double BLSSS::log_model_prob(const Selector &g)const{
// borrowed from MLVS.cpp
double num = vpri_->logp(g);
if(num==BOOM::negative_infinity() || g.nvars() == 0) {
// If num == -infinity then it is in a zero support point in the
// prior. If g.nvars()==0 then all coefficients are zero
// because of the point mass. The only entries remaining in the
// likelihood are sums of squares of y[i] that are independent
// of g. They need to be omitted here because they are omitted
// in the non-empty case below.
return num;
}
SpdMatrix ivar = g.select(pri_->siginv());
num += .5*ivar.logdet();
if(num == BOOM::negative_infinity()) return num;
Vector mu = g.select(pri_->mu());
Vector ivar_mu = ivar * mu;
num -= .5*mu.dot(ivar_mu);
bool ok=true;
ivar += g.select(suf().xtx());
Matrix L = ivar.chol(ok);
if(!ok) return BOOM::negative_infinity();
double denom = sum(log(L.diag())); // = .5 log |ivar|
Vector S = g.select(suf().xty()) + ivar_mu;
Lsolve_inplace(L,S);
denom-= .5*S.normsq(); // S.normsq = beta_tilde ^T V_tilde beta_tilde
return num-denom;
}
bool SelectRegion(Selector& sel, const typename TAAPos::ValueType& p, TAAPos& aaPos,
typename Grid::traits<typename TGeomObj::side>::callback cbRegionBoundary)
{
typedef typename Grid::traits<TGeomObj>::iterator TIter;
if(!sel.grid())
return false;
Grid& g = *sel.grid();
// first try to find the element which contains p
TGeomObj* startElem = NULL;
for(TIter iter = g.begin<TGeomObj>(); iter != g.end<TGeomObj>(); ++iter){
if(ContainsPoint(*iter, p, aaPos)){
startElem = *iter;
break;
}
}
if(!startElem)
return false;
sel.clear<TGeomObj>();
sel.select(startElem);
SelectionFill<TGeomObj>(sel, cbRegionBoundary);
return true;
}
//======================================================================
Vector &impute_mvn(Vector &observation,
const Vector &mean, const SpdMatrix &variance,
const Selector &observed, RNG &rng) {
if (observed.nvars() == observed.nvars_possible()) {
return observation;
} else if (observed.nvars() == 0) {
observation = rmvn_mt(rng, mean, variance);
return observation;
}
if (observation.size() != observed.nvars_possible()) {
report_error("observation and observed must be the same size.");
}
// The distribution we want is N(mu, V), with
// V = Sig11 - Sig12 Sig22.inv Sig.21
// and
// mu = mu1 - Sig12 Sig22.inv (y2 - mu2)
// The 1's are missing, and the 2's are observed.
Selector missing = observed.complement();
Matrix cross_covariance = missing.select_rows(
observed.select_cols(variance));
SpdMatrix observed_precision = observed.select_square(variance).inv();
Vector mu = missing.select(mean) + cross_covariance * observed_precision
* (observed.select(observation) - observed.select(mean));
SpdMatrix V = missing.select_square(variance)
- sandwich(cross_covariance, observed_precision);
Vector imputed = rmvn_mt(rng, mu, V);
observed.fill_missing_elements(observation, imputed);
return observation;
}
void BLSSS::draw_beta() {
Selector g = m_->coef().inc();
if(g.nvars() == 0) {
m_->drop_all();
return;
}
SpdMatrix ivar = g.select(pri_->siginv());
Vector ivar_mu = ivar * g.select(pri_->mu());
ivar += g.select(suf().xtx());
ivar_mu += g.select(suf().xty());
Vector b = ivar.solve(ivar_mu);
b = rmvn_ivar_mt(rng(), b, ivar);
// If model selection is turned off and some elements of beta
// happen to be zero (because, e.g., of a failed MH step) we don't
// want the dimension of beta to change.
