ValuedAction DESPOT::Evaluate(VNode* root, vector<State*>& particles,
RandomStreams& streams, POMCPPrior* prior, const DSPOMDP* model) {
double value = 0;
for (int i = 0; i < particles.size(); i++) {
particles[i]->scenario_id = i;
}
for (int i = 0; i < particles.size(); i++) {
State* particle = particles[i];
VNode* cur = root;
State* copy = model->Copy(particle);
double discount = 1.0;
double val = 0;
int steps = 0;
while (!streams.Exhausted()) {
int action =
(cur != NULL) ?
OptimalAction(cur).action : prior->GetAction(*copy);
assert(action != -1);
double reward;
OBS_TYPE obs;
bool terminal = model->Step(*copy, streams.Entry(copy->scenario_id),
action, reward, obs);
val += discount * reward;
discount *= Discount();
if (!terminal) {
prior->Add(action, obs);
streams.Advance();
steps++;
if (cur != NULL && !cur->IsLeaf()) {
QNode* qnode = cur->Child(action);
map<OBS_TYPE, VNode*>& vnodes = qnode->children();
cur = vnodes.find(obs) != vnodes.end() ? vnodes[obs] : NULL;
}
} else {
break;
}
}
for (int i = 0; i < steps; i++) {
streams.Back();
prior->PopLast();
}
model->Free(copy);
value += val;
}
return ValuedAction(OptimalAction(root).action, value / particles.size());
}
double AEMS::AEMS2Likelihood(QNode* qnode) {
VNode* vnode = qnode->parent();
QNode* qstar = NULL;
for (int action = 0; action < vnode->children().size(); action++) {
QNode* child = vnode->Child(action);
if (qstar == NULL || child->upper_bound() > qstar->upper_bound())
qstar = child;
}
return qstar == qnode;
}
