public virtual void UpdateQ_SARSA_Lambda(int prevActIndex, int curActIndex)
{
double reward = EnvironmentModeler.GetReward(m_prevState, m_curState,
SoccerAction.GetActionTypeFromIndex(prevActIndex, Params.MoveKings));
UpdateQ_SARSA_Lambda(reward, m_prevState, m_curState, prevActIndex, curActIndex);
}
public virtual void UpdateQ_QL_Watkins(int prevActIndex, int curActIndex, bool isNaive)
{
double reward = EnvironmentModeler.GetReward(m_prevState, m_curState,
SoccerAction.GetActionTypeFromIndex(prevActIndex, Params.MoveKings));
UpdateQ_QL_Watkins(reward, m_prevState, m_curState, prevActIndex, curActIndex, isNaive);
}
protected virtual int RLThink()
{
int greedyActIndex = m_qTable.GetCurrentGreedyActionIndex();
int actIndex = greedyActIndex;
if (Params.Epsillon > 0.0)
{
actIndex = ChooseActionEpsilonGreedy(
Params.Epsillon,
SoccerAction.GetActionCount(Params.MoveKings, TeammatesCount),
greedyActIndex);
}
//SoccerAction act = SoccerAction.GetActionFromIndex(actIndex, Params.MoveKings, MyUnum);
if (Cycle > 0) // because in cycle 0 there is no prev state
{
double reward = EnvironmentModeler.GetReward(m_prevState, m_curState,
SoccerAction.GetActionTypeFromIndex(m_prevActionIndex, Params.MoveKings));
switch (Params.RLMethod)
{
case Params.RLMethods.Evolutionary:
m_qTable.UpdateQ_Evolutionary(m_prevActionIndex, reward);
break;
case Params.RLMethods.Q_Zero:
m_qTable.UpdateQ_QLearning(m_prevActionIndex);
break;
case Params.RLMethods.SARSA_Zero:
m_qTable.UpdateQ_SARSA(m_prevActionIndex, actIndex);
break;
case Params.RLMethods.SARSA_Lambda:
m_qTable.UpdateQ_SARSA_Lambda(m_prevActionIndex, actIndex);
break;
case Params.RLMethods.Q_Lambda_Watkins:
m_qTable.UpdateQ_QL_Watkins(m_prevActionIndex, actIndex, false);
break;
case Params.RLMethods.Q_Lambda_Naive:
m_qTable.UpdateQ_QL_Watkins(m_prevActionIndex, actIndex, true);
break;
default:
break;
}
if (m_performanceLogger.Enabled)
{
m_performanceLogger.Log(Cycle, reward, OurScore, OppScore);
}
}
return(actIndex);
}