Computers & Industrial Engineering 94 (2016) 269279

Contents lists available at ScienceDirect

Computers & Industrial Engineering

journal homepage: www.elsevier.com/locate/caie

Asymptotic optimality of the queue service probability for

the radial basis function network-based queue selection rule
Mu-Song Chen , Hao-Wei Yen
Department of Electrical Engineering, Da-Yeh University, Changhua 51591, Taiwan, ROC

a r t i c l e i n f o a b s t r a c t

Article history: A queueing model is generally designed with sufficient capacity or resources to ensure that the system is
Received 25 January 2015 stable, while preserving quality of service. However, the multi-queue system with finite capacity and
Received in revised form 18 September timing constraints in an overload condition are more often encountered and discussed in a variety of
2015
real-life problems. In such a situation, waiting time is usually an important performance metric quanti-
Accepted 4 February 2016
Available online 22 February 2016
fying the effectiveness and efficiency of the system. The concerned issue is still an open research topic
and is not fully addressed and investigated. Since an exact analysis is practically infeasible owing to
the complexity of such systems, emphasis has been concentrated on the approximate analysis. This paper
Keywords:
Radial basis function network
is thus intended to estimate the upper bound of waiting times of a multi-queue system with a specialized
Queue service probability scheduling paradigm, extending from a series of our research on message scheduling. Without resorting
Waiting time to complex statistical approaches, the study provides a machine learning methodology to resolve this
Queue selection rule subject. With the learning capability of the radial basis function network (RBFN) as the queue selection
Overload rule, this paper particularly focuses on deriving the asymptotic optimality of the queue service probabil-
ity, under the conditions of multi-queue, finite capacity, and timing constraints in the overload situation.
In fact, the RBFN is incorporated with two novel types of learning which lead to develop the support the-
orem and to obtain the closed-form of queue service probability as well as waiting time. Importantly, the
learning feature is definitely essential in providing optimal queue service probability with dynamical
scheduling scheme. Several existing queue selection rules are also evaluated and compared with the
RBFN-based queue selection rule. Simulation results illustrate the feasibility and accuracy of the pro-
posed strategy.
2016 Elsevier Ltd. All rights reserved.

1. Introduction
Queueing systems are popular and widely used tools for analyz-
ing the performance of communication networks, transportation
system (Sahba & Balcioglu, 2011), job-shop manufacturing system
(Georgiadis & Michaloudis, 2012; Shen & Buscher, 2012), and
machine-repairman system (Yuan, 2012). Usually, many queueing
systems are generally designed with sufficient capacity or
resources to ensure that the system is stable, while preserving
quality of service (QoS). However, variability and constraints, e.g.
heavy traffic, timing constraints, finite queue capacity, etc., in a
queueing system may have significant impacts on its performance.
The study of heavy traffic began in the 1960s by Kingman (1962).
Kingmans bound implies that even small variations in service
times or inter-arrival times can lead to significant delay. Apart
from this, it is inevitable that a system can also encounter a peri-
odic or temporary overload,1 either due to an unpredictable
increase in load, unavailable service resources, or system break-
down. Another practical situation occurs on message scheduling in
the controller area network (Mubeen, Mki-Turja, & Sjdin, 2014).
In such a context, the maximum normal utilization of the controller
area network is about 30%. Under heavy disturbance, it rises to about
40% due to retransmission of corrupted messages (Johansson,
Trngren, & Nielsen, 2005). The direct consequence can possibly
bring about a fatal catastrophe or even a matter of life and death.
The aforementioned problems become worse when the queueing
system is conditioned by finite capacity and/or imposed on stringent
timing constraints.
The finite-capacity queueing systems with timing constraints
have been widely studied by many researchers (Osorio &

Corresponding author. 1
The load is a crucial indicator of whether the queueing system will grow beyond
E-mail addresses: chenms@mail.dyu.edu.tw (M.-S. Chen), chenmsac1510@ its capacity. The occurrence in a heavy load situation can temporarily make the
yahoo.com.tw (H.-W. Yen). system unstable.

