Queuing Processes

Abstract.

Queuing Processes is a subject that have been overlooked by research in

economic applications. The objective of this paper is to give some economic insights of

the queuing applications that have been tried over the years, as well as to propose some

new ways of tackling the problem.

Keywords: Review, Poisson Process, Markov Process, Design and Control of queues.

1. Introduction

This paper is meant to be review of the applications of the queuing theory in

economic literature, as well as (hopefully) give a new start in this domain. The first part

of this paper is meant to give a short history of development of queuing theory and to

familiarize the reader with the specific jargon of this language.

The second part is devoted to economic applications of queuing theory. The

classic customer service model is presented and some extensions of this model.

Part three is trying to present some queuing applications from a new perspective.

There are two main development areas in recent queuing literature; unfortunately none of

them have been fully implemented in economic applications. Consider a resource

allocation problem, for example, using linear programming to determine how much of

each product should be made, within the limitations of the resources needed to make the

products. This is a prescriptive model since it prescribes the optimal course of action to

1
follow. This type of approach to the problem has the generic name of Design and Control

of queues.

The other approach to the problem is to observe a real live situation, and to try to

figure out the parameters involved. This sounds like an old song because all the queuing

theory is doing just that: put a real live situation into a known mathematical pattern and

apply formulas. Simulation isn’t just as simple though, because most of the real live

problems don’t have a definite mathematical model just for that case, and even if there is

one it might not have closed form solutions. Also, the next step for simulation is to try to

improve the model by trying diverse “laboratory situations” and see which one improves

the model the best.

Part I. Basics of queuing theory.

1.1 History

The theory of queues was initiated by the Danish mathematician A. K. Erlang,

who in 1909 published “The theory of Probabilities and Telephone Conversation”. He

observed that a telephone system was generally characterized by either (1) Poisson input

(the number of calls), exponential holding (service) time, and multiple channels (servers),

or (2) Poisson input, constant holding time and a single channel. Erlang was also

responsible in his later works for the notion of stationary equilibrium and for the first

consideration of the optimization of a queuing system.

Applications of the theory to the telephony were soon appearing. In 1927, E. C.

Molina published “Application of the Theory of Probability to Telephone Trunking

2
Problems”, and one year later Thornton Fry printed “Probability and its Engineering

Uses” which expand much of Erlang’s earlier work. In the early 1930’s Felix Pollaczeck

did some further pioneering work on Poisson input, arbitrary output, and single and

multiple channel problems. Other names working in the same field during that period

included Kolmogorov and Khintchine in Russia, Crommelin in France and Palm in

Sweden. The work in queuing theory picked up momentum rather slowly in its early

days, but in 1950 started to accelerate and there have been a great deal of work in the area

since then.

1.2 Early applications

There are many valuable applications of the theory, most of which have been well

documented in the literature of probability, operations research, management science and

industrial engineering. Examples include: traffic flow (vehicles, aircraft, people,

communications), scheduling (patients at the doctor, programs on a computer), and

facility design (banks, post offices, fast-food restaurants).

Queuing theory originated as a very practical subject but much of literature up to

middle 1980’s was of little direct practical value. Since then the emphasis in literature on

finding the exact solution of queuing problems with clever mathematical tricks is now

becoming secondary to model building and the direct use of these techniques in decision

making. Most real problems do not correspond exactly to a mathematical model and do

not always have closed-form solutions, but most of the time we are able to conduct

computational analysis and find approximate solutions. We have to thank for this to our

every day companion, the computer.

3
 1.3 Characteristics

The mechanism of the queuing process is very simple. Customers (not necessarily

human customers) are arriving for service, waiting for service if it is not immediate, and

leaving the system as soon as they are served.

There are six basic characteristics of queuing processes which provide an

adequate description of a queuing system: (1) arrival pattern of customers, (2) service

pattern of servers, (3) number of service channels, (4) system capacity and (5) queue

discipline.

