void report_multiplication_error(const SparseKalmanMatrix *T, const SparseVector &Z, bool new_time, double fraction_in_initial_period, const VEC &v) { ostringstream err; int state_dim = T->nrow(); err << "incompatible sizes in AccumulatorTransitionMatrix multiplication" << endl << "T.nrow() = " << state_dim << endl << "Z.size() = " << Z.size() << endl << "v.size() = " << v.size() << endl << "The first two should match. The last should be two more " << "than the others" << endl; report_error(err.str()); }
Vec RQR_Multiply(const VECTOR &v, const SparseKalmanMatrix &RQR, const SparseVector &Z, double H) { int state_dim = Z.size(); if(v.size() != state_dim + 2) { report_error("wrong sizes in RQR_Multiply"); } // Partition v = [eta, epsilon, 0] ConstVectorView eta(v, 0, state_dim); double epsilon = v[state_dim]; // Partition this Vec RQRZ = RQR * Z.dense(); double ZRQRZ_plus_H = Z.dot(RQRZ) + H; Vec ans(v.size()); VectorView(ans, 0, state_dim) = (RQR * eta).axpy(RQRZ, epsilon); ans[state_dim] = RQRZ.dot(eta) + ZRQRZ_plus_H * epsilon; return ans; }
{ TEST_EQUAL(sv2.size(), 8) } END_SECTION START_SECTION((void resize(size_type newsize))) { sv2.resize(10); TEST_EQUAL(sv2.size(),10) } END_SECTION START_SECTION((SparseVector& operator=(const SparseVector &source))) { sv2=sv; TEST_EQUAL(sv2.size(), 8) } END_SECTION START_SECTION((bool operator==(const SparseVector &rhs) const )) { SparseVector<double> sv3(sv); TEST_EQUAL((sv3==sv), true) } END_SECTION START_SECTION((bool operator<(const SparseVector &rhs) const )) { SparseVector<double> sv3(sv); sv3[0]=-1.23;
void SoftmaxPolicyPlayout::updateProbabilitiesBeforeAction(const Go::Board *board, Color player) { // check the last move size_t historySize = board->getHistoryCount(); const Board::MoveChangeEntry *lastMove = board->getHistory(historySize-1); // enemy's move const Board::MoveChangeEntry *secondLastMove = board->getHistory(historySize-2); // last my move const Board::MoveChangeEntry *thirdLastMove = board->getHistory(historySize-3); double *table = NULL; SparseVector *featureTable = NULL; if (player == BLACK) { table = m_probTableBlack; featureTable = m_featureTableBlack; } else { table = m_probTableWhite; featureTable = m_featureTableWhite; } if (lastMove && lastMove->m_putPos != PASS) table[lastMove->m_putPos] = 0; if (secondLastMove && secondLastMove->m_putPos != PASS) table[secondLastMove->m_putPos] = 0; PointSet updatedMoves; // reset static features which are set in the last time of player PointSet &previousSetMoves = player == BLACK ? m_toResetStaticFeaturesMovesBlack : m_toResetStaticFeaturesMovesWhite; for (size_t i=0; i<previousSetMoves.size(); i++) { //cerr << previousSetMoves[i] << ":" << featureTable[previousSetMoves[i]].toString() << endl; Point p = previousSetMoves[i]; StandardFeatureExtractor::clearStaticFeatures(featureTable[p]); if (board->isColor(p, FREE) && (board->getNeighborEmptyCount(p) >= 1 || board->checkLegalHand(p, player, Board::flipColor(player)) == Board::PUT_LEGAL)) { updatedMoves.insert(p,p); } } previousSetMoves.clear(); // enumerate pattern update moves PointSet patternUpdateMoves; enumerateMovesOfPatternChanged(patternUpdateMoves, board, player, lastMove, secondLastMove, false); // enumerate moves that their status will change PointSet