jPhase Users Guide

Juan F. P erez
[email protected]

Germ an Ria no
[email protected]

Centro de Optimizaci on y Probabilidad Aplicada (COPA) Departamento de Ingenier a Industrial Universidad de los Andes Bogot a, Colombia

Contents
1 Phase-Type Distributions 1.1 Continuous Phase-Type Distributions . . 1.2 Discrete Phase-Type Distributions . . . . 1.3 Closure Properties . . . . . . . . . . . . . 1.4 Further Closure Properties for Continuous 1.5 Phase-Type Random Variates Generation 1.6 Fitting Algorithms . . . . . . . . . . . . . 2 jPhase: the object-oriented 2.1 General Structure . . . . . 2.2 Interfaces . . . . . . . . . 2.3 Abstract Classes . . . . . 2.4 Implementing Classes . . 2.5 jPhase Examples . . . . . Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 2 3 3 5 6 6 6 6 7 7 7 8

3 jPhaseGenerator: the variates generator Module 10 3.1 PhaseGenerator Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 3.2 Implementing Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 3.3 jPhaseGenerator Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 4 jPhaseFit: the tting module 4.1 Abstract Classes . . . . . . . . . . . . . . . . . . . . . 4.2 Maximum Likelihood Algorithms . . . . . . . . . . . . 4.2.1 General PH Distribution EM Algorithm . . . . 4.2.2 Hyper-exponential Distribution EM Algorithm 4.2.3 Hyper-Erlang Distribution EM Algorithm . . . 4.3 Moment Matching Algorithms . . . . . . . . . . . . . . 4.3.1 Acyclic Continuous order-2 Distributions . . . 4.3.2 Erlang-Coxian Distributions . . . . . . . . . . . 4.3.3 Acyclic Continuous Distributions . . . . . . . . 4.4 jPhaseFit Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 13 14 14 14 14 14 15 15 15 15

Introduction
Phase-Type (PH) distributions are a powerful tool in stochastic models of real systems. jPhase [19] is a Java-based framework that allows the representation of PH distributions through computational objects. The developed structure induces a formal representation of a Phase-type distribution and a set of properties that it should have. In this Manual, we illustrate the use of jPhase through several examples that are included in each section. This document is organized as follows: the rst section gives an introductory background in PH distributions. Section 2 shows how the core module jPhase if structured and gives some examples. Sections 3 and 4 illustrate the structure of the packages jPhaseFit and jPhaseGenerator, and also give several examples of how to use the services provided by the package. More examples and applications for stochastic modeling with jMarkov can be found in the release that can be downloaded from copa.uniandes.edu.co.

Phase-Type Distributions

In this section, we review the denition and some properties of PH distributions. We follow the treatment presented in [14] and [10], and therefore, we do not include proofs in this section since the interested reader can nd them in those books.

1.1

Continuous Phase-Type Distributions

A Continuous Phase-Type distribution can be dened as the time until absorption in a Continuous Markov Chain, with one absorbing state and all others transient. The generator matrix of that process can be written as: Q= 0 0 , a A

where the rst entry in the state space represents the absorbing state. As the sum of the elements on each row must equal zero, a is determined by a = A1, where 1 is a column vector of ones. In order to completely determine the process, the initial probability distribution is dened and can be partitioned in the same way of the generator matrix 0 , where 0 is the probability of starting the process in the state 0, and the sum of all the components in the vector must be equal to 1. Therefore, 0 is determined by the following relationship 0 = 1 1. In this way, the distribution of a Continuous Phase-Type variable X is completely determined by the parameters (, A), and its cumulative distribution function (CDF) is F (x) = 1 eAx 1, x 0,

which has a clear connection to the well known exponential distribution. Furthermore, if there is just one transient phase with associate rate and it is selected with probability one, then the distribution is exactly the exponential case. From the previous expression, the probability density function (PDF) of the continuous part can be computed as f (x) = eAx a, 2 x > 0.

And similarly, the Laplace-Stieltjes transform of F (), is given by 0 + (sI A)1 a, Re(s) 0,

from which, the non-centered moments can be calculated as E [X k ] = k !(A1 )k 1, k 1.

1.2

Discrete Phase-Type Distributions

A Discrete Phase-Type distribution can be seen as an analogous case to the continuous distribution. In this case, the distribution can be dened as the number of steps until absorption in a Discrete Markov Chain, with one absorbing state and all other transient. The transition probability matrix of that process may be dened as: P= 1 0 , a A

where the rst row in the matrix represents the absorbing state. As the sum of the elements in every row of the matrix must equal one (in order to be a probability mass function), a is determined by: a = 1 A1. Similarly, the initial probability distribution is dened as 0 , where 0 = 1 1 is the probability of starting the process in the absorbing state, i.e. the number of steps in that case would be equal to zero. As before, the distribution of a discrete Phase-Type variable X is completely determined by the parameters (, A) and its probability mass function is dened as P {X = k } = 0 Ak a ,k = 0 ,k 1

This last denition makes natural the denition of the cumulative probability function of the discrete Phase-Type variable P {X k } = 1 Ak 1, Also, the generating function can be calculated as 0 + z (I z A)1 a, |z | 1. k 0.

from which, the factorial moments of the distribution can be computed E [X (X 1) . . . (X k + 1)] = k !(I A)k Ak1 1, k 1.

