void InvertNChannelControllerHebbH::learnHebb(const matrix::Matrix& context_sensors, const matrix::Matrix& h_update){ // preprocess context sensors Matrix c_sensors = context_sensors; for (int i=0;i<number_context_sensors;i++){ if (c_sensors.val(i,0)<0.15) { c_sensors.val(i,0)=0; // IR's should only have positive values } } // adapt hebbian weights for (uint i=0; i<number_motors; i++){ for (uint j=0; j<(uint)number_context_sensors; j++){ if (i==j){ // TODO: remove (it is just for testing) double dp= eps_hebb* h_update.val(i,0) * c_sensors.val(j,0) *(1 - pow(p.val(i,j),2)); // std::cout<<eps_hebb<<"*"<<h_update.val(i,0)<<" * "<<c_sensors.val(j,0)<<std::endl; p.val(i,j)+=dp; } } } /* // remove this !!! (just a test) for (int i=0; i<number_motors; i++){ for (int j=0; j<number_context_sensors; j++){ if ((j==0) || (j==1)){ p.val(i,j)=-0.1; } else { p.val(i,j)=0.1; } } } */ }
/** * predict the update of h based on the actual context sensors * @param context_sensors prediction is based on these sensors */ matrix::Matrix InvertNChannelControllerHebbH::predictHebb(const matrix::Matrix& context_sensors){ // preprocess context sensors Matrix c_sensors = context_sensors; for (int i=0;i<number_context_sensors;i++){ if (c_sensors.val(i,0)<0.15) { c_sensors.val(i,0)=0; // IR's should only have positive values } } Matrix pred_h_update(number_motors,1) ; for (unsigned int k = 0; k < number_motors; k++) { pred_h_update.val(k,0)=0; } for (uint i=0; i<number_motors; i++){ for (uint j=0; j<(uint)number_context_sensors; j++){ pred_h_update.val(i,0)+= p.val(i,j) * context_sensors.val(j,0); } } return pred_h_update; }
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; } }
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; };