In usual queuing systems the arrival pattern of customers is stochastic and it is

thus necessary to know the probability distribution describing the time between

successive customer arrivals (interarrival times). Also the arrival pattern can change with

time so we differentiate between stationary and nonstationary arrival patterns. The same

discussion applies to the service pattern of servers, a probability distribution is needed to

describe the sequence of customer service times. Queue discipline refers to the manner in

which customers are selected for service when a queue has formed. The most common

discipline is first come, first served (FCFS), but there are many others like last come, first

served (LCFS) which is applicable in many inventory systems as it is easier to reach the

nearest item; randomly selecting for service (RSS) independent of the arrival time of the

customer; and a variety of priority schemes, the customers with higher priority being

served ahead of the lower priority customers regardless of the order in which they arrived

to the system.

4
 1.4 Notation

As shorthand for describing queuing processes, a notation has evolved, due to Kendall

(1953), which is now standard throughout the queuing literature. A queuing process is

described as A/B/X/Y/Z, where A indicates the interarrival-time distribution, B the service

pattern described by the probability distribution for service time, X the number of servers,

Y the restriction on system capacity and Z the queue discipline. Table 1 bellow

summarizes some of the most common symbols.

Characteristic Symbol Explanation

Interarrival time M Exponential
Distribution (A) D Deterministic
and Ek Erlang type k (k=1,2,…)
Service-time Hk Mixture of k exponentials
Distribution (B) G General distribution
Number of servers (X) 1,2,…,∞
Restriction on system capacity 1,2,…,∞
FCFS First come first served
Queue LCFS Last come, first served
discipline RSS Random selection for service
PR Priority
GD General discipline

In many situations only the first three symbols are used. Current practice is to

omit the service capacity symbol if no restriction is imposed (Y=∞), and to omit the

queue discipline if it is first come first served (Z=FCFS). The symbol G represents a

general probability distribution; all we know is that the interarrival times are independent

and identically distributed.

5
 Part II. Review of literature. Economic Examples.

Surprisingly there are not so many economic applications studied yet. Some of the

latest applications include:

2.1 Customer service

The classical model of clients waiting for service while incurring waiting cost to get

reward from service have been implemented extensively but for the practical

applications it has not surpass the model M/M/1. There have been attempts to construct

more complex models. For example Kleinrock (1968) propose a model M/M/1 with a

little different service discipline. Instead of FCFS discipline customers are using bribe to

gain a better position in line. A customer paying a bribe will be served before the ones

who paid smaller bribes in the queue, but after the people in the queue who paid larger

bribes. This model is appropriate to some undeveloped countries where for gaining an

audience you are encouraged to pay a fee to the clerk doing registration. The model is

also appropriate to an auction process. Leff (1970) suggested that such a model may

have beneficial effects serving as a catalizator for an otherwise sluggish economy.

Myrdal (1968, chap. 20), argues that the corrupt official may deliberately cause delays

in order to attract more bribes. If this is the case then the efficiency argument is gone.

This opinion is answered by Lui, who in 1985 showed that, if the server does not decide

the amounts of bribe payments, but the customers themselves are doing that, there exists

Nash equilibrium. Under the extra condition, that the server is interested in speeding up

the process in order to gain more bribes, the outcome is also socially optimal.

6
 Luski (1976) discusses the notion of equilibrium in a queuing system with two

servers (M/M/2 model). Specifically there are two firms offering the same kind of

service, each fixing its price as to maximize profit. The decision of customers in joining

one of the queues or not joining at all is made according to their cost of waiting, the

expected waiting time and the price of the service in each firm. He showed that if the

firms are not cooperating, and if the service time for each firm μ < λ/2, each firm will

sell at a different price in order that an equilibrium exists. In the alternate case μ > λ/2

the firms will sell at the same price (the same mechanism that governs the real life). Not

only that but he showed that if the firms are cooperating it is possible to raise joint

profits by playing with prices (again logical: more or less like a monopolistic model).