1.3

Closure Properties

An important issue of Phase-Type distributions is that they are closed under some operations, which can be useful in the analysis of some systems. The following closure properties are valid for both discrete and continuous distributions.

1. Convolution of a nite number of Phase-Type distributions If X P H (, A) and Y P H ( , B) (independent of X ), with n and m phases respectively, then the convolution is P H ( , C) with m + n phases and = [, 0 ] and C= A a . 0 B

2. Convex mixture of a nite number of Phase-Type distributions If X P H (, A) and Y P H ( , B) (independent of X ), with n and m phases respectively, and distribution functions F () and G(). Then, the convex mixture F () + (1 )G(), with 0 1, has representation P H ( , C) with m + n phases, where = [, (1 ) ] and C= A 0 . 0 B

3. Convolution of a discrete Phase-Type number of Phase-Type distributions N If Xi are i.i.d. continuous P H (, A) and N is a discrete P H ( , B), then k=0 Xi is P H ( , C), with = and C = A I + a B.
1

The function denotes the Kronecker product and the Kronecker sum.

As the geometric distribution is a particular case of Discrete Phase-Type distributions, this property also holds for the geometric case. If Xi are i.i.d. continuous P H (, A) and N follows a geometric distribution with parameter p, then N k=0 Xi is P H ( , C), with = and C = A + (1 p)a.

4. The minimum of a set of Phase-Type distributions If X P H (, A) and Y P H ( , B) (independent of X ), with n and m phases respectively, then min(X, Y ) P H ( , C) with mn phases and = . In this case, the matrix C has a dierent denition if the process is discrete or continuous. In the discrete case, the resulting probability transition matrix is given by C = A B. For the continuous case, the generator matrix is given by C = A B. 5. The maximum of a set of Phase-Type distributions If X P H (, A) and Y P H ( , B) (independent of X ), with n and m phases respectively, then max(X, Y ) P H ( , C) with mn + n + m phases and = [ , 0 , 0 ] AB Ib aI A 0 . C= 0 0 0 B
a12 B a22 B . . . am2 B ... ... .. . ... a 1n B a 2n B . . . amn B

The Kronecker product of matrices A and B is dened as a11 B a21 B AB= . . . am1 B

And the Kronecker sum of matrices A and B is dened as A B = A I + I B.

1.4

Further Closure Properties for Continuous Distributions

There are some other important closure properties that only apply for the case of Continuous Phase-Type distributions, which are listed below. 1. Waiting time in a M/P H/1 queue If X P H (, A) is the service time distribution in a M/G/1 queue, then the distribution of the waiting time W () is P H ( , C), with = (1 ) and C = A + a ,

where = m is the trac coecient, is the arrival rate and m is the expected value of the service time. is the stationary probability vector of A + a, i.e. = (A1 1)A1 . 2. Residual time distribution If X P H (, A), then the residual time distribution G(x) = P(X x|X > ) has representation P H ( , C), with = 1 eA . 1 F ( )

3. Equilibrium Residual time distribution If X P H (, A), then the equilibrium residual time distribution G(x) = 1 E [X ]
x

(1 F (u))du,
0

has representation P H ( , A), where has the same meaning as stated above. 4. Termination time of a Phase-Type process with Phase-Type failures [15] Consider a process where the service time is determined by Phase-Type distribution with m phases and representation P H (, A), and it is subject to failures that occur according to a Poisson process with rate . If the duration of the failure is P H ( , B) with n phases, then the total completion time has distribution G() with representation P H ( , C). Two dierent cases must be dierentiated: if the service must be restarted after the failure, or if the task can begin from the point where it was left before the failure. In the rst case, the resulting distribution has m + n phases and = [, 0] and C= A I 1 . b B

In the second case, the representation has m + mn phases and

= [, 0]

and

C=

A I I . Ib IB

The Continuous and Discrete Phase-Type distributions have the important property of being dense in [0, ) and the non-negative integers, respectively (the proof of this property can be found in [5]). This implies that any distribution with support on those sets can be approximated by a Phase-Type distribution with the appropriate number of phases and parameters and A. 5

1.5

Phase-Type Random Variates Generation

In many large applications, simulation is the appropriate tool to model the system because of the complex relations between dierent stochastic variables. This makes that a random number generator become an important tool to model a wide range of non-deterministic systems. Neuts and Pagano [13] developed two similar algorithms to generate random variates from discrete and continuous Phase-Type distributions. These algorithms are supported on the alias method [12] to generate variates from discrete distributions in order to simulate the process of selecting an initial state and then jump to the next one according to random vectors.

