List of usage examples for org.apache.commons.math3.optim.nonlinear.scalar.gradient BoundedNonLinearConjugateGradientOptimizer BoundedNonLinearConjugateGradientOptimizer
public BoundedNonLinearConjugateGradientOptimizer(final Formula updateFormula, ConvergenceChecker<PointValuePair> checker)
From source file:gdsc.smlm.fitting.nonlinear.MaximumLikelihoodFitter.java
public FitStatus fit(int n, double[] y, double[] y_fit, double[] a, double[] a_dev, double[] error, double noise) { numberOfFittedPoints = n;/* w ww. j a v a2 s . co m*/ LikelihoodWrapper maximumLikelihoodFunction; // We can use different likelihood wrapper functions: switch (likelihoodFunction) { case POISSON_GAMMA_GAUSSIAN: // Poisson-Gamma-Gaussian - EM-CCD data if (alpha > 0 && sigma > 0) { maximumLikelihoodFunction = new PoissonGammaGaussianLikelihoodWrapper(f, a, y, n, alpha, sigma); break; } case POISSON_GAUSSIAN: // Poisson-Gaussian - CCD data if (sigma > 0) { maximumLikelihoodFunction = new PoissonGaussianLikelihoodWrapper(f, a, y, n, sigma); break; } case POISSON: default: // Poisson - most counting data maximumLikelihoodFunction = new PoissonLikelihoodWrapper(f, a, y, n); } // Check if the method requires the gradient but it cannot be computed if (searchMethod.usesGradient && !maximumLikelihoodFunction.canComputeGradient()) { maximumLikelihoodFunction = new PoissonLikelihoodWrapper(f, a, y, n); } try { double[] startPoint = getInitialSolution(a); PointValuePair optimum = null; if (searchMethod == SearchMethod.POWELL || searchMethod == SearchMethod.POWELL_BOUNDED) { // Non-differentiable version using Powell Optimiser // This is as per the method in Numerical Recipes 10.5 (Direction Set (Powell's) method) // I could extend the optimiser and implement bounds on the directions moved. However the mapping // adapter seems to work OK. final boolean basisConvergence = false; // Perhaps these thresholds should be tighter? // The default is to use the sqrt() of the overall tolerance //final double lineRel = FastMath.sqrt(relativeThreshold); //final double lineAbs = FastMath.sqrt(absoluteThreshold); //final double lineRel = relativeThreshold * 1e2; //final double lineAbs = absoluteThreshold * 1e2; // Since we are fitting only a small number of parameters then just use the same tolerance // for each search direction final double lineRel = relativeThreshold; final double lineAbs = absoluteThreshold; CustomPowellOptimizer o = new CustomPowellOptimizer(relativeThreshold, absoluteThreshold, lineRel, lineAbs, null, basisConvergence); OptimizationData maxIterationData = null; if (getMaxIterations() > 0) maxIterationData = new MaxIter(getMaxIterations()); if (searchMethod == SearchMethod.POWELL) { if (powellFunction == null) { // We must map all the parameters into the same range. This is done in the Mortensen MLE // Python code by using the sqrt of the number of photons and background. if (mapGaussian) { Gaussian2DFunction gf = (Gaussian2DFunction) f; // Re-map signal and background using the sqrt int[] indices = gf.gradientIndices(); int[] map = new int[indices.length]; int count = 0; // Background is always first if (indices[0] == Gaussian2DFunction.BACKGROUND) { map[count++] = 0; } // Look for the Signal in multiple peak 2D Gaussians for (int i = 1; i < indices.length; i++) if (indices[i] % 6 == Gaussian2DFunction.SIGNAL) { map[count++] = i; } if (count > 0) { powellFunction = new MappedMultivariateLikelihood(maximumLikelihoodFunction, Arrays.copyOf(map, count)); } } if (powellFunction == null) { powellFunction = new MultivariateLikelihood(maximumLikelihoodFunction); } } // Update the maximum likelihood function in the Powell function wrapper powellFunction.fun = maximumLikelihoodFunction; OptimizationData positionChecker = null; // new org.apache.commons.math3.optim.PositionChecker(relativeThreshold, absoluteThreshold); if (powellFunction.isMapped()) { MappedMultivariateLikelihood adapter = (MappedMultivariateLikelihood) powellFunction; optimum = o.optimize(maxIterationData, new MaxEval(getMaxEvaluations()), new ObjectiveFunction(powellFunction), GoalType.MINIMIZE, new InitialGuess(adapter.map(startPoint)), positionChecker); double[] solution = adapter.unmap(optimum.getPointRef()); optimum = new