void PiMax::setSensorTeaching(const matrix::Matrix& teaching){ assert(teaching.getM() == number_sensors && teaching.getN() == 1); // calculate the a_teaching, // that belongs to the distal teaching value by the inverse model. a_teaching = (A.pseudoInverse() * (teaching-b)).mapP(0.95, clip); intern_isTeaching=true; }
void PiMax::setMotorTeaching(const matrix::Matrix& teaching){ assert(teaching.getM() == number_motors && teaching.getN() == 1); // Note: through the clipping the otherwise effectless // teaching with old motor value has now an effect, // namely to drive out of the saturation region. a_teaching= teaching.mapP(0.95,clip); intern_isTeaching=true; }
void SosAvgGrad::setC(const matrix::Matrix& _C){ assert(C.getM() == _C.getM() && C.getN() == _C.getN()); C=_C; }
void SosAvgGrad::setA(const matrix::Matrix& _A){ assert(A.getM() == _A.getM() && A.getN() == _A.getN()); A=_A; }
void SosAvgGrad::setS(const matrix::Matrix& _S){ assert(S.getM() == _S.getM() && S.getN() == _S.getN()); S=_S; }
void PiMax::seth(const matrix::Matrix& _h){ assert(h.getM() == _h.getM() && h.getN() == _h.getN()); h=_h; }
void PiMax::setA(const matrix::Matrix& _A){ assert(A.getM() == _A.getM() && A.getN() == _A.getN()); A=_A; }
void PiMax::setC(const matrix::Matrix& _C){ assert(C.getM() == _C.getM() && C.getN() == _C.getN()); C=_C; }
void RandomDyn::setC(const matrix::Matrix& _C){ assert(C.getM() == _C.getM() && C.getN() == _C.getN()); C=_C; }
void RandomDyn::seth(const matrix::Matrix& _h){ assert(h.getM() == _h.getM() && h.getN() == _h.getN()); h=_h; }
static void keepMatrixTraceUp(matrix::Matrix& m){ int l = std::min((short unsigned int)2,std::min(m.getM(), m.getN())); for(int i=0; i<l; i++){ if(m.val(i,i)<0.8) m.val(i,i)+=0.001; } }
virtual void seth(const matrix::Matrix& _h){ assert(h.getM() == _h.getM() && h.getN() == _h.getN()); h=_h; }
virtual void setC(const matrix::Matrix& _C){ assert(C.getM() == _C.getM() && C.getN() == _C.getN()); C=_C; }
double CuriosityLoop::updatePrediction(const matrix::Matrix& smHist, const matrix::Matrix& s, const matrix::Matrix& m, int phase){ matrix::Matrix sm = s.above(m); matrix::Matrix f; f.set(1,1); f.val(0,0) = 1; sm = sm.above(f); //1. Go through the predictions of this predictor determining the prediction errors at each dimension. matrix::Matrix error; error.set(smHist.getM(), 1); prediction_error = 0; for(int i = 0; i < prediction.getM(); i++){ if(pOutput.val(i,0) == 1){ error.val(i,0) = prediction.val(i,0) - sm.val(i,0); prediction_error = prediction_error + pow(error.val(i,0),2); // cout << error << "predictionError\n"; } else{ // cout << "This dimension is not predicted, and does not count towards the error\n"; error.val(i,0) = 0; //prediction_error = prediction_error + error.val(i,0); } } parent_error.val(phase,0) = prediction_error; //2. Change the weights by the delta rule. for(int i = 0; i < prediction.getM(); i++){//to for(int j = 0; j < predictorWeights.getN(); j++){//from // predictorWeights.val(i,j) = predictorWeights.val(i,j) - 0.00001*error.val(i,0)*smHist.val(j,0); predictorWeights.val(i,j) = predictorWeights.val(i,j) - 0.0001*error.val(i,0)*smHist.val(j,0); if(predictorWeights.val(i,j) > 10) predictorWeights.val(i,j) = 10; else if(predictorWeights.val(i,j) < -10) predictorWeights.val(i,j) = -10; } } prediction_error_time_average = 0.9999*prediction_error_time_average + (1-0.9999)*prediction_error; //Update the fitness of this predictor based on the instantaneous reduction / increase in prediction error. this->fitness = 0.1 + 100*(prediction_error_time_average - old_prediction_error_time_average); old_prediction_error_time_average = prediction_error_time_average; //cout << fitness << " "; //Improve the method of determining this gradient later! //UPDATE THE UNRESTRICTED PREDICTOR NOW AS WELL, ALWAYS... //1. Go through the predictions of this UNRESTRICTED predictor determining the prediction errors at each dimension. matrix::Matrix uError; uError.set(smHist.getM(), 1); uPrediction_error = 0; for(int i = 0; i < uPrediction.getM(); i++){ if(uPOutput.val(i,0) == 1){ uError.val(i,0) = uPrediction.val(i,0) - sm.val(i,0); uPrediction_error = uPrediction_error + pow(uError.val(i,0),2); // cout << error << "predictionError\n"; } else{ // cout << "This dimension is not predicted, and does not count towards the error\n"; uError.val(i,0) = 0; //prediction_error = prediction_error + error.val(i,0); } } //cout << "phase = " << phase << "\n"; offspring_error.val(phase,0) = uPrediction_error; //2. Change the weights by the delta rule. for(int i = 0; i < uPrediction.getM(); i++){ for(int j = 0; j < uPredictorWeights.getN(); j++){ uPredictorWeights.val(i,j) = uPredictorWeights.val(i,j) - 0.0001*uError.val(i,0)*smHist.val(j,0); if(uPredictorWeights.val(i,j) > 10) uPredictorWeights.val(i,j) = 10; else if(uPredictorWeights.val(i,j) < -10) uPredictorWeights.val(i,j) = -10; } } //************************UNRESTRICTED PREDICTOR CODE **************************** return this->fitness; };
void SeMoX::setSensorTeaching(const matrix::Matrix& teaching){ assert(teaching.getM() == number_sensors && teaching.getN() == 1); // calculate the y_teaching, that belongs to the distal teaching value by the inverse model. y_teaching = (A.pseudoInverse(0.001) * (teaching-B)).mapP(0.95, clip); intern_useTeaching=true; }