1.6

Fitting Algorithms

In the last twenty years, the problem of tting the parameters of a Phase-Type distribution has received great attention from the applied probability community. These dierent approaches can be classied in two major groups: maximum likelihood methods and moment matching techniques, as noted in [11]. Nevertheless, almost all the algorithms designed for this task have an important characteristic in common: they reduce the set of distributions to be tted, from the whole PhaseType set to a special subset. In section 4, those algorithms included in the computational package will be revisited and further explained.

2
2.1

jPhase: the object-oriented Framework

General Structure

The jPhase package is supported on a set of interfaces, abstract classes and implementing classes. The interfaces determine the characteristics of an object and have no implementation of any method. As can be seen in the simple Class Diagram of Figure 1, there are three interfaces in the jPhase package: PhaseVar, ContPhaseVar, and DiscPhaseVar. These interfaces determine the behavior of a PH distribution in both the continuous and discrete cases.

Figure 1: Simple jPhase Package Class Diagram The abstract classes AbstractContPhaseVar and AbstractDiscPhaseVar implements the corresponding interface (discrete or continuous), in order to develop some of the methods determined by the interfaces. Finally, the implementing classes extends the corresponding abstract class, and thus they make use of the already implemented methods. These methods are useful for any user that wants to develop his own implementing class, because he does not need to get worried about the whole set of distribution properties, but only needs to implement a little set of simple methods.

In the next sections, the properties of these interfaces, abstract and implementing classes will be explained.

2.2

Interfaces

As it was said above, the jPhase package consists of three interfaces, that determine the behavior of any PH distribution as shown next. PhaseVar This interface denes properties that are common to both discrete and continuous Phasetype distributions. Since this is the core interface in the framework, it has the major quantity of methods and all other interfaces ans classes have fewer. The methods that the interface force to implement for any distribution can be divided in three groups: access, moments and distribution methods. DiscPhaseVar and ContPhaseVar This interfaces determine some of the closure properties valid for discrete and continuous PH variables, as those discussed in section 1. The methods dened by each one of this interfaces can be partitioned in two groups: distribution and closure methods. The closure properties can only be dened at this level because each one of the discrete and continuous sets are closed under these properties, but not the whole set of PH distributions.

2.3

Abstract Classes

As shown in Figure 1, the ContPhaseVar interface is implemented by the abstract class AbstractContPhaseVar, which implements almost all the methods dened by PhaseVar and ContPhaseVar. In particular, none of the methods implemented by this class depends on the formal representation of the matrices and vectors involved. This means that all the operations are executed using solvers and preconditioners that apply for both sparse and dense representations of matrices and vectors. Moreover the probably most dicult routines are solved by this abstract class, such as the computation of the probability density function, that implies the use of uniformization methods for solve a set of dierential equations[10]. The same arguments apply for the abstract class AbstractDiscPhaseVar, that implements the interface DiscPhaseVar. This way, the only methods that the user must implement when developing an implementing class are: getMatrix and setMatrix; getVector and setVector; newVar and copy. As can be seen, this methods depend on the particular representation of the distribution, e.g. if the matrix is represented by a particular sparse pattern, then the only one class of matrices that can be set must have the same pattern. Since jPhase is built over the Matrix Toolkit for Java (MTJ) library [8], the user could develop his own representations over those oered by MTJ. Also the newVar and copy methods must return a variable that belongs to the same class of the original one. The implementing classes explained in the next section are themselves examples of classes that extend the abstract ones.

2.4

Implementing Classes

The developed implementing classes are those that a nal user will utilize. They have been designed as general PH representations for the continuous and discrete cases, and with dense and sparse storage. The DenseContPhaseVar and DenseDiscPhaseVar classes represent continuous and discrete PH distributions, using the DenseMatrix and DenseVector classes dened by MTJ [8]. These classes are useful for many applications, where the number of phases is not large and the memory is not a problem. They also have constructors for many simple distributions such as exponential or Erlang in the continuous case, and geometric or negative binomial in the discrete case. 7

Nevertheless, the use of matrices with dense representation can be a problem because of the large number of phases. The SparseContPhaseVar and SparseDiscPhaseVar classes are built over the FlexCompRowMatrix and SparseVector MTJ classes, which give a good alternative when the number of phases is large but the number of entries is little relative to the total number of n2 entries. It is important to note that the FlexCompRowMatrix allows a exible sparse pattern stored by rows, that makes of this class a general sparse representation. Other specic representation could be developed by using a particular sparse pattern, e.g. upper-diagonal matrices.