PointValuePair(solution, optimum.getValue()); } else { optimum = o.optimize(maxIterationData, new MaxEval(getMaxEvaluations()), new ObjectiveFunction(powellFunction), GoalType.MINIMIZE, new InitialGuess(startPoint), positionChecker); } } else { // Try using the mapping adapter for a bounded Powell search MultivariateFunctionMappingAdapter adapter = new MultivariateFunctionMappingAdapter( new MultivariateLikelihood(maximumLikelihoodFunction), lower, upper); optimum = o.optimize(maxIterationData, new MaxEval(getMaxEvaluations()), new ObjectiveFunction(adapter), GoalType.MINIMIZE, new InitialGuess(adapter.boundedToUnbounded(startPoint))); double[] solution = adapter.unboundedToBounded(optimum.getPointRef()); optimum = new PointValuePair(solution, optimum.getValue()); } iterations = o.getIterations(); evaluations = o.getEvaluations(); } else if (searchMethod == SearchMethod.BOBYQA) { // Differentiable approximation using Powell's BOBYQA algorithm. // This is slower than the Powell optimiser and requires a high number of evaluations. int numberOfInterpolationPoints = this.getNumberOfFittedParameters() + 2; BOBYQAOptimizer o = new BOBYQAOptimizer(numberOfInterpolationPoints); optimum = o.optimize(new MaxEval(getMaxEvaluations()), new ObjectiveFunction(new MultivariateLikelihood(maximumLikelihoodFunction)), GoalType.MINIMIZE, new InitialGuess(startPoint), new SimpleBounds(lower, upper)); iterations = o.getIterations(); evaluations = o.getEvaluations(); } else if (searchMethod == SearchMethod.CMAES) { // TODO - Understand why the CMAES optimiser does not fit very well on test data. It appears // to converge too early and the likelihood scores are not as low as the other optimisers. // CMAESOptimiser based on Matlab code: // https://www.lri.fr/~hansen/cmaes.m // Take the defaults from the Matlab documentation double stopFitness = 0; //Double.NEGATIVE_INFINITY; boolean isActiveCMA = true; int diagonalOnly = 0; int checkFeasableCount = 1; RandomGenerator random = new Well19937c(); boolean generateStatistics = false; // The sigma determines the search range for the variables. It should be 1/3 of the initial search region. double[] sigma = new double[lower.length]; for (int i = 0; i < sigma.length; i++) sigma[i] = (upper[i] - lower[i]) / 3; int popSize = (int) (4 + Math.floor(3 * Math.log(sigma.length))); // The CMAES optimiser is random and restarting can overcome problems with quick convergence. // The Apache commons documentations states that convergence should occur between 30N and 300N^2 // function evaluations final int n30 = FastMath.min(sigma.length * sigma.length * 30, getMaxEvaluations() / 2); evaluations = 0; OptimizationData[] data = new OptimizationData[] { new InitialGuess(startPoint), new CMAESOptimizer.PopulationSize(popSize), new MaxEval(getMaxEvaluations()), new CMAESOptimizer.Sigma(sigma), new ObjectiveFunction(new MultivariateLikelihood(maximumLikelihoodFunction)), GoalType.MINIMIZE, new SimpleBounds(lower, upper) }; // Iterate to prevent early convergence int repeat = 0; while (evaluations < n30) { if (repeat++ > 1) { // Update the start point and population size data[0] = new InitialGuess(optimum.getPointRef()); popSize *= 2; data[1] = new CMAESOptimizer.PopulationSize(popSize); } CMAESOptimizer o = new CMAESOptimizer(getMaxIterations(), stopFitness, isActiveCMA, diagonalOnly, checkFeasableCount, random, generateStatistics, new SimpleValueChecker(relativeThreshold, absoluteThreshold)); PointValuePair result = o.optimize(data); iterations += o.getIterations(); evaluations += o.getEvaluations(); //System.out.printf("CMAES [%d] i=%d [%d], e=%d [%d]\n", repeat, o.getIterations(), iterations, // o.getEvaluations(), totalEvaluations); if (optimum == null || result.getValue() < optimum.getValue()) { optimum = result; } } } else if (searchMethod == SearchMethod.BFGS) { // BFGS can use an approximate line search minimisation where as Powell and conjugate gradient // methods require a more accurate line minimisation. The BFGS search does not do a full // minimisation but takes appropriate steps in the direction of the current gradient. // Do not use the convergence checker on the value of the function. Use the convergence on the // point coordinate