Have you celebrated Baegil yet?

September 25th, 2016

A Korean friend asked my wife if our son had celebrated Baegil or 100 days yet. Not being familiar with the celebration, we said that we hadn’t kept track of the days since his birth. I decided to write a little program in Go in order to check what day of one’s life one is currently on.

The program should take the subject’s date of birth as a string from the first command argument. This was simple enough for me to decide against using the flags package and check the command line arguments directly:

if len(os.Args) == 1 {
    os.Stderr.WriteString("Please tell me your date of birth!\n")
} else {
...
}

I wanted to save programming effort and avoid reinventing the square wheel by using a library for handling the dates. The time package from the standard library can represent times and parse strings to time structs.

The time package has a peculiar (but easy to grok) way of specifying date layouts for formatting. Rather than %Y, %h, %s and so on from the standard C library and its descendants, in Go a fixed date where days are 1 and months are 2 and so on is used. So, to parse dates such as “2016-06-20” or “1980-10-28”, we need:

var layout = "2006-01-02"
var dateOfBirth, dOBParseError = time.Parse(layout, os.Args[1])

Full details can be found on this blog post:

https://pauladamsmith.com/blog/2011/05/go_time.html

Of course, a user of the program might not enter a parseable date string, so we need a little bit more error handling:

if dOBParseError != nil {
    fmt.Fprintf(os.Stderr, "Unable to parse your date of birth: %s\n", dOBParseError)
} else {
...
}

Once we have a date of birth, we can perform our arithmetic. The time package has a Since(...) method that can give you the duration of time since a given time struct:

var age = time.Since(dateOfBirth)

For a simple program like this, methods like Since are more convenient than creating a date object of the current time and comparing the given time to it. However, this does create an issue when testing code that does anything to do with time. If your code makes direct use of the system’s clock, your functions are impure as the same input won’t always return the same output. How do you test that a given input produces the correct output? Such tests would only pass at specific times on a given date. The purely functional way to do this is to write your business logic handling functions to accept all the used time objects already created, so that appropriate mocks can be passed in during tests without calling any methods that get data from the system’s clock. Another approach is the override the methods that call the system’s clock so that they return a set time in the test. This is called monkey patching. I’ve heard of people even setting the system clock during the test, but this is likely to have unforeseen consequences, especially on a shared build server.

The time package defines an Hours function for Duration structs but not one for days. A quick and dirty solution to this is the divide the number of hours by 24 and floor the result:

var ageInDays = math.Floor(age.Hours() / 24)

This makes the assumption that every day has exactly 24 hours. What about leap seconds? A person might have a few leap seconds during a lifetime. What if this program is run within a few seconds of midnight? The result might be different from the true result. As we are not using the precise time of birth and are flooring the number of days anyway, this is not too much of a concern. However, if really precise maths is needed or you using the components of one time struct to create a new time struct, inaccuracies can creep into your code. The fun really begins when you need to handle issues like leap years, time zones and seasonal time adjustments. If I ran this code in San Francisco in Winter for someone who was born in Sydney during Daylight Saving Time, would the result be accurate? A more fully developed version of this program would probably use a library to handle the mind boggling complexity of time tracking in the real world.

Another issue to consider is off by one errors. I’m flooring the number of days in order to say, “days since”. If I enter today’s date, I will get 0 as the subject’s age has not reached one day yet. This might be confusing to someone expecting 1 on the first day and so on. How ages are reckoned changes among different cultures. How Koreans consider their ages is surprising to many foreigners.

https://en.wikipedia.org/wiki/East_Asian_age_reckoning#Korean

Koreans celebrate Baegil on the 100th day, so we would expect the program to return 99 on the appropriate day. A programmatic solution to this might be add 1 to the result and change the wording. Programmers need to consider the audience of their programs carefully and gather requirements fully.

Obviously, you can’t be born in the future and still use the program, so we need a bit more user input checking:

if ageInDays < 0 {
    os.Stderr.WriteString("Wow! I must be talking to a foetus!\n")
} else {
    ...
}

If we’ve made it this far, we can now print the age in days:

fmt.Printf("Days old: %.0f\n", ageInDays)