2.5

jPhase Examples

In this section, the use of jPhase is illustrated through relevant examples. All of these examples should be included in a class with a main class that calls them to be executed. In Figure 2, two PH variables: V1 is a exponential variable with parameter = 3, and V2 is an Erlang variable with parameters = 1.5 and k = 2. Once this is done, a third variable (V3 ) is created as the maximum between V1 and V2 . Finally, the probability that V3 takes a value not greater than 2.0 is computed through the cumulative density function. This value is printed and is shown in Figure 3.
ContPhaseVar v1 = DenseContPhaseVar . expo ( 3 ) ; ContPhaseVar v2 = DenseContPhaseVar . Erlang ( 1 . 5 , 2 ) ; ContPhaseVar v3 = v1 . max( v2 ) ; System . out . p r i n t l n ( P( v3 < =2.0): \ t +v3 . c d f ( 2 . 0 ) ) ;

Figure 2: jPhase: Example 1

P( v3 < =2.0): 0.7988666135682244

Figure 3: jPhase: Result for Example 1

ContPhaseVar v1 = DenseContPhaseVar . Erlang ( 0 . 8 , 3 ) ; ContPhaseVar v2 = DenseContPhaseVar . Erlang ( 1 . 5 , 2 ) ; ContPhaseVar v3 = v1 . sum ( v2 ) ; System . out . p r i n t l n ( v3 : \ n+v3 . t o S t r i n g ( ) ) ;

Figure 4: jPhase: Example 2 Figure 4 shows again the creation of two particular PH variables: V1 is an Erlang variable with parameters = 0.8 and k = 3, and V2 is an Erlang variable with parameters = 1.5 and k = 2. Then, the variable V3 is created as the sum of V1 and V2 . This variable is printed, and the result is shown in Figure 5. The printed version of the variable includes the initial probability vector and the transition matrix A, as explained in section 1. As shown in Figure 6, the distributions can be created from arrays of doubles, that represent the initial probability vector and the generator matrix of the transient states (as specied in section 1). Once the distributions are created, they can be manipulated through the use of closure properties, as shown in Figure 6, where the sum of the variables v 1 and v 2 is calculated. The resulting variable is shown next in Figure 7, where the calculated variable is printed. Since jPhase is built over the Matrix Toolkit for Java (MTJ) library [8], it is also possible to construct PH distributions from matrices and vectors dened in that library. As can be seen in Figure 8, the matrix and the vector of the PH distribution are rst built as DenseMatrix and 8

v3 : Phase Type D i s t r i b u t i o n Number o f Phases : 5 Vector : 1.00 0.00 0.00 Matrix : 0.80 0.80 0.00 0.00 0.80 0.80 0.00 0.00 0.80 0.00 0.00 0.00 0.00 0.00 0.00

0.00 0.00 0.00 0.80 1.50 0.00

0.00 0.00 0.00 0.00 1.50 1.50

Figure 5: jPhase: Result for Example 2

double [ ] [ ] A = new double [ ] [ ] { { 2 ,2 } , { 2 , 5 } } ; double [ ] a l p h a = new double [ ] { 0 . 2 , 0 . 4 } ; DenseContPhaseVar v1 = new DenseContPhaseVar ( alpha , A ) ; double [ ] [ ] B = new double [ ] [ ] { { 4 ,2 ,1 } , { 1 , 3 ,1 } , { 2 , 1 , 5 } } ; double [ ] b e t a = new double [ ] { 0 . 1 , 0 . 2 , 0 . 2 } ; DenseContPhaseVar v2 = new DenseContPhaseVar ( beta , B ) ; ContPhaseVar v3 = v1 . sum ( v2 ) ; System . out . p r i n t l n ( v3 : +v3 . t o S t r i n g ( ) ) ;

Figure 6: jPhase: Example 3

v3 : Phase Type D i s t r i b u t i o n Number o f Phases : 5 Vector : 0.2000 0.4000 0.0400 0.0800 0.0800 Matrix : 2.0000 2 . 0 0 0 0 0 . 0 0 0 0 0 . 0 0 0 0 0 . 0 0 0 0 2 . 0 0 0 0 5.0000 0 . 3 0 0 0 0 . 6 0 0 0 0 . 6 0 0 0 0 . 0 0 0 0 0 . 0 0 0 0 4.0000 2 . 0 0 0 0 1 . 0 0 0 0 0 . 0 0 0 0 0 . 0 0 0 0 1 . 0 0 0 0 3.0000 1 . 0 0 0 0 0 . 0 0 0 0 0 . 0 0 0 0 2 . 0 0 0 0 1 . 0 0 0 0 5.0000

Figure 7: jPhase: Result for Example 3

DenseMatrix A = new DenseMatrix ( new double [ ] [ ] { { 4 ,2 ,1 } , { 1 , 3 ,1 } , { 2 , 1 , 5 } } ) ; DenseVector a l p h a = new DenseVector ( new double [ ] {0.1 , 0.2 , 0.2}); DenseContPhaseVar v1 = new DenseContPhaseVar ( alpha , A ) ; double rho = 0 . 5 ; PhaseVar v2 = v1 . waitingQ ( rho ) ; System . out . p r i n t l n ( v2 : \ n+v2 . t o S t r i n g ( ) ) ;

Figure 8: jPhase: Example 4

v2 : Phase Type D i s t r i b u t i o n Number o f Phases : 3 Vector : 0.1500 0.2250 0.1250 Matrix : 3.8500 2 . 2 2 5 0 1 . 1 2 5 0 1 . 1 5 0 0 2.7750 1 . 1 2 5 0 2 . 3 0 0 0 1 . 4 5 0 0 4.7500

