1305.
[7] I. Wechsung and K. De Moor, “Quality of experience versus user experience,” in Quality of Experience: Advanced Concepts,
Applications and Methods, S. Möller and A. Raake, Eds. Springer International Publishing, 2014, pp. 35–54.
[8] T. Zhao, Q. Liu, and C. W. Chen, “Qoe in video transmission: A user experience-driven strategy,” IEEE Communications
Surveys & Tutorials, vol. 19, no. 1, pp. 285–302, First quarter 2017.
[9] Y. Chen, K. Wu, and Q. Zhang, “From QoS to QoE: A tutorial on video quality assessment,” IEEE Communications
Surveys & Tutorials, vol. 17, no. 2, pp. 1126–1165, Second quarter 2015.
[10] M. Alam, D. Yang, K. Huq, F. Saghezchi, S. Mumtaz, and J. Rodriguez, “Towards 5G: Context aware resource allocation
for energy saving,” Jour. of Signal Proc. Systems, vol. 83, no. 2, pp. 279–291, May 2016.
[11] Y. Lin, R. Zhang, C. Li, L. Yang, and L. Hanzo, “Graph-based joint user-centric overlapped clustering and resource
allocation in ultra dense networks,” to appear in IEEE Transactions on Vehicular Technology, 2017.
 30

[12] J. Zhao, Y. Liu, K. K. Chai, M. Elkashlan, and Y. Chen, “Matching with peer effects for context-aware resource allocation
in D2D communications,” IEEE Communications Letters, vol. 21, no. 4, pp. 837–840, April 2017.
[13] M. Zalghout, S. Abdul-Nabi, A. Khalil, M. Helard, and M. Crussiere, “Optimizing context-aware resource and network
assignment in heterogeneous wireless networks,” in Proc. of IEEE Wireless Communications and Networking Conference
(WCNC), San Francisco, CA, USA, March 2017, pp. 1–6.
[14] E. Bastug, M. Bennis, and M. Debbah, “Living on the edge: The role of proactive caching in 5G wireless networks,” IEEE
Communications Magazine, vol. 52, no. 8, pp. 82–89, Aug 2014.
[15] P. Makris, D. N. Skoutas, and C. Skianis, “A survey on context-aware mobile and wireless networking: On networking
and computing environments’ integration,” IEEE Communications Surveys & Tutorials, vol. 15, no. 1, pp. 362–386, First
2013.
[16] C. Perera, A. Zaslavsky, P. Christen, and D. Georgakopoulos, “Context aware computing for the Internet of Things: A
survey,” IEEE Communications Surveys & Tutorials, vol. 16, no. 1, pp. 414–454, First quarter 2014.
[17] M. Proebster, M. Kaschub, and S. Valentin, “Context-aware resource allocation to improve the quality of service of
heterogeneous traffic,” in Proc. of IEEE International Conference on Communications, Kyoto, Japan, Jul. 2011.
[18] L. Huang, “System intelligence: Model, bounds and algorithms,” in Proc. of the 17th ACM International Symposium on
Mobile Ad Hoc Networking and Computing. New York, NY, USA: ACM, 2016, pp. 171–180.
[19] C. Bishop, Pattern Recognition and Machine Learning. New York, NY, USA: Springer Information Science and Statistics,
2007.
[20] M. Chen, U. Challita, W. Saad, C. Yin, and M. Debbah, “Machine learning for wireless networks with artificial intelligence:
A tutorial on neural networks,” arXiv preprint arXiv:1710.02913, Oct. 2017.
[21] S. Petkoski, A. Spiegler, T. Proix, and V. Jirsa, “Effects of multimodal distribution of delays in brain network dynamics,”
BMC Neuroscience, vol. 16, no. Suppl 1, p. P109, 2015.
[22] Y. Yang, L. T. Park, N. B. Mandayam, I. Seskar, A. L. Glass, and N. Sinha, “Prospect pricing in cognitive radio networks,”
IEEE Transactions on Cognitive Communications and Networking, vol. 1, no. 1, pp. 56–70, Oct. 2015.
[23] T. K. Moon, “The expectation-maximization algorithm,” IEEE Signal Processing Magazine, vol. 13, no. 6, pp. 47–60, Nov
1996.
[24] G. Mclachlan and T. Krishnan, “The EM algorithm and extensions,” vol. 382, 03 1998.
[25] M. J. Neely, “Stochastic network optimization with application to communication and queueing systems,” Synthesis Lectures
on Communication Networks, vol. 3, no. 1, pp. 1–211, 2010.
[26] M. J. Neely, E. Modiano, and C. P. Li, “Fairness and optimal stochastic control for heterogeneous networks,” IEEE/ACM
Transactions on Networking, vol. 16, no. 2, pp. 396–409, April 2008.
[27] K. Seong, M. Mohseni, and J. M. Cioffi, “Optimal resource allocation for ofdma downlink systems,” in Proc. of IEEE
International Symposium on Information Theory, Seattle, WA, USA, July 2006, pp. 1394–1398.
[28] W. Yu and R. Lui, “Dual methods for nonconvex spectrum optimization of multicarrier systems,” IEEE Transactions on
Communications, vol. 54, no. 7, pp. 1310–1322, July 2006.
[29] S. Boyd and A. Mutapcic, “Subgradient methods,” Lecture notes of EE364b, Stanford University, Winter Quarter, vol.
2007, 2006.
[30] J. Friedman, T. Hastie, and R. Tibshirani, The elements of statistical learning. Springer series in statistics New York,
2001, vol. 1.
[31] J. Berger, “A robust generalized bayes estimator and confidence region for a multivariate normal mean,” The Annals of
Statistics, pp. 716–761, 1980.
[32] A. Papoulis and S. U. Pillai, Probability, random variables, and stochastic processes. Tata McGraw-Hill Education, 2002.