Putting it all together, this is my program:

package main
 
import (
    "fmt"
    "math"
    "os"
    "time"
)
 
func main() {
    if len(os.Args) == 1 {
        os.Stderr.WriteString("Please tell me your date of birth!\n")
    } else {
        var layout = "2006-01-02"
        var dateOfBirth, dOBParseError = time.Parse(layout, os.Args[1])
 
        if dOBParseError != nil {
            fmt.Fprintf(os.Stderr, "Unable to parse your date of birth: %s\n", dOBParseError)
        } else {
            var age = time.Since(dateOfBirth)
            var ageInDays = math.Floor(age.Hours() / 24)
 
            if ageInDays < 0 {
                os.Stderr.WriteString("Wow! I must be talking to a foetus!\n")
            } else {
                fmt.Printf("Days old: %.0f\n", ageInDays)
            }
        }
    }
}

Entering my son’s date of birth, I see that we still have a couple of days to go:

$ ./daysOld 2016-06-20
Days old: 97

The source code can be found at:

https://github.com/robert-impey/CodingExperiments/blob/master/Go/daysOld.go

Coin Flips with Go

September 15th, 2016

Following on from my little experiment with flipping coins millions of times, I thought that it would be interesting to write the same program in Go for comparison.

func main() {
    rand.Seed(time.Now().UTC().UnixNano())
 
    var competitionSize, runs int
    flag.IntVar(&competitionSize, "competitionSize", 10, "The size of each coin tossing competition.")
    flag.IntVar(&runs, "runs", 10, "The number of runs of the competition")
 
    flag.Parse()
 
    fmt.Println("heads, tails, flips")
    for i := 0; i < runs; i++ {
        countHeads := 0
        for j := 0; j < competitionSize; j++ {
            countHeads += rand.Intn(2)
        }
 
        fmt.Printf("%d, %d, %d\n", countHeads, competitionSize-countHeads, competitionSize)
    }
}

The basic idea of how to write the program is unchanged in this language. The output showed a similar distribution to the C# program.

However, Go has straightforward command line arguments parsing built into the standard library (like Python or Perl). In C#, there are a number of third party libraries that do this that can be installed via NuGet. I went with Command Line Parser.

To use this, I needed to add a class with properties with attributes to describe my expected options:

class Options
{
    // Taken from http://www.bbc.co.uk/news/politics/eu_referendum/results
    [Option('s', "competitionSize", DefaultValue = 33551983, HelpText = "The size of each coin tossing competition.")]
    public int CompetitionSize { get; set; }
 
    [Option('r', "runs", DefaultValue = 1000, HelpText = "The number of runs of the competition")]
    public int Runs { get; set; }
}

The command line arguments are then parsed in the main method:

var commandLineArgs = new Options();
if (Parser.Default.ParseArguments(args, commandLineArgs))
{
    var runs = commandLineArgs.Runs;
    var flips = commandLineArgs.CompetitionSize;
    ...
}
else
{
    WriteLine("Unable to parse the command line arguments!");
}

While this is not difficult to understand, it is more complicated than it needs to be and requires quite a bit more typing than in the Go program.

10 runs of the referendum simulation in both programs on Windows 10 with an Intel i7 4500U running at 1.8GHz took 7 seconds for the C# program and 13 seconds for the Go program. I wouldn’t take such simple programs very seriously for drawing conclusions about the relative speed of these languages.

The updated C# program and the complete Go program are on GitHub:
https://github.com/robert-impey/CodingExperiments/blob/master/C-Sharp/CoinFlipsAreRandom/CoinFlipsAreRandom/Program.cs
https://github.com/robert-impey/CodingExperiments/blob/master/Go/coinToss.go

Was the EU Referendum Random?

September 5th, 2016

One of the claims that I saw on social media in the aftermath of the recent EU referendum here in the UK was that the result (52% to 48%) was so close that it was little different from tossing a coin.

Without getting bogged down in the politics of that referendum, or the various campaigns that led up to it; I want to consider whether this claim holds any water. How similar to millions people each tossing a coin and voting accordingly was the result?

According to the BBC, there were 17,410,742 votes to leave and 16,141,241 votes to remain, giving a total of 33,551,983 votes. If we were to make each of these people toss a coin and count up the results, the ratio of heads/tails or remains/leaves could be anywhere between all heads and all tails. However, we would expect the counts to be about equal if the coins were all fair. Of course, any ratio is possible, but if we were to run the coin tossing game repeatedly, we would expect to mean ratio to converge on 1:1. How likely would a 52:48 ratio be?

The leave side got a share of the vote equal to 0.51891842. Therefore, their absolute deviation from the expectation (the mean, or 0.5) is 0.01891842. The difference between the two counts is 1,269,501. Would we expect to see a deviation of this magnitude in a coin tossing competition?

Being a computer programmer rather than a mathematician, I’m going to look at this using a simple program.

