The Python numpy.random package

Elizabeth Daly

HDip Data Analytics 2019 Programming for Data Analysis Assessment

Git-hub repository at: https://github.com/elizabethdaly/numpy-random.git

Jupyter notebook: numpy-random.ipynb

Table of contents

Introduction
The purpose of the numpy.random package
Simple random data and Permutations
1. Simple random data functions
2. Permutations functions
Examples of distributions functions
The use of seeds in generating pseudorandom numbers
Conclusion
References

1. Introduction

This README describes work done on the numpy.random package for the Programming for Data Analysis module assessment, due 11 November 2019. Resources used include Python and associated packages Jupyter, matplotlib, and NumPy.
The analysis takes the form of a single Jupyter notebook of filename given above. To view this file, download it from this repository and start Jupyter notebook from the folder containing the file. Jupyter notebook comes as part of the Anaconda distribution of Python (as do the other packages listed above).
Alternatively, view a static version of the notebook (by providing its GitHub url) using Jupyter Nbviewer.
All images intended for inclusion in this README are located in the images subdirectory of this repository.
I have tried to structure the Jupyter notebook and this README so that they have corresponding sections. However, I do not wish to merely repeat here what has been stated in the notebook. Therefore, I will endeavour to have this README complement the analyses performed in the notebook.

https://commons.wikimedia.org/w/index.php?curid=4213838 Image credit: Antoine Taveneaux - Own work, CC BY-SA 3.0 This is an example of a bean machine, a device invented by Sir Francis Galton (in the days before numpy.random) to demonstrate that, for large sample sizes, the binomial distribution approximates a normal distribution.

2. The purpose of the numpy.random package

There is a Python module called random which can be used to generate pseudorandom numbers, so why the need for numpy.random? When researching the difference between random and numpy.random I found some useful stackoverflow posts. It seems that Python's built-in random module generates single numbers, while numpy.random can easily generate large arrays (or just single numbers) containing sample values from lots of different kinds of probability distribution. It is based on the NumPy ndarray, which allows for fast mathematical operations on arrays without the need to write loops. Arrays are essential for representing matrices, which are in turn very important in modelling real world scenarios. For example, in the past, I have performed the calculations for control systems and done image processing using matrix algebra. numpy.random also contains functions to perform random shuffling of the contents of arrays and others which provide some information on, and control of, the seed used to generate the random numbers.

In probability theory and statistics, the probability distribution of a variable describes the probability of occurrence of each possible value of the variable over a number of trials. The mathematical function describing the probability that a particular value will occur is called the probability density function or PDF. It describes the likelihood of obtaining the possible values that the random variable can take on.

For example, imagine an experiment to measure the temperature in Galway every day over the summer. We know that the values will fall within a fairly consistent range, and the most commonly measured value will be the average or mean. The variable (temperature) can theoretically take on any value in a continuous range of values, but certain values are more likely to occur than others, so there will be some spread in the measured values around the mean. At the end of the summer a histogram of temperature values would result in a bell-shaped curve. This experiment could be simulated with a probability distribution of a continuous variable such as the normal distribution, of given mean and standard deviation.

There are other experiments where the output is discrete rather than continuous. The simplest example of such an experiment is to imagine flipping a coin a number of times and attempting to predict how often each of the possible outcomes, heads (H) or tails (T), will occur. Here, the possible outcomes are discrete (H or T) and each is as likely to occur as the other with a fair coin. The coin toss experiment could be simulated using a discrete PDF such as the binomial distribution. The numpy.random package provides the ability to simulate any random process as it can be used to generate sample values from many types of probability distribution.

The cumulative probability distribution (CDF) of a random variable X, with a given PDF, is the probability that X will have values less than or equal to x. The CDF is obtained by integrating the PDF up to x; it is the area under the PDF curve up to x. So, if we know the distribution of some random variable X, we can make predictions about the probability that X will lie inside a specified range using its CDF. I demonstrate its use in the notebook for some of the distributions discussed.

