private void SetupNewQFunction()
{
// Iterate through all states and create Q function entry for each potential action
// Have to make assumption that all tiles are equally spaced
qFunction = new QFunction(discountFactor, kActionProb);
foreach (KeyValuePair <State, Tile> pair in states)
{
// Don't add transitions for goal state
if (pair.Value.isGoal)
{
continue;
}
float[] upMove = { pair.Key.position[0] + tileDistance, pair.Key.position[1], pair.Key.position[2] };
float[] downMove = { pair.Key.position[0] - tileDistance, pair.Key.position[1], pair.Key.position[2] };
float[] leftMove = { pair.Key.position[0], pair.Key.position[1], pair.Key.position[2] + tileDistance };
float[] rightMove = { pair.Key.position[0], pair.Key.position[1], pair.Key.position[2] - tileDistance };
// Add action if there is a tile that is walkable by moving up, down, left, or right
AddStateIfCan(pair.Key,
new AgentAction(AgentAction.Move.Up),
upMove);
AddStateIfCan(pair.Key,
new AgentAction(AgentAction.Move.Down),
downMove);
AddStateIfCan(pair.Key,
new AgentAction(AgentAction.Move.Left),
leftMove);
AddStateIfCan(pair.Key,
new AgentAction(AgentAction.Move.Right),
rightMove);
}
}
public override Tensor GetNextAction(Tensor state)
{
if (Rand.NextDouble() < Epsilon)
{
return(base.GetNextAction(state));
}
return(QFunction.GetOptimalAction(state));
}
public void QFunctionTest()
{
var function = new QFunction(x, n, 1, k);
Assert.AreEqual(new Fraction(1, 2), function.Calculate());
var function1 = new QFunction(x, n, 2, k);
Assert.AreEqual(new Fraction(0), function1.Calculate());
var function2 = new QFunction(x, n, 3, k);
Assert.AreEqual(new Fraction(0), function2.Calculate());
var function3 = new QFunction(x, n, 4, k);
Assert.AreEqual(new Fraction(0), function3.Calculate());
}
// case if s = 0;
private Fraction StartValueCalculate()
{
var qFunction = new QFunction(_x, _n, _t, _k);
return(qFunction.Calculate());
}