WriteLine("heads, tails, flips, heads share");
 
var runs = 1000;
 
// Taken from http://www.bbc.co.uk/news/politics/eu_referendum/results
var flips = 33551983;
 
var randomNumberGenerator = new Random();
 
for (var run = 0; run < runs; run++)
{
    var heads = 0;
 
    for (var coinToss = 0; coinToss < flips; coinToss++)
    {
        var toss = randomNumberGenerator.Next(2);
 
        if (toss > 0)
        {
            heads++;
        }
    }
 
    WriteLine($"{heads}, {flips - heads}, {flips}, {(0.0 + heads) / flips}");
}

This program simulates 1,000 coin tossing competitions with 33,551,983 players and writes the counts as comma separated values.

Putting the output into Excel, the largest deviation was 0.000275081 (or a difference between counts of 18,459) and the smallest was 1.49022E-08 (which was 16,775,992 heads and 16,775,991 tails. This happened twice in 1,000 runs!) The largest difference between heads and tails was 69 times smaller than the result from the referendum.

Plotting the shares in decreasing order, we see how quickly larger deviations fall off:

deviation

Putting the deviations into bins and counting the competitions by deviation from the expectation, we see that smaller deviations are more common:

bins

Whatever else we might say about the result, we cannot seriously claim that the result was random.

The full program can be found here:

https://github.com/robert-impey/CodingExperiments/blob/master/C-Sharp/CoinFlipsAreRandom/CoinFlipsAreRandom/Program.cs

The Excel file can be found here:
http://www.reversing-entropy.com/wp-content/uploads/2016/09/EuRef.xlsx

Does Declarative Programming generally help?

April 17th, 2016

A recent post compared different ways of approaching the same problem in Scala and Go:

https://www.quora.com/Scala-vs-Go-Could-people-help-compare-contrast-these-on-relative-merits-demerits/answer/Nick-Snyder-1?srid=dsGW&share=191eaf13

The Go approach is longer than the one in Scala, but every line is dead simple to understand. We might dismiss it as repetitive, boiler plate code rather than abstract, business logic. However, setting break points for debugging or changing the behaviour of the code when the new requirements arrive, six months down the line, is trivial. This can be a mixed blessing, as there’s really nothing to stop code like that ballooning in size as new conditional statements get added to cope with the changing needs of the software. Every 5 kLoC ball of mud that any programmer ever wrestled with started out as something dead simple like that.

The following article investigates how non-programmers approach programming problems and whether programming languages make the task harder because they do not match how an untrained mind thinks.

http://alumni.cs.ucr.edu/~ratana/PaneRatanamahatanaMyers00.pdf

More children developed rules to describe the behaviour of Pacman than instructions for the sprites to follow. People apparently start out thinking in a more declarative style and then develop the habit of thinking in the imperative style.

Is this an argument for using more declarative languages? If declarative programming is more natural, surely it would be easier. However, many programmers start out with imperative code and perhaps only try declarative code later. Is this just a historical or social accident caused by the greater availability of imperative languages compared to declarative ones? It is possible that this is a greater issue for novice programmers than for experienced ones.

Often, we cannot move between languages at whim. Few programmers have the privilege of being able to write part of their system in Go in the morning, then head over to Scala land after lunch. However, .Net has Linq, which affords a C# developer, like your author, greater freedom to move freely back and forth between imperative and declarative styles of programming.

Imagine a book stall in a market. There is a table with books with different numbers of pages with covers that are red or blue.

The types in this world are an enum for the colours and a class for the book:

enum Colour
{
    Red,
    Blue
}
 
class Book
{
    internal Book(string name, int numberOfPages, Colour colour)
    {
        Name = name;
        NumberOfPages = numberOfPages;
        Colour = colour;
    }
 
    internal string Name { get; }
    internal int NumberOfPages { get; }
    internal Colour Colour { get; }
}

We can model our bookstall with a list:

var books = new List<Book>
{
    new Book ("War and Peace", 1123, Colour.Red),
    new Book ("Doctor Zhivago", 803, Colour.Blue),
    new Book ("Jeeves and the Feudal Spirit", 154, Colour.Red),
    new Book ("Shogun", 1213, Colour.Red)
};

You want to write down the names of the books that are more than 500 pages long and have red covers. Here are three ways to solve this problem.

The imperative approach:

foreach (var book in books)
{
    if (book.Colour == Colour.Red && book.NumberOfPages > 500)
    {
        WriteLine(book.Name);
    }
}

This way, we are putting ourselves in the role of the person actually sifting through the books one by one and writing the names of the books that meet our criteria as we go. However, the presentation code (the WriteLine) is mixed up with the filtering logic in the loop.