http://dx.doi.org/10.1016/j.cie.2016.02.005
0360-8352/ 2016 Elsevier Ltd. All rights reserved.
270 M.-S. Chen, H.-W. Yen / Computers & Industrial Engineering 94 (2016) 269279
Nomenclature
si queue cycle of queue Q i
ci queue service probability of queue Q i
W i; W i waiting time and mean waiting time of queue Q i
ki , li arrival rate and service rate of queue Q i
Bierlaire, 2009; Sakuma & Inoie, 2014; Wu, Lin, & Chie, 2012) and
our recent publications (Chen, 2015; Chen & Yen, 2011, 2012a,
2012b, 2013, 2014). A queueing system with these restrictions
occurs frequently in our real life. For instance, there may be a surge
in demand for the limited hospital emergency rooms during a
catastrophic event (Lakshmi & Sivakumar, 2013). In manufacturing
systems, there are limited waiting rooms between workstations in
assembly lines. In semiconductor wafer fabrication manufacturing,
queue-time constraints occur when a set of consecutive process
steps must be completed within a fixed time window (Tsai,
2008). Violations of these constraints can result in rework, scrap,
and longer cycle times. To ensure product quality, the queue time
between consecutive operations is often required to be shorter
than a pre-specified duration. The effect is similar to imposing a
limited buffer size between operations. On highways, there are
rush hours during which drivers have to queue for the use of high-
ways. Regardless of how many lanes already exist, it has finite road
capacities and is vulnerable to possible congestion during weekend
or holidays. In all these circumstances, the system can have exces-
sive demands than it can handle and possibly leads to unstable
levels or causes unpredictable performance degradation.
Regarding aforementioned problems, three primary considera-
tions are (1) how the system can be stabilized, (2) how the system
throughput is maximized, and (3) how the upper bounds of wait-
ing times can be estimated. Pertaining to issues (1) and (2), several
related literatures aim to avoid the degradation of throughput of
the server as well as to maintain the system at its steady state.
Schemes such as admission control (Cherkasova & Phaal, 2002)
or specialized scheduling policies (Chen & Mohapatra, 2002), or a
combination of both (Elnikety, Nahum, Tracey, & Zwaenepoel,
2004) have been studied. Furthermore, Solar (Stolyar, 2004) ana-
lyzed the parallel server queueing systems, under the assumption
that the system is in heavy-traffic. That is, the system is stable
but operates close to the boundary of the stability region.
Shakkottai, Srikant, and Stolyar (2004) and Bell and Williams
(2005) also used the complete resource pooling condition to estab-
lish heavy-traffic optimality of other resource allocation models.
Other researchers (Andradottir & Ayhan, 2005; Andradottir,
Ayhan, & Down, 2001, 2003; Tassiulas & Bhattacharya, 2000) con-
cerned with the dynamic assignment of servers to maximize the
system throughput in queueing networks. In addition,
Venkataraman and Lin (2007) proposed scheduling algorithms that
can stabilize the network at given offered loads, which also ensures
that the long term average and service rate are no less than the
arrival rate of each user.
Regarding issue (3), let us consider another example of the hos-
pital queueing systems. Derlet, Richards, and Kravitz (2001) and
Carter and Lapierre (1991) pointed out that the causes of hospital
overcrowding include low staff availability, bed shortage, insuffi-
cient hospital space, and so forth. Actually, long waiting time is
symptomatic of inefficiency and indicative of the systems inability
to satisfy patients demand. To ensure QoS, several criteria can be
established for a satisfactory level of service. One possible gauge
might be that the upper limit of waiting time should not exceed
a certain value. Therefore, patients satisfactions are often related
to the amount of time they have to wait before receiving treat-
ments. For the mentioned criterion, the upper bound of waiting
qi message load of queue Q i
Di deadline of messages in queue Q i
Ki maximum number of messages in queue Q i
time needs to be known first for this purpose. This problem is prin-
cipally related to issue (3) and is vital when the queueing system is
imposed by timing constraints and with restrictions of finite capac-
ity. In fact, a search of available literatures indicates that relatively
few studies have been devoted to fully address this problem. Fur-
thermore, due to the growth of traffic with the increasing amount
of data exchanged between electronic control units in the con-
troller area network, our motivation of this study aims at estimat-
ing the message waiting times in an overloaded circumstance.
Based on our research findings, this paper presents a machine
learning approach with the radial basis function network (RBFN)
(Chen & Yen, 2011) in dealing with the aforementioned issue.
The RBFNs have demonstrated their usefulness in a variety of
applications, including classification, prediction, and system mod-
eling. In fact, the RBFN is a variant of the multilayer perceptron
networks (Haykin, 1994) with single hidden layer and makes use
of locally supported functions to calculate the Euclidian distance
between the input vector and the centers of RBFs. The latest stud-
ies and applications on the RBFNs can be refereed to Dehghan and
his colleagues (Dehghan & Mohammadi, 2014, 2015; Dehghana,
Abbaszadeha, & Mohebbib, 2014, 2015; Ilati & Dehghan, 2015).
With the radial basis function network as the queue selection rule
(QSR), messages in queues are selected for admission into the ser-
vice facility in accordance with the decision made by the RBFN. To
the best of our knowledge, this study is the first attempt to com-
bine machine learning strategy with dynamical scheduling to esti-
mate waiting times under the overload situation. Especially, the
proposed RBFN-based QSR directly leads to derive the asymptotic
optimality of the queue service probability (QSP) as well as waiting
time in the time domain directly.
The remaining part of the paper is organized as follows. Sec-
tion 2 introduces the definition of queue cycle, the effective arrival
rate, and the mean queue length for subsequent analysis. In Sec-
tion 3, relationship between waiting time and mean waiting time
is established. Section 4 is devoted to derive the queue service
probability of the RBFN-based queue selection rule with support
theorem. Section 5 reports and validates the results of simulation
experiments. Finally, Section 6 concludes the paper and suggests
our future research directions.
2. The queue cycle
Typically, the queueing system consists of a number of parallel
queues attended by a single server. Thus, the control policy of the
QSR focuses on determining the queue and events in that queue for
admission into the service facility. Once a queue is selected for ser-
vice, events within that queue are scheduled using first-come-first-
served order. At this point the word event is a generic expression
that represents many real-world phenomena, such as airplanes
arriving to an airport, shoppers in a grocery store, or messages in
a queue waiting to be executed. In this study, we focus on the issue
of message scheduling. We also consider a multi-class2 model, con-
sisting of q finite-capacity queues, labeled as Q 1 ; Q 2 ,. . .,Q q . Usually,
2
Multi-class queuing model can be used to model complex service systems which
have different QoS requirements.
 M.-S. Chen, H.-W. Yen / Computers & Industrial Engineering 94 (2016) 269279 271

RBFN-based
Queue Selection Rule

Queue Decision
status making

Q1 buffer size K1 with deadline D1

1 1

Poisson Arrivals

Q2 buffer size K2 with deadline D2

2 2
service
mechanism

    
Qq buffer size Kq with deadline Dq
q q

Message Queue Pool

Fig. 1. A multi-queue system with the RBFN-based queue selection rule.

messages arrive at Q i ; i 1; 2; . . . ; q, according to a Poisson arrival with the allocation of limited resources to messages over time. It
process with rate ki . The service time is also a random variable with is a decision-making process with the goal of optimizing one or
mean l1i . The load to Q i is qi ki =li and the total offered load is cal- more objectives with competition. Although the QSR may have dif-
P
culated as q 8i qi . To comply with the real queueing model, each ferent options, it always depends on several criterions. Neverthe-
queue is also limited with finite buffer size K i ; 8i. In general, the less, the QSR always affects the queue cycle as well as waiting
aforementioned queueing model with q queues is also referred as time significantly. In this sense, it is reasonably to notice this
the q  M/G/1/K i model, using Kendall notation (Yen, 2011). A typi- dependence in the queue cycle.
cal multi-queue system with the RBFN-based queue selection rule is According to Fig. 2, we can calculate the effective arrival rate
depicted in Fig. 1. and the mean queue length, respectively. In Fig. 2, there are
Assuming that each queue is dedicated to a particular class of K i  1 messages in the interarrival interval, where K i is the maxi-
messages with the same priority, it is thus assigned with a class- mum capacity of Q i . Until the end of service for the head message
specific deadline Di ; 8i. The deadline of a message is the point in in Q i , one vacancy is available for the new arriving message to join
time before which the message must complete its service or trans- the queue. After that, Q i reaches its maximum limit K i with zero
action. In our convention, it is supposed that the importance of any arrival rate and the queue cycle repeats again. In this sense, the
queue is measured by deadlines with Q i favored over Q j when j > i. effective arrival rate ki;eff of Q i is
In this sense, Q 1 has a stringent or the smallest deadline than the k1 si  k1 1
others. In real-time system, deadlines of messages are very critical. ki;eff i
ki i
0 2
si si si
The order how the queues are selected for service is a significant
factor for providing QoS and better behaviors to time critical appli- Likewise, the mean queue length Li is also written as
cations. If timing constraints cannot be guaranteed, messages
k1 si  k1 k1 l1
could miss their deadlines, causing unpredictable performance Li i
K i  1 i
K i  ki;eff l1 Ki  i i
3
degradations. Definitely, the queue selection rule is devoted to si si i
si
support proper paradigm for these concerns. When the queue load or workload is large, all messages are almost
To proceed with our goal, the queue cycle is first described for always queued and si becomes large. Thus, the expected queue
subsequent study. In the following discussion, it is assumed that length in Eq. (3) is approximately equal to the maximum capacity
the model is in the overload condition.3 The queue cycle si is of the corresponding queue, i.e. Li  K i , for a large si . This is partic-
defined as ularly consistent with intuition and experience. Using Little formula
(Little, 1961), Li ki;eff W i , the results of Eqs. (2) and (3) help us to
si k1
i xi li
1
1 determine the waiting times. In formula
The queue cycle consists of (1) an interarrival interval, (2) a waiting W i K i si  k1
i  li
1
4
interval, and (3) a service interval. The timing diagram of the queue
cycle is also demonstrated in Fig. 2 for reference. With known arri- Certainly, Eq. (4) makes sense since any message in Q i has to expe-
val rate ki and service rate li , the queue cycle mainly depends on rience K i queue cycles, i.e. K i si , to receive service. If we consider
the waiting time in queue. It should be noted that the QSR deals waiting time in queue (thus excluding k1
i and l1
i ), then W i is
determined as K i si   l which is completely in agreement
k1
i
1
i
3
If q > 1, the system is overloaded since the arrival rate is higher than the service with Eq. (4). From the aforementioned discussions, W i can be
rate. In these critical situations, the service requests exceed the available resources. decided explicitly if si is known in advance. However, it is not easy
272 M.-S. Chen, H.-W. Yen / Computers & Industrial Engineering 94 (2016) 269279

