public override float[] Decide(List <float> vectorObs, List <Texture2D> visualObs, List <float> heuristicAction, List <float> heuristicVariance = null)
{
var result = base.Decide(vectorObs, visualObs, heuristicAction, heuristicVariance);
if (log)
{
logStep++;
logger.AddData("Average MAES iteration", optimizer.Iteration);
logger.AddData("Average MAES best Score", (float)optimizer.BestScore);
if (logStep % logInterval == 0)
{
logger.LogAllCurrentData(logStep);
}
}
return(result);
}
public override Dictionary <Agent, TakeActionOutput> TakeAction(Dictionary <Agent, AgentInfo> agentInfos)
{
var result = new Dictionary <Agent, TakeActionOutput>();
var agentList = new List <Agent>(agentInfos.Keys);
if (agentList.Count <= 0)
{
return(result);
}
float[,] vectorObsAll = CreateVectorInputBatch(agentInfos, agentList);
var visualObsAll = CreateVisualInputBatch(agentInfos, agentList, BrainToTrain.brainParameters.cameraResolutions);
float[,] actions = null;
var evalOutput = modelSL.EvaluateAction(vectorObsAll, visualObsAll);
actions = evalOutput.Item1;
var vars = evalOutput.Item2;
int i = 0;
int agentNumWithDecision = 0;
float actionDiff = 0; //difference between decision and from networknetwork
foreach (var agent in agentList)
{
var info = agentInfos[agent];
var agentDecision = agent.GetComponent <AgentDependentDecision>();
if (agentDecision != null && agentDecision.useDecision)
{
//if this agent will use the decision, use it
float[] action = null;
if (vars != null)
{
action = agentDecision.Decide(info.stackedVectorObservation, info.visualObservations, new List <float>(actions.GetRow(i)), new List <float>(vars.GetRow(i)));
}
else
{
action = agentDecision.Decide(info.stackedVectorObservation, info.visualObservations, new List <float>(actions.GetRow(i)));
}
var tempAction = new TakeActionOutput();
tempAction.outputAction = action;
result[agent] = tempAction;
if (BrainToTrain.brainParameters.vectorActionSpaceType == SpaceType.continuous)
{
actionDiff += Enumerable.Zip(action, actions.GetRow(i), (a, b) => Mathf.Sqrt((a - b) * (a - b))).Aggregate((a, v) => a + v);
}
else
{
actionDiff += Enumerable.Zip(action, actions.GetRow(i), (a, b) => (Mathf.RoundToInt(a) == Mathf.RoundToInt(b))?0:1).Aggregate((a, v) => a + v);
}
agentNumWithDecision++;
}
else
{
//use result from neural network directly
var tempAction = new TakeActionOutput();
tempAction.outputAction = actions.GetRow(i);
result[agent] = tempAction;
}
i++;
}
if (agentNumWithDecision > 0)
{
stats.AddData("action difference", actionDiff / agentNumWithDecision);
}
return(result);
}