Figure 9: jPhase: Result for Example 4 DenseVector (MTJ objects), and then the continuous PH distribution is constructed. In this example, the distribution of the waiting time in queue is computed taking the variable V1 as the service time distribution and assuming that the trac coecient of the M/PH/1 queue is equal to 0.5. The resulting distribution is then printed and the output is shown in Figure 9. Another way to do the former calculations is through the use of the Graphic User Interface (GUI). This can be used to build PH variables from direct input, or from a data set that can be t the parameters of the distribution. It also allows to compute closure properties and has the capabilities to show graphically the probability density function or the cumulative probability distribution of a specied PH distribution. As can be seen, the developed framework is an easy way to deal with PH distributions and can be used as a supporting tool in several practical researches, where the main point is to build a probabilistic model that describes the system, and the PH distributions are an important tool to do it. Thus, the researcher can focus on the modeling issue based on the computational representation developed in this work.

jPhaseGenerator: the variates generator Module

This package was developed in order to dene the behavior of any PH random variates generator. This behavior is specied by the abstract class PhaseGenerator, which is the core the package. As can be seen in Figure 10, this abstract class is extended by the implementing classes NeutsContPHGenerator and NeutsDiscPHGenerator. Those classes implement the algorithms proposed by Neuts and Pagano [13].

10

Figure 10: Simple jPhaseGenerator Package Class Diagram

3.1

PhaseGenerator Interface

This abstract class denes the basic methods that a PH random variate generator should have. The class includes an attribute, that belongs to the PhaseVar class, and is the distribution from which, the random variates will be generated. This distribution can only be specied in the constructor method, because the variable must be persistent in time for a particular PhaseGenerator object. This means that if the user wants to generate variates from another distribution, he must create a new PhaseGenerator. In the constructor method, the variable is assigned and the initialization() method is called. It is expected that the user make use of this method in order to eectively initialize the algorithm, and then a random variate can be generated after the construction of the PhaseGenerator. Another method dened by the abstract class is getVar(), which always returns the PH variable that remain under the PhaseGenerator and is already implemented. The last two methods that a PhaseGenerator must implement are getRandom() and getRandom(k ). The rst one must return a variate that follows the distribution specied at the construction, and the second must return k independent variates with the same characteristic.

3.2

Implementing Classes

Currently, there are two implementing classes that extend the previously explained PhaseGenerator abstract class. These are NeutsContPHGenerator and NeutsDiscPHGenerator, which implement the methods proposed by Neuts and Pagano [13]. The rst one implements the continuous case and the second the discrete one. The continuous algorithm has a rst step, in which a discrete chain is biult over the continuous one, using the well-known embedded chain. Thereafter, the main algorithm (for discrete distributions) can be used for both cases. The algorithm simulates the whole process in the chain: it rst choose an initial state from the distribution given by the initial probability vector; then it selects a next state to visit using the discrete distribution associated with the present state, given by the associated row in the transition matrix; the selection of the next state is repeated until the chosen state is the absorbing one. In the discrete case, the value of the random variate is the number of steps (selections) made until absorption. For the continuous case, the number of visits to each state is stored and an Erlang variate is generated for each state with non-zero number of visits. The parameters of the Erlang distributions are the associated rate of the state and the number of visits carried out. For example, if the state i was visited ni times and has an associated rate of i , an Erlang(i ,ni ) random variate must be generated. The sum of these variates over all the states is the value of the PH random variate. 11

Two important issues of this algorithm must be emphasized. The rst one is the use of discrete distributions to generate the variates, which can be done eciently through the alias method [12]. The second issue is that for the continuous case, in addition to the discrete variates, only Erlang variates must be generated. In the case of many visits to the same state, these variates can also be eciently generated by multiplying a gamma variate with parameters (ni , 1) times i , that will be an Erlang variate with the required parameters [13]. The algorithms implemented in these classes are supported by the utilities class GeneratorUtils, that have several procedures useful for the generators. Particularly, it has a general implementation of the alias method used to generate variates from discrete distributions [12]. It also has an implementation of the polynomial-time algorithm proposed by Gonzalez et. al. [6] to perform a Kolmogorov-Smirnov test, that can be useful to test the goodness-of-t of the generated numbers in relation to the theoretic PH distribution.

