public override ReinforcementLearningAction ActionForState(ReinforcementLearningState state, ReinforcementLearningQStore qFunction)
{
Random r = new Random();
if (r.NextDouble() > EPSILON)
{
return new InvertedPendulumAction2((float)r.NextDouble());
}
else
{
ReinforcementLearningAction ret;
float value;
qFunction.GetBestActionAndUtilityForState(state, out ret, out value);
return ret;
}
}
public virtual ReinforcementLearningAction ActionForState(ReinforcementLearningState state, ReinforcementLearningQStore qFunction)
{
ReinforcementLearningAction[] actions;
float[] probs;
ActionProbabilities(state, qFunction, out actions, out probs);
float rand = (float)r.NextDouble();
int i = 0;
float sum = 0;
do
{
sum += probs[i]; ++i;
}
while ((i < actions.Length) && (sum < rand));
return actions[i - 1];
}
public override void ActionProbabilities(ReinforcementLearningState state, ReinforcementLearningQStore qFunction, out ReinforcementLearningAction[] actions, out float[] probabilities)
{
ReinforcementLearningAction action;
float utility;
actions = state.GetActions();
probabilities = new float[actions.Length];
qFunction.GetBestActionAndUtilityForState(state, out action, out utility);
for (int i = 0; i < actions.Length; ++i)
{
if (actions[i].Equals(action))
{
probabilities[i] = Epsilon;
}
else probabilities[i] = (1 - Epsilon) / (actions.Length - 1);
}
}
public new ReinforcementLearningAction ActionForState(ReinforcementLearningState state, ReinforcementLearningQStore qFunction)
{
ReinforcementLearningAction action;
float utility;
qFunction.GetBestActionAndUtilityForState(state, out action, out utility);
return action;
}
public abstract void ActionProbabilities(ReinforcementLearningState state, ReinforcementLearningQStore qFunction, out ReinforcementLearningAction[] actions, out float[] probabilities);
public void Initialize(ReinforcementLearningEnvironment env)
{
qFunction = CreateQStore(env);
}
public override void ActionProbabilities(ReinforcementLearningState state, ReinforcementLearningQStore qFunction, out ReinforcementLearningAction[] actions, out float[] probabilities)
{
actions = state.GetActions();
probabilities = new float[actions.Length];
float maxq = float.MinValue;
for(int i=0; i<actions.Length; ++i)
{
float q = qFunction.Evaluate(state, actions[i]);
probabilities[i] = q;
if (q > maxq) maxq = q;
}
float sum = 0;
for (int i = 0; i < actions.Length; ++i)
{
probabilities[i] = (float)Math.Exp((probabilities[i] - maxq) / Temperature);
sum += probabilities[i];
}
for (int i = 0; i < actions.Length; ++i)
{
probabilities[i] /= sum;
}
}
public override void ActionProbabilities(ReinforcementLearningState state, ReinforcementLearningQStore qFunction, out ReinforcementLearningAction[] actions, out float[] probabilities)
{
throw new NotImplementedException();
}
