Converting a method to use any Enum

Question

My Problem:

I want to convert my randomBloodType() method to a static method that can take any enum type. I want my method to take any type of enum whether it be BloodType, DaysOfTheWeek, etc. and perform the operations shown below.

Some Background on what the method does:

The method currently chooses a random element from the BloodType enum based on the values assigned to each element. An element with a higher value has a higher probability to be picked.

Code:

    public enum BloodType
    {
        // BloodType = Probability
        ONeg = 4,
        OPos = 36,
        ANeg = 3,
        APos = 28,
        BNeg = 1,
        BPos = 20,
        ABNeg = 1,
        ABPos = 5
    };

    public BloodType randomBloodType()
    {
        // Get the values of the BloodType enum and store it in a array
        BloodType[] bloodTypeValues = (BloodType[])Enum.GetValues(typeof(BloodType));
        List<BloodType> bloodTypeList = new List<BloodType>();

        // Create a list where each element occurs the approximate number of 
        // times defined as its value(probability)
        foreach (BloodType val in bloodTypeValues)
        {
            for(int i = 0; i < (int)val; i++)
            {
                bloodTypeList.Add(val);
            }
        }

        // Sum the values
        int sum = 0;
        foreach (BloodType val in bloodTypeValues)
        {
            sum += (int)val;
        }

        //Get Random value
        Random rand = new Random();
        int randomValue = rand.Next(sum);

        return bloodTypeList[randomValue];

    }

What I have tried so far:

I have tried to use generics. They worked out for the most part, but I was unable to cast my enum elements to int values. I included a example of a section of code that was giving me problems below.

    foreach (T val in bloodTypeValues)
    {
        sum += (int)val; // This line is the problem.
    }

I have also tried using Enum e as a method parameter. I was unable to declare the type of my array of enum elements using this method.

Answer 1

Assuming your enum values are all of type int (you can adjust this accordingly if they're long, short, or whatever):

static TEnum RandomEnumValue<TEnum>(Random rng)
{
    var vals = Enum
        .GetNames(typeof(TEnum))
        .Aggregate(Enumerable.Empty<TEnum>(), (agg, curr) =>
        {
            var value = Enum.Parse(typeof (TEnum), curr);
            return agg.Concat(Enumerable.Repeat((TEnum)value,(int)value)); // For int enums
        })
        .ToArray();

    return vals[rng.Next(vals.Length)];
}

Here's how you would use it:

var rng = new Random();
var randomBloodType = RandomEnumValue<BloodType>(rng);

---

People seem to have their knickers in a knot about multiple indistinguishable enum values in the input enum (for which I still think the above code provides expected behavior). Note that there is no answer here, not even Peter Duniho's, that will allow you to distinguish enum entries when they have the same value, so I'm not sure why this is being considered as a metric for any potential solutions.

Nevertheless, an alternative approach that doesn't use the enum values as probabilities is to use an attribute to specify the probability:

public enum BloodType
{
    [P=4]
    ONeg,
    [P=36]
    OPos,
    [P=3]
    ANeg,
    [P=28]
    APos,
    [P=1]
    BNeg,
    [P=20]
    BPos,
    [P=1]
    ABNeg,
    [P=5]
    ABPos
}

Here is what the attribute used above looks like:

[AttributeUsage(AttributeTargets.Field, AllowMultiple = false)]
public class PAttribute : Attribute
{
    public int Weight { get; private set; }

    public PAttribute(int weight)
    {
        Weight = weight;
    }
}

and finally, this is what the method to get a random enum value would like:

static TEnum RandomEnumValue<TEnum>(Random rng)
{
    var vals = Enum
        .GetNames(typeof(TEnum))
        .Aggregate(Enumerable.Empty<TEnum>(), (agg, curr) =>
        {
            var value = Enum.Parse(typeof(TEnum), curr);

            FieldInfo fi = typeof (TEnum).GetField(curr);
            var weight = ((PAttribute)fi.GetCustomAttribute(typeof(PAttribute), false)).Weight;

            return agg.Concat(Enumerable.Repeat((TEnum)value, weight)); // For int enums
        })
        .ToArray();

    return vals[rng.Next(vals.Length)];
}

(Note: if this code is performance critical, you might need to tweak this and add caching for the reflection data).

Answer 2

(Note: My apologies in advance for the lengthy answer. My actual proposed solution is not all that long, but there are a number of problems with the proposed solutions so far and I want to try to address those thoroughly, to provide context for my own proposed solution).

In my opinion, while you have in fact accepted one answer and might be tempted to use either one, neither of the answers provided so far are correct or useful.

Commenter Ben Voigt has already pointed out two major flaws with your specifications as stated, both related to the fact that you are encoding the enum value's weight in the value itself:

1. You are tying the enum's underlying type to the code that then must interpret that type.
2. Two enum values that have the same weight are indistinguishable from each other.