m_->set_included_coefficients(b, g);
}
void BLSSS::draw_beta() {
Selector g = model_->coef().inc();
if (g.nvars() == 0) {
model_->drop_all();
return;
}
SpdMatrix precision = g.select(slab_->siginv());
Vector scaled_mean = precision * g.select(slab_->mu());
precision += g.select(suf().xtx());
Cholesky precision_cholesky_factor(precision);
scaled_mean += g.select(suf().xty());
Vector posterior_mean = precision_cholesky_factor.solve(scaled_mean);
Vector beta = rmvn_precision_upper_cholesky_mt(
rng(), posterior_mean, precision_cholesky_factor.getLT());
// If model selection is turned off and some elements of beta
// happen to be zero (because, e.g., of a failed MH step) we don't
// want the dimension of beta to change.
model_->set_included_coefficients(beta, g);
}
int main(int argc, const char * argv[])
{
int port = 4567; // Default port number
int protocol = 1; // Default protocol number
if (argc >= 2) port = atoi(argv[1]);
if (argc >= 3) protocol = atoi(argv[2]);
fprintf(stderr, "Usage %s [port number] [protocol number]\n", argv[0]);
fprintf(stderr, "Starting server at port %d, using protocol %d\n", port, protocol);
try {
Selector serverSelector;
DatagramSocket serverSocket(port);
UDPMessageBox<FastMessage> messageBox(&serverSocket);
IgpMessageListener listener(messageBox);
messageBox.addListener(&listener);
messageBox.addSessionListener(&listener);
serverSelector.addSelectable(&serverSocket);
IgpVirtualPeerMessageBox<FastMessage> igpMBox(listener, 1);
PuyoIgpNatTraversal natPuncher(igpMBox, listener);
igpMBox.addListener(&natPuncher);
PuyoServer *responder;
switch (protocol) {
case 1: responder = new PuyoServerV1(igpMBox); break;
case 2: responder = new PuyoServerV2(igpMBox); break;
default:
fprintf(stderr, "Erros: valid protocols are 1 and 2");
}
igpMBox.addListener(responder);
while (true) {
serverSelector.select(10);
try {
messageBox.idle();
// Pb de timeout par la
responder->idle();
natPuncher.idle();
}
catch (ios_fc::Exception e) {
e.printMessage();
}
}
}
catch (ios_fc::Exception e) {
e.printMessage();
}
return 0;
}
//----------------------------------------------------------------------
double LSB::log_model_prob(const Selector &g)const{
double num = vs_->logp(g);
if(num==BOOM::negative_infinity()) return num;
Ominv = g.select(pri_->siginv());
num += .5*Ominv.logdet();
if(num == BOOM::negative_infinity()) return num;
Vec mu = g.select(pri_->mu());
Vec Ominv_mu = Ominv * mu;
num -= .5*mu.dot(Ominv_mu);
bool ok=true;
iV_tilde_ = Ominv + g.select(suf()->xtx());
Mat L = iV_tilde_.chol(ok);
if(!ok) return BOOM::negative_infinity();
double denom = sum(log(L.diag())); // = .5 log |Ominv|
Vec S = g.select(suf()->xty()) + Ominv_mu;
Lsolve_inplace(L,S);
denom-= .5*S.normsq(); // S.normsq = beta_tilde ^T V_tilde beta_tilde
return num-denom;
}
// When dimensions are small the updates are trivial, and careful
// optimization is not necessary.
void Marginal::low_dimensional_update(
const Vector &observation,
const Selector &observed,
const SparseKalmanMatrix &transition,
const SparseKalmanMatrix &observation_coefficient_subset) {
set_prediction_error(
observed.select(observation)
- observation_coefficient_subset * state_mean());
SpdMatrix forecast_variance =
observed.select_square(model_->observation_variance(time_index())) +
observation_coefficient_subset.sandwich(state_variance());
SpdMatrix forecast_precision = forecast_variance.inv();
set_forecast_precision_log_determinant(forecast_precision.logdet());
set_scaled_prediction_error(forecast_precision * prediction_error());
set_kalman_gain(transition * state_variance() *
observation_coefficient_subset.Tmult(forecast_precision));
}
virtual void run() {
Selector selector;
socket = new DatagramSocket(1900);
socket->setReuseAddr();
socket->setBroadcast();
socket->joinGroup("239.255.255.250");
socket->registerSelector(selector);
while (!interrupted()) {
if (selector.select(1000) > 0) {
char buffer[1024] = {0,};
int len = socket->recv(buffer, sizeof(buffer));
if (len > 0) {
buffer[len - 1] = 0;
cout << "[RECV] >> " << buffer << endl;
}
} else {
idle(10);
}
}
socket->close();
}
// Effects:
// 4 things are set.