Linq makes use of the Where extension method to allow the so-called fluent syntax:

var longRedBooks = books.Where(b => b.Colour == Colour.Red && b.NumberOfPages > 500);
 
foreach (var longRedBook in longRedBooks)
{
    WriteLine(longRedBook.Name);
}

Here, we first ask for the books that match our criteria, defined in the lambda expression passed to the Where method. Then, we print the books that matched. The filtering logic and presentation logic are now separate.

Using Linq’s query syntax below, the code looks quite different from the fluent syntax used above, but the meaning is the same. This approach is perhaps more familiar to developers with a background in SQL.

var longRedBooks = from book in books
                   where book.Colour == Colour.Red && book.NumberOfPages > 500
                   select book;
 
foreach (var longRedBook in longRedBooks)
{
    WriteLine(longRedBook.Name);
}

The complete source can be found on GitHub:
https://github.com/robert-impey/CodingExperiments/blob/master/C-Sharp/BookFinding/Program.cs

Which is easiest to understand? If the requirements change, let’s say that we now only want one book, which should be the longest book, shorter than 500 pages, with a blue cover, which approach will be easiest to update and understand with the new logic? With the Linq approach, we can take advantage of the type system to see that the query now returns a Book (which may be null) rather than an IEnumerable<Book> and handle the object appropriately. Will the change in meaning be as obvious in the imperative code?

For what it’s worth, I’ve experienced plenty of resistance in code reviews to Linq queries, but almost none to writing the same thing in a for loop. Sometimes the reasoning is to do with performance, but I doubt there’s much in that. Anyway, a read-only code review is a terrible place to make assumptions about relative performance. However, it’s very easy to reason about the performance of the for loop, but more advanced knowledge of lazy evaluation, the iterator pattern and its implementation is needed to understand what gets executed where and how many times for the linq approaches.

I would be very interested to see an analysis of open source projects that compared the frequency on these approaches to see what is the approach that people tend to settle on. Do pull request with Linq get accepted more or less often? Is code generally refactored from one to the other? In which direction?

How should this influence job interviews and hiring? If we want to hire experienced programmers, we could check for signs that their thinking has been moulded to the structures of industrial programming languages. Perhaps we should pity such creatures. Should we look for signs of skills in imperative programming in hands-on programmers and for a declarative approach in tech leads and more senior roles, where coordinating the efforts of several developers is the required skill?

Comprised of

February 4th, 2015

I have just read an article about an extremely dedicated and efficient WikiGnome who has edited tens of thousands Wikipedia articles to change a single grammatical error – “is comprised of”.

https://medium.com/backchannel/meet-the-ultimate-wikignome-10508842caad

My initial reaction was that the phrase was perfectly acceptable and the journalist’s dread of having published work that included the phrase was overblown.

I checked a dictionary, which list examples of the phrase:

http://dictionary.cambridge.org/dictionary/british/comprise

It is sufficiently common that it doesn’t jar.

However, the Wikipedian’s own explanation of why he opposes the phrase is logically sound:

https://en.wikipedia.org/wiki/User:Giraffedata/comprised_of

As is this note from the dictionary:

http://dictionary.cambridge.org/grammar/british-grammar/consist-comprise-or-compose

These sentences are standard English:

“Air consists of nitrogen, oxygen and other gases.”

“Air is composed of nitrogen, oxygen and other gases.”

“Air comprises nitrogen, oxygen and other gases.”

Sentences of this sort are commonly accepted but, I now accept, illogical:

“Air is comprised of nitrogen, oxygen and other gases.”

The phrase “is comprised of” merges the correct uses of consist and compose. It attempts to mean the same thing in the passive and active voice. It’s similar to “irregardless” being used to mean “regardless”: an opposite with the same meaning.

Whether we should use “is comprised of” is a contentious question, and many people feel that it’s fine. However, why do we instantly reject this sentence?

“Air is consisted of nitrogen, oxygen and other gases.”

This sentence is logically equivalent to one with “is comprised of”. Searching for the phrase “is consisted of” in the English Wikipedia turns up only 69 instances, mainly in user talk pages rather than the actual articles:

https://www.google.co.uk/#q=site:en.wikipedia.org+%22is+consisted+of%22

It seems likely that our acceptance of sentences as grammatical or not is based more on recognition and familiarity than on a strict logical evaluation of the words.

