private void calculateHiddenLayerErrorRates()
{
for (int i = 0; i < numberOfHiddenOrContextNeurons; i++)
{
hiddenLayer[i].ErrorRate = 0d;
for (int j = 0; j < numberOfOutputNeurons; j++)
{
hiddenLayer[i].ErrorRate = hiddenLayer[i].ErrorRate + outputLayer[j].ErrorRate * outputLayer[j].Weights[i];
}
hiddenLayer[i].ErrorRate = hiddenLayer[i].ErrorRate * MathematicalFunction.sigmoidDerivative(hiddenLayer[i].OutputSignal);
}
}
public void uniformDistributionTest()
{
double[] setOfNumbers = MathematicalFunction.uniformDistribution(10);
double sum = 0.0d;
foreach (double number in setOfNumbers)
{
sum += number;
}
if (sum > 1.0d)
{
Assert.Fail();
}
}
static ScenarioValueFunctions()
{
ClearRunAtBaseScenarioValueFunction
= new ReverseLogisticFunction(leftAsymptoticX: -50, rightAsymptoticX: 30, minAsymptoticY: 0, maxAsymptoticY: 100000); // one hundred thousand
LockDownEnemyTankForOtherTankToDestroyValueFunction
= new ReverseLogisticFunction(leftAsymptoticX: -120, rightAsymptoticX: 120, minAsymptoticY: 0, maxAsymptoticY: 10000); // ten thousand
AvoidBlockingFriendlyTankFunction
= new RampFunction(leftX: 4, rightX: 8, minY: -100, maxY: 0);
// Note that the function falls off very slowly, meaning that there is a slight incentive to stay closer together
AvoidWalkingIntoABulletFunction
= new ReverseLogisticFunction(leftAsymptoticX: 0, rightAsymptoticX: 6, minAsymptoticY: 1, maxAsymptoticY: -VALUE_OF_A_TANK); // fifty thousand
// slack is number of ticks until bullet collides, 6 ticks is enough time to cross the path of the bullet
ShootBulletHeadOnFunction
= new ReverseLogisticFunction(leftAsymptoticX: 0, rightAsymptoticX: 1, minAsymptoticY: 0, maxAsymptoticY: VALUE_OF_A_TANK); // fifty thousand
// slack is number of ticks until bullet collides
DodgeBulletFunction
= new ReverseLogisticFunction(leftAsymptoticX: 0, rightAsymptoticX: 1, minAsymptoticY: 0, maxAsymptoticY: VALUE_OF_A_TANK); // fifty thousand
// slack is number of ticks till reaching a survival point less number of ticks until bullet collides
ProlongEnemyDisarmamentFunction
= new TriangularFunction(startX: -500, modeX: -81, endX: 0, maxY: 30000); // thirty thousand
// use this to fake a linear decline, rather than the logistic function decline
// was: ReverseLogisticFunction(leftAsymptoticX: -80, rightAsymptoticX: 0, minAsymptoticY: 0, maxAsymptoticY: 30000);
AttackEnemyBaseFunction
= new ReverseLogisticFunction(leftAsymptoticX: 0, rightAsymptoticX: 300, minAsymptoticY: 0, maxAsymptoticY: 100000); // hundred thousand
// this is a low value, designed purely to get the tanks away from their base and into the game
AttackEnemyBaseAttackActionFunction
= new ReverseLogisticFunction(leftAsymptoticX: 0, rightAsymptoticX: 300, minAsymptoticY: 0, maxAsymptoticY: 30000); // thirty thousand
// This function boosts the action along the calculated shortest path.
// This helps to break deadlocks between adjacent choices, probably caused by logistic curve not being granular enough
GrappleWithEnemyTankAttackDiffFunction
= new ReverseLogisticFunction(leftAsymptoticX: -5, rightAsymptoticX: -1, minAsymptoticY: 0, maxAsymptoticY: 20000); // ten thousand
GrappleWithEnemyTankAttackFunction
= new ReverseLogisticFunction(leftAsymptoticX: 0, rightAsymptoticX: 40, minAsymptoticY: 0, maxAsymptoticY: 5000);
GrappleWithEnemyTankAttackActionFunction
= new ReverseLogisticFunction(leftAsymptoticX: 0, rightAsymptoticX: 4, minAsymptoticY: 0, maxAsymptoticY: 10000);
AttackDisarmedEnemyTankSlackUntilRearmedFunction
= new ReverseLogisticFunction(leftAsymptoticX: -25, rightAsymptoticX: 0, minAsymptoticY: 0, maxAsymptoticY: 30000); // thirty thousand
AttackDisarmedEnemyTankAttackActionFunction
= new ReverseLogisticFunction(leftAsymptoticX: -25, rightAsymptoticX: 0, minAsymptoticY: 0, maxAsymptoticY: 5000); // five thousand
AttackLockedDownEnemyTankFunction
= new ReverseLogisticFunction(leftAsymptoticX: 0, rightAsymptoticX: 200, minAsymptoticY: 10000, maxAsymptoticY: 40000); // ten thousand to forty thousand
}
private void createLayers()
{
hiddenLayer = new HiddenNeurone[numberOfHiddenOrContextNeurons];
for (int i = 0; i < numberOfHiddenOrContextNeurons; i++)
{
double[] weights = MathematicalFunction.uniformDistribution(inputLength + numberOfHiddenOrContextNeurons);
hiddenLayer[i] = new HiddenNeurone(weights);
}
contextLayer = new ContextNeurone[numberOfHiddenOrContextNeurons];
for (int i = 0; i < numberOfHiddenOrContextNeurons; i++)
{
contextLayer[i] = new ContextNeurone();
}
outputLayer = new OutputNeurone[numberOfOutputNeurons];
for (int i = 0; i < numberOfOutputNeurons; i++)
{
double[] weights = MathematicalFunction.uniformDistribution(numberOfHiddenOrContextNeurons);
outputLayer[i] = new OutputNeurone(weights);
}
}
protected override void calculateOutputSignal()
{
outputSignal = MathematicalFunction.sigmoid(state);
}