/// <summary>
/// Calculate an approximate gradient of a given function, at a number of discrete sample points.
/// </summary>
/// <param name="paramSamplingInfo">Sampling metadata.</param>
/// <param name="yArr">The function output/result at a number of discrete sample points.</param>
/// <param name="gradientArr">An array to store the calculated gradients within.</param>
public static void CalcGradients(
ParamSamplingInfo paramSamplingInfo,
double[] yArr,
double[] gradientArr)
{
// Notes.
// The gradient at a sample point is approximated by taking the gradient of the line between the two
// sample points either side of that point. For the first and last sample points we take the gradient
// of the line between the sample point and its single adjacent sample point (as an alternative we could
// sample an additional point at each end that doesn't get used for the function regression evaluation.
//
// This approach is rather crude, but fast. A better approach might be to do a polynomial regression on
// the sample point and its nearest two adjacent samples, and then take the gradient of the polynomial
// regression at the required point; obviously that would required more computational work to do so may
// not be beneficial in the overall context of an evolutionary algorithm.
//
// Furthermore, the difference between this gradient approximation and the true gradient decreases with
// increases sample density, therefore this is a reasonable approach *if* the sample density is
// sufficiently high.
// Handle the end points as special cases.
// First point.
double[] xArr = paramSamplingInfo.XArr;
gradientArr[0] = CalcGradient(xArr[0], yArr[0], xArr[1], yArr[1]);
// Intermediate points.
int width = Vector <double> .Count;
int i = 1;
for (; i < xArr.Length - width - 1; i += width)
{
// Calc a block of x deltas.
var vecLeft = new Vector <double>(xArr, i - 1);
var vecRight = new Vector <double>(xArr, i + 1);
var xVecDelta = vecRight - vecLeft;
// Calc a block of y deltas.
vecLeft = new Vector <double>(yArr, i - 1);
vecRight = new Vector <double>(yArr, i + 1);
var yVecDelta = vecRight - vecLeft;
// Divide the y's by x's to obtain the gradients.
var gradientVec = yVecDelta / xVecDelta;
gradientVec.CopyTo(gradientArr, i);
}
// Calc gradients for remaining intermediate points (if any).
for (; i < xArr.Length - 1; i++)
{
gradientArr[i] = CalcGradient(xArr[i - 1], yArr[i - 1], xArr[i + 1], yArr[i + 1]);
}
// Last point.
gradientArr[i] = CalcGradient(xArr[i - 1], yArr[i - 1], xArr[i], yArr[i]);
}
private static BlackBoxProbe CreateBlackBoxProbe(
Func <double, double> fn,
ParamSamplingInfo paramSamplingInfo)
{
// Determine the mid output value of the function (over the specified sample points) and a scaling factor
// to apply the to neural network response for it to be able to recreate the function (because the neural net
// output range is [0,1] when using the logistic function as the neuron activation function).
FuncRegressionUtils.CalcFunctionMidAndScale(fn, paramSamplingInfo, out double mid, out double scale);
return(new BlackBoxProbe(paramSamplingInfo, mid, scale));
}
/// <summary>
/// Probe the given function by taking samples of it at a number of discrete sample points.
/// </summary>
/// <param name="fn">The function to probe/sample.</param>
/// <param name="paramSamplingInfo">Sampling metadata.</param>
/// <param name="responseArr">An array to store the sample results within.</param>
public static void Probe(
Func <double, double> fn,
ParamSamplingInfo paramSamplingInfo,
double[] responseArr)
{
Debug.Assert(responseArr.Length == paramSamplingInfo.SampleResolution);
double[] xArr = paramSamplingInfo.XArr;
for (int i = 0; i < xArr.Length; i++)
{
responseArr[i] = fn(xArr[i]);
}
}
/// <summary>
/// Construct a new instance.
/// </summary>
/// <param name="fn">The target function.</param>
/// <param name="paramSamplingInfo">Sampling (defines the x range and sampling density).</param>
/// <param name="gradientMseWeight">The fitness weighting to assign to the gradient mean squared error (MSE) score.</param>
public FuncRegressionEvaluationScheme(
Func <double, double> fn,
ParamSamplingInfo paramSamplingInfo,
double gradientMseWeight)
{
_paramSamplingInfo = paramSamplingInfo;
_gradientMseWeight = gradientMseWeight;
// Alloc arrays.
int sampleCount = _paramSamplingInfo.SampleResolution;
_yArrTarget = new double[sampleCount];
_gradientArrTarget = new double[sampleCount];
// Predetermine target responses.