Finding Reversed Words

November 23rd, 2014

Following on from coding experiments with using the decorator pattern to reverse strings in Java and with palindromes and string reversal in C#, I decided to search for the words in the English language that are the reverse of other words. The Devil lived, but Black Sabbath gave us Live Evil. To beat Sega takes ages. You’re upping the ante living near Etna.

I had a copy of a dictionary file from Ubuntu on my hard drive. To make life a little simpler, I used some PowerShell magic to remove the many words that end in “‘s” (there are no words that start “s'”) and to make the list all lower case:

PS M:data> Get-Content .\british-english.csv | ?{ $_ -NotMatch "'s$" } | %{ $_.ToLower() } | Sort-Object | Get-Unique | Set-Content british-english-without-apostrophe-s.txt

The simplest algorithm for finding words that are the reverse of other words is to simply go through the whole list and then check whether any of the other words are the reverse of that word. To do this, we need a method to test if two strings are a reversed pair. Pushing the characters of the first string onto a stack and then popping them off as you check them against each character of the second string can achieve this:

public bool AreReversedPair(string firstString, string secondString)
{
    if (firstString.Length != secondString.Length)
    {
        return false;
    }
 
    var firstStringStack = new Stack<char>();
 
    foreach (var c in firstString)
    {
        firstStringStack.Push(c);
    }
 
    foreach (var currentCharFromSecond in secondString)
    {
        var currentCharFromFirst = firstStringStack.Pop();
 
        if (currentCharFromFirst != currentCharFromSecond)
        {
            return false;
        }
    }
 
    return true;
}

Next, the code for finding the pairs:

public IEnumerable<ReversedWordPair> FindReversedWords(IQueryable<string> allWords)
{
    var reversedWords = new LinkedList<ReversedWordPair>();
    var reversedStringChecker = new StackReversedStringChecker();
 
    foreach (var firstWord in allWords)
    {
        foreach (var secondWord in allWords)
        {
            if (reversedStringChecker.AreReversedPair(firstWord, secondWord))
            {
                reversedWords.AddLast(new ReversedWordPair(firstWord, secondWord));
            }
        }
    }
 
    return reversedWords;
}

This code is very straightforward, but not fast at all. The time complexity of the algorithm is O(n^2) because for each word, we need to check every other word. For a list with almost 73,000 words, this takes a long time. In spite of this, it does finish. I left the program running, went to dinner and found a list waiting for me when I came back.

reversed-word-pairs.txt

As the list of words is not going to change very often and I only need to run the algorithm once, I could just leave the program as it is. However, it would be nice to reduce the time complexity of the algorithm. The simplest way to do this would be to reverse each word and then check whether the list of all the words contains the reversed word (readers who are familiar with .Net may have noticed that I passed the list of words as an IQueryable rather than an IEnumerable for the list of words). Depending on the data structure, checking whether the reversed string is a member of the set of words should be less than O(n). For example, retrieval from a Trie takes O(m) where m is the length of the string. This would leave us with O(nm) time. As with all optimisations, you’ve got to time the code to see whether you get a boost or not.

My next task is to come up with a meaningful sentence with all words the reverse of the corresponding word at the other end of the sentence, with a palindrome at the centre. There must be something that can be made of Dennis, who sinned in the straw, only to get warts to stun his nuts.

See if you can make a mirror sentence of your own.

The complete code, along with some tests, can be found at:

https://github.com/robert-impey/CodingExperiments/tree/master/C-Sharp/FindReversedWords

Fizz Buzz with Functional Programming

October 11th, 2014

Fizz Buzz is a counting game where players count upwards saying either the number, “Fizz!” if the number is a multiple of 3, “Buzz!” if a multiple of 5, or “Fizz! Buzz!” if the a multiple of 3 and 5.

It’s also a simple programming problem that can be solved in a number of ways.

In C#, the solution could be as simple as:

var max = 30;
 
Console.WriteLine("With a for loop:");
 
for (var i = 1; i <= max; i++)
{
    Console.Write(i + ": ");
 
    if (i % 3 == 0)
    {
        Console.Write("Fizz! ");
    }
 
    if (i % 5 == 0)
    {
        Console.Write("Bang!");
    }
 
    Console.WriteLine();
}

href=”https://github.com/robert-impey/CodingExperiments/blob/master/C-Sharp/FizzBang/FizzBang/Program.cs

I’m going to show a way of solving this problem using F#. This is not the simplest way to solve this problem. However, the F# solution below does demonstrate the use of higher order functions, an important concept in functional programming.

