/// Uses the stored information to run the tensorflow graph and generate
/// the actions.
public void DecideAction(Dictionary <Agent, AgentInfo> newAgentInfos)
{
int currentBatchSize = newAgentInfos.Count();
List <Agent> newAgentList = newAgentInfos.Keys.ToList();
List <Agent> recordableAgentList = newAgentList.Where((a) => currentInfo != null && currentInfo.ContainsKey(a) && prevActionOutput.ContainsKey(a)).ToList();
/*if (currentBatchSize == 0)
* {
* return;
* }*/
//get the datas only for the agents in the agentInfo input
var prevInfo = GetValueForAgents(currentInfo, recordableAgentList);
var prevActionActions = GetValueForAgents(prevActionOutput, recordableAgentList);
var newInfo = GetValueForAgents(newAgentInfos, recordableAgentList);
if (recordableAgentList.Count > 0 && trainerInterface.IsTraining() && trainerInterface.GetStep() <= trainerInterface.GetMaxStep())
{
trainerInterface.AddExperience(prevInfo, newInfo, prevActionActions);
trainerInterface.ProcessExperience(prevInfo, newInfo);
}
if (trainerInterface.IsTraining() && trainerInterface.GetStep() <= trainerInterface.GetMaxStep())
{
trainerInterface.IncrementStep();
}
//update the info
UpdateInfos(ref currentInfo, newAgentInfos);
var actionOutputs = trainerInterface.TakeAction(GetValueForAgents(currentInfo, newAgentList));
UpdateActionOutputs(ref prevActionOutput, actionOutputs);
//TODO Update the agent's other info if there is
foreach (Agent agent in newAgentList)
{
if (actionOutputs.ContainsKey(agent) && actionOutputs[agent].outputAction != null)
{
agent.UpdateVectorAction(trainerInterface.PostprocessingAction(actionOutputs[agent].outputAction));
}
}
if (trainerInterface.IsReadyUpdate() && trainerInterface.IsTraining() && trainerInterface.GetStep() <= trainerInterface.GetMaxStep())
{
trainerInterface.UpdateModel();
}
//clear the prev record if the agent is done
foreach (Agent agent in newAgentList)
{
if (newAgentInfos[agent].done || newAgentInfos[agent].maxStepReached)
{
currentInfo.Remove(agent);
}
}
}