double FlatMCTS::SimulateQ(STATE &state, QNODE &qnode, int action, int depth) {
int observation;
double immediateReward;
double delayedReward = 0.0;
bool terminal = Simulator.Step(state, action, observation, immediateReward);
if (Params.ThompsonSampling) {
qnode.Update(observation, immediateReward, 1);
}
History.Add(action, observation);
if (Params.Verbose >= 3) {
Simulator.DisplayAction(action, cout);
Simulator.DisplayObservation(state, observation, cout);
Simulator.DisplayReward(immediateReward, cout);
Simulator.DisplayState(state, cout);
}
VNODE *&vnode = qnode.Child(observation);
if (!terminal) {
if (vnode) {
delayedReward = SimulateV(state, vnode, depth + 1);
} else {
vnode = ExpandNode(&state, History);
STATE *copy = Simulator.Copy(state);
delayedReward = Rollout(*copy, depth + 1);
Simulator.FreeState(copy);
if (Params.ThompsonSampling) {
vnode->GetCumulativeReward(state).Add(delayedReward);
}
else {
vnode->UCB.Value.Add(delayedReward);
}
}
} else {
if (!vnode) {
vnode = ExpandNode(&state, History);
}
if (Params.ThompsonSampling) {
vnode->GetCumulativeReward(state).Add(0.0);
}
else {
vnode->UCB.Value.Add(0.0);
}
}
double totalReward = immediateReward + Simulator.GetDiscount() * delayedReward;
if (!Params.ThompsonSampling) {
qnode.UCB.Value.Add(totalReward);
}
return totalReward;
}
double FlatMCTS::QValue(QNODE &qnode, bool sampling, int depth) const
{
double qvalue = 0;
{
const std::vector<std::pair<int, double>> &observations =
qnode.TS.Observation.ThompsonSampling(sampling);
for (std::vector<std::pair<int, double>>::const_iterator it =
observations.begin();
it != observations.end(); ++it) {
qvalue += it->second * HValue(qnode.Child(it->first), sampling, depth);
}
}
qvalue *= Simulator.GetDiscount();
{
const std::vector<std::pair<double, double>> &rewards =
qnode.TS.ImmediateReward.ThompsonSampling(sampling);
for (std::vector<std::pair<double, double>>::const_iterator it =
rewards.begin();
it != rewards.end(); ++it) {
qvalue += it->second * it->first;
}
}
return qvalue;
}
double BAUCT::SimulateQ(uint state, QNODE& qnode, uint action)
{
uint observation;
double immediateReward, delayedReward = 0;
uint ii = state*SA+action*S;
Sampler* nextSSampler = SampFact.getTransitionSampler(pcounts+ii,state,action,S);
observation = nextSSampler->getNextStateSample();
immediateReward = getReward(state,action,observation);
delete nextSSampler;
pcounts[state*SA+action*S+observation] += 1;
bool terminal = false; //FIXME Assumes non-episodic tasks...
History.Add(action, observation);
VNODE*& vnode = qnode.Child(observation);
if (!vnode && !terminal && qnode.Value.GetCount() >= Params.ExpandCount)
vnode = ExpandNode(); //&state);
if (!terminal)
{
TreeDepth++;
if (vnode)
delayedReward = SimulateV(observation, vnode);
else{
delayedReward = Rollout(observation);
}
TreeDepth--;
}
double totalReward = immediateReward + Simulator.GetDiscount() * delayedReward;
qnode.Value.Add(totalReward);
return totalReward;
}
double MCTS::SimulateQ(STATE &state, QNODE &qnode, int action) {
int observation;
double immediateReward, delayedReward = 0;
if( Simulator.HasAlpha())
Simulator.UpdateAlpha(qnode, state);
bool terminal = Simulator.Step(state, action, observation, immediateReward);
assert(observation >= 0 && observation < Simulator.GetNumObservations());
History.Add(action, observation);
if( Params.Verbose >= 3 ) {
Simulator.DisplayAction(action, cout);
Simulator.DisplayObservation(state, observation, cout);
Simulator.DisplayReward(immediateReward, cout);
Simulator.DisplayState(state, cout);
}
VNODE *&vnode = qnode.Child(observation);
if( !vnode && !terminal && qnode.Value.GetCount() >= Params.ExpandCount )
vnode = ExpandNode(&state);
if( !terminal ) {
TreeDepth++;
if( vnode )
delayedReward = SimulateV(state, vnode);
else
delayedReward = Rollout(state);
TreeDepth--;
}
double totalReward = immediateReward + Simulator.GetDiscount() * delayedReward;
qnode.Value.Add(totalReward);
return totalReward;
}
double BAMCP::SimulateQ(const SIMULATOR* mdp, uint state, QNODE& qnode, uint action)
{
uint observation;
double immediateReward, delayedReward = 0;
bool terminal = mdp->Step(state, action, observation, immediateReward);
//assert(observation >= 0 && observation < mdp.GetNumObservations());
History.Add(action, observation);
if (Params.Verbose >= 3)
{
mdp->DisplayAction(action, cout);
mdp->DisplayObservation(state, observation, cout);
mdp->DisplayReward(immediateReward, cout);
mdp->DisplayState(state, cout);
}
VNODE*& vnode = qnode.Child(observation);
if (!vnode && !terminal && qnode.Value.GetCount() >= Params.ExpandCount)
vnode = ExpandNode(); //&state);
if (!terminal)
{
TreeDepth++;
if (vnode)
delayedReward = SimulateV(mdp, observation, vnode);
else {
if(Params.RB < 0){
delayedReward = Rollout(mdp, observation);
}else{
//Warning this will fail with lazy sampling factories
assert(mdp->T != 0);
MDPutils::valueIterationRmax ( S,
A,
Simulator.rsas,
mdp->T,
Simulator.R,
Simulator.GetDiscount(),
0.0001,
RLPI,
V,counts,Params.RB);
delayedReward = V[observation];
}
}
TreeDepth--;
}
double totalReward = immediateReward + mdp->GetDiscount() * delayedReward;
qnode.Value.Add(totalReward);
return totalReward;
}