void singlePeCalMin() {
vector<inputValues> b; // Blank runs
vector<inputValues> t; // Total runs
vector<outputValues> o;
vector<string> cutString(cut_b.size(), "Min > ");
// Create strings to perform cuts
for (int i = 0; i < cut_b.size(); i++) { cutString[i].append(to_string(cut_b[i])); }
// Get input and output values from .root spectra
for (int i = 0; i < fileNames_b.size(); i++) {
b.push_back(getInputValues(fileNames_b[i], pulseWidths_b[i], cutString[i]));
for (int j = 0; j < fileNames_t.size(); j++) {
t.push_back(getInputValues(fileNames_t[j], pulseWidths_t[j], cutString[i]));
o.push_back(getOutputValues(b[i], t[j]));
}
}
makePlots(b, t, o, cut_b[0]);
// Debugging
cout << "BG E[T}: " << b[0].e << " BG a/n: " << b[0].a / b[0].n << endl;
cout << setw(12) << "a/n " << setw(12) << "E[L] " << setw(12)
<< "E[Psi]" << setw(12) << "E[T]" << endl;
for ( int i = 0; i < t.size(); i++) {
cout << setw(12) << (float)t[i].a / t[i].n << setw(12) << o[i].eL << setw(12)
<< o[i].ePsi << setw(12) << t[i].e << endl;
}
}
void singlePeVaryMode() {
vector<inputValues> b; // Blank runs
vector<inputValues> t; // Total runs
vector<outputValues> o;
vector<string> cutString(cut_b.size(), "Charge > ");
// Create strings to perform cuts - I would be able to use to_string if C++11 worked
for (int i = 0; i < cut_b.size(); i++) {
ostringstream ss;
ss << cut_b[i];
cutString[i].append(ss.str());
}
// Get input and output values from .root spectra
for (int i = 0; i < fileNames_b.size(); i++) {
b.push_back(getInputValues(fileNames_b[i], pulseWidths_b[i], useChargeMode_b[i], cutString[i]));
for (int j = 0; j < fileNames_t.size(); j++) {
if (useChargeMode_t[j] == useChargeMode_b[i]) {
t.push_back(getInputValues(fileNames_t[j], pulseWidths_t[j], useChargeMode_t[j], cutString[i]));
o.push_back(getOutputValues(b[i], t[j]));
}
}
}
makePlots(b, t, o);
}
double strategytdexp4::boardValue( const board& brd, const hash_map<string,int>* context ) const
{
// get the inputs from the board, assuming the player holds the dice
vector<double> inputs = getInputValues( brd, brd.perspective() );
// calculate the middle layer node values
vector<double> middles = getMiddleValues( inputs );
// calculate the output node values from the middles
double probWin = getOutputProbValue( middles );
double probCondGammonWin = getOutputGammonWinValue( middles, brd );
double probCondGammonLoss = getOutputGammonLossValue( middles, brd );
double probCondBgWin = getOutputBackgammonWinValue( middles, brd );
double probCondBgLoss = getOutputBackgammonLossValue( middles, brd );
// calculate the expected number of points the player will win. probWin
// corresponds to the probability of a win (any win); probCondGammon
// corresponds to the probability of a gammon conditional on a win;
// probCondGammonLoss corresponds to the probability of a gammon loss
// conditional on a loss. Ditto for backgammon win/loss conditional probs.
double ppg = probWin * ( 1 * ( 1 - probCondGammonWin ) + 2 * probCondGammonWin * ( 1 - probCondBgWin ) + 3 * probCondGammonWin * probCondBgWin )
- ( 1 - probWin ) * ( 1 * ( 1 - probCondGammonLoss ) + 2 * probCondGammonLoss * ( 1 - probCondBgLoss ) + 3 * probCondGammonLoss * probCondBgLoss );
// the ppg is always from player 0's perspective. But we want it from the board
// perspective.
if( brd.perspective() == 0 )
return ppg;
else
return -ppg;
}
MainWindow::MainWindow(QWidget *parent) :
QMainWindow(parent),
ui(new Ui::MainWindow)
{
ui->setupUi(this);
QObject::connect(this->ui->minimizeButton,SIGNAL(clicked()),SLOT(getInputValues()));
QObject::connect(this,SIGNAL(inputComplete()),SLOT(printResult()));
}
void strategytdexp4::update( const board& oldBoard, const board& newBoard )
{
// get the values from the old board
vector<double> oldInputs = getInputValues( oldBoard, oldBoard.perspective() );
vector<double> oldMiddles = getMiddleValues( oldInputs );
double oldProbOutput = getOutputProbValue( oldMiddles );
double oldGammonWinOutput = getOutputGammonWinValue( oldMiddles, oldBoard );
double oldGammonLossOutput = getOutputGammonLossValue( oldMiddles, oldBoard );
double oldBgWinOutput = getOutputBackgammonWinValue( oldMiddles, oldBoard );
double oldBgLossOutput = getOutputBackgammonLossValue( oldMiddles, oldBoard );
// calculate all the partial derivatives we'll need (of output node values
// to the various weights)
int i, j;
// then do derivs of the prob nodes to each of the middle->input weights (that's a 2d array), and the derivs of each of
// the middle nodes to its weights->inputs.