F# is a strongly typed language. This means that the compiler checks that the values that we bind and pass to and return from functions conform to its expectations. This catches a huge number of programmer errors almost as soon as they are made. It also allow the programmer to state his or her intentions by defining types. In this program, we are dealing with checking numbers that may or may not have the special property of causing us to say a message. We are going to need a function that can check numbers, so we define a type that describes such functions:

type NumberChecker = int -> string option

An example of such a function is

let isEven x = 
    if x % 2 = 0 then Some "That number's even!"
    else None

This takes x and checks whether the remainder after dividing by 2 is 0. If so, it returns something (the message). Otherwise, it returns nothing.

We can test this function using a match statement. We call isEven with a number. If the function returns a message (the case where there is Some message) we can print it. If the function returned None, then we’ll just print the number.

let number = 4
match isEven number with
| Some message -> printfn "%s" message
| None -> printfn "%d" number

This code is fine if we only want to check one number with one number checker. We want to be able to check numbers with different checker functions.

For example, we might also define:

let isOdd x =
    if x % 2 = 1 then Some "That number's odd!"
    else None

To test this, we could duplicate our code:

match isOdd number with
| Some message -> printfn "%s" message
| None -> printfn "%d" number

Nobody likes duplicated effort. The only difference between these lines of code and the isEven printing code is the checker function. This is where a higher order functions come in. A higher order function takes a function as an input or returns a function as the output.

let numberCheckerPrinter (numberChecker : NumberChecker) number =
    match numberChecker number with
    | Some message -> printfn "%s" message
    | None -> printfn "%d" number

The first argument to this function is annotated as a NumberChecker, which we defined at the top of our program. We can pass into this function any number checker function we like and it will either print the message (if one is returned) or the number on its own.

numberCheckerPrinter isEven 2
numberCheckerPrinter isEven 3
numberCheckerPrinter isOdd 2
numberCheckerPrinter isOdd 3

This has this output:

That number's even!
3
2
That number's odd!

Note that the first argument (which happens to be a function) is passed to the function in the same way as the second argument (which happens to be an integer).

We could go on defining number functions in the way, but we’re already beginning to repeat ourselves. The definition of the higher order function was a function that took a function as an input or returned a function. It would nice if the programming language could some of the work for us and make number checkers. All that stays the same from number checker to number checker is that they all must accept an integer and return either a message or nothing.

We might have a number of number checkers that vary only in that they check a different divisor and have a different message.

let divisorNumberChecker (divisor, message) number = 
    if number % divisor = 0 then
        Some message
    else    
        None

Note that this function does not conform to the NumberChecker type that is defined at the top as it takes two arguments: a tuple of the divisor and the message and the number. We can use this function to make new number checkers by fixing the first input with partial function application.

let fizz = divisorNumberChecker (3, "Fizz!")

Even though divisorNumberChecker takes two arguments, we only supplied the first. This causes F# to return a new function which accepts the remaining argument and has fixed the first argument.

Running this in F# interactive gives the type of fizz, which matches NumberChecker.

> 

val fizz : (int -> string option)

Making the “Buzz!” checker is simply:

let buzz = divisorNumberChecker (5, "Buzz!")

We have functions for the “Fizz!” and the “Buzz!” part of the game, we now need to deal with the case when the number is divisible by 3 and 5. We could define a function to do this, but this a duplicated effort that we want to avoid. After all we have fizz and buzz, we want to reuse them.

let andNumberChecker (f : NumberChecker, g : NumberChecker) x  = 
    match f x, g x with
    | Some messageF, Some messageG -> messageF + " " + messageG |> Some
    | _ -> None

The andNumberChecker lets us combine two number checker functions into one. If both functions return a message for the given number, then it joins the messages together and returns them. Otherwise, it returns None.

Again, I’ve use partial function application to fix the fizz and buzz functions to make fizzBuzz.

let fizzBuzz = andNumberChecker (fizz, buzz)

So far, I’ve been passing single functions as arguments to other functions. Functions can be treated as other objects in F#. Therefore, we can put them into a list and pass that list to a function. This way we can keep checking the same number with different checkers and use the message from the first one that returns something rather than nothing:

let doManyChecks (numberCheckers : NumberChecker list) number =
    let message = 
        List.fold (fun message numberChecker ->
            match message with 
            | Some _ -> message
            | None -> numberChecker number
        ) None numberCheckers
    match message with
    | Some message' -> printfn "%s" message'
    | None -> printfn "%d" number   
 
let doFizzBuzzChecks = doManyChecks [ fizzBuzz; fizz; buzz ]

We can use this to play the game:

doFizzBuzzChecks 3
doFizzBuzzChecks 4
doFizzBuzzChecks 5
doFizzBuzzChecks 15

In FSI:

Fizz!
4
Buzz!
Fizz! Buzz!