3.3

jPhaseGenerator Example

A simple example of the use jPhaseGenerator is shown in Figure 11. First, a new PH variable V1 is created form its initial condition vector and the generator matrix. Then, a new generator (gen) is created using V1 as parameter, since the variates to be generated must have this distribution. These variates are going to me stored in the variates array, that is declared. The next step is ask a set of ten independent random variates from the the generator. Finally, these variates are printed and the result is shown in Figure 12.
DenseMatrix A = new DenseMatrix ( new double [ ] [ ] { { 4 ,2 ,1 } , { 1 , 3 ,1 } , { 2 , 1 , 5 } } ) ; DenseVector a l p h a = new DenseVector ( new double [ ] { 0 . 3 , 0 . 3 , 0 . 4 } ) ; DenseContPhaseVar v1 = new DenseContPhaseVar ( alpha , A ) ; NeutsContPHGenerator gen = new NeutsContPHGenerator ( v1 ) ; double [ ] v a r i a t e s = new double [ 1 0 ] ; v a r i a t e s = gen . getRandom ( 1 0 ) ; f or ( int i = 0 ; i < 1 0 ; i ++) System . out . p r i n t l n ( var [ +i+ ] : +v a r i a t e s [ i ] ) ;

Figure 11: jPhaseGenerator: Example 1

var var var var var var var var var var [0]: [1]: [2]: [3]: [4]: [5]: [6]: [7]: [8]: [9]: 0.8280372072645692 0.20203766144304366 0.8894691713002715 2.1955965497565346 0.21362619583588516 0.5985719384190227 0.05454947718195429 0.9746203239373876 2.1989174446014657 0.4107918940269194

Figure 12: jPhaseGenerator: Result for Example 1

12

jPhaseFit: the tting module

This package contains the structure that denes the behavior of the classes that implement algorithms to t the parameters of a PH distribution. As shown in Figure 13, the interface PhaseFitter is in the top of the package and denes the basic method that any PhaseFitter should have: fit(). This method has no parameters and must return a PH variable as the result of the tting process.

Figure 13: Simple jPhaseFit Package Class Diagram

4.1

Abstract Classes

In the next level, there are two abstract classes that implement the PhaseFitter interface: ContPhaseFitter and DiscPhaseFitter, for the continuous and discrete case, respectively. These classes have two additional issues: a constructor method from a data set in array format; and a method to compute the log-likelihood of the tted distribution in relation to the data set (getLoglikelihood()). This is done because the log-likelihood is a usual way to compare the performance of tting algorithms. In addition, these classes specify the continuous or discrete nature of the variable to be tted in two dierent ways: the rst one is the inclusion of the var attribute, where the tted variable must be stored (a ContPhaseVar object for the continuous case or a DiscPhaseVar for the discrete case); the other way is the use of a data array as attribute, that in the continuous case is a double array, and in the discrete case is an integer array. In the next level of abstract classes, a further division is done between classes that implement Maximum Likelihood (ML) algorithms and those related to Moment Matching techniques. This is done for both continuous and discrete cases. For the ML classes (MLContPhaseFitter and MLDiscPhaseFitter), there is a new attribute called logLH, that stores the log-likelihood value in order to take advance of the usual computation of the log-likelihood in the tting process. For the MomentMatching related classes (MomentsContPhaseFitter and MomentsDiscPhaseFitter), a new set of attributes is dened: m1, m2, and m3. These are the moments to be matched and are specied with a new constructor that receives only the three moments to be matched. An alternative way is the use of the redened constructor that receives the data trace and calculates its moments. It must be said that there is not alternative to change the data, moments of log-likelihood attributes from outside the class, implying a safe tting process.

13

4.2

Maximum Likelihood Algorithms

The set of classes that implement maximum likelihood algorithms are almost all for Continuous PH distributions, because the most of the eorts have been done in that direction. For each one of the following algorithms, there is an associated class that executes the procedures to t the parameters of a distribution. 4.2.1 General PH Distribution EM Algorithm

In 1996 Asmussen, Nerman and Olsson [1] presented a specialized version of the EM algorithm in order to t the parameters of the whole set of continuous PH distributions, without reducing the target distribution to a restricted subset. The EM algorithm is a general statistical technique that was rst introduced by Dempster et. al. [4] to deal with the problem of incomplete data (for a review, see [7]). The idea behind this algorithm is that a complete sample from PH realizations should include the selected initial state, the whole path of states followed until absorption, and the time spent in each of these states. With this complete sample, its easy to estimate the parameters of the distribution. 4.2.2 Hyper-exponential Distribution EM Algorithm

The hyper-exponential distribution is a very special case of PH distributions, since the initial probability vector denes the probability of choosing the exponential phase to visit, and the generator matrix have diagonal representation with the rates of the i-th phase in the position (i, i). Thus the number of parameter to t a n-phase hyper-exponential distribution are 2n. The algorithm proposed by Khayari et. al. [9] is also an EM algorithm like the explained above. It begins with an initial guess of the parameters, that can be random or related to the properties of the trace (e.g. the expected value). The authors propose an easy way to select the initial parameters. Then a function to evaluate the quality of the parameters is calculated in the E-step through the probability density function of the data trace given the parameters. In the M-step, the new set of parameters is computed using estimators for the rates and the probabilities but not for the number of phases, that is taken as a given parameter. 4.2.3 Hyper-Erlang Distribution EM Algorithm

In 2005, Th ummler et. al. presented a method that ts the parameters of a hyper-Erlang distribution [21], which is a very interesting subset of the PH distributions since they are also dense in [0, ). In some results provided by them, the EM algorithm developed for this special class has a better behavior in terms of likelihood than the one designed for the complete Phase family [1]. The algorithm receives as a parameter the number of Erlang branches in the distribution as well as the total number of exponential phases in the distribution. With this information, the algorithm determines all the possible congurations of the Erlang branches and executes a version of the EM algorithm for each case. Finally, the conguration with the greatest likelihood is selected as the result of the algorithm.