interarrival waiting service

interval interval interval

i1 i1 i i1

prob.=1- i prob.= i

arrival rate=i arrival rate=0

# of messages=Ki-1 # of messages=Ki
queue cycle, i

Fig. 2. Timing diagram of the queue cycle si in an overload situation. In fact, k1
i and l1
i are considerably smaller than the waiting interval. Thus, the waiting interval plays a
significant role in estimating the queue cycle.

Message at the rightmost is

ready for service from the sever
i-1

Wi-(Ki-1)i Wi-(Ki-2)i Wi-i Wi

i-1
(a) All messages in queue shift one
(e)
i-1 position to the right and a tagged
All messages in queue increase message is ready to join Qi
their waiting times by i-1 and
stay on their positions
TM is ready empty Wi-(Ki-1)i+i-1 Wi-2i+i-1 Wi-i+i-1
TM
Wi-Kii+3i-1 Wi-(Ki-1)i+3i-1 Wi-2i+3i-1 Wi-i+3i-1 All messages in queue increase (b)
their waiting times by i-1 and
stay on their positions i-1
(d)
TM
i-1
Wi-Kii+2i-1 Wi-(Ki-1)i+2i-1 Wi-2i+2i-1 Wi-i+2i-1

(c)

Fig. 3. The movement of messages in queue. In (b), a tagged message (TM) joins the tail of Q i .

to obtain si directly. Rationally, the queue cycle strongly depends on
the QSR. It is obvious that si becomes smaller when Q i is selected
for service more often than the others. In other words, si is inversely
dependent on the queue service probability (QSP) ci . Without
resorting to solve queue cycle straightforwardly, there are other
alternatives to express si in terms of ci and available quantities.
Since each queue cycle also contains a service time l1 i , the queue
cycle can be reasonably expressed as
1
si
li ci
Eq. (5) is extremely important in developing the upper bounds
of waiting times. Consequently, the results of the queue cycle
turn out to be the derivation of the queue service probability.
In Yen (2011), QSPs for several other queue selection rules, e.g.
Earliest deadline first (EDF) (Kargahi & Movaghar, 2006), First-
Come First-Served (FCFS) (Kramer & Assad, 2012), and Priority
Queueing (PQ) (Ndreca & Scoppola, 2008), are investigated. In
the following, we focus on using the radial basis function network
as our queue selection rule (QSR). The RBFN is essential to pro-
vide a machine learning strategy to optimize certain objectives.
Moreover, the RBFN-based QSR presents two novel types of learn-
ing to ensure a low and stable message waiting time variance (or
a uniform distribution of waiting time) and lower untimely ser-
vice ratios (Chen & Yen, 2011). This consequence can directly lead
to develop the support theorem as well as a closed form of QSP.
In the next section, we discuss the measurements of waiting time
W i in the overload condition and establish the relationship
between W i and the mean waiting time W i for succeeding
discussions.
3. Waiting time W i and mean waiting time W i
In the following analysis, we have assumed that the server is
immediately available when the service is completed. Further-
5 more, preemption is not allowed. In other words, once a service
operation has started, it cannot be interrupted by any other
requests. As long as the queue is not full, the arriving message joins
the tail of the dedicated queue immediately and the waiting time
of this message starts to be counted. Then, any message can be
ready for service until it reaches the head of the queue. Let us
observe the movement of a message in Q i . It is worth mentioning
that the movement of any message depends on its own position
in queue and the number of messages before it for service. Fig. 3
illustrates this process in three steps.
Step 1. In Fig. 3(a), Q i is filled with messages. Let W i denote the
waiting time of the head message of Q i . The waiting time
of the messages immediately behind the head message is
W i  si . Accordingly, the waiting times of all queued mes-
sages in Q i can be arranged in an increasing order from left
to right as
 M.-S. Chen, H.-W. Yen / Computers & Industrial Engineering 94 (2016) 269279 273

fW i  K i  1si ; . . . ; W i  2si ; W i  si ; W i g 6 nonlinearly transformed input vector x with the following

expression
Step 2. If Q i is selected for service, the head message gains the ser-
vice. After the completion of service, all messages move X
yj wi /i x 13
forward one position in queue and wait to accommodate
i
a newly arriving message (Fig. 3(b)). In this stage, there
are only (K i -1) messages left in queue. Since the service or
time is l1 i , the waiting times of all queued messages in
P
i wi /i x
Q i also form an increasing ordered list as
yj P 14
  i /i x
W i  K i  1si l1
i ; . . . ; . . . ; W i  2si l i ; W i  si l i
1 1

7 where yj is the output of the jth output node and wi are the output
Thus, the sum of waiting times in Eq. (7) is written as T1 and linear combining weights. The /i (x) is the ith radial basis function
  (RBF) in the hidden layer. The output of the network is a simple lin-
K i si ear combination of the radial basis functions (or hidden neurons) of
T 1 K i  1 W i l1
i  8
2 Eq. (13) or the weighted average of the output associated with each
receptive field of Eq. (14). Basically, the radial basis function is a
When a message joins the tail of Q i , all other messages real-valued function whose output depends on the distance of the
before this new tagged message remain in their positions. input vector from a given center vector. Although many kinds of
However, all messages waiting times increase l1 i (as RBFs can be utilized, a multivariate Gaussian function is adopted
examples shown in Fig. 3(c) and (d)). In this case, the total in this study as
waiting times are calculated as !
kx  ci k2
X
Ki
   1 /i x exp  15
2r2i
T2 i  m li  j  1si
W i  c1
which is characterized by a center vector ci and a width ri . These
j1
 