Both of these issues can be addressed. Indeed, while the answer you accepted (why?) fails to address the first issue, the one provided by Dweeberly does address this through the use of Convert.ToInt32() (which can convert from long to int just fine, as long as the values are small enough).

But the second issue is much harder to address. The answer from Asad attempts to address this by starting with the enum names and parsing them to their values. And this does indeed result in the final array being indexed containing the corresponding entries for each name separately. But the code actually using the enum has no way to distinguish the two; it's really as if those two names are a single enum value, and that single enum value's probability weight is the sum of the value used for the two different names.

I.e. in your example, while the enum entries for e.g. BNeg and ABNeg will be selected separately, the code that receives these randomly selected value has no way to know whether it was BNeg or ABNeg that was selected. As far as it knows, those are just two different names for the same value.

Now, even this problem can be addressed (but not in the way that Asad attempts to…his answer is still broken). If you were, for example, to encode the probabilities in the value while still ensuring unique values for each name, you could decode those probabilities while doing the random selection and that would work. For example:

enum BloodType
{
    // BloodType = Probability
    ONeg = 4 * 100 + 0,
    OPos = 36 * 100 + 1,
    ANeg = 3 * 100 + 2,
    APos = 28 * 100 + 3,
    BNeg = 1 * 100 + 4,
    BPos = 20 * 100 + 5,
    ABNeg = 1 * 100 + 6,
    ABPos = 5 * 100 + 7,
};

Having declared your enum values that way, then you can in your selection code divide the enum value by 100 to obtain its probability weight, which then can be used as seen in the various examples. At the same time, each enum name has a unique value.

But even solving that problem, you are still left with problems related to the choice of encoding and representation of the probabilities. For example, in the above you cannot have an enum that has more than 100 values, nor one with weights larger than (2^31 - 1) / 100; if you want an enum that has more than 100 values, you need a larger multiplier but that would limit your weight values even more.

In many scenarios (maybe all the ones you care about) this won't be an issue. The numbers are small enough that they all fit. But that seems like a serious limitation in what seems like a situation where you want a solution that is as general as possible.

And that's not all. Even if the encoding stays within reasonable limits, you have another significant limit to deal with: the random selection process requires an array large enough to contain for each enum value as many instances of that value as its weight. Again, if the values are small maybe this is not a big problem. But it does severely limit the ability of your implementation to generalize.

So, what to do?

I understand the temptation to try to keep each enum type self-contained; there are some obvious advantages to doing so. But there are also some serious disadvantages that result from that, and if you truly ever try to use this in a generalized way, the changes to the solutions proposed so far will tie your code together in ways that IMHO negate most if not all of the advantage of keeping the enum types self-contained (primarily: if you find you need to modify the implementation to accommodate some new enum type, you will have to go back and edit all of the other enum types you're using…i.e. while each type looks self-contained, in reality they are all tightly coupled with each other).

In my opinion, a much better approach would be to abandon the idea that the enum type itself will encode the probability weights. Just accept that this will be declared separately somehow.

Also, IMHO is would be better to avoid the memory-intensive approach proposed in your original question and mirrored in the other two answers. Yes, this is fine for the small values you're dealing with here. But it's an unnecessary limitation, making only one small part of the logic simpler while complicating and restricting it in other ways.

I propose the following solution, in which the enum values can be whatever you want, the enum's underlying type can be whatever you want, and the algorithm uses memory proportionally only to the number of unique enum values, rather than in proportion to the sum of all of the probability weights.

In this solution, I also address possible performance concerns, by caching the invariant data structures used to select the random values. This may or may not be useful in your case, depending on how frequently you will be generating these random values. But IMHO it is a good idea regardless; the up-front cost of generating these data structures is so high that if the values are selected with any regularity at all, it will begin to dominate the run-time cost of your code. Even if it works fine today, why take the risk? (Again, especially given that you seem to want a generalized solution).

Here is the basic solution:

static T NextRandomEnumValue<T>()
{
    KeyValuePair<T, int>[] aggregatedWeights = GetWeightsForEnum<T>();
    int weightedValue =
            _random.Next(aggregatedWeights[aggregatedWeights.Length - 1].Value),

        index = Array.BinarySearch(aggregatedWeights,
            new KeyValuePair<T, int>(default(T), weightedValue),
            KvpValueComparer<T, int>.Instance);

    return aggregatedWeights[index < 0 ? ~index : index + 1].Key;
}

static KeyValuePair<T, int>[] GetWeightsForEnum<T>()
{
    object temp;

    if (_typeToAggregatedWeights.TryGetValue(typeof(T), out temp))
    {
        return (KeyValuePair<T, int>[])temp;
    }

    if (!_typeToWeightMap.TryGetValue(typeof(T), out temp))
    {
        throw new ArgumentException("Unsupported enum type");
    }

    KeyValuePair<T, int>[] weightMap = (KeyValuePair<T, int>[])temp;
    KeyValuePair<T, int>[] aggregatedWeights =
        new KeyValuePair<T, int>[weightMap.Length];
    int sum = 0;

    for (int i = 0; i < weightMap.Length; i++)
    {
        sum += weightMap[i].Value;
        aggregatedWeights[i] = new KeyValuePair<T,int>(weightMap[i].Key, sum);
    }