Or to play the game with many numbers:

[ 1 .. 30 ] |> List.iter doFizzBuzzChecks

The complete code for this can be found here:

https://github.com/robert-impey/CodingExperiments/blob/master/F-Sharp/Loose/FizzBuzz.fsx

Functional programming has been described as a non-solution to a non-problem. Compared to the for loop in C#, this looks like a lot of work. If you have a problem as simple as Fizz! Buzz!, then it’s best to stick to the simplest solution. However, many interesting problems are much more complicated.

Higher order functions allow for abstracting away the details and generalising code. We might add a rule of saying “Boom!” for primes. The code such a checker might be complicated and require extensive testing. The code that makes use of the checkers does not need to change in order to use this new checker, we simply add the function to the list of checkers. We can also combine it with other checkers, enabling code reuse. Functional programming is getting increasing attention because it allows programmers to closely focus their attention on just their problem and decouple that code from the part of the program that will make use of it.

People don’t sit on Park Benches at Random

March 16th, 2014

As I was approaching Berkeley Square at lunchtime on Thursday, I noticed a pattern in the way that the people were sitting on the park benches.

2014-03-13 12.45.16-small

Even though it was a sunny day, the square was not very full, and there were about twice as many bench seats as there were people. Nobody wanted to sit too close to another stranger (this is London, after all), so they sat the maximum distance apart that they could. This resulted in evenly distributed people – Empty, Person, Empty, Person, and so on. The sequence cannot be called random because the people are actively deciding where to sit.

I decided to look at a simple random sequence in F#.

To start off with, I decided to define a type for my program. This might not sound like a promising starting point for a program, but in F# it’s so painless that it is a good way to begin thinking about the domain and what states make sense in your domain. For more on this, see Scott Wlaschin’s slides on Domain Driven Design in F#.

My domain is coin tossing. Therefore, I created the following discriminated union in F# interactive:

> type CoinToss = Tails | Heads;;
 
type CoinToss =
  | Tails
  | Heads

This allows a coin toss to be either heads or tails but not both or neither. Static typing allows the program to make definite statements about the world.

I also needed a random number generator:

> let r = System.Random();;
 
val r : System.Random

F# allows you to define infinite sequences that can be evaluated lazily. The following is a potentially infinite sequence of random coin tosses.

> let coins = Seq.initInfinite (fun _ -> if r.Next(2) = 0 then Heads else Tails);;
 
val coins : seq<CoinToss>

r.Next(2) means the next random integer less than 2 but greater than or equal to 0, that is a random 0 or 1.

If the idea of an infinite sequence of coin tosses sounds weird and unruly, here’s how we can make use of them. The take function allows us to tame infinity. Or at least take a little bit of it.

The following says “Take 10 of the randomly generated coins and pass them along (|>) to a function that iterates through them one at a time and prints them out”

> Seq.take 10 coins |> Seq.iter (printfn "%A");;
Heads
Heads
Tails
Heads
Heads
Tails
Tails
Heads
Tails
Tails
val it : unit = ()

That sequence is not very dissimilar from the lunchers on the park benches in the photo above. There are no long clusters of one thing (person, heads, etc.) or another (empty seat, tails, etc.).

However, this isn’t always the case:

> Seq.take 10 coins |> Seq.iter (printfn "%A");;
Tails
Tails
Heads
Heads
Heads
Tails
Heads
Tails
Heads
Heads
val it : unit = ()

or even:

> Seq.take 10 coins |> Seq.iter (printfn "%A");;
Tails
Tails
Tails
Tails
Tails
Heads
Heads
Heads
Heads
Heads
val it : unit = ()

Five tails followed by five heads looks completely fixed. In a casino, you might start to get worried. But it happened. Trust me.

One way to check if the coin is being tossed fairly, is to count how many times each face lands. This function does that:

> let rec countCoins coins (heads, tails) = 
    match coins with 
    | [] -> heads, tails
    | hd :: tl ->
        let newCounts =
            match hd with
            | Heads -> heads + 1, tails
            | Tails -> heads, tails + 1
        countCoins tl newCounts;;
 
val countCoins : coins:CoinToss list -> heads:int * tails:int -> int * int

The way to read this function is as follows.

If the argument coins matches an empty list, return the counts for heads and tails as they are.