  1 K i  1
i  m li  si
K i W i  c1 9 values along with the connection weight wi in Eq. (13) can affect
2 the system performance significantly. Even though RBFNs belong
to forward network models because of their structure, the method
In Eq. (9), m = 2 (or m = 3) corresponds to Fig. 3(c) (or Fig. 3
of initializing these parameters is different from the backpropaga-
(d)), etc.
tion model, in which parameters are initialized randomly. Usually,
Step 3. When Q i is selected again for service, the status return to
the corresponding parameters of the RBFN can be determined
Fig. 3(a) and the aforementioned process is repeated again.
according to the distribution of sample data. Moreover, the RBFN
with too few RBFs may be inadequate to achieve the desired perfor-
Until now, we can calculate the mean waiting time W i of Q i . As mance (i.e. high bias and low variance). Too many RBFs, on the
a consequence other hand, can also induce model overfitting (i.e. low bias but high
Pc1 variance) and a long processing time. Depending on the model to be
T 1 m2 i
T2 optimized, many parameter learning as well as structure learning
Wi 10
K i c1
i  1 strategies can be applied simultaneously as long as the model can
smoothly and quickly converge. Related to this issue, the corre-
By substituting T1 (Eq. (8)) and T2 (Eq. (9)) into Eq. (10) and applying
sponding parameter/structure learning schemes for the queueing
simple calculations, the final formulation of W i is problems can be referred to our recent publications (Chen, 2015;
W i l1  k1 Chen & Yen, 2011).
Wi i i
11 In fact, the RBFN provides a promising and novel learning
2 2
scheme in dealing with resource-constrained scheduling problems.
The last equation clearly gives a useful insight to understand the Particularly, the learning algorithms include two hybrid learning
relationship between W i and W i (Yen, 2011). If we neglect l1 i
for handling different queue load variations. In Chen and Yen
and k1 (2011), we have concluded that the RBFN-based QSR can maintain
i ; W i is approximately equal to half of W i , i.e. W i 0:5W i .
Indeed, this fact can also be verified from Fig. 3 by averaging the a fair bandwidth sharing between queues as well as prevent possi-
maximum and the minimum of waiting time, namely W i and ble untimely service ratios. These distinctive learning features also
W i  K i  1si l1 lead us to derive a closed form of the queue service probability
i . In other words,
with known quantities. In the next section, the RBFN-based QSR
1  1 is illustrated.
Wi W i W i  K i  1si l1
i  Wi 12
2 2
Noticeably, Eq. (12) is nearly the same as what we had in Eq. (11). 4. Queue service probability of the RBFN-based QSR
Starting from the previous considerations in Section 2, we realize
that an essential step for estimating the waiting time in the over- Although the relationship between W i and W i is established in
load situation through a variety of queue selection mechanisms is the last section, either of them still cannot be obtained analytically.
to determine the queue service probability in advance. Therefore, The difficulty is due to different control strategies of QSRs we con-
the preceding results are applied together with the specific RBFN- cerned with. In this paper, we focus on using the RBFN as our QSR.
based QSR to derive the closed form of queue service probability. Theorem 1 is present to illustrate how the corresponding QSP is
The RBFN belongs one of the most important models of artificial derived. Other possible QSRs and their QSPs can be referred to
neural networks that are structured based on the theory of func- Yen (2011).
tional approximation and pattern classification. It has a three layer
structure (Fig. 1) like a perceptron network. The RBFN accom- Theorem 1. The queue service probability of using the RBFN-based
plishes an inputoutput nonlinear mapping by a combination of QSR is
274 M.-S. Chen, H.-W. Yen / Computers & Industrial Engineering 94 (2016) 269279

Ki ai
ci 16 ci 22
2li cDi k1 bi c
i
P
i li =2 and bi Di ki
where ai K i D1 and the sum of ci , i.e. c
1 1 1
where c is any arbitrary positive constant and i i,
for all queues is unity, i.e.
Wi
c 17 Xq
ai
Di 1 23
b
i1 i
c
Proof. There are two types of learning schemes, type I and type II Qq
We then multiple i1 bi c on both sides of Eq. (23) to reform the
learning, for the RBFN-based QSR (Chen & Yen, 2011). Type I learn- expression as
ing is a prior learning to prevent possible causes of non-uniform
bandwidth allocation among requesting queues. On the other Y Y Y Y
q
a1 bi c a2 bi c    aq bi c bi c
hand, type II learning is intended to reduce possibilities of unti-
8i; i1 8i; i2 8i; iq i1
mely service ratios. According to our investigation, the occurrences
of type I learning are more often than type II learning. Thus, we can
24
nearly neglect the influence of type II learning. Typically, type I Through simple computations, a polynomial of degree q with
learning is activated whenever the mean waiting time variance unknowns c is formulated as
(MWTV) is greater than a predefined threshold. The MWTV can ! !
also be measured from the between-queue mean waiting time in X
q X
q Y
q X
q Y
cq bi  ai cq1    bi  bj c 0 0 25
the following sequence i1 i1 i1 i1 8ji
( )
W1 W2 Wq Since all quantities in parenthesis are known in advance, the c val-
; ;...; 18 ues can be found by solving the roots of Eq. (25). h
D1 D2 Dq
The results of Corollary 1 are indeed important. With Eq. (25),
In general, the smaller the MWTV is, the more uniform the bandwidth we can solve c as well as ci and other interesting queue parameters,
is used. In an ideal case, we can expect the MWTV to be zero. From Eq.
such as si (Eq. (5)), W i (Eqs. (4) or (20)), and W i (Eq. (11)). Accord-
(18), the ratio of W i to Di should be nearly a constant c to retain a ing to the general forms of waiting time and queue cycle in Eqs. (4)
lower MWTV. In other words, W i is approximately equal to cDi , e.g. and (5), QSPs for different kinds of QSRs, including RBFN, EDF
W i  cDi ; 8i. In accordance with Eqs. (11) and (17), we conclude (Kargahi & Movaghar, 2006), FCFS (Kramer & Assad, 2012), and
! PQ (Ndreca & Scoppola, 2008), in the overload condition are sum-
Wi 1 W i l1  k1 marized in the following formulations (Eq. (26)) for comparisons.
i i
c 19
Di Di 2 2
Ki
cRBFN
2li cDi k1
i
i
Thus,
Ki
i  li
W i 2cDi k1 1
20 cEDF
i
li W EDF
1 i l1 DDi;1
k1 1
26
To comply with the Little formula Li ki;eff W i and Eqs. (2), (3), and
(5), we can arrive at the following formulation
cFCFS
i lim cEDF
i
DDi;1 0

Ki  k1
i l 1
c li ci li 2cDi
i i k1
i l 1
i 21 cPQ
i lim cEDF
i
DDi;1 !1

Finally, the closed form of Eq. (16) is obtained. This completes our
In Eq (26), DDi;1 is the difference of deadlines of Q i and Q 1 , i.e.
proof. h
DDi;1 Di D1 . Moreover, detail derivations of cEDFi ; cFCFS
i , and cPQ
i
can be referred to our report in Yen (2011). In order to discriminate
Remark 1. The QSP in Eq. (16) is equivalent to the activation fre- different kinds of QSRs, superscripts of the used queue selection
quency in the output layer of the RBFN. If the total simulation time rule are added whenever they are necessary in the subsequent dis-
is T and the output neuron i is activated ni times, then ci  ni l1
i =T. cussions. For instance, notations of QSP, queue cycle, and waiting
The obtained QSP in Eq. (16) illustrates a general dependency times are expressed as cRBFN
i ; sRBFN
i , and W RBFN
i for the RBFN-based
with respect to K i and Di . Since we consider an overload situation, QSR.
Q i should be selected more frequently to avoid possible untimely
services when K i becomes larger. It simply implies that ci is pro-
5. Simulation
portional to K i . Likewise, the lower the value of the deadline Di ,
the higher the probability of Q i being served. Significantly, the
In this section, two experiments are conducted to verify the cor-
inclusion of c in the QSP involves the learning features of the RBFN
rectness of the proposed method, especially when the system is
to preserve QoS and deadline constraints. Until now, there is still
one more unknown c to be resolved in Eq. (16). In the next Corol-
lary, we discuss the method to solve the values of c from a q-th
degree polynomial, where q is the number of queues. The coeffi- Table 1
cients of the given polynomial are expressed as the functions of Parameter settings.