and gradient //BFGSOptimizer o = new BFGSOptimizer(new SimpleValueChecker(rel, abs)); BFGSOptimizer o = new BFGSOptimizer(); // Configure maximum step length for each dimension using the bounds double[] stepLength = new double[lower.length]; for (int i = 0; i < stepLength.length; i++) { stepLength[i] = (upper[i] - lower[i]) * 0.3333333; if (stepLength[i] <= 0) stepLength[i] = Double.POSITIVE_INFINITY; } // The GoalType is always minimise so no need to pass this in OptimizationData positionChecker = null; //new org.apache.commons.math3.optim.PositionChecker(relativeThreshold, absoluteThreshold); optimum = o.optimize(new MaxEval(getMaxEvaluations()), new ObjectiveFunctionGradient(new MultivariateVectorLikelihood(maximumLikelihoodFunction)), new ObjectiveFunction(new MultivariateLikelihood(maximumLikelihoodFunction)), new InitialGuess(startPoint), new SimpleBounds(lowerConstraint, upperConstraint), new BFGSOptimizer.GradientTolerance(relativeThreshold), positionChecker, new BFGSOptimizer.StepLength(stepLength)); iterations = o.getIterations(); evaluations = o.getEvaluations(); } else { // The line search algorithm often fails. This is due to searching into a region where the // function evaluates to a negative so has been clipped. This means the upper bound of the line // cannot be found. // Note that running it on an easy problem (200 photons with fixed fitting (no background)) the algorithm // does sometimes produces results better than the Powell algorithm but it is slower. BoundedNonLinearConjugateGradientOptimizer o = new BoundedNonLinearConjugateGradientOptimizer( (searchMethod == SearchMethod.CONJUGATE_GRADIENT_FR) ? Formula.FLETCHER_REEVES : Formula.POLAK_RIBIERE, new SimpleValueChecker(relativeThreshold, absoluteThreshold)); // Note: The gradients may become unstable at the edge of the bounds. Or they will not change // direction if the true solution is on the bounds since the gradient will always continue // towards the bounds. This is key to the conjugate gradient method. It searches along a vector // until the direction of the gradient is in the opposite direction (using dot products, i.e. // cosine of angle between them) // NR 10.7 states there is no advantage of the variable metric DFP or BFGS methods over // conjugate gradient methods. So I will try these first. // Try this: // Adapt the conjugate gradient optimiser to use the gradient to pick the search direction // and then for the line minimisation. However if the function is out of bounds then clip the // variables at the bounds and continue. // If the current point is at the bounds and the gradient is to continue out of bounds then // clip the gradient too. // Or: just use the gradient for the search direction then use the line minimisation/rest // as per the Powell optimiser. The bounds should limit the search. // I tried a Bounded conjugate gradient optimiser with clipped variables: // This sometimes works. However when the variables go a long way out of the expected range the gradients // can have vastly different magnitudes. This results in the algorithm stalling since the gradients // can be close to zero and the some of the parameters are no longer adjusted. // Perhaps this can be looked for and the algorithm then gives up and resorts to a Powell optimiser from // the current point. // Changed the bracketing step to very small (default is 1, changed to 0.001). This improves the // performance. The gradient direction is very sensitive to small changes in the coordinates so a // tighter bracketing of the line search helps. // Tried using a non-gradient method for the line search copied from the Powell optimiser: // This also works when the bracketing step is small but the number of iterations is higher. // 24.10.2014: I have tried to get conjugate gradient to work but the gradient function // must not behave suitably for the optimiser. In the current state both methods of using a // Bounded Conjugate Gradient Optimiser perform poorly relative to other optimisers: // Simulated : n=1000, signal=200, x=0.53, y=0.47 // LVM : n=1000, signal=171, x=0.537, y=0.471 (1.003s) // Powell : n=1000, signal=187, x=0.537, y=0.48 (1.238s) // Gradient based PR (constrained): n=858, signal=161, x=0.533, y=0.474 (2.54s) // Gradient based PR (bounded): n=948, signal=161, x=0.533, y=0.473 (2.67s) // Non-gradient based : n=1000, signal=151.47, x=0.535, y=0.474 (1.626s) // The conjugate optimisers are slower, under predict the signal by the most and in the case of // the gradient based optimiser, fail to converge on some problems. This is worse when constrained // fitting is used and not tightly bounded fitting. // I will leave the code in as an option but would not recommend using it. I may remove it in the // future. // Note: It is strange that the non-gradient based line minimisation is more successful. // It may be that the gradient function is not accurate (due to round off error) or that it is // simply wrong when far from the optimum. My JUnit tests only evaluate the function within the // expected range of the answer. // Note the default step size on the Powell optimiser is 1 but the initial directions are unit vectors. // So our bracketing step should be a minimum of 1 / average length of the first gradient vector to prevent // the first step being too large when bracketing. final double gradient[] = new double[startPoint.length]; maximumLikelihoodFunction.likelihood(startPoint, gradient); double l = 0; for (double d : gradient) l += d * d; final double bracketingStep = FastMath.min(0.001, ((l > 1) ? 1.0 / l : 1)); //System.out.printf("Bracketing step = %f (length=%f)\n", bracketingStep, l); o.setUseGradientLineSearch(gradientLineMinimisation); optimum = o.optimize(new MaxEval(getMaxEvaluations()), new ObjectiveFunctionGradient(new MultivariateVectorLikelihood(maximumLikelihoodFunction)), new ObjectiveFunction(new MultivariateLikelihood(maximumLikelihoodFunction)), GoalType.MINIMIZE, new InitialGuess(startPoint), new SimpleBounds(lowerConstraint, upperConstraint), new BoundedNonLinearConjugateGradientOptimizer.BracketingStep(bracketingStep)); iterations = o.getIterations(); evaluations = o.getEvaluations(); //maximumLikelihoodFunction.value(solution, gradient); //System.out.printf("Iter = %d, %g @ %s : %s\n", iterations, ll, Arrays.toString(solution), // Arrays.toString(gradient)); } final double[] solution = optimum.getPointRef(); setSolution(a, solution); //System.out.printf("Iter = %d, Eval = %d, %g @ %s\n", iterations, evaluations, optimum.getValue(), // java.util.Arrays.toString(solution)); // Compute residuals for the FunctionSolver interface if (y_fit == null || y_fit.length < n) y_fit = new double[n]; f.initialise(a); residualSumOfSquares = 0; for (int i = 0; i < n; i++) { y_fit[i] = f.eval(i); final double residual = y[i] - y_fit[i]; residualSumOfSquares += residual * residual; } if (a_dev != null) { // Assume the Maximum Likelihood estimator returns the optimum fit (achieves the Cramer Roa // lower bounds) and so the covariance can be obtained from the Fisher Information Matrix. final int[] gradientIndices = f.gradientIndices(); final int nparams = gradientIndices.length; GradientCalculator calculator = GradientCalculatorFactory.newCalculator(nparams); final double[] I = calculator.fisherInformationDiagonal(n, a, f); for (int i = 0; i < gradientIndices.length; i++) a_dev[gradientIndices[i]] = 1.0 / Math.sqrt(I[i]); } error[0] = NonLinearFit.getError(residualSumOfSquares, noise, n, f.gradientIndices().length); totalSumOfSquares = getSumOfSquares(n, y); } catch (TooManyIterationsException e) { //System.out.printf("Too many iterations = %d\n", e.getMax()); //e.printStackTrace(); return FitStatus.FAILED_TO_CONVERGE; } catch (TooManyEvaluationsException e) { //System.out.printf("Too many evaluations = %d\n", e.getMax()); //e.printStackTrace(); return FitStatus.FAILED_TO_CONVERGE; } catch (ConvergenceException e) { // Occurs when QR decomposition fails - mark as a singular non-linear model (no solution) //System.out.printf("Singular non linear model = %s\n", e.getMessage()); return FitStatus.SINGULAR_NON_LINEAR_MODEL; } catch (BFGSOptimizer.LineSearchRoundoffException e) { //System.out.println("BFGS error: " + e.getMessage()); //e.printStackTrace(); return FitStatus.FAILED_TO_CONVERGE; } catch (Exception e) { //System.out.printf("Unknown error = %s\n", e.getMessage()); e.printStackTrace(); return FitStatus.UNKNOWN; } return FitStatus.OK; }