2.2 Derivatives of the Customer Service Model: Bottleneck model.

This model is an extension of the previous M/M/1 model. Customers are arriving at a

facility and are faced with the decision of entering or not in the system (balking

phenomenon). Each customer has a cost per unit of service and waiting time, c, and

receives a benefit R if he is served by the facility. In addition to the previous models the

system charges a fee θ for the service and this toll determines a critical queue size. This

model is pertinent with some real life situations, like today’s Internet providers or toll

paying highways. P. Naor (1969) has shown that the critical length of the queue, which

maximizes the social welfare function is greater then the one which maximizes the

expected revenue per unit time. In other words, the revenue-maximizing toll exceeds the

socially optimal toll. Edelson and Hildebrand (1975) have shown that this is not the case

if no balking is permitted and the same toll is Nash and Pareto Optimal.

7
 R. Arnott, A. de Palma and R. Lindsey (1995-1996), basing their work on a model

initially developed by Vickrey (1963,1969) and similar with the previous one,

generalized this problem. They allowed multiple servers (service time is deterministic),

and also the customer, instead of just deciding whether or not to enter the system, now

has the extra choice of when to use the facility. The decision is made accordingly to

information that each client has, information that is stochastic with some distribution. The

authors define better information from the social efficiency of the model point of view.

The model is very applicable. Roads with today’s severe congestion problems and with

the information that each driver can obtain (listening radio, information panels etc…) are

the best example of such application. Arnott et all have shown that, if the customers

decide only whether or not to use the facility, better information is more efficient. They

have shown also that if the customers decide both whether and when to use the facility

then the above result doesn’t hold anymore and better information can be harmful when

the congestion is unpriced. However if efficient tolling (different toll prices during the

congestion hours) is applied the above result holds.

Another extension of the M/M/1 model is an article by Joseph Daniel (1995). He

applies this queue to an airport, modeling arrivals and service time until departures as

stochastic. Also he is applying his thus constructed model to a real life case using data

from the Minneapolis-St. Paul Airport during the firs week of May 1990. He allows the

arrival rate to be modeled by a time dependent Poisson Process, in fact dividing the time

into equal intervals during which the arrival rate is constant. The costs incurred are the

usual ones: queuing cost and service cost, where by service cost he understands the price

8
paid for the difference from the scheduled time of departure. The social planner can

impose a congestion fee, which will depend of the scheduled time of operation.

What is remarkable about this article is that it estimates the parameters of the model

and then uses those parameters to simulate the data. Using simulation allows him to find

the optimal congestion toll and the optimal number of servers (i.e. by building additional

runways).

Part III. Examples in mathematical form

3.1 Insurance risk

We make the following assumptions regarding the business of an insurance company:

(i) The number of claims arising in a time interval (0,t] has a Poisson Distribution

with parameter λ, 0≤ λ≤ ∞.

(ii) The amounts of successive claims are independent random variables with

common distribution B(x) (-∞< x< ∞), where x can take negative values in the

case of ordinary whole-life annuities.

We can consider this model as a M/G/1 queuing system where the input process A(t) is

just a simple Poisson process with parameter λ giving us the moments of claims. We will

consider the service as being the amount of each claim, and let X(t) the total amount of

claims arising in (0,t]. Assuming X(t)=0 we have:

(3.1) X(t) = ζ1 + ζ2 + … + ζA(t) ,

where ζ1 , ζ2 ,… are the successive claims. Because A(t) is a Poisson process, the

distribution function of X(t) is given by:

9
 ∞
( λt ) n
(3.2) χ(x,t) = P{ X(t) ≤ x }= ∑e λ
n =0
− t

n!
Bn ( x )

This is the compound Poisson distribution. We have:

∞ ∞
(3.3) EX (t ) = λt ∫ xdB( x) , and VarX (t ) = λt ∫ x 2 dB ( x)
−∞ −∞

(iii) The totality of policyholders pays the company premiums at a constant rate β. We

call β the gross risk premium rate. We will assume that α and β have the same

sign.