several knowns, such as K i ;Di , ki , and li . # of queues (q) 3

Utilization (q) 0.55
Corollary 1. The values of c in Eq. (16) can be solved from a qth q1 :q2 :q3 0.3:0.2:0.5
Service time (l1
i )
[1 2 3]
degree polynomial.
Queue capacity (K i ) [25 20 45]
Deadline (Di ) [100 200 300]
Simulation times 5000
# of repeated experiments 20
Proof. Let us rewrite Eq. (16) as
 M.-S. Chen, H.-W. Yen / Computers & Industrial Engineering 94 (2016) 269279 275

Queue Service Probability Waiting Time

100
0.3 88.5
0.277 80
RBFN
W1 =88.5
1 0.25 RBFN
1 =0.277
W
1
60
RBFN(sim.)
W1
0.2
RBFN(sim.)
1 40

20
(a) (b)
0.15
0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5

200
0.221 177.6

0.2 150
RBFN
W2
2 2
RBFN
=0.221 W
2
=177.6
100 RBFN(sim.)
W2
RBFN(sim.)
0.15 2
50
(c) (d)
0.1
0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5

0.55 300
0.502 265.0
0.5 250
0.45 200 RBFN
3 3
RBFN
=0.502 W
3
W3 =265.0
0.4 150 RBFN(sim.)
0.35
RBFN(sim.)
3 W3
100
0.3
(e) 50 (f)
0.25
0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5

Fig. 4. Queue service probabilities ((a), (c), (e)) and waiting times ((b), (d), (f)) of the RBFN-based queue selection rule. (The blue curve is the theoretical value and the red one
is the simulation result.) (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

overloaded. The first experiment fixed the queue capacity (K i ), becomes sufficiently large, the resource competition becomes
while the second one is simulated with a constant load intensity more evident and the system turns out to be overloaded quickly.
(q). The former one examines the limiting behaviors of the QSP In general, ci
RBFNsim:
approaches its theoretical bound cRBFN (hori-
i
as well as waiting times, while the latter case further validates
zontal line) asymptotically, i.e. limq!1 ci cRBFN
RBFNsim:
i . Fig. 4(a),
the proposed paradigm. Although our simulations focus on three-
(c), (e) confirms our claims. Another important finding is that
queue problem, the closed form expressions of QSP and waiting
times can be generalized for any number of queues. Since several
cRBFN
1 of Q 1 is not necessarily larger than the others, even Q 1 has
the stringent deadline.4 Although it is difficult to analyze the mutual
tunable parameters of the RBFN are initialized with random seeds,
effects between queues, several hints can still be readily drawn from
each experiment is thus repeated 20 times and the final results are
taken from their average values. Eq. (16). Certainly, the derived cRBFN
i is intended to balance the dual
constraints, namely QoS and deadline constraints, with its best
efforts. Indeed, Eq. (16) can be approximated as
5.1. First experiment cRBFN
i K i =cli Di li k1i  K i =cli Di . Apparently, there are three
crucial factors, i.e. c; K i , and Di , dominating cRBFNi . The inclusion of c
In this experiment, the limiting behaviors of QSP and waiting originates from the proposed learning features of the RBFN. The
times are investigated when the values of q changes from normal machine learning paradigm does provide a fair sharing of bandwidth
load to overload situations, i.e. from 0.5 to 5. Parameters of various as well as lower untimely service ratios. As a result, it immediately
simulation scenarios are also tabulated in Table 1. Simulation leads to an optimal cRBFN in this sense. This characteristic partially
i
results of ci and ci ; ci , and ci
RBFNsim: EDFsim: FCFSsim: PQ sim:
are also ana- explains the observed truth. Since we consider an overload model,
lyzed and compared with their theoretical bounds cRBFN
i and each queue is always fully occupied. Thus, K i appearing in the
cEDF
i ; cFCFS
i , and cPQ
i , respectively.
numerator of cRBFN i can truly reflect this fact. In order to preserve
To obtain cRBFN i for the RBFN-based QSR, the c values are first QoS, cRBFN
i become large for a large K i . On the other hand, Di in the
calculated based on Corollary 1 and available quantities in Table 1. denominator also reveals the deadline constraints. In this sense, a
Although the resulting 3-degree polynomial (q = 3) consists of conclusion thus can be drawn that either a large K i or a small Di
three real roots of c, i.e. 0.4449, 0.0075, and 0.0047, the valid can always result in a large cRBFN i .
positive one is 0.4449. With this value, the QSP can be calculated
with the aid of Eq. (16), accordingly. The obtained results are ana-
4
lyzed as follows. In this experiment, Q 1 is assigned with the smallest deadline (100) than the
others (200 and 300), as shown in Table 1. It simply implies that messages in Q 1 are
When q is less than one, c
RBFNsim:
always increase consistently
i expected to have the highest service priority, in order to meet the strict deadlines.
in order to increase service frequency, as shown in Fig. 4. When q However, it is not always true for different queue selection paradigms.
276 M.-S. Chen, H.-W. Yen / Computers & Industrial Engineering 94 (2016) 269279

Queue Service Probability

0.22 0.218 0.5

0.307
0.3 0.475
0.2 EDF 0.45
EDF 2 =0.218
1 =0.307 EDF
3 =0.475
1 0.25 2 0.18 3 0.4
EDF(sim.)
EDF(sim.)
1 0.16 2 EDF(sim.)
0.35 3
0.2 0.14
0.3
0.12
(a) (b) 0.25
(c)
0.15 0.1
1 2 3 4 5 1 2 3 4 5 1 2 3 4 5

0.22
0.3 0.287 0.7 0.675
0.2 0.199
0.197
0.6 FCFS
3 =0.675
0.25 0.18 FCFS
2 =0.199
1 2 3 FCFS(sim.)
0.16
FCFS(sim.)
0.5 0.499 3
0.2 FCFS 2
1 =0.125 0.14
0.4
FCFS(sim.) 0.12
1
0.15 (d) (e) 0.3 (f)
0.125 0.1

1 2 3 4 5 1 2 3 4 5 1 2 3 4 5

0.4 0.5
1 1.0
PQ
1 =1.0 0.3 0.4
0.8
1 PQ(sim.)
1 2 3 0.3
0.6 0.2
PQ PQ
2 =0.0 0.2 3 =0.0
0.4 0.1 PQ(sim.)
2 0.1
3
PQ(sim.)
(g) (h) (i)
0.2
0 0
1 2 3 4 5 1 2 3 4 5 1 2 3 4 5

Fig. 5. Queue service probabilities of the EDF ((a), (b), (c)), FCFS ((d), (e), (f)), and PQ ((g), (h), (i)).

