%function [ total_reward,i,Q,Model,last_actions,last_states,last_reward,last_Q,lastMaxK,lastMaxVar,lastDreward,lastMA_noise_n,last_maxD,last_meanD ] =...
function Results =...
Episode_WithReset( maxsteps,initDrugStartSteps,therapyStartSteps,therapyEndSteps,simulatedTherapy,resetModelFactor,resetPolicyFactor,Environment,start,parametersMF,parametersMBFW,parametersMBBW)
% Episode do one episode of the mountain car with sarsa learning
% maxstepts: the maximum number of steps per episode
% Q: the current QTable
% alpha: the current learning rate
% gamma: the current discount factor
% epsilon: probablity of a random action
% statelist: the list of states
% actionlist: the list of actions
% Maze
% based on the code of:
% Jose Antonio Martin H. <jamartinh@fdi.ucm.es>
%
%
QTablePerm.mean=zeros(Environment.Num_States,Environment.Num_Actions);
QTablePerm.time=zeros(Environment.Num_States,Environment.Num_Actions);
QTablePerm.var=1.5*eye(Environment.Num_States*Environment.Num_Actions);
last_actions=zeros(1,maxsteps);
last_states=zeros(1,maxsteps);
last_reward=zeros(1,maxsteps);
lastMaxK=zeros(1,maxsteps);
lastMaxVar=lastMaxK;
lastDreward=lastMaxK;
lastMA_noise_n=lastMaxK;
last_maxD=lastMaxK;
last_meanD=lastMaxK;
last_Q=zeros(maxsteps,Environment.Num_States,Environment.Num_Actions);
priorCounts=4;
parametersMBFW.knownTransitions
copyTransitionsFromEnvironment=(parametersMBFW.knownTransitions);
Model =CreateModel(Environment,priorCounts,copyTransitionsFromEnvironment);
HealthyModel=Model;
AddictedModel=Model;
HealthyQ=QTablePerm;
AddictedQ=QTablePerm;
HealedModel=Model;
HealedQ=QTablePerm;
currentState = start;
total_reward = 0;
reset=1;
% selects an action using the epsilon greedy selection strategy
%a = e_greedy_selection(Q,s,epsilon);
stateActionVisitCounts=zeros(Model.Num_States,Model.Num_Actions);
stateActionVisitCounts2=stateActionVisitCounts;
Environment=changeToTherapyReward(Environment);
j=0;
for i=1:maxsteps
%% change environment phase (initial,drug,therapy,postDrug)
if i==initDrugStartSteps
Environment=changeToBaseReward(Environment);
HealthyQ=QTablePerm;
HealthyModel=Model;
elseif i==therapyStartSteps
Environment=changeToTherapyReward(Environment);
AddictedModel=Model;
AddictedQ=QTablePerm;
if simulatedTherapy
QTablePerm=combinePolicies(QTablePerm,HealthyQ,resetPolicyFactor,parametersMF);
if(resetModelFactor>=0 & resetModelFactor<=1)
Model=combineModels(AddictedModel,HealthyModel,resetModelFactor);
elseif resetModelFactor<0
Model=punishDrugModel(HealthyModel,Environment,resetModelFactor);
end
end
end
if (i==therapyEndSteps)
Environment=changeToBaseReward(Environment);
HealedModel=Model;
HealedQ=QTablePerm;
end
%% execute agent
%run internal simulations
if parametersMBFW.runInternalSimulation
% display('parametersMBFW.runInternalSimulation')
[QTablePermOut,~]=runInternalSimulation(QTablePerm,currentState,Model,parametersMBFW,reset);
reset=0;
end
if parametersMBFW.computePolicyWithDP
if j==0
% display('parametersMBFW.computePolicyWithDP')
QTablePerm.mean = DP( parametersMBFW,Environment);
save (strcat('Environment',num2str(i,'%06i'),'.mat'),'Environment')
j=100;
else
j=j-1;
end
end
%Qtable_Integrated=QTablePerm;
%select action a
if (parametersMBFW.mixMFMBPolicies )
QtableSelect=QTablePerm;
QtableSelect.mean=parametersMBFW.mf_factor*QTablePerm.mean+parametersMBFW.mb_factor*QTablePermOut.mean;
action=actionSelection(currentState,Model,parametersMF, QtableSelect,stateActionVisitCounts);
else
QTablePerm.mean=parametersMBFW.mf_factor*QTablePerm.mean+parametersMBFW.mb_factor*QTablePermOut.mean;
action=actionSelection(currentState,Model,parametersMF, QTablePerm,stateActionVisitCounts);
end
%do the selected action and get the next car state
[reward, new_state] = DoAction( action , currentState, Environment );
%QTablePerm=Qtable_Integrated;
if parametersMBFW.updateModel
% display('parametersMBFW.updateModel')
decay=0.995;
Model=updateModel(reward, new_state,action , currentState,Model,decay,parametersMBFW.knownTransitions);
end
%display('actualInteraction')
if parametersMF.updateQTablePerm
% display('parametersMF.updateQTablePerm')
[QTablePerm,maxK,maxVar,dreward,MA_noise_n]=updateQTablePerm(QTablePerm, reward, new_state, action , currentState, stateActionVisitCounts, parametersMF,reset);
reset=0;
end
if parametersMBBW.internalReplay
% display('parametersMBBW.internalReplay')
[QTablePerm,maxdiffQp,meanQp]=internalReplay(QTablePerm,Model,parametersMBBW,stateActionVisitCounts2,reset);
reset=0;
end
last_actions(i)=action;
last_states(i)=currentState;
last_reward(i)=reward;
%lastMaxK(i)=maxK;
%lastMaxVar(i)=maxVar;
%lastDreward(i)=max(abs(dreward));
%lastMA_noise_n(i)=MA_noise_n;
% last_maxD(i)=maxdiffQp;
% last_meanD(i)=meanQp;
last_Q(i,:,:)=QTablePerm.mean;
currentState=new_state;
Q=QTablePerm.mean;
end
Results.total_reward=total_reward;
Results.i=i;
Results.Q=Q;
Results.Model=Model;
Results.last_actions=last_actions;
Results.last_states=last_states;
Results.last_reward=last_reward;
Results.last_Q=last_Q;
Results.lastMaxK=lastMaxK;
Results.lastMaxVar=lastMaxVar;
Results.lastDreward=lastDreward;
Results.lastMA_noise_n=lastMA_noise_n;
Results.last_maxD=last_maxD;
Results.last_meanD=last_meanD;
Results.HealthyModel=HealthyModel;
Results.HealthyQ=HealthyQ;
Results.AddictedModel=AddictedModel;
Results.AddictedQ=AddictedQ;
Results.HealedModel=HealedModel;
Results.HealedQ=HealedQ;