// 1) prediction_error
// 2) scaled_prediction_error
// 3) forecast_precision_log_determinant
// 4) kalman_gain
void Marginal::high_dimensional_update(
const Vector &observation,
const Selector &observed,
const SparseKalmanMatrix &transition,
const SparseKalmanMatrix &observation_coefficient_subset) {
Vector observed_subset = observed.select(observation);
set_prediction_error(observed_subset
- observation_coefficient_subset * state_mean());
// At this point the Kalman recursions compute the forecast precision Finv
// and its log determinant. However, we can get rid of the forecast
// precision matrix, and replace it with the scaled error = Finv *
// prediction_error.
//
// To evaluate the normal likelihood, we need the quadratic form:
// error * Finv * error == error.dot(scaled_error).
// We also need the log determinant of Finv.
//
// The forecast_precision can be computed using a version of the binomial
// inverse theorem:
//
// (A + UBV).inv =
// A.inv - A.inv * U * (I + B * V * Ainv * U).inv * B * V * Ainv.
//
// When applied to F = H + Z P Z' the theorem gives
//
// Finv = Hinv - Hinv * Z * (I + P Z' Hinv Z).inv * P * Z' * Hinv
//
// This helps because H is diagonal. The only matrix that needs to be
// inverted is (I + PZ'HinvZ), which is a state x state matrix.
//
// We don't compute Finv directly, we compute Finv * prediction_error.
// observation_precision refers to the precision of the observed subset.
DiagonalMatrix observation_precision(observed.select(
1.0 / model_->observation_variance(time_index()).diag()));
// Set inner_matrix to I + P * Z' * Hinv * Z
SpdMatrix ZTZ = observation_coefficient_subset.inner(
observation_precision.diag());
// Note: the product of two SPD matrices need not be symmetric.
Matrix inner_matrix = state_variance() * ZTZ;
inner_matrix.diag() += 1.0;
LU inner_lu(inner_matrix);
// inner_inv_P is inner.inv() * state_variance. This matrix need not be
// symmetric.
Matrix inner_inv_P = inner_lu.solve(state_variance());
Matrix HinvZ = observation_precision *
observation_coefficient_subset.dense();
set_scaled_prediction_error(
observation_precision * prediction_error()
- HinvZ * inner_inv_P * HinvZ.Tmult(prediction_error()));
// The log determinant of F.inverse is the negative log of det(H + ZPZ').
// That determinant can be computed using the "matrix determinant lemma,"
// which says det(A + UV') = det(I + V' * A.inv * U) * det(A)
//
// https://en.wikipedia.org/wiki/Matrix_determinant_lemma#Generalization
//
// With F = H + ZPZ', and setting A = H, U = Z, V' = PZ'.
// Then det(F) = det(I + PZ' * Hinv * Z) * det(H)
set_forecast_precision_log_determinant(
observation_precision.logdet() - inner_lu.logdet());
// The Kalman gain is: K = T * P * Z' * Finv.