Fig. 5 also illustrates QSPs for EDF, FCFS, and PQ, respectively. (2) In Fig. 5(d), (e) and (f ), the First-Come First-Served (FCFS)
Similarly, these plots behave similarly as those QSPs in Fig. 4(a), (Kramer & Assad, 2012) discipline is applied as the QSR.
(c), and (e). These plots are discussed, respectively. The FCFS is a queuing policy, in which head messages in
queues are served in the order of their arrivals without other
(1) The earliest-deadline first (EDF) discipline (Chen & Yen, biases or preferences. In order to maintain the same waiting
2013) takes into account the deadlines of messages with times for all queued messages (Chen & Yen, 2013), ci
FCFSsim:

dynamical priorities. The EDF strategy has shown that it reaches 0.287, 0.197, and 0.499 when q is approximately
can offer substantial performance with respect to schedula- one. These values are individually marked on the plots (d),
bility and the desired quality of service over generalized (e) and (f ) and are also roughly proportional to the queues
bandwidth sharing in both deterministic and stochastic set- maximum capacity K i .5 The preceding results are in agree-
tings. In this regard, the EDF policy minimizes the maximum ment with our recent publication (Chen & Yen, 2013). Finally,
lateness and the maximum tardiness. Three facts can be
cFCFSsim:
i quickly converges to its theoretical bound cFCFS
i when
readily observed in Fig. 5(a), (b) and (c). Firstly, the asymp-
q is larger than one.
totic behavior of ci approaches cEDF
EDFsim:
i as was expected. (3) In Fig. 5(g), (h), and (i), priority queuing (PQ) (Ndreca &
Secondly, cEDF1 is not necessarily to be always larger than Scoppola, 2008) is applied. PQ is also known as Head-of-
cEDF
2 and c EDF
3 . In fact, c EDF
3 is the largest one in the current Line priority and works with the principle of non-
parameter setting. This phenomenon is the similar to those preemptive strict priority. In PQ, a particular queue is sched-
QSPs, using the RBFN-based QSR. The learning capability of uled only if queues of higher priority are empty. This scheme
the RBFN and dynamic feature of the EDF, however, offer dif- provides a simple method for preferential treatment and
ferent scheduling scenarios but lead to similar conclusions. delay bounds for high priority messages. With strict priority
Detail discussions of this issue can be referred to our recent queuing, it can cause bandwidth starvation of other queues.
study (Chen, 2015). Thirdly, it is observed that the condition In this experiment, Q 1 is always given the highest service
m  D m lm  W i
W EDF  Di l1
i ; 8 m; i, is satisfied in the
1 EDF
order at the expense of scarifying the service of low priority
overload circumstance. This is in consistence with (Yen, queues Q 2 and Q 3 . In the long run, messages in Q 2 and Q 3
2011), which showed that the EDF scheduling reaches a
fair relative deadline expiry for queues at this specific
The ratios of ci
5 FCFSsim:
are 0.287:0.197:0.499, which illustrates a linear depen-
situation. dency on K i , i.e. 25:20:45.
 M.-S. Chen, H.-W. Yen / Computers & Industrial Engineering 94 (2016) 269279 277

Waiting Time
200 300
80 80 180 280
250
150
60 W1
EDF EDF EDF
W =80 W W2 =180 W 200 W3 =280
1 2 3
EDF(sim.) 100 150
40 W1 EDF(sim.)
W2
EDF(sim.)
W3
100
20 50
(a) (b) 50 (c)
1 2 3 4 5 1 2 3 4 5 1 2 3 4 5

200 197.9 200 196.9 200 195.9

150 FCFS 150 150

W W1 =197.9 W
FCFS
W2 =196.9 W
FCFS
W3 =195.9
1 2 3
100 FCFS(sim.)
W1 100 FCFS(sim.) 100
W2 W3
FCFS(sim.)

50 50 50
(d) (e) (f)
1 2 3 4 5 1 2 3 4 5 1 2 3 4 5

25
23.4 350 4000

20 300
3000
250
W PQ W W
1 15 W1 =23.4 2 PQ 3 W3
PQ
200 W2
2000
10 PQ(sim.) 150
W1 PQ(sim.) W3
PQ(sim.)
W2
100 1000
5
(g) 50 (h) (i)
1 2 3 4 5 1 2 3 4 5 1 2 3 4 5

Fig. 6. Waiting times of the EDF ((a), (b), (c)), FCFS ((d), (e), (f)), and PQ ((g), (h), (i)).

deadline. The queue selection rule can recognize this truth and result
Table 2
Theoretical values of queue cycles. in a large queue cycle. In contrast, a large value of K i can also give
rise to a large queue service probability and thus a small queue cycle.
Queue selection rule Queue cycle
However, the joint effect of Di and K i affects each other as well as the
s1 s2 s3 concerned model. The choice of these parameters becomes a tradeoff
RBFN 3.60 9.07 5.98 when considering QoS and timing constraints. In Table 2, the ratios
EDF 3.26 9.18 6.31 of sRBFN
i , i.e. 3.6:9.07:5.98, are nearly consistent with the ratios of
FCFS 7.98 10.03 4.44
Di /K i , i.e. 100/25:200/20:300/45, from the parameter setting in
PQ 1 N.A. N.A.
Table 1. Likewise, the analysis of sEDF i is similar. The results of sFCFS
i
are also related to the ratios of Di =K i . In Yen (2011) and Eq. (26),
we conclude that FCFS scheduling can be obtained from the EDF
can undergo infinite waiting times. In Fig. 5g; Q 1 always scheduling in the condition of identical deadlines Di for all queues.
incurs increasing ci
PQ sim:
and dominates the resources. Even- Accordingly, sFCFS i is only dependent on 1/K i .6 Finally, sPQ
1 is the small-
tually, starvations occur in Q 2 and Q 3 , as shown in (h) and (i). est queue cycle at the expense of completely scarifying the service
quality of Q 2 and Q 3 .
In addition, the simulation results and the theoretical bounds of
waiting times for several QSRs are demonstrated. As long as the 5.2. Second experiment
QSPs are properly calculated, the waiting times are expected to
be accurate. A common trend in Fig. 4(b), (d), and (f ) and Fig. 6 In the second experiment, the theoretical values of W RBFNi are
are that the waiting times increase rapidly and converge to their investigated further for different queue capacity. This assessment
theoretical bounds consistently. These waiting times are absolutely conducts the simulations with varying queue capacity K from 3
affected by the corresponding QSPs. Actually, these waiting times to 10 and K 1 K 2 K 3 K, with a fixed q at 10. The other queue
are inversely proportional to their QSPs, e.g. W RBFN i =W EDF
i  parameters are the same as in Table 1. Related parameters of
ci =ci ; 8i. Consequently, it is obvious that the larger the queue
EDF RBFN the RBFN, including centers ci and widths ri . (Eq. (15)) and
service probability, the smaller the waiting time. weights wi (Eq. (13)), are also summarized in Appendix for
In Table 2, the theoretical results of queue cycle si are also pre-
sented for reference. From Eqs. (16) and (5), sRBFN
i is approximately
dependent on Di =K i . In fact, a large Di implies less stringent 6
The ratios of sFCFS
i , 7.98:10.03:4.44, are proportional to 1/K i , i.e. 1/25:1/20:1/45.
278 M.-S. Chen, H.-W. Yen / Computers & Industrial Engineering 94 (2016) 269279

