Is using Random and OrderBy a good shuffle algorithm? [closed]

It's not a way of shuffling that I like, mostly on the grounds that it's O(n log n) for no good reason when it's easy to implement an O(n) shuffle. The code in the question "works" by basically giving a random (hopefully unique!) number to each element, then ordering the elements according to that number.

I prefer Durstenfeld's variant of the Fisher-Yates shuffle which swaps elements.

Implementing a simple Shuffle extension method would basically consist of calling ToList or ToArray on the input then using an existing implementation of Fisher-Yates. (Pass in the Random as a parameter to make life generally nicer.) There are plenty of implementations around... I've probably got one in an answer somewhere.

The nice thing about such an extension method is that it would then be very clear to the reader what you're actually trying to do.

EDIT: Here's a simple implementation (no error checking!):

public static IEnumerable<T> Shuffle<T>(this IEnumerable<T> source, Random rng)
{
    T[] elements = source.ToArray();
    // Note i > 0 to avoid final pointless iteration
    for (int i = elements.Length-1; i > 0; i--)
    {
        // Swap element "i" with a random earlier element it (or itself)
        int swapIndex = rng.Next(i + 1);
        T tmp = elements[i];
        elements[i] = elements[swapIndex];
        elements[swapIndex] = tmp;
    }
    // Lazily yield (avoiding aliasing issues etc)
    foreach (T element in elements)
    {
        yield return element;
    }
}

EDIT: Comments on performance below reminded me that we can actually return the elements as we shuffle them:

public static IEnumerable<T> Shuffle<T>(this IEnumerable<T> source, Random rng)
{
    T[] elements = source.ToArray();
    for (int i = elements.Length - 1; i >= 0; i--)
    {
        // Swap element "i" with a random earlier element it (or itself)
        // ... except we don't really need to swap it fully, as we can
        // return it immediately, and afterwards it's irrelevant.
        int swapIndex = rng.Next(i + 1);
        yield return elements[swapIndex];
        elements[swapIndex] = elements[i];
    }
}

This will now only do as much work as it needs to.

Note that in both cases, you need to be careful about the instance of Random you use as:

  • Creating two instances of Random at roughly the same time will yield the same sequence of random numbers (when used in the same way)
  • Random isn't thread-safe.

I have an article on Random which goes into more detail on these issues and provides solutions.


This is based on Jon Skeet's answer.

In that answer, the array is shuffled, then returned using yield. The net result is that the array is kept in memory for the duration of foreach, as well as objects necessary for iteration, and yet the cost is all at the beginning - the yield is basically an empty loop.

This algorithm is used a lot in games, where the first three items are picked, and the others will only be needed later if at all. My suggestion is to yield the numbers as soon as they are swapped. This will reduce the start-up cost, while keeping the iteration cost at O(1) (basically 5 operations per iteration). The total cost would remain the same, but the shuffling itself would be quicker. In cases where this is called as collection.Shuffle().ToArray() it will theoretically make no difference, but in the aforementioned use cases it will speed start-up. Also, this would make the algorithm useful for cases where you only need a few unique items. For example, if you need to pull out three cards from a deck of 52, you can call deck.Shuffle().Take(3) and only three swaps will take place (although the entire array would have to be copied first).

public static IEnumerable<T> Shuffle<T>(this IEnumerable<T> source, Random rng)
{
    T[] elements = source.ToArray();
    // Note i > 0 to avoid final pointless iteration
    for (int i = elements.Length - 1; i > 0; i--)
    {
        // Swap element "i" with a random earlier element it (or itself)
        int swapIndex = rng.Next(i + 1);
        yield return elements[swapIndex];
        elements[swapIndex] = elements[i];
        // we don't actually perform the swap, we can forget about the
        // swapped element because we already returned it.
    }

    // there is one item remaining that was not returned - we return it now
    yield return elements[0]; 
}

Starting from this quote of Skeet:

It's not a way of shuffling that I like, mostly on the grounds that it's O(n log n) for no good reason when it's easy to implement an O(n) shuffle. The code in the question "works" by basically giving a random (hopefully unique!) number to each element, then ordering the elements according to that number.

I'll go on a little explaining the reason for the hopefully unique!

Now, from the Enumerable.OrderBy:

This method performs a stable sort; that is, if the keys of two elements are equal, the order of the elements is preserved

This is very important! What happens if two elements "receive" the same random number? It happens that they remain in the same order they are in the array. Now, what is the possibility for this to happen? It is difficult to calculate exactly, but there is the Birthday Problem that is exactly this problem.

Now, is it real? Is it true?

As always, when in doubt, write some lines of program: http://pastebin.com/5CDnUxPG

This little block of code shuffles an array of 3 elements a certain number of times using the Fisher-Yates algorithm done backward, the Fisher-Yates algorithm done forward (in the wiki page there are two pseudo-code algorithms... They produce equivalent results, but one is done from first to last element, while the other is done from last to first element), the naive wrong algorithm of http://blog.codinghorror.com/the-danger-of-naivete/ and using the .OrderBy(x => r.Next()) and the .OrderBy(x => r.Next(someValue)).

Now, Random.Next is

A 32-bit signed integer that is greater than or equal to 0 and less than MaxValue.

so it's equivalent to

OrderBy(x => r.Next(int.MaxValue))

To test if this problem exists, we could enlarge the array (something very slow) or simply reduce the maximum value of the random number generator (int.MaxValue isn't a "special" number... It is simply a very big number). In the end, if the algorithm isn't biased by the stableness of the OrderBy, then any range of values should give the same result.

The program then tests some values, in the range 1...4096. Looking at the result, it's quite clear that for low values (< 128), the algorithm is very biased (4-8%). With 3 values you need at least r.Next(1024). If you make the array bigger (4 or 5), then even r.Next(1024) isn't enough. I'm not an expert in shuffling and in math, but I think that for each extra bit of length of the array, you need 2 extra bits of maximum value (because the birthday paradox is connected to the sqrt(numvalues)), so that if the maximum value is 2^31, I'll say that you should be able to sort arrays up to 2^12/2^13 bits (4096-8192 elements)