double TwoStageDynamicBayesianNetwork::
GetYProbabilityGeneral(
const Scope& Xscope,
const vector<Index>& X,
const Scope& Ascope,
const vector<Index>& A,
const Scope& YIIscope,
const vector<Index>& YII,
const Scope& Yscope,
const vector<Index>& Y
) const
{
double p = 1.0;
for(Index Y_index=0; Y_index < Y.size(); Y_index++)
{
//Y_index is index in vector Y
Index y = Yscope.at(Y_index); // the index to the variable we look at
Index yVal = Y.at(Y_index); // the value of that variable
vector<Index> X_restr(GetXSoI_Y(y).size());
IndexTools::RestrictIndividualIndicesToNarrowerScope(X, Xscope, GetXSoI_Y(y), X_restr );
vector<Index> A_restr(GetASoI_Y(y).size());
IndexTools::RestrictIndividualIndicesToNarrowerScope(A, Ascope, GetASoI_Y(y), A_restr );
vector<Index> YII_restr(GetYSoI_Y(y).size());
IndexTools::RestrictIndividualIndicesToNarrowerScope(YII, YIIscope, GetYSoI_Y(y), YII_restr );
Index iiI = IndividualToJointYiiIndices(y, X_restr, A_restr, YII_restr);
double p_y = _m_Y_CPDs[y]->Get(yVal, iiI);
p *= p_y;
}
return(p);
}
double FSAOHDist_NECOF::GetOHProb_relT(
const Scope& agSc,
const std::vector<Index>& ohI)
const
{
double p = 1.0;
for(Index i=0; i < agSc.size(); i++)
{
Index agI = agSc.at(i);
Index ohI_agI = ohI.at(i);
double p_this_agent = _m_oHistMarginals.at(agI).at(ohI_agI);
p *= p_this_agent;
}
return p;
}
string MADPComponentFactoredStates::SoftPrintPartialState(
const Scope& sfSc,
const std::vector<Index>& sfValues) const
{
if(sfSc.size() != sfValues.size())
throw E("MADPComponentFactoredStates::SoftPrintPartialState: sfSc.size() != sfValues.size()");
stringstream ss;
ss << "<";
for(Index scI=0; scI < sfSc.size(); scI++)
{
Index sfI = sfSc.at(scI);
Index sfValI = sfValues.at(scI);
StateFactorDiscrete* sfac = _m_stateFactors.at(sfI);
string name = sfac->GetName();
string value = sfac->GetStateFactorValue(sfValI);
if(scI > 0)
ss << ",";
ss << name << "=" << value;
}
ss << ">";
return(ss.str());
}
double FSAOHDist_NECOF::GetXOHProb_relT_SufficientSFscope(
const Scope& sfSc, const std::vector<Index>& sfacIndices,
const Scope& agSc, const std::vector<Index>& ohIndices
) const
{
double p = GetXProb(sfSc, sfacIndices);
for(Index i=0; i < agSc.size(); i++)
{
Index agI=agSc.at(i);
Index ohI=ohIndices.at(i);
vector<Index> sfacIrestr(_m_sfacSoI.at(agI).size());
try{
IndexTools::RestrictIndividualIndicesToNarrowerScope(
sfacIndices, sfSc, _m_sfacSoI.at(agI), sfacIrestr );
}catch(ENoSubScope& e)
{
stringstream ss;
ss <<"Error FSAOHDist_NECOF::GetXOHProb_relT should be called with all state factors on which the observation histories (of the agents in agSc) depend\n";
ss <<"sfSc="<<sfSc<<", agSc="<<agSc<<endl;
string s(ss.str());
cerr << s;
throw ENoSubScope(s);
}
Index x = IndexTools::IndividualToJointIndicesStepSize(sfacIrestr,
_m_stepsize.at(agI) );
double p_i = _m_oHistConditional.at(agI)->Get(ohI, x);
p *= p_i;
}
return p;
}
void BGCG_SolverMaxPlus::Construct_JointType_Factors_CGBG(
const boost::shared_ptr<const BayesianGameCollaborativeGraphical> &bgcg,
const vector< vector<Index> >& var_indices,
const vector< libDAI::Var >& vars,
vector<libDAI::Factor>& facs,
int verbosity)
{
size_t nrLRFs = bgcg->GetNrLRFs();
//compute nrFacs:
size_t nrFacs = 0;
for(Index e=0; e < nrLRFs; e++)
{
const BayesianGameIdenticalPayoff * bgip_e = bgcg->GetBGIPforLRF (e);
size_t nrJT_e = bgip_e->GetNrJointTypes();
if(verbosity >= 2)
cout << "Edge "<<e<<" has " << nrJT_e <<
" local joint types (i.e. factors)" <<endl;
nrFacs += nrJT_e;
}
if(verbosity >= 1)
cout << "FactorGraph has "<<nrFacs<<" factors"<<endl;
facs.reserve(nrFacs);
for(Index e=0; e < nrLRFs; e++)
{
//construct the factors for LRF component e.