    _typeToAggregatedWeights[typeof(T)] = aggregatedWeights;

    return aggregatedWeights;
}

readonly static Random _random = new Random();

// Helper method to reduce verbosity in the enum-to-weight array declarations
static KeyValuePair<T1, T2> CreateKvp<T1, T2>(T1 t1, T2 t2)
{
    return new KeyValuePair<T1, T2>(t1, t2);
}

readonly static KeyValuePair<BloodType, int>[] _bloodTypeToWeight =
{
    CreateKvp(BloodType.ONeg, 4),
    CreateKvp(BloodType.OPos, 36),
    CreateKvp(BloodType.ANeg, 3),
    CreateKvp(BloodType.APos, 28),
    CreateKvp(BloodType.BNeg, 1),
    CreateKvp(BloodType.BPos, 20),
    CreateKvp(BloodType.ABNeg, 1),
    CreateKvp(BloodType.ABPos, 5),
};

readonly static Dictionary<Type, object> _typeToWeightMap =
    new Dictionary<Type, object>()
    {
        { typeof(BloodType), _bloodTypeToWeight },
    };

readonly static Dictionary<Type, object> _typeToAggregatedWeights =
    new Dictionary<Type, object>();

Note that the work of actually selecting a random value is simply a matter of choosing a non-negative random integer less than the sum of the weights, and then using a binary search to find the appropriate enum value.

Once per enum type, the code will build the table of values and weight-sums that will be used for the binary search. This result is stored in a cache dictionary, _typeToAggregatedWeights.

There are also the objects that have to be declared and which will be used at run-time to build this table. Note that the _typeToWeightMap is just in support of making this method 100% generic. If you wanted to write a different named method for each specific type you wanted to support, that could still used a single generic method to implement the initialization and selection, but the named method would know the correct object (e.g. _bloodTypeToWeight) to use for initialization.

Alternatively, another way to avoid the _typeToWeightMap while still keeping the method 100% generic would be to have the _typeToAggregatedWeights be of type Dictionary<Type, Lazy<object>>, and have the values of the dictionary (the Lazy<object> objects) explicitly reference the appropriate weight array for the type.

In other words, there are lots of variations on this theme that would work fine. But they will all have essentially the same structure as above; semantics would be the same and performance differences would be negligible.

One thing you'll notice is that the binary search requires a custom IComparer<T> implementation. That is here:

class KvpValueComparer<TKey, TValue> :
    IComparer<KeyValuePair<TKey, TValue>> where TValue : IComparable<TValue>
{
    public readonly static KvpValueComparer<TKey, TValue> Instance =
        new KvpValueComparer<TKey, TValue>();

    private KvpValueComparer() { }

    public int Compare(KeyValuePair<TKey, TValue> x, KeyValuePair<TKey, TValue> y)
    {
        return x.Value.CompareTo(y.Value);
    }
}

This allows the Array.BinarySearch() method to correct compare the array elements, allowing a single array to contain both the enum values and their aggregated weights, but limiting the binary search comparison to just the weights.

Answer 3

Some of this you can do and some of it isn't so easy. I believe the following extension method will do what you describe.

static public class Util {
    static Random rnd = new Random();
    static public int PriorityPickEnum(this Enum e) {
        // The approved types for an enum are byte, sbyte, short, ushort, int, uint, long, or ulong
        // However, Random only supports a int (or double) as a max value.  Either way
        // it doesn't have the range for uint, long and ulong.
        //
        // sum enum 
        int sum = 0;
        foreach (var x in Enum.GetValues(e.GetType())) {
            sum += Convert.ToInt32(x);
            }

        var i = rnd.Next(sum); // get a random value, it will form a ratio i / sum

        // enums may not have a uniform (incremented) value range (think about flags)
        // therefore we have to step through to get to the range we want,
        // this is due to the requirement that return value have a probability
        // proportional to it's value.  Note enum values must be sorted for this to work.
        foreach (var x in Enum.GetValues(e.GetType()).OfType<Enum>().OrderBy(a => a)) {
            i -= Convert.ToInt32(x);
            if (i <= 0) return Convert.ToInt32(x);
            }
        throw new Exception("This doesn't seem right");
        }
    }

Here is an example of using this extension:

        BloodType bt = BloodType.ABNeg;
        for (int i = 0; i < 100; i++) {
            var v = (BloodType) bt.PriorityPickEnum();
            Console.WriteLine("{0}:  {1}({2})", i, v, (int) v);
            }

This should work pretty well for enum's of type byte, sbyte, ushort, short and int. Once you get beyond int (uint, long, ulong) the problem is the Random class. You can adjust the code to use doubles generated by Random, which would cover uint, but the Random class just doesn't have the range to cover long and ulong. Of course you could use/find/write a different Random class if this is important.