4.3

Moment Matching Algorithms

The distribution moments usually play an important role in the performance analysis of real systems [18]. This has been an important motivation for the improvement of moment matching techniques, and the attention given by dierent research communities (Operations Research, Computer Science and Telecommunication Networks, among others). Some of the most recent advances have been implemented in the jPhaseFit module, as will be explained in this section.

14

4.3.1

Acyclic Continuous order-2 Distributions

In 2002 Telek and Heindl [20] proposed an algorithm to t the parameters of an acyclic PH distributions of second order (two phases). Acyclic distributions have been extensively studied since they have some important properties, as a canonic form developed by Cumani [3] and an upper triangular transition or generator matrix. In that paper, they establish bounds on the set of rst three moments representable by acyclic distributions of second order, for the discrete and continuous cases. Over the characterization of these bounds, they build the algorithm that matches three moments with the three parameters of this distribution: the rates of each phase and the absorption probability after the rst phase (the initial probability is all in the rst phase as in the Coxian distribution). 4.3.2 Erlang-Coxian Distributions

The next step in moment-matching techniques was given by Osogami and Harchol in a series of papers [17, 16, 18]. This extension consists on the characterization of the bounds imposed over the rst three moments representable by a PH distribution with n phases. They also introduce Erlang-Coxian distributions, named because they can be represented as the convolution of an Erlang and a Coxian distribution of second order. They present an algorithm to t the parameters of a Erlang-Coxian distribution with or without mass at zero, an important issue in constructing matrix-geometric models from phase type distributions. An important issue is that the algorithm itself determines the number of phases needed to represent the set of moments, making easier the use of the algorithm since the user doesnt need to try with dierent congurations. The resulting distributions are not large in the number of phases but are not strictly minimal. 4.3.3 Acyclic Continuous Distributions

One of the most recent eort done in this area was made in 2005 by Bobbio, Horvath and Telek [2], who present an algorithm to match a set of rst three moments with acyclic PH distributions (APH). They show the possible sets that can be represented by an acyclic distribution of order n. Then they show how to match the rst three moments in a minimal way, i.e. using the minimal number of phases needed to do it. It is done by determining the region representable by an APH of n phases but not with an APH with n 1. This region is then partitioned in ve areas that represent dierent distribution congurations, such as the Erlang-Exp structure that represents and n 1 Erlang distribution with an additional exponential phase after it. The algorithm proposed by the authors for the positive case is implemented in the MomentsACPHFit class. There the algorithm begins with the rst three non-central moments and computes the rst two normalized moments. With this information, the required number of phases is computed and the moment set is evaluated in order to nd in which region it falls. When it is determined, the parameters are tted according to the equations presented by the authors.

4.4

jPhaseFit Examples

In order to illustrate the use of jPhaseFit, two examples are given in this section. The rst one is shown in Figure 14, where the rst step is the loading of a data set that comes from a text le. The data set is stored in an array of doubles and it is used as the parameter to create the fitter object. This tter is of the EMHyperErlangfFit type, that implements the method proposed in [21]. After the tter initialization, a PH variable (V1 ) is created as the result of tting procedure applied over the data set, and the specication of the number of phases desired for the tted distribution (4). Finally, the parameters of the tted variable are printed, as well as the log-likelihood reached by the method in the last iteration. The result is shown in Figure 15.

15

double [ ] data = r e a d T e x t F i l e ( data /W2. t x t ) ; EMHyperErlangFit f i t t e r = new EMHyperErlangFit ( data ) ; ContPhaseVar v1 = f i t t e r . f i t ( 4 ) ; System . out . p r i n t l n ( v1 : \ n+v1 . t o S t r i n g ( ) ) ; System . out . p r i n t l n ( logLH : \ t + f i t t e r . g e t L o g L i k e l i h o o d ( ) ) ;

Figure 14: jPhaseFit: Example 1

v1 : Phase Type D i s t r i b u t i o n Number o f Phases : 4 Vector : 0.220344 Matrix : 0.152183 0.000000 0.000000 0.000000

0.491673 0.000000 0.916394 0.000000 0.000000

0.205378 0.000000 0.000000 9.177850 0.000000

0.083606 0.000000 0.000000 0.000000 233.160991

logLH :

