+

Generating Random
Numbers
+
THE GENERATION OF PSEUDO-
RANDOM NUMBERS
 Agenda
 generating random number uniformly distributed

 Why they are important in simulation


+
Why important in General

 Numerical analysis,
 random numbers are used in the solution of complicated integrals.

 In computer programming,
 random numbers make a good source of data for testing the effectiveness
of computer algorithms.

 Cryptography
 Random numbers also play an important role in cryptography.
+
Why it is important in Simulation
 random numbers are used in order to introduce randomness in the model.
 In Machine interference Simulation model we assumed constant for most of
TIMES (repair time, operational time) that is not the case in general
 If we are able to observe the operational/repair times of a machine over a
reasonably long period, we will find that they are typically characterized by a
theoretical or an empirical probability distribution.
 THEREFORE we should randomly fallow a theoretical or an empirical
probability distribution
 Hence we need uniformly distributed random numbers, to produce random
variates or stochastic variates
+
Recap.. Uniform distribution

 each member of the family, all intervals of the same length on the
distribution's support are equally probable. FIGURE SHOWS THE
PDF
+
Pseudo-Random Numbers
 there is no such a thing as a single random number

 we speak of a sequence of random numbers that follow a specified

theoretical or empirical distribution

 To generate
 physical phenomenon (true random)
 Unpredictable non reproducible, natural software hardware
 Thermal noise of a diode, radioactive decay
 ideal for cryptography, not good for simulation WHY?

 We need to reproduce the same random numbers

+
To Generate (cont.)

 TO mathematical algorithms
 produce numbers in a deterministic fashion.
 Start value Seed
 these numbers appear to be random since they pass a number of statistical
tests designed to test various properties of random numbers
 Therefore they are referred to as pseudo-random numbers.
 DEBUG
 on demand generation no need to store
 uniformly distributed in the space [0, 1]

 ALTERNATIVE to pseudo-random numbers???

+
Rule of thumb for accepting random
numbers
 generating random numbers must yield sequences of numbers or bits
that are:
 uniformly distributed
 statistically independent
 reproducible, and
 non-repeating for any desired length.
+
Methods

 The congruential method

 This is a very popular method and most of the available computer code for the
generation of random numbers use some variation of this method.
 Simple Fast Acceptable

 xi, a, c and m are all non-negative numbers.

 mixed congruential method.
between 0, m-1, Normalizing them [0, 1] divide by m

 if x0 = 0, a = c = 7, and m = 10 then we can obtain the following sequence of

numbers: 7, 6, 9, 0, 7, 6, 9, 0,...
+
Period
 The number of successively generated pseudo-random numbers after
which the sequence starts repeating itself is called the period.

 If the period is equal to m, then the generator is said to have a full

period

 Theorems from number theory show that the period depends on m.

The larger the value of m, the larger is the period.
+
guarantee a full period:

 In particular, the following conditions on a, c, and m guarantee a full

period:
 m and c have no common divisor.
 a=1(mod r)if r is a prime factor of m. That is, if r is a prime
number(divisible only by itself and 1) that divides m, then it divides a-1.
 a = 1 (mod 4) if m is a multiple of 4.

 Example Set of
 A set of such values is: a = 314, 159, 269, c = 453, 806, 245, and m = 2^32
(for a 32 bit machine).
+
A question ??????

 pseudo-random number generator involves a multiplication, an

addition and a division.

 How can we avoid division??

 by setting m equal to the size of the computer word

 What if an overflow occurs

 We lose only the significant bit which is multiple of m remaining is the
remainder
 To demonstrate m=100 max number Is 99
 If a=8, x=2 c=10 result is 26
 If x=20 result is axi+c=170 but remainder 70
+
General congruential methods

 General congruential

 f(.) is a function of previously generated pseudo-random numbers. A

special case of the above general congruential method is the
quadratic congruential generator.

 Special Case a1=a2=1 c=0 m is power of 2

+
Composite generators

 We can develop composite generators by combining two separate

generators (usually congruential generators). By combining separate
generators, one hopes to achieve better statistical behavior than either
individual generator.

 One of the good examples for this type of generator uses the second
congruential generator to shuffle the output of the first congruential
generator
+
Tausworthe generators

 Tausworthe generators are additive congruential generators obtained

when the modulus m is equal to 2 to produce bıt streams

 ai are binary

 They are too slow since they only produce bits

 trinomial-based Tausworthe generator

+
The lagged Fibonacci generators

 based on the well-known Fibonacci sequence,

 General form of LFG

 where 0<j<k, and appropriate initial conditions have been made.

 combining two previously calculated elements that lag behind the current element
utilizing an algebraic operation O.

 addition, or a subtraction, or a multiplication as well as it can be a binary operation

XOR.

 additive LFG (ALFG). multiplicative LFG (MLFG).

+
Mersanne Twister
+
Statistical tests for pseudo-random
number generators
 Not all random-number generators available through programming
languages and other software packages have a good statistical
behavior.

RANDOMNESS OF SEQUENCE OF BITS

 frequency test,
 serial test,
 autocorrelation test,

GOODNESS OF FIT
 runs test, and
 chi-square test for goodness of fit.
+
Frequency test (Monobit test)

 the most fundamental test

 In a true random sequence, the number of 1’s and 0’s should be about
the same

 This test checks whether this is correct or not

 complementary error function (erfc) MATLAB , JAVA

+
HOW TO

 Generate m pseudo-random numbers and concatenate them into a

string of bits. Let the length of this string be n. The 0’s are first
converted into –1’s and then the obtained set of -1 and 1 values are all
added up. Let

 Compute the test statistic

 Compute the P-value Pvalue<0.01 reject

+
Example

 let us consider the bit string: 1011010101.

 Then, Sn = 1-1 +1+1- 1+1-1+1-1+1 = 2 and

 Sobs=0.6324.

 For a level of significance α = 0.01,

 we obtain that P- value = erfc(0.6324) = 0.5271 > 0.01.

 Thus, the sequence is accepted as being random.

+
Serial Test

 There are 2k different ways of combining k bits. Each of these

combinations has the same chance of occurring, if the sequence of the
k bits is random. The serial test determines whether the number of
times each of these combinations occurs is uniformly distributed.

 k=1 it is frequency test

 e =sequence of n bits created by a pseudo-random number

generator. (recommended min n 100)

 The serial statistical test checks the randomness of overlapping

blocks of k, k-1, k-2 bits found in e
+

 A statistical testing for pseudorandom number generators for

cryptographic applications”, NIST special publication 800-22.
+
Serial Test
 Step1– obtain e1,e2,e3
 Augment e by appending its first k-1 bits to the end of the bit string e, and
let e1 be the resulting bit string.
 Likewise, augment e by appending its first k-2 bits to the end of the bit
string e, and let e2 be the resulting bit string.
 Finally, augment e by appending its first k-2 bits to the end of the bit string
e, and let e3 be the

 Step 2 – compute frequency

 For sequences e1, e2, e3, in overlap k, k-1, k. frequencies
 fi1,fi2 fi3 represent the frequency of occurrence of the ith k, k-1, k-2
overlapping bit combination

 Step 3 compute statistics

 +
Serial Test
 Step 4 –Compute difference statistics

 Step 5 Compute P values

 Step 6
 If P-value1 or P-value2 < 0.01 then we reject the null hypothesis that the
sequence is random, else we accept the null hypothesis that the sequence is
random. (0.01 is significance level)
As in the frequency test, if the overlapping bit combinations are not uniformly
distributed, ΔS 2 and Δ2 S 2 become large that causes the P-values to be small.
+
Example Serial Test
 e = 0011011101, with n=10, and k=3
        Frequency  
C1 0 0 0 f1 0
 Step 1 C2 0 0 1 f2 1
 Append the k-1=2 bits e1 = 001101110100. C3 0 1 0 f3 1
C4 0 1 1 f4 2
 appending the first k-2=1 e2=00110111010 C5 1 0 0 f5 1
C6 1 0 1 f6 2
 k-3=0 e3 = e = 0011011101. C7 1 1 0 f7 2
C8 1 1 1 f8 0
 Step 2 frequency calculation
 For the 3-bit blocks we use e1
 For the 2- bit blocks we use e2     Frequency  
 For the 1-bit blocks we use e=e3 C1 0 0 f1 1
  Frequency   C2 0 1 f2 3
C1 0 f1 4
C2 1 f2 6 C3 1 0 f3 3
C4 1 1 f4 3
+

 Step3 Compute difference statistics

 Step 4 calculate P values

 Since both P-value1 and P-value2 are greater than 0.01, this sequence
passes the serial test.
+
Auto Corelations

 Let e denote a sequence of n bits created by a pseudo-random number

generator. The intuition behind the autocorrelation test is that if the
sequence of bits in e is random, then it will be different from another
bit string obtained by shifting the bits of e by d positions.
+
Runs Test

 The runs test can be used to test the assumption that the pseudo-
random numbers are independent of each other. We start with a
sequence of pseudo-random numbers in [0,1]. We then look for
unbroken subsequences of numbers, where the numbers within each
subsequence are monotonically increasing. Such a subsequence is
called a run up, and it may be as long as one number.
+
Chi-square test for goodness of fit

 This test checks whether a sequence of pseudo-random numbers in

[0,1] are uniformity distributed. The chi-square test, in general, can
be used to check whether an empirical distribution follows a specific
theoretical distribution. In our case, we are concerned about testing
whether the numbers produced by a generator are uniformly
distributed.
+
+
BIBLIOGRAPHY

 http://en.wikipedia.org/wiki/Uniform_distribution_(continuous)

 COMPUTER SIMULATION TECHNIQES HARRY PERROS