The company’s reserve fund at time t is given by Z(t) = x + βt – X(t) , with initial

value Z(0) = x ≥ 0. Here Z(t) may assume positive or negative values.

The objective of the company is to choose the initial reserve x large enough to

avoid ruin over a finite or infinite horizon.

So, the company is concerned with the stopping time:

(3.4) T(x) = inf { t > 0 : x + βt – X(t) < 0 },

and wants to evaluate the probability of avoiding ruin over a finite or an infinite

horizon, that is,

(3.5) P{T(x) > t} (for some 0 < t <∞) or P{ T(x) = ∞}.

Prabhu (1998) showed that the storage process Z(t) is a Markov process and he

computed the limit probabilities (3.5) when t goes to ∞.

In the model described above we have seen the basic process, which is the

compounded Poisson Process. An important feature of this model is that at any moment

of time only a finite number of events have occurred (namely customer claims). The

10
resulting property of the resulting process is that its sample function takes only a finite

number of jumps in each finite interval. However this description of the input is very

unrealistic in certain situations. Which brings us to the next model:

3.2 Storage Model

Consider a model of a granary with a large enough (effectively infinite) capacity.

Let X(t) denote the input of grain into it during a time interval (0,t].

In order to formulate such a process let’s suppose that X(t) has stationary and

independent increments i.e.:

(i) For 0≤t1≤…≤tn with n≥2 we have X(t1), X(t2)- X(t1),…, X(tn)-X(tn-1) are

independent

(ii) Distribution of increment X(tp)-X(tp-1) depends only on tp-tp-1

Suppose furthermore that the process X is right continuous, the quantity X(t)-X(t-) being

the jump at time t with only a finite number of jumps allowed.

Under these regularity conditions X the input process is a Levy process (a broad

family containing Brownian Motion and Compounded Poisson Process among others). In

t
particular Z(t)=Z(0)+X(t)+ ∫ 1{Z ( s ) =0} ds is a semimartingale so all the theory of Stochastic
0

integration can be applied (see Protter 1988).

Again we are concerned with T(x)= inf{t : Z(t)=0}the period until the first

shortage appears (here x=Z(0))and again Prabhu (1998) found the distribution of T under

the assumption T(0)=0.

11
 3.3 Livestock problem:

Suppose that we have an animal farm that grows pigs. We want to put this into a

queuing model. To do this we will make the following assumptions:

Each time a new animal is born we have to pay a fixed amount of money β for doctor and

other costs. In addition each animal costs us money until we service it. C denotes this cost

per unit time. We will consider this model as a M/M/1/K queue where the last number K

is the finite capacity of the model.

As usual arrival rate is λ service rate is μ. We want to find the optimal truncation

value k, which maximizes the expected profit rate. Mathematically this problem is to

find:

max(βλ (1 − p k ) − cλ (1 − p k )W )
1≤ k ≤ K

, where the quantities pk and W are given by:

k
⎛λ⎞
pk= ⎜⎜ ⎟⎟ for : 0 ≤ k ≤ K and W=
⎝μ⎠

3.3.1 Another option for livestock model (M/M/1 queue with controlled

service rate):

Suppose now that we have an unlimited storage capacity but we can control the

rate of service μ in an interval [0, &μ&& ]. We can change the rate of service after servicing

each pig. Let S(μ) be the cost per unit time for using rate μ and let C(i) be the cost per

unit time when there are i pigs in the system (queue plus service). We want to cut down C

by choosing a faster service rate, which costs more; in other words we want to find

optimal tradeoff between service cost and system cost.

12
Assumptions:

1. There is a μ such that μ > λ

2. C is positive, nondecreasing and “convex”: C(i+2) - C(i+1) ≥ C(i+1) – C(i)

3. S is nonnegative and continuous on [0, &μ&& ]; S(0)=0

Here the state variable is the number of customers in the system. The control

variable is the choice of service rate after a customer departs.