Otherwise, coins must be a list with a first element (hd) and the remaining part of the part of the list (tl), which might be an empty list. We need to find the new counts of heads or tails by adding 1 to either count. Finally, we call the countCoins function again with the remainder of the list (tl). This will continue until we have stripped all the first elements from the list and tl is an empty list. The function then terminates as we have satisfied the first condition above.

For convenience, I defined a function that gets a number of coin tosses and puts them into a list to be counted:

> let getCoins numberOfCoins = Seq.take numberOfCoins coins |> Seq.toList;;
 
val getCoins : numberOfCoins:int -> CoinToss list

For a short list, the counts for each face might not be equal:

> countCoins (getCoins 10) (0, 0);;
val it : int * int = (6, 4)

Note that the counts for heads and tails start off as zeroes.

However, with a larger number of coins (in this case, pown 10 7 or 10 to the power of 7 or 10 million), the counts should both approach half the number of tosses:

> countCoins (getCoins (pown 10 7)) (0, 0);;
val it : int * int = (4997646, 5002354)

Walking on Hampstead Heath

March 16th, 2014

Living in London certainly makes one appreciate the few sunny days that we do get.

Half of London seemed to be on Hampstead Heath this afternoon.

2014-03-16 12.55.51-small

It’s unseasonably warm at the moment. It is certainly very strange to see Londoners dressed in flip flops and shorts when the leaves haven’t even come on on the trees yet.

Panorama

I just hope that the good weather lasts!

Using F# to minimise a function

August 22nd, 2013

Finding the minimum of functions is at the heart of optimisation. Mathematicians, engineers and programmers have come up with a large number of approaches to solving this problem, including differentiation, genetic algorithms and even exhaustive search.

Consider a quadratic function such that could be written in F# as

let f x = (x ** 2.0) - (2.0 * x) + 1.0

Finding the minimum means finding the input value for the function that returns to lowest value. If we plot the curve, then the minimum is the lowest point of the curve.

One way to find this is find the derivative of the function:

(2.0 * x) - 2.0

and solve the equation where the derivative is equal to 0, which is when x is 1. This probably the best way to solve this problem in practice.

However, not all functions have derivatives that are easy to find. Therefore, an alternative way to minimise a function is an exhaustive search of the inputs in some range. This is not an elegant or generally efficient solution, but it works.

In F#, we might proceed as follows.

We define the range that we want to search and the step between candidate inputs:

let min = 0.0
let max = 10.0
let step = 0.01

We know where to start searching- with the lowest candidate input:

let firstCandidate = f min, min

This creates a tuple of two floats, the first being the output and the second being the lowest candidate input.

We also want a sequence of tuples of two floats that are the remaining candidate solutions:

let remainingCandidates = seq { for c in min + step .. step .. max do yield f c, c }

Note that at this point the sequence has not been enumerated and the function has not been run with each of the candidate inputs. Because sequences are evaluated lazily, the function will not be run for each of the candidate inputs until it is asked for.

We are trying to minimise the function, therefore we need a function that can compare the first elements of two candidate solutions:

let findMin currentMinSln candidateSln =
    match fst currentMinSln < fst candidateSln with
    | true -> currentMinSln
    | false -> candidateSln

Running this code in fsi shows its type to be:

val findMin : 'a * 'b -> 'a * 'b -> 'a * 'b when 'a : comparison

The F# compiler can infer that the function takes in two tuples, each with two elements. The tuples must be of the same type and the first element of each tuple must be comparable. Way to go, type inference!

Everything has been set up at this point. We just need to run the code:

let minSln = Seq.fold findMin firstCandidate remainingCandidates

The fold function is to reduce a sequence to a single value, starting with a given accumulator. In this case, the first candidate solution that we created earlier is our starting point. The output of the function for each candidate input is compared to that accumulator.

Running the code finds the same answer (1) that we found with calculus.

This code runs pretty quickly, but this approach is generally slow. Our range is only 10 wide and the step is 0.01, so there are only 1000 candidate inputs. Increasing the size of the range or decreasing the step would increase the number of inputs. More worryingly, the size of the search space increases exponentially as the number of inputs to the function increases. So a function that took two inputs for the same range for each input would have a million candidate inputs, three inputs would take a billion and so on.

However, it is also quite straightforward to parallelise this approach. The sequence of candidate solutions could be split up into partitions and each partition could be sent off to a different computer. Each batch would find a local minimum for the range it was given. Finding the minimum for the whole of our range of inputs is simply a matter of find the minimum in the returned list.

The complete code for this can be found here:

https://github.com/robert-impey/CodingExperiments/blob/master/F%23/Loose/MinimiseQuadratic.fs