// Substituting the expression for Finv from above gives
//
// K = T * P * Z' *
// (Hinv - Hinv * Z * (I + P Z' Hinv Z).inv * P * Z' * Hinv)
Matrix ZtHinv = HinvZ.transpose();
set_kalman_gain(transition * state_variance() *
(ZtHinv - ZTZ * inner_inv_P * ZtHinv));
}
int main(int argc, char *args[])
{
int n = 8;
cerr << "argc:"<<argc<<endl;
if(argc!=n){
hyperdatasetUsage(args[0]);
exit(1);
}
int rands = 0;
rands = atoi(args[1]);
if(rands!=0)
srand(rands);
else{
rands = time(0);
srand(rands);
cout << "ny seed: " << rands << endl;
}
//1. load the example file
NEATsettings * s = new NEATsettings();
ifstream ifs(args[2],ios::in);
ifs>>s;
ifs.close();
//2. generate the pop
TransferFunctions * tfs = new TransferFunctions(s);
Population * pop = makePopulation(args[3],s,tfs);
//3. then the selector
Selector * sel = makeSelector(args[4]);
int g = atoi(args[5]);
int iter = atoi(args[6]);
string dfile = args[7];
DataSet * set = new DataSet(false,dfile,0.0);
FitnessEvaluator * de = new DatasetHyperNEAT(s,tfs,set);
Evaluator * ev = new Evaluator(de);
LocalReproducer * rp = new LocalReproducer();
int gc = 0; double sum=0; double sum2=0; double sum3=0;
Phenotype * cbest = NULL;
Phenotype * sbest = NULL;
Phenotype * best = NULL;
int osize=pop->getMembers()->size();
int ss=0;
double ocomp=0;
time_t startt;
double timesum = 0;
for(int i2=0;i2<iter;i2++){
gc = 0;
best = NULL;
startt = time(0);
for(int i=0;i<g;i++){
gc++;
// cout << "before eval.." << endl;
ev->evaluate(pop->getMembers(),pop->getMembers()->size());
// cout << "after eval.." << endl;
if((unsigned int)pop->getOriginalSize()!=pop->getMembers()->size())
cout << "size not right after eval.." << endl;
ss=pop->getSpecies()->size();
// cout << "before pop juggling.." << endl;
pop->updateSpeciesStats();
pop->sortmembers();
pop->sortspecies();
cbest = pop->getCopyOfCurrentBest();
de->f(cbest);
// cout << "before best juggling.." << endl;
if(best==NULL){
best = cbest;
}else if(cbest->getFitness()>best->getFitness()){
delete best;
best = cbest;
}else{
delete cbest;
}
// cout << "after best juggling.." << endl;
cout << "gen: "<<i<<" best: " << best->getFitness() << "worst: " << pop->getMembers()->at(pop->getMembers()->size()-1)->getFitness() << endl;
if(((DatasetHyperNEAT*)de)->done(best)){
i = g;
// cout << best;
((DatasetHyperNEAT*)de)->runTest(best);
}
// cout << "after xor testing.." << endl;
sel->select(pop,0);
rp->reproduce(pop);
// cout << "after select and reproduce.." << endl;
}
if(sbest==NULL)
sbest = best;
else if(best->getFitness()>sbest->getFitness())
sbest = best;
time_t ft = time(0)-startt;
timesum += ft;
cout << "run: "<<i2+1<<" gc: "<<gc<<" maxfitness: " << pop->getHighestFitness()
<< " sbest fitness: " << sbest->getFitness() << " species: " << ss
<< " time: " << ft << " time/gen: " << (double)ft/(double)gc << endl;
sum += gc;
sum2 += best->getGenome()->extrons();
sum3 += best->getGenome()->countHidden();
if(i2!=iter){
if(pop->getOriginalSeed()!=NULL){
Genome * oseed = pop->getOriginalSeed();
int oelitism = pop->getOriginalInitialElitism();
ocomp = pop->getOcomp();
delete pop;
pop = new Population(s,ocomp,tfs);
// oseed->setTfs(pop->getTfs());
pop->genesis(oseed,osize,oelitism);
}else{
pop->resetSpawn();
}
}
}
((DatasetHyperNEAT*)de)->runTest(sbest);
cout << ((DatasetHyperNEAT*)de)->output(sbest);
}
Vector NeRegSuf::xty(const Selector &inc)const{
return inc.select(xty_);}
SpdMatrix NeRegSuf::xtx(const Selector &inc)const{
reflect();
return inc.select(xtx_);
}
Vector QrRegSuf::xty(const Selector &inc)const{
// if(!current) refresh_qr();
return inc.select(xty());
}
SpdMatrix QrRegSuf::xtx(const Selector &inc)const{
// if(!current) refresh_qr();
return inc.select(xtx());
}