double mid, input, v1, v2, v3, v4, v5;
for( i=0; i<nMiddle; i++ )
{
mid = oldMiddles.at(i);
v1 = outputProbWeights.at(i);
v2 = outputGammonWinWeights.at(i);
v3 = outputGammonLossWeights.at(i);
v4 = outputBackgammonWinWeights.at(i);
v5 = outputBackgammonLossWeights.at(i);
probDerivs.at(i) = mid * oldProbOutput * ( 1 - oldProbOutput );
gamWinDerivs.at(i) = mid * oldGammonWinOutput * ( 1 - oldGammonWinOutput );
gamLossDerivs.at(i) = mid * oldGammonLossOutput * ( 1 - oldGammonLossOutput );
bgWinDerivs.at(i) = mid * oldBgWinOutput * ( 1 - oldBgWinOutput );
bgLossDerivs.at(i) = mid * oldBgLossOutput * ( 1 - oldBgLossOutput );
for( j=0; j<198; j++ )
{
input = oldInputs.at(j);
probInputDerivs.at(i).at(j) = v1 * input * oldProbOutput * ( 1 - oldProbOutput ) * mid * ( 1 - mid );
gamWinInputDerivs.at(i).at(j) = v2 * input * oldGammonWinOutput * ( 1 - oldGammonWinOutput ) * mid * ( 1 - mid );
gamLossInputDerivs.at(i).at(j) = v3 * input * oldGammonLossOutput * ( 1 - oldGammonLossOutput ) * mid * ( 1 - mid );
bgWinInputDerivs.at(i).at(j) = v4 * input * oldBgWinOutput * ( 1 - oldBgWinOutput ) * mid * ( 1 - mid );
bgLossInputDerivs.at(i).at(j) = v5 * input * oldBgLossOutput * ( 1 - oldBgLossOutput ) * mid * ( 1 - mid );
}
probInputDerivs.at(i).at(198) = v1 * oldProbOutput * ( 1 - oldProbOutput ) * mid * ( 1 - mid );
gamWinInputDerivs.at(i).at(198) = v2 * oldGammonWinOutput * ( 1 - oldGammonWinOutput ) * mid * ( 1 - mid );
gamLossInputDerivs.at(i).at(198) = v3 * oldGammonLossOutput * ( 1 - oldGammonLossOutput ) * mid * ( 1 - mid );
bgWinInputDerivs.at(i).at(198) = v4 * oldBgWinOutput * ( 1 - oldBgWinOutput ) * mid * ( 1 - mid );
bgLossInputDerivs.at(i).at(198) = v5 * oldBgLossOutput * ( 1 - oldBgLossOutput ) * mid * ( 1 - mid );
}
probDerivs.at(nMiddle) = oldProbOutput * ( 1 - oldProbOutput );
gamWinDerivs.at(nMiddle) = oldGammonWinOutput * ( 1 - oldGammonWinOutput );
gamLossDerivs.at(nMiddle) = oldGammonLossOutput * ( 1 - oldGammonLossOutput );
bgWinDerivs.at(nMiddle) = oldBgWinOutput * ( 1 - oldBgWinOutput );
bgLossDerivs.at(nMiddle) = oldBgLossOutput * ( 1 - oldBgLossOutput );
// now calculate the next estimate of the outputs. That's known if the game is over; otherwise we use the network's
// estimate on the new board as a proxy. Note that the update fn is only ever called by the game when the player wins, not when
// the player loses, just because the winner is the last one to play. But we need to train on prob of losing a gammon too,
// so we flip board perspective and train again based on that.
