/** * Modifies this map through a single backpropagation iteration using the given error values on * the output nodes. * * @param error */ public void train(List<Double> error, double learningRate) { RealVector eOut = new ArrayRealVector(error.size()); for (int i : series(error.size())) eOut.setEntry(i, error.get(i)); // * gHidden: delta for the non-bias nodes of the hidden layer gHidden.setSubVector(0, stateHidden.getSubVector(0, n)); // optimize for (int i : Series.series(gHidden.getDimension())) gHidden.setEntry(i, activation.derivative(gHidden.getEntry(i))); eHiddenL = weights1.transpose().operate(eOut); eHidden.setSubVector(0, eHiddenL.getSubVector(0, h)); for (int i : series(h)) eHidden.setEntry(i, eHidden.getEntry(i) * gHidden.getEntry(i)); weights1Delta = MatrixTools.outer(eOut, stateHidden); weights1Delta = weights1Delta.scalarMultiply(-1.0 * learningRate); // optimize weights0Delta = MatrixTools.outer(eHidden, stateIn); weights0Delta = weights0Delta.scalarMultiply(-1.0 * learningRate); weights0 = weights0.add(weights0Delta); weights1 = weights1.add(weights1Delta); }
/** * Solve an estimation problem using a least squares criterion. * * <p>This method set the unbound parameters of the given problem starting from their current * values through several iterations. At each step, the unbound parameters are changed in order to * minimize a weighted least square criterion based on the measurements of the problem. * * <p>The iterations are stopped either when the criterion goes below a physical threshold under * which improvement are considered useless or when the algorithm is unable to improve it (even if * it is still high). The first condition that is met stops the iterations. If the convergence it * not reached before the maximum number of iterations, an {@link EstimationException} is thrown. * * @param problem estimation problem to solve * @exception EstimationException if the problem cannot be solved * @see EstimationProblem */ @Override public void estimate(EstimationProblem problem) throws EstimationException { initializeEstimate(problem); // work matrices double[] grad = new double[parameters.length]; ArrayRealVector bDecrement = new ArrayRealVector(parameters.length); double[] bDecrementData = bDecrement.getDataRef(); RealMatrix wGradGradT = MatrixUtils.createRealMatrix(parameters.length, parameters.length); // iterate until convergence is reached double previous = Double.POSITIVE_INFINITY; do { // build the linear problem incrementJacobianEvaluationsCounter(); RealVector b = new ArrayRealVector(parameters.length); RealMatrix a = MatrixUtils.createRealMatrix(parameters.length, parameters.length); for (int i = 0; i < measurements.length; ++i) { if (!measurements[i].isIgnored()) { double weight = measurements[i].getWeight(); double residual = measurements[i].getResidual(); // compute the normal equation for (int j = 0; j < parameters.length; ++j) { grad[j] = measurements[i].getPartial(parameters[j]); bDecrementData[j] = weight * residual * grad[j]; } // build the contribution matrix for measurement i for (int k = 0; k < parameters.length; ++k) { double gk = grad[k]; for (int l = 0; l < parameters.length; ++l) { wGradGradT.setEntry(k, l, weight * gk * grad[l]); } } // update the matrices a = a.add(wGradGradT); b = b.add(bDecrement); } } try { // solve the linearized least squares problem RealVector dX = new LUDecompositionImpl(a).getSolver().solve(b); // update the estimated parameters for (int i = 0; i < parameters.length; ++i) { parameters[i].setEstimate(parameters[i].getEstimate() + dX.getEntry(i)); } } catch (InvalidMatrixException e) { throw new EstimationException("unable to solve: singular problem"); } previous = cost; updateResidualsAndCost(); } while ((getCostEvaluations() < 2) || (Math.abs(previous - cost) > (cost * steadyStateThreshold) && (Math.abs(cost) > convergence))); }