1180.4890525003095

Figure 15: jPhaseFit: Result for Example 1

16

The next example is shown in Figure 16. It shows the creation of a fitter of the MomentsACPHFit type, which implements the algorithm proposed in [2]. This fitter is created by specifying a set of moments that the tted variable should have, in this case m1 = 2, m2 = 6 and m3 = 25. As this information is enough for the method, it is not necessary to specify any other parameter and the variable V1 is created as the tted variable coming from running the method for the moments set. Finally, the variable is printed, as well as its moments, which naturally correspond to the asked ones. This result is shown in Figure 17.
MomentsACPHFit f i t t e r = new MomentsACPHFit ( 2 , 6 , 2 5 ) ; ContPhaseVar v1 = f i t t e r . f i t ( ) ; System . out . p r i n t l n ( v1 : \ n+v1 . t o S t r i n g ( ) ) ; System . out . p r i n t l n ( m1: \ t +v1 . moment ( 1 ) ) ; System . out . p r i n t l n ( m2: \ t +v1 . moment ( 2 ) ) ; System . out . p r i n t l n ( m3: \ t +v1 . moment ( 3 ) ) ;

Figure 16: jPhaseFit: Example 2

v1 : Phase Type D i s t r i b u t i o n Number o f Phases : 4 Vector : 0.567595 Matrix : 0.737313 0.000000 0.000000 0.000000

0.432405 0.737313 2.438659 0.000000 0.000000

0.000000 0.000000 2.438659 2.438659 0.000000

0.000000 0.000000 0.000000 2.438659 2.438659

m1 : m2 : m3 :

2.000000000000001 6.00000000000035 24.999999999983892

Figure 17: jPhaseFit: Result for Example 2

17

References
[1] S. Asmussen, O. Nerman, and M. Olsson, Fitting Phase Type distributions via the EM algorithm, Scandinavian Journal of Statistics, 23 (1996), pp. 419,441. [2] A. Bobbio, A. Horvath, and M. Telek, Matching three moments with minimal acyclic Phase Type distributions, Stochastic Models, 21 (2005), pp. 303326. [3] A. Cumani, On the canonical representation of homogeneous Markov processes modeling failure-time distributions, Microelectronics and Reliability, 22 (1982), pp. 583602. [4] A. Dempster, N. Laird, and R. D.B., Maximum likelihood from incomplete data via the EM algorithm, Journal of the Royal Statistical Society. Series B, 39 (1977), pp. 138. [5] Y. Fang and I. Chlamtac, Teletrac analysis and mobility modeling for PCS networks, IEEE Transactions on Communications, 47 (1999), pp. 10621072. [6] T. Gonzalez, S. Sahni, and W. Franta, An ecient algorithm for the KolmogorovSmirnov and Lilliefors tests, ACM Transactions on Mathematical Software, 3 (1977), pp. 60 64. [7] C. Gourieroux and A. Monfort., Statistics and Econometric Models, vol. 1, Cambridge University Press, 1995, ch. 13 - Numerical Procedures, pp. 443491. [8] B. Heimsund, Matrix Toolkits for Java (MTJ). Webpage: http://rs.cipr.uib.no/mtj/, December 2005. Last checked on 05-Dec-2005. [9] R. Khayari, R. Sadre, and B. Haverkort, Fitting world-wide web request traces with the EM-algorithm, Performance Evaluation, 52 (2003), pp. 175191. [10] G. Latouche and V. Ramaswami, Introduction to matrix analytic methods in stochastic modeling, Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 1999. [11] A. Lang and J. Arthur, Parameter approximation for Phase-Type distributions, in Matrix Analytic methods in Stochastis Models, S. Chakravarty, ed., Marcel Dekker, Inc., 1996. [12] A. Law and D. Kelton, Simulation, Modeling and Analysis, McGraw-Hill Higher Education, 2000. [13] M. Neuts and M. Pagano, Generating random variates from a distribution of Phase-Type, in Proceedings of the Winter Simulation Conference, 1981. [14] M. F. Neuts, Matrix-geometrix solutions in stochastic models, The John Hopkins University Press, 1981. [15] M. F. Neuts, Two further closure properties of PH-distributions, Asia-Pacic J. Oper. Res., 9 (1992), pp. 7785. [16] T. Osogami and M. Harchol, A closed form solution for mapping general distributions to minimal PH distributions, in Proceedings of the TOOLS 2003, 2003. [17] [18] , Necessary and sucient conditions for representing general distributions by coxians, in Proceedings of the TOOLS 2003, 2003. , Closed form solutions for mapping general distributions to quasi-minimal PH distributions, Performance Evaluation, 63 (2006), pp. 524552. 18

rez and G. Rian o, jPhase: an object- oriented tool for modeling Phase-Type distri[19] J. F. Pe butions, in Proceedings of the SMCtools06, October 2006. Pisa, Italy. [20] M. Telek and A. Heindl, Matching moments for acyclic discrete and continuous PhaseType distributions of second order., I.J. of Simulation, 3 (2002), pp. 4757. mmler, P. Buchholz, and M. Telek, A novel approach for tting probability [21] A. Thu distributions to real trace data with the EM algorithm, in Proceedings of the International Conference on Dependable Systems and Networks, 2005.

19