const BayesianGameIdenticalPayoff * bgip_e = bgcg->GetBGIPforLRF (e);
Scope agentScope = bgcg->GetScope(e);
const size_t nrJA_e = bgip_e->GetNrJointActions();
for(Index jtI_e=0; jtI_e < bgip_e->GetNrJointTypes(); jtI_e++)
{
vector<Index> indTypes_e =
bgip_e->JointToIndividualTypeIndices(jtI_e);
if(verbosity > 5)
cout << "Adding factor for joint type jtI_e="<<jtI_e<<
" ( indtypes_e="<<SoftPrintVector(indTypes_e) << " )"<<endl;
//compute the agent-type (i.e., factor) indices atI, get the
//relevant variables and put them in a VarSet
libDAI::VarSet vs;
for(Index agI_e=0; agI_e < bgip_e->GetNrAgents(); agI_e++)
{
//translate agI_e to global agent index:
Index agI = agentScope.at(agI_e);
Index atI = var_indices[agI][indTypes_e[agI_e]];
vs = vs | vars[atI];
if(verbosity > 5)
cout << "agent "<<agI<<
" contributes through variable atI="<<atI<<endl;
}
libDAI::Real factor_vals[nrJA_e];
for(Index jaI_e=0; jaI_e < nrJA_e; jaI_e++)
{
vector<Index> indAcs =
bgip_e->JointToIndividualActionIndices(jaI_e);
vector<size_t> indAcsST(indAcs.size());
for(Index i=0; i<indAcs.size(); i++)
indAcsST[i] = (size_t) indAcs[i];
if(verbosity > 5)
cout << "Computing joint factor index (JFI) for vs="<<vs
<<", (action)indices="<< SoftPrintVector(indAcs) <<endl;
Index jointFactorIndex =
libDAI::Factor::IndividualToJointFactorIndex(vs, indAcsST);
if(verbosity > 5)
cout << "JFI="<<jointFactorIndex<< " Utility "
<< bgip_e->GetUtility(jtI_e, jaI_e) << " Prob "
<< bgip_e->GetProbability(jtI_e) << endl;
double perturbation=(1e-6*rand())/RAND_MAX;
factor_vals[jointFactorIndex] =
(bgip_e->GetUtility(jtI_e, jaI_e) + perturbation) *
bgip_e->GetProbability(jtI_e);
}
libDAI::Factor f(vs, factor_vals);
if(verbosity > 4)
{
cout << "Added factor f:"<< f << endl;
}
facs.push_back(f);
}// <- end for jtI_e
}// <- end for e
}
double FSAOHDist_NECOF::GetXOHProb_relT(
const Scope& sfSc, const std::vector<Index>& sfacIndices,
const Scope& agSc, const std::vector<Index>& ohIndices
) const
{
//we use a running example to illustrate the code:
// ->we want to compute P(o1o2x1x2)
//However, o1 additionally depends on x3 and o2 on x4
//first check if sfSc contains all state factors that can influence
//the observation (and thus the OHs) of the agents in agentSc, if
//not, we will need to marginalize over the state factors that are not
//in sfSc (but are in the SoI).
Scope all_relevant_sfacIs(sfSc);
for(Index i=0; i < agSc.size(); i++)
all_relevant_sfacIs.Insert( _m_puf->GetFDPOMDPD()->
Get2DBN()->GetYSoI_O( agSc[i] ) );
Scope to_marginalize_over_sfacIs(all_relevant_sfacIs);
to_marginalize_over_sfacIs.Remove(sfSc);
//p(x1x2) = P(x1)*P(x2)
double p1 = GetXProb(sfSc, sfacIndices);
vector<Index> marg_vec(to_marginalize_over_sfacIs.size(), 0);
vector<size_t> marg_nrSFvals( to_marginalize_over_sfacIs.size());
IndexTools::RestrictIndividualIndicesToScope(
_m_puf->GetFDPOMDPD()->GetNrValuesPerFactor(),
to_marginalize_over_sfacIs,
marg_nrSFvals);
//this will hold P(o1o2|x1x2)
double p_ohIndices_given_sfacIndices = 0.0;
do {
//the probability of this instantiation of marg_vec
//P(x3x4)
double p2 = GetXProb(to_marginalize_over_sfacIs, marg_vec);
if(p2>0) // optimalization: if p2 is zero the rest will be zero as well
{
//construct the 'full' state vector (i.e., containing all_relevant_sfacIs)
//the vector x1x2x3x4
vector<Index> all_relevant_vals(sfacIndices);
all_relevant_vals.insert(all_relevant_vals.end(),
marg_vec.begin(),
marg_vec.end());
//P(o1o2|x1x2x3x4)
double p_ohIndices_given_all_relevant_vals = 1.0;
for(Index i=0; i < agSc.size(); i++)
{
Index agI=agSc.at(i);
Index ohI=ohIndices.at(i);
vector<Index> sfacIrestr(_m_sfacSoI.at(agI).size());
try{
IndexTools::RestrictIndividualIndicesToNarrowerScope(
all_relevant_vals,
all_relevant_sfacIs,
_m_sfacSoI.at(agI),
sfacIrestr);
}catch(ENoSubScope& e)
{ throw ENoSubScope("bah"); }
Index x = IndexTools::IndividualToJointIndicesStepSize(
sfacIrestr, _m_stepsize.at(agI) );
double p_OHi_given_all_relevant_vals =
_m_oHistConditional.at(agI)->Get(ohI, x);
//P(o1o2|x1x2x3x4) = P(o1|x1x2x3)P(o2|x1x2x4)
p_ohIndices_given_all_relevant_vals *=
p_OHi_given_all_relevant_vals;
}
//P(o1o2x3x4|x1x2) = P(o1o2|x1x2x3x4)P(x3x4)
double p_ohIndices_marg_vec_given_sfacIndices =
p2 * p_ohIndices_given_all_relevant_vals;
//P(o1o2|x1x2) = sum_{x3x4} P(o1o2x3x4|x1x2)
p_ohIndices_given_sfacIndices +=
p_ohIndices_marg_vec_given_sfacIndices;
}
}while (! IndexTools::Increment( marg_vec, marg_nrSFvals ) );
//finally P(o1o2x1x2):
double result = p1 * p_ohIndices_given_sfacIndices;
return(result);
}