RBFN(sim.)
35 W1
RBFN
30 W1
W1 25
20
RMSE(W RBFN(sim.)
1
,W RBFN
1
)=5.48x10-1
15
(a)
10
3 4 5 6 7 8 9 10
K1

RBFN(sim.)
W2
50
RBFN
W2
W2 40

30
RMSE(W RBFN(sim.)
2
,W RBFN
2
) =3.6x10-1
20 (b)
3 4 5 6 7 8 9 10
K2

80 W3
RBFN(sim.)

70 RBFN
W3
W3 60
50
40
RMSE(W RBFN(sim.)
3
,W RBFN
3
) =5.19x10-1
30
(c)
20
3 4 5 6 7 8 9 10
K3

Fig. 7. Root-mean-squared-error (RMSE) of waiting times between theoretical and simulation results of the RBFN-based QSR.

Table A1
Center vectors and widths for 7 RBFs.
6. Conclusion

Center vector Width The motivation of this study originates from supporting the
RBF1 0.875 0.778 0.587 0.409 0.403 0.362 0.174 industrial Controller Area Network (CAN) as the backbone of the
RBF2 0.496 0.471 0.403 0.409 0.470 0.245 0.612 Intelligent Vehicle System (IVS). The primary role of the CAN is
RBF3 0.672 0.783 0.711 0.435 0.417 0.523 0.352
RBF4 1.210 1.061 0.640 0.505 0.608 0.508 0.517
to provide the data communications infrastructure for controlling
RBF5 0.624 1.072 0.851 0.530 0.509 0.388 0.439 various devices and subsystems within an IVS. Under heavy or
RBF6 1.046 0.936 0.744 0.594 0.487 0.472 0.215 overload fluctuations, the CAN is very vulnerable to network load
RBF7 0.769 0.766 0.860 0.415 0.455 0.351 0.184 and may lead to increase in the delay of the messages exchanged
between the controller, actuators, and sensors, etc. In the worst
case, this phenomenon may jeopardize the correct behavior of
the whole system. Indeed, this paper is an extension of a series
Table A2 of our previous works of message scheduling related to the CAN,
Weight vectors from RBFs to outputs. including theoretical analysis, hardware implementation, and inte-
grations of both issues. This paper thus particularly deals with the
Weight vector
issue of multi-queue, finite capacity, timing constraint scheduling
RBF1 0.004 0.018 0.010 problems in overload conditions. Under the specific RBFN-based
RBF2 0.174 0.151 0.004
RBF3 0.019 0.028 0.039
queue selection rule, the closed form expressions of the optimal
RBF4 0.226 0.069 0.097 queue service probabilities not only provide particular insight for
RBF5 0.139 0.127 0.064 normal load and overload problems, but yield an explicit and
RBF6 0.018 0.013 0.046 detailed characterization of the message waiting times and other
RBF7 0.089 0.086 0.043
interesting quantities. Experimental results show that the pro-
posed methods work satisfactorily on the simulation examples.
Several existing traditional queue selection rules, i.e. EDF, FCFS,
reference. The performance metrics are evaluated with the and PQ, are also evaluated and compared with the RBFN-based
root-mean-squared-error (RMSE). The purposes of this experiment QSR.
are twofold. Firstly, the influence of change of queue capacity is In summary, the advantages of using RBFN-based QSR over
examined with respect to waiting times. Secondly, we validate other conventional scheduling scenarios are its learning capability
RBFNsim: and novelty to adapt its behavior to the changing environments
the derived W RBFN
i and the measured W i . In the overload sit-
under a variety of uncertain or constrained conditions. Specially,
uation, each queue is instantly full. If K i (or K) is large, messages cooperation with the learning-based approach in the queueing
must experience larger waiting times. In Fig. 7, waiting times models is one of the pioneer researches in this area. The difficulties
RBFNsim:
increase with a large K i , as was expected. Furthermore, W i (or disadvantages) are the design and establishing of the novel
closely resembles and the RMSE (in the order of 101 )
W RBFN
i learning strategies to comply with different queueing circum-
between them is also shown in Fig. 7 for reference. stances. It is inevitable that the design of systems with learning
 M.-S. Chen, H.-W. Yen / Computers & Industrial Engineering 94 (2016) 269279
mechanisms becomes complex and their efficient realizations can
be possibly accomplished by those with expertise both in the
selected learning algorithms and the application domain. Even
when such expertise is available, implementations of conceptually
simple learning-based programs can be time consuming. In this
sense, improving the computational efficiency can be achieved by
combining integrated hardware frameworks. Our future work
should incorporate all of our research activities (Chen, 2015;
Chen & Yen, 2011, 2012a, 2012b, 2013, 2014) together to realize
a practical industrial Controller Area Network prototype system,
which can provide sound theoretical supports for practical
implementations.
Acknowledgment
This research was supported by the Ministry of Science and
Technology, Taiwan, R.O.C., under contract numbers MOST 103-
2221-E-212-011 and MOST 104-2221-E-212-011.
Appendix A
The structure of the RBFN is 6-7-3, which corresponds to the
number of inputs, number of RBFs, and number of queues.
Table A1 illustrates center vectors ci and widths ri for seven RBFs.
The last column is the widths of these RBFs. On the other hand,
Table A2 shows the connection weight vectors w from RBFs to
three output nodes.
References
Andradottir, S., & Ayhan, H. (2005). Throughput maximization for tandem lines
with two stations and flexible servers. Operations Research, 53, 516531.
Andradottir, S., Ayhan, H., & Down, D. G. (2001). Server assignment policies for
maximizing the steady-state throughput of finite queueing systems.
Management Science, 47, 14211439.
Andradottir, S., Ayhan, H., & Down, D. G. (2003). Dynamic server allocation for
queueing network with flexible servers. Operations Research, 51, 952968.
Bell, S. L., & Williams, R. J. (2005). Dynamic scheduling of a parallel server system in
heavy traffic with complete resource pooling: Asymptotic optimality of a
threshold policy. Electronic Journal of Probability, 10441115.
Carter, M. & Lapierre, S. (1991). Scheduling emergency room physicians. Technical
Report, Mechanical and Industrial Engineering, University of Toronto.
Chen, M.-S. (2015). Neuro-fuzzy approach for online message scheduling.
Engineering Applications of Artificial Intelligence, 38(February), 5969.
Chen, H. & Mohapatra, P. (2002). Session-based overload control in QoS-aware Web
servers. In Proceedings of the twenty-first annual joint conference of the IEEE
computer and communications societies (INFOCOM 2002) (vol. 2, pp. 516524).
Chen, M.-S., & Yen, H.-W. (2011). Applications of machine learning approach on