toBeLegal, toBeIllegal; PointSet legalMovesSet; toBeLegal.clear(); toBeIllegal.clear(); enumerateToBeLegalAndIllegalMoves(board, table, player, toBeLegal, toBeIllegal, legalMovesSet, false); PointSet &staticFeatureUpdateMoves = previousSetMoves; if (lastMove == NULL) return; // no further old moves // update static features m_featureExtractor.updateStaticFeaturesForAllMovesWithoutClearOldFeatures(board, player, legalMovesSet, featureTable, staticFeatureUpdateMoves); // update patterns PointSet::ListConstIterator it, end = patternUpdateMoves.end(); for (it = patternUpdateMoves.begin(); it!=end; it++) { SparseVector &targetFeatures = featureTable[*it]; assert (board->isColor(*it, FREE)); m_featureExtractor.updatePatternFeature(targetFeatures, board, *it, player); //if (board->checkLegalHand(*it, player, Board::flipColor(player)) == Board::PUT_LEGAL) { if ((!toBeIllegal.contains(*it) && table[*it] != 0) || toBeLegal.contains(*it)) { updatedMoves.insert(*it,*it); } } end = toBeLegal.end(); for (it = toBeLegal.begin(); it!=end; it++) { updatedMoves.insert(*it,*it); // check capturing, atari features } end = staticFeatureUpdateMoves.end(); for (it=staticFeatureUpdateMoves.begin(); it!=end; it++) { //table[*it] = m_expFeatureWeights.multiplyAll(featureTable[*it]); updatedMoves.insert(*it,*it); } end = updatedMoves.end(); for (it=updatedMoves.begin(); it!=end; it++) { //table[*it] = m_expFeatureWeights.multiplyAll(featureTable[*it]); table[*it] = fmath::expd(m_featureWeights.dot(featureTable[*it])); //table[*it] = exp(m_featureWeights.dot(featureTable[*it])); } end = toBeIllegal.end(); for (it = toBeIllegal.begin(); it!=end; it++) { table[*it] = 0; StandardFeatureExtractor::clearStaticFeatures(featureTable[*it]); } #ifdef STRICT_CHECK for (int y=0; y<board->getSize(); y++) { for (int x=0; x<board->getSize(); x++) { Point p = board->xyToPoint(x,y); Board::PutType err = board->checkLegalHand(p, player, Board::flipColor(player)); assert (!(err == Board::PUT_LEGAL && table[p] == 0) && !(err != Board::PUT_LEGAL && table[p] != 0)); if (err == Board::PUT_LEGAL) { double v = m_expFeatureWeights.multiplyAll(featureTable[p]); if (fabs(v-table[p]) >= 0.001) { cerr << "v = " << v << endl; cerr << "table[p] = " << table[p] << endl; } assert (fabs(v-table[p]) < 0.001); SparseVector vec; m_featureExtractor.extractFromStateAndAction(vec, board, p, player); // for (int i=0; i<vec.size(); i++) { // if (vec[i] > 0 && vec[i] < StandardFeatureExtractor::STATIC_FEATURE_SIZE) { // vec.erase(i);i--; // } // } for (int i=0; i<featureTable[p].size(); i++) { if (featureTable[p][i] > 0 && featureTable[p][i] < 9) { featureTable[p].erase(i);i--; } } if (vec.size() != featureTable[p].size()) { // cerr << "updated moves:" << endl; // for (int i=0; i<staticFeatureUpdateMoves.size(); i++) { // cerr << staticFeatureUpdateMoves[i] << ":" << featureTable[staticFeatureUpdateMoves[i]].toString() << endl; // } cerr << vec.toString() << endl; cerr << featureTable[p].toString() << endl; cerr << "point " << p << endl; cerr << "turn = " << player << endl; cerr << "last move = " << board->getLastMove() << endl; cerr << "second last move = " << secondLastMove->m_putPos << endl; board->printToErr(); printProbabilityTableToErr(board); } assert (vec.size() == featureTable[p].size()); } } } #endif }