Search in sources :

Example 1 with CustomPowellOptimizer

use of uk.ac.sussex.gdsc.smlm.math3.optim.nonlinear.scalar.noderiv.CustomPowellOptimizer in project GDSC-SMLM by aherbert.

the class EMGainAnalysis method fit.

/**
	 * Fit the EM-gain distribution (Gaussian * Gamma)
	 * 
	 * @param h
	 *            The distribution
	 */
private void fit(int[] h) {
    final int[] limits = limits(h);
    final double[] x = getX(limits);
    final double[] y = getY(h, limits);
    Plot2 plot = new Plot2(TITLE, "ADU", "Frequency");
    double yMax = Maths.max(y);
    plot.setLimits(limits[0], limits[1], 0, yMax);
    plot.setColor(Color.black);
    plot.addPoints(x, y, Plot2.DOT);
    Utils.display(TITLE, plot);
    // Estimate remaining parameters. 
    // Assuming a gamma_distribution(shape,scale) then mean = shape * scale
    // scale = gain
    // shape = Photons = mean / gain
    double mean = getMean(h) - bias;
    // Note: if the bias is too high then the mean will be negative. Just move the bias.
    while (mean < 0) {
        bias -= 1;
        mean += 1;
    }
    double photons = mean / gain;
    if (simulate)
        Utils.log("Simulated bias=%d, gain=%s, noise=%s, photons=%s", (int) _bias, Utils.rounded(_gain), Utils.rounded(_noise), Utils.rounded(_photons));
    Utils.log("Estimate bias=%d, gain=%s, noise=%s, photons=%s", (int) bias, Utils.rounded(gain), Utils.rounded(noise), Utils.rounded(photons));
    final int max = (int) x[x.length - 1];
    double[] g = pdf(max, photons, gain, noise, (int) bias);
    plot.setColor(Color.blue);
    plot.addPoints(x, g, Plot2.LINE);
    Utils.display(TITLE, plot);
    // Perform a fit
    CustomPowellOptimizer o = new CustomPowellOptimizer(1e-6, 1e-16, 1e-6, 1e-16);
    double[] startPoint = new double[] { photons, gain, noise, bias };
    int maxEval = 3000;
    String[] paramNames = { "Photons", "Gain", "Noise", "Bias" };
    // Set bounds
    double[] lower = new double[] { 0, 0.5 * gain, 0, bias - noise };
    double[] upper = new double[] { 2 * photons, 2 * gain, gain, bias + noise };
    // Restart until converged.
    // TODO - Maybe fix this with a better optimiser. This needs to be tested on real data.
    PointValuePair solution = null;
    for (int iter = 0; iter < 3; iter++) {
        IJ.showStatus("Fitting histogram ... Iteration " + iter);
        try {
            // Basic Powell optimiser
            MultivariateFunction fun = getFunction(limits, y, max, maxEval);
            PointValuePair optimum = o.optimize(new MaxEval(maxEval), new ObjectiveFunction(fun), GoalType.MINIMIZE, new InitialGuess((solution == null) ? startPoint : solution.getPointRef()));
            if (solution == null || optimum.getValue() < solution.getValue()) {
                double[] point = optimum.getPointRef();
                // Check the bounds
                for (int i = 0; i < point.length; i++) {
                    if (point[i] < lower[i] || point[i] > upper[i]) {
                        throw new RuntimeException(String.format("Fit out of of estimated range: %s %f", paramNames[i], point[i]));
                    }
                }
                solution = optimum;
            }
        } catch (Exception e) {