multi-queue message scheduling. Expert Systems With Applications, 38(4),
33233335.
Chen, M.-S., & Yen, H.-W. (2012a). A state diagram analysis of the multi-queue M/
M/1 model with finite lengths. Journal of the Chinese Institute of Engineers, 35(2),
165179.
Chen, M.-S., & Yen, H.-W. (2012b). An online RBF network approach for adaptive
message scheduling on controller area networks. Journal of Information Science
and Engineering, 28(3), 503519.
Chen, M.-S., & Yen, H.-W. (2013). A probabilistic approach to estimate the mean
waiting times in the EDF polling. Computers & Industrial Engineering, 64(3),
817826.
Chen, M.-S., & Yen, H.-W. (2014). A two-stage approach in solving the state
probabilities of the multi-queue M/G/1 model. International Journal of Systems
Science (June). http://dx.doi.org/10.1080/00207721.2014.919427.
Cherkasova, L., & Phaal, P. (2002). Session-based admission control: A mechanism
for peak load management of commercial web sites. IEEE Transactions on
Computers, 51(6), 669685.
Dehghana, M., Abbaszadeha, M., & Mohebbib, A. (2014). The numerical solution of
nonlinear high dimensional generalized Benjamin BonaMahonyBurgers
equation via the meshless method of radial basis functions. Computers &
Mathematics with Applications, 68(3), 212237.
279
Dehghan, M., Abbaszadeh, M., & Mohebbi, A. (2015). A meshless technique based on
the local radial basis functions collocation method for solving parabolic
parabolic PatlakKellerSegel chemotaxis model. Engineering Analysis with
Boundary Elements, 56, 129144.
Dehghan, M., & Mohammadi, V. (2014). The numerical solution of FokkerPlanck
equation with radial basis functions (RBFs) based on the meshless technique of
Kansas approach and Galerkin method. Engineering Analysis with Boundary
Elements, 47, 3863.
Dehghan, M., & Mohammadi, V. (2015). The numerical solution of CahnHilliard
(CH) equation in one, two and three-dimensions via globally radial basis
functions (GRBFs) and RBFs-differential quadrature (RBFs-DQ) methods.
Engineering Analysis with Boundary Elements, 51, 74100.
Derlet, R., Richards, J., & Kravitz, R. (2001). Frequent overcrowding in US emergency
departments. Academic Emergency Medicine, 8(2), 151155.
Elnikety, S., Nahum, E., Tracey, J., & Zwaenepoel, W. (2004). A method for
transparent admission control and request scheduling in E-commerce Web
sites. In Proceedings of the 13th international conference on World Wide Web
(pp. 276286). ACM Press.
Georgiadis, P., & Michaloudis, C. (2012). Real-time production planning and control
system for job-shop manufacturing: A system dynamics analysis. European
Journal of Operational Research, 216(1), 94104.
Haykin, S. (1994). Neural networks: A comprehensive foundation. New York:
Macmillan College Publishing Company.
Ilati, M., & Dehghan, M. (2015). The use of radial basis functions (RBFs) collocation
and RBF-QR methods for solving the coupled nonlinear sine-Gordon equations.
Engineering Analysis with Boundary Elements, 52, 99109.
Johansson, K. H., Trngren, M., & Nielsen, L. (2005). Vehicle applications of
controller area network. Handbook of Networked and Embedded Control
Systems (2005), (pp. 741766). Birkhuser.
Kargahi, M., & Movaghar, A. (2006). A method for performance analysis of earliest-
deadline-first scheduling policy. Journal of Supercomputing, 37(2), 197222.
Kingman, J. F. C. (1962). The effect of queue discipline on waiting time variance.
Mathematical Proceedings of the Cambridge Philosophical Society, 58(1), 163164.
Kramer, S. B., & Assad, A. A. (2012). Alternating priority versus FCFS scheduling in a
two-class queueing system. Operations Research Letters, 40(6), 506509.
Lakshmi, C., & Sivakumar, A. I. (2013). Application of queueing theory in health care:
A literature review. Operations Research for Health Care, 2(12), 2539.
Little, J. D. C. (1961). A proof for the queueing formula L = kW. Operations Research, 9
(3), 383387.
Mubeen, S., Mki-Turja, J., & Sjdin, M. (2014). Extending worst case response-time
analysis for mixed messages in controller area network with priority and FIFO
queues. IEEE Access, 2, 365380.
Ndreca, S., & Scoppola, B. (2008). Discrete time GI/Geom/1 queueing system with
priority. European Journal of Operational Research, 189(3), 14031408.
Osorio, C., & Bierlaire, M. (2009). An analytic finite capacity queueing network
model capturing the propagation of congestion and blocking. European Journal
of Operational Research, 196(3), 9961007.
Sahba, P., & Balcioglu, B. (2011). The impact of transportation delays on repairshop
capacity pooling and spare part inventories. European Journal of Operational
Research, 214(3), 674682.
Sakuma, Y., & Inoie, A. (2014). An approximation analysis for an assembly-like
queueing system with time-constraint items. Applied Mathematical Modelling,
38(24), 58705882.
Shakkottai, S., Srikant, R., & Stolyar, A. L. (2004). Pathwise optimality of the
exponential scheduling rule for wireless channels. Advances in Applied
Probability, 10211045.
Shen, L. J., & Buscher, U. (2012). Solving the serial batching problem in job shop
manufacturing systems. European Journal of Operational Research, 221(1), 1426.
Stolyar, A. L. (2004). Maxweight scheduling in a generalized switch: State space
collapse and workload minimization in heavy traffic. Annals of Applied
Probability, 14(1), 153.
Tassiulas, L., & Bhattacharya, L. L. (2000). Allocation of interdependent resources for
maximal throughput. Stochastic Models, 16(1), 2748.
Tsai, C. -C. (2008). Dispatching rules with queue-time-limit consideration for furnace in
wafer fabrication factory (Master thesis). National Tsing Hua University.
Venkataraman, V. J. & Lin, X. (2007). Structural properties of LDP for queue-length
based wireless scheduling algorithms. In Proceedings of the 45th annual allerton
conference on communication, control, and computing, Monticello, IL.
Wu, C.-H., Lin, J. T., & Chie, W.-C. (2012). Dynamic production control in parallel
processing systems under process queue time constraints. Computers &
Industrial Engineering, 63(1), 192203.
Yen, H. W. (2011). Research and analysis of queue selection criterions in the multi-
queue systems (Ph.D. thesis). Da-Yeh University.
Yuan, L. (2012). Reliability analysis for a k-out-of-n: G system with redundant
dependency and repairmen having multiple vacations. Applied Mathematics and
Computation, 218(24), 1195911969.