Take ν = λ + &μ&& the uniform transition rate. Bertsekas(1987) shows that the

problem can be reduced to this transition rate. The cost per stage is: C(i) + S(μ). Then the

Bellman equation is:

J(0)= min [C(0) + (ν- λ)J(0) + λJ(1)]

μ

J(i)= min [C(i)+S(μ) + μJ(i-1) + (ν- λ -μ)J(i) + λJ(i+1)]

μ

So we get the optimal policy: use at state i the service rate that minimizes:

S(μ) - μ [J(i) - J(i-1)]

3.3.2 M/M/1 model with controlled arrival rate

We will look to the same problem from yet another perspective. Suppose that the

&&& ]. The
service rate is fixed λ but we can control now the birth rate into an interval [0, λ

only thing that modifies from the previous model is that we have now a cost S(λ) per unit

time associated with the specific rate of birth chosen instead of the cost specific to

different service rates.

&&& + μ, the uniform transition rate. In this case Bellman equation

Again let ν = λ

becomes

13
 J(0) = min [C(0) + S(λ) + (ν- λ)J(0) + λJ(1)]
λ

J(i)= min [C(i)+S(μ) + μJ(i-1) + (ν- λ -μ)J(i) + λJ(i+1)]

λ

Maximizing this time with respect to λ gives us the optimal policy: use at state i

the birth rate that minimizes S(λ) + λ [ J(i) – J(i-1)].

Bibliography:

• Arnott R., de Palma A., Lindsey R., “Economics of a Bottleneck” Journal of Urban

Economics 27, (January 1990), 111-130

• Arnott R., de Palma A., Lindsey R., “Does Providing Information to Drivers Reduce

Traffic Congestion?”, Transportation Research 25, (no.5 1991), 309-318

• Arnott R., de Palma A., Lindsey R., “Information and Usage of Free-Access Congestible

Facilities with Stochastic Capacity and Demand”, International Economic Review 37,

(January 1996), 181-203

• Arnott R., de Palma A., Lindsey R., “Information and Time-of-Usage Decisions in the

Bottleneck Model with Stochastic Capacity and Demand”, European Economic Review

43, (March 1999), 525-548

• Bertzekas, Dimitri P. Dynamic programming and stochastic control. New York:

Academic Press, 1976

• Daniel Joseph, “Congestion Pricing and Capacity of Large Hub Airports: A Bottleneck

Model with Stochastic Queues”, Econometrica 63, (March 1995), 327-370

• Edelson N., Hildebrand D. “Congestion Tools for Poisson Queuing Processes”,

Econometrica 43, (January 1975) 81-92

14
• Gross, D., Harris, C. Fundamentals of Queueing Theory. New York: Wiley, 1998

• Kleinrock, Leonard. “Optimum Bribing for Queue Position”. Operations Res. 15

(March/April 1967): 304-318

• Leff, Nathaniel H. “Economic Development through Bureaucratic Corruption”. In

Political Corruption: Readings in Comparative Analysis, edited by Arnold

Heidenheimer. New York: Holt, Rinehart, & Winston, 1970.

• Lui, Francis T. “An Equilibrium Queuing Model of Bribery”. Journal of Political

Economy 93 (August 1985): 760-781

• Luski, Israel “On Partial Equilibrium in a Queuing System with Two Servers” Review of

Economic Studies 43 (October 1976) 519-525

• Myrdal, Gunnar. Asian Drama: An inquiry into the Poverty of Nations. New York:

Pantheon, 1969.

• Naor, P. “The regulation of Queue Size by Levying Tolls”. Econometrica 37 (January

1969): 15-24

• Prabhu N.U. Foundations of Queuing Theory. Boston: Kluwer Academic Publishers,

1997.

• Prabhu N.U. Stochastic storage processes: queues, insurance risk, dams, and data

communication. New York: Springer, 1998

15