bool trainGammonLoss = true;
bool trainGammonWin = true;
bool trainBgLoss = true;
bool trainBgWin = true;
double newProbOutput, newGammonWinOutput, newGammonLossOutput, newBgWinOutput, newBgLossOutput;
if( newBoard.bornIn0Raw() == 15 )
{
trainGammonLoss = false; // can't train the conditional prob of a gammon loss if there isn't a loss
trainBgLoss = false; // ditto for backgammon loss
newProbOutput = 1.;
if( newBoard.bornIn1Raw() == 0 ) // gammon or backgammon
{
newGammonWinOutput = 1.;
vector<int> checks( newBoard.checkers1Raw() );
bool foundOne = newBoard.hit1Raw() > 0;
if( !foundOne )
{
for( int i=0; i<6; i++ )
if( checks.at(i) > 0 )
{
foundOne = true;
break;
}
}
newBgWinOutput = foundOne ? 1 : 0;
}
else
{
newGammonWinOutput = 0.;
trainBgWin = false; // no gammon win so can't train conditional bg win prob
}
}
else if( newBoard.bornIn1Raw() == 15 )
{
trainGammonWin = false;
trainBgWin = false;
newProbOutput = 0.;
if( newBoard.bornIn0Raw() == 0 ) // gammon loss or backgammon loss
{
newGammonLossOutput = 1;
vector<int> checks( newBoard.checkers0Raw() );
bool foundOne = newBoard.hit0Raw() > 0;
if( !foundOne )
{
for( int i=18; i<24; i++ )
if( checks.at(i) > 0 )
{
foundOne = true;
break;
}
}
newBgLossOutput = foundOne ? 1 : 0;
}
else
{
newGammonLossOutput = 0;
trainBgLoss = false;
}
}
else
{
// estimate from the new board's outputs, remembering that after the move is done,
// the other player gets the dice.
vector<double> midVals( getMiddleValues( getInputValues( newBoard, !newBoard.perspective() ) ) );
newProbOutput = getOutputProbValue( midVals );
newGammonWinOutput = getOutputGammonWinValue( midVals, newBoard );
newGammonLossOutput = getOutputGammonLossValue( midVals, newBoard );
newBgWinOutput = getOutputBackgammonWinValue( midVals, newBoard );
newBgLossOutput = getOutputBackgammonLossValue( midVals, newBoard );
}
// train the nodes as appropriate
for( i=0; i<nMiddle; i++ )
{
outputProbWeights.at(i) += alpha * ( newProbOutput - oldProbOutput ) * probDerivs.at(i);
if( trainGammonWin )
outputGammonWinWeights.at(i) += alpha * ( newGammonWinOutput - oldGammonWinOutput ) * gamWinDerivs.at(i);
if( trainGammonLoss )
outputGammonLossWeights.at(i) += alpha * ( newGammonLossOutput - oldGammonLossOutput ) * gamLossDerivs.at(i);
if( trainBgWin )
outputBackgammonWinWeights.at(i) += alpha * ( newBgWinOutput - oldBgWinOutput ) * bgWinDerivs.at(i);
if( trainBgLoss )
outputBackgammonLossWeights.at(i) += alpha * ( newBgLossOutput - oldBgLossOutput ) * bgLossDerivs.at(i);
for( j=0; j<199; j++ )
{
middleWeights.at(i).at(j) += beta * ( newProbOutput - oldProbOutput ) * probInputDerivs.at(i).at(j);
if( trainGammonWin )
middleWeights.at(i).at(j) += beta * ( newGammonWinOutput - oldGammonWinOutput ) * gamWinInputDerivs.at(i).at(j);
if( trainGammonLoss )
middleWeights.at(i).at(j) += beta * ( newGammonLossOutput - oldGammonLossOutput ) * gamLossInputDerivs.at(i).at(j);
if( trainBgWin )
middleWeights.at(i).at(j) += beta * ( newBgWinOutput - oldBgWinOutput ) * bgWinInputDerivs.at(i).at(j);
if( trainBgLoss )
middleWeights.at(i).at(j) += beta * ( newBgLossOutput - oldBgLossOutput ) * bgLossInputDerivs.at(i).at(j);
}
}
outputProbWeights.at(nMiddle) += alpha * ( newProbOutput - oldProbOutput ) * probDerivs.at(nMiddle);
if( trainGammonWin )
outputGammonWinWeights.at(nMiddle) += alpha * ( newGammonWinOutput - oldGammonWinOutput ) * gamWinDerivs.at(nMiddle);
if( trainGammonLoss )
outputGammonLossWeights.at(nMiddle) += alpha * ( newGammonLossOutput - oldGammonLossOutput ) * gamLossDerivs.at(nMiddle);
if( trainBgWin )
outputBackgammonWinWeights.at(nMiddle) += alpha * ( newBgWinOutput - oldBgWinOutput ) * bgWinDerivs.at(nMiddle);
if( trainBgLoss )
outputBackgammonLossWeights.at(nMiddle) += alpha * ( newBgLossOutput - oldBgLossOutput ) * bgLossDerivs.at(nMiddle);
}