FuncRegressionUtils.Probe(fn, paramSamplingInfo, _yArrTarget);
FuncRegressionUtils.CalcGradients(paramSamplingInfo, _yArrTarget, _gradientArrTarget);
// Create blackbox probe.
_blackBoxProbe = CreateBlackBoxProbe(fn, paramSamplingInfo);
}
/// <summary>
/// Construct a new instance.
/// </summary>
/// <param name="paramSamplingInfo">Parameter sampling info.</param>
/// <param name="gradientMseWeight">Fitness weighting to apply to the gradient fitness score.</param>
/// <param name="yArrTarget">Array of target y values (function output values).</param>
/// <param name="gradientArrTarget">Array of target gradient values.</param>
/// <param name="blackBoxProbe">Black box probe. For obtaining the y value response array from an instance of <see cref="IBlackBox{T}"/>.</param>
internal FuncRegressionEvaluator(
ParamSamplingInfo paramSamplingInfo,
double gradientMseWeight,
double[] yArrTarget,
double[] gradientArrTarget,
IBlackBoxProbe blackBoxProbe)
{
_paramSamplingInfo = paramSamplingInfo;
_gradientMseWeight = gradientMseWeight;
_yMseWeight = 1.0 - gradientMseWeight;
_yArrTarget = yArrTarget;
_gradientArrTarget = gradientArrTarget;
// Alloc working arrays for receiving black box outputs.
int sampleCount = _paramSamplingInfo.SampleResolution;
_yArr = new double[sampleCount];
_gradientArr = new double[sampleCount];
_blackBoxProbe = blackBoxProbe;
}
/// <summary>
/// Determine the mid output value of the function (over the specified sample points) and a scaling factor
/// to apply the to neural network response for it to be able to recreate the function (because the neural net
/// output range is [0,1] when using the logistic function as the neuron activation function).
/// </summary>
/// <param name="fn">The function to be sampled.</param>
/// <param name="paramSamplingInfo">Parameter sampling info.</param>
/// <param name="mid">Returns the mid value of the function (halfway between min and max).</param>
/// <param name="scale">Returns the scale of the function.</param>
public static void CalcFunctionMidAndScale(
Func <double, double> fn,
ParamSamplingInfo paramSamplingInfo,
out double mid, out double scale)
{
double[] xArr = paramSamplingInfo.XArr;
double min = fn(xArr[0]);
double max = min;
for (int i = 0; i < xArr.Length; i++)
{
double y = fn(xArr[i]);
min = Math.Min(y, min);
max = Math.Max(y, max);
}
// TODO: explain this (0.8 is logistic function range, 0.5 is the logistic function output value when its input is zero).
double range = max - min;
scale = range / 0.8;
mid = (min + max) / 2.0;
}
public static void CalcGradients(
ParamSamplingInfo paramSamplingInfo,
double[] yArr,
double[] gradientArr)
{
// TODO: Can this be vectorized?
// Handle the end points as special cases.
// First point.
double[] xArr = paramSamplingInfo.XArr;
gradientArr[0] = CalcGradient(xArr[0], yArr[0], xArr[1], yArr[1]);
// Intermediate points.
for (int i = 1; i < xArr.Length - 1; i++)
{
gradientArr[i] = CalcGradient(xArr[i - 1], yArr[i - 1], xArr[i + 1], yArr[i + 1]);
}
// Last point.
int lastIdx = xArr.Length - 1;
gradientArr[lastIdx] = CalcGradient(xArr[lastIdx - 1], yArr[lastIdx - 1], xArr[lastIdx], yArr[lastIdx]);
}
/// <summary>
/// Construct a new instance.
/// </summary>
/// <param name="paramSamplingInfo">Parameter sampling info.</param>
/// <param name="offset">Offset to apply to each neural network output response.</param>
/// <param name="scale">Scaling factor to apply to each neural network output response.</param>
public BlackBoxProbe(ParamSamplingInfo paramSamplingInfo, double offset, double scale)
{
_paramSamplingInfo = paramSamplingInfo;
_offset = offset;
_scale = scale;
}