3. Simple random data and Permutations

3.i Simple random data functions

The simple random data functions in the numpy.random package allow for the generation of arrays of random numbers of different types (floats, integers), within specified ranges, or from a provided array. The easiest way to explain the range of functionality provided is to illustrate with a few functions. Some functions on the documentation page appear to have identical explanations, but it turns out that some are deprecated (allowed but discouraged) and others are aliases to a single function.

3.ii Permutations functions

The permutations functions provide for random shuffling of data.

4. Examples of distributions functions

I will discuss five probability distributions here and illustrate their use in the Jupyter notebook.

4.i The normal distribution

This function is used to generate random samples from a normal (Gaussian) distribution. It is a symmetric probability density function with a characteristic bell shaped curve. It is characterised primarily by its mean and variance. Normal distributions occur often in nature. The numpy.random documentation states that it describes the distribution of samples influenced by a large number of tiny, random disturbances, each with its own unique distribution. Uses include:

IQ scores in a population are normally distributed.
In biology the log of variables tend to be normally distributed, for example heights, weights, blood pressure of adult humans, lengths of hair, nails, and teeth.
errors in physical experiments can be simulated with the normal distribution.
bell-curve grading can assign grades assuming a normal distribution.

Image: https://www.compensationcafe.com

4.ii The binomial distribution

A binomial distribution can be thought of as simply the probability of a success or failure outcome in an experiment that is repeated multiple times. The binomial is a type of distribution that has two possible outcomes (the prefix “bi” means two, or twice), so the random variable is discrete. This distribution can be used to model a fixed number of trials n, each with a constant probability of success p. Our example above of flipping a fair coin (50% chance of getting a head) n times can be simulated using a binomial distribution. It can be used to simulate any experiment where the possible outcomes are success or failure such as:

The chance of obtaining a certain number of heads on flipping a coin a number of times.
The chance of winning (or not) a lottery.
The success or failure of a drug in a clinical trial.
The chances of finding oil (or not) when digging a well (example in numpy.random documentation).
The probability that a call centre employee will make a sale.

The Poisson distribution (also available in numpy.random) is a special case of the binomial when the number of trials becomes very large and the probability of success in each one is very small.

Image: https://www.spss-tutorials.com

4.iii The exponential distribution

The exponential distribution is a continuous probability distribution used to model the time we must wait before some given event occurs. One interesting thing I learned is that exponential distributions are the only continuous distributions that are memoryless. This means that the probability distribution is independent of its history: any time may be marked down as time zero, so the likelihood of something happening in the future has no relation to whether or not it has happened in the past. This is also true of coin flips, but they follow the discrete binomial distribution. The lumen website in the references below has lots of examples of real world situations which can be simulated with the exponential distribution. These include:

The amount of time until an earthquake occurs, beginning now.
The amount of time, in months, a car battery lasts.
The value of change in your pocket over time.
The time between page requests on Wikipedia (example in numpy.random documentation)
The size of raindrops (which grow with time) in rainstorms (example in numpy.random documentation)
The number of days ahead of travel that people purchase flights.
Failure of computer parts over time.

Image: https://www.individual.utoronto.ca

4.iv The Rayleigh distribution

The Rayleigh distribution is a non-symmetric continuous probability distribution for positive random variables. It is named after Lord Rayleigh, a British physicist who provided the first theoretical explanation for why the sky is blue, among other things. A Rayleigh distribution is often observed when the magnitude of a vector is related to its directional components. For example, when wind speed is analysed in two directions (say east and west components), the overall wind speed will follow a Rayleigh distribution. Each component must be normally distributed, with zero mean and equal variances, and there must be no linear relationship between the two constituent components. Knowledge of the wind speed distribution is very important when selecting sites for wind turbines for example. The mean and variance of this distribution are calculated from its scale (see the Jupyter notebook). The Rayleigh distribution can be used:

to model the wind distribution for land-based wind turbine sites (the most common real-world application I found).
to account for the distribution of background noise in magnetic resonance magnitude images.
to describe the distribution of wave heights under certain conditions (Science Direct reference).
in communications to model the paths taken by a signal traveling through a dense scatterer towards a receiver.

It is a special case of the Weibul distribution, which is also available with numpy.random.

