/// <summary>
/// /// Creates a new ExponentiallyDecayingReservoir
/// </summary>
/// <param name="size">The number of samples to keep in the sampling reservoir</param>
/// <param name="alpha">The exponential decay factor; the higher this is, the more biased the sample will be towards newer values</param>
/// <param name="clock">the clock used to timestamp samples and track rescaling</param>
public ExponentiallyDecayingReservoir(int size, double alpha, Clock clock)
{
_values = new ConcurrentDictionary<double, WeightedSample>();
_lock = new ReaderWriterLockSlim(LockRecursionPolicy.SupportsRecursion);
_alpha = alpha;
_size = size;
this.clock = clock;
this._count = new AtomicLong(0);
this._startTime = CurrentTimeInSeconds();
this._nextScaleTime = new AtomicLong(clock.getTick() + RESCALE_THRESHOLD);
}
private VolatileLong(VolatileLong value) : this()
{
Set(value);
}
private static VolatileLong Add(VolatileLong left, VolatileLong right)
{
left.Set(left.Get() + right.Get());
return(left.Get());
}
/// <summary>
/// "A common feature of the above techniques—indeed, the key technique that
/// allows us to track the decayed weights efficiently—is that they maintain
/// counts and other quantities based on g(ti − L), and only scale by g(t − L)
/// at query time. But while g(ti −L)/g(t−L) is guaranteed to lie between zero
/// and one, the intermediate values of g(ti − L) could become very large. For
/// polynomial functions, these values should not grow too large, and should be
/// effectively represented in practice by floating point values without loss of
/// precision. For exponential functions, these values could grow quite large as
/// new values of (ti − L) become large, and potentially exceed the capacity of
/// common floating point types. However, since the values stored by the
/// algorithms are linear combinations of g values (scaled sums), they can be
/// rescaled relative to a new landmark. That is, by the analysis of exponential
/// decay in Section III-A, the choice of L does not affect the final result. We
/// can therefore multiply each value based on L by a factor of exp(−α(L′ − L)),
/// and obtain the correct value as if we had instead computed relative to a new
/// landmark L′ (and then use this new L′ at query time). This can be done with
/// a linear pass over whatever data structure is being used."
/// </summary>
/// <param name="now"></param>
/// <param name="next"></param>
private void Rescale(long now, long next)
{
if (_nextScaleTime.CompareAndSet(next, now + RESCALE_THRESHOLD))
{
lockForRescale();
try
{
var oldStartTime = _startTime;
_startTime = CurrentTimeInSeconds();
double scalingFactor = Math.Exp(-_alpha * (_startTime - oldStartTime));
var keys = new List<double>(_values.Keys);
foreach (double key in keys)
{
WeightedSample sample = null;
if (_values.TryRemove(key, out sample))
{
WeightedSample newSample = new WeightedSample(sample.value, sample.weight * scalingFactor);
_values.AddOrUpdate(key * scalingFactor, newSample, (k, v) => v);
}
}
}
finally
{
unlockForRescale();
}
}
}
private VolatileLong(VolatileLong value): this()
{
Set(value);
}
private static VolatileLong Add(VolatileLong left, VolatileLong right)
{
left.Set(left.Get() + right.Get());
return left.Get();
}