Image: https://www.slideplayer.com

4.v The logistic distribution

The logistic distribution is a symmetric continuous probability distribution. It is very similar to the normal distribution but with higher tails (so that it falls away to zero a little more slowly). For this reason it is useful for predicting the likelihood of extreme events and for taking into account the underlying or base data. It is also used because its CDF (the logistic function) has a simple formula which approximates the normal distribution very well. It has applications in:

the modelling of population growth,
logistic regression for classification purposes,
world chess rankings where the performance of each player is a logistically distributed random variable,
medicine to model the growth of tumours.

Image: https://www.vosesoftware.com

5. The use of seeds in generating pseudorandom numbers

The Real Python reference below asks the question: How random is random? Pseudorandom numbers (as opposed to truly random ones) are generated in the computer via algorithms called pseudorandom number generators. However, the behaviour of a generator can be predicted if the seed used to initialize it is known. Obviously, that's not a good situation for security-sensitive applications such as cryptography. So it's important to know a little bit about how these numbers are generated in the computer.

Image: https://www.sharpsightlabs.com

Python uses the Mersenne Twister as the core generator. It was developed in 1997 and is now the most widely-used pseudorandom number generator. The period length of the generator is chosen to be a Mersenne Prime, hence its name. A Mersenne prime is a prime number that is one less than a power of two; it has the form 2ⁿ-1 where n is an integer. The algorithm itself generates numbers in the range [0, 2^w-1] where w is the word size in bits. The state of the pseudorandom number generator must be initialized or seeded by providing a w-bit seed value. The last group of functions in the numpy.random package is concerned with this seed. The Jupyter notebook associated with this README contains some examples of their use.

As both Python random and numpy.random are completely deterministic if some information about the seed of the generator is known, neither are suitable for security or cryptographic applications. There is a module called secrets which seems to be recommended for those types of application.

6. Conclusions

The purpose of the numpy.random package is to make the generation of large arrays of random data, containing sample values from lots of different kinds of probability distribution, very efficient. The ability to generate samples of random data is important when simulating experiments where it is just not possible to measure a whole population. The package could also be used to simulate an experiment (that might be practically difficult to perform) given certain experimental parameters. So, simulations can help to design better experiments. The accompanying Jupyter notebook demonstrates how some of the functions in the numpy.random package can be used. Random numbers are also very important for machine learning in data analytics, as discussed in the machine learning mastery reference below. That reference lists tasks such as:

random initialization of weights in an artificial neural network,
random splitting of data into train/test sets,
random shuffling of the training data set in stochastic gradient descent,

as being important ones for which a good understanding of randomness is essential.

7. References

General:

[1] Anaconda Distribution https://www.anaconda.com/
[2] Python Software Foundation https://www.python.org/
[3] Project Jupyter https://jupyter.org/
[4] Sharing Jupyter notebooks https://nbviewer.jupyter.org/
[5] matplotlib: Python plotting library htts://matplotlib.org/
[6] NumPy https://numpy.org/
[7] GitHub https://github.com/
[8] Python for Data Analysis: Data Wrangling with Pandas, NumPy, and IPython. Wes McKinney. ISBN-13: 978-1491957660 ISBN-10: 1491957662
[9] Random sampling (numpy.random) https://docs.scipy.org/doc/numpy-1.16.0/reference/routines.random.html

Related to random number generation in Python:

[10] Python random module https://docs.python.org/3/library/random.html
[11] Python Random Number Generation tutorial https://pynative.com/python-random-module/
[12] Generating Random Data in Python (Guide) https://realpython.com/python-random/
[13] Real Python: Generating Random Data in Python (Guide) https://realpython.com/python-random/
[14] Differences between numpy.random and random.random in Python https://stackoverflow.com/questions/7029993/differences-between-numpy-random-and-random-random-in-python
[15] Performance difference between numpy.random and random.random in Python https://stackoverflow.com/questions/57220804/performance-difference-between-numpy-random-and-random-random-in-python
[16] Machine Learning Mastery: How to Generate Random Numbers in Python https://machinelearningmastery.com/how-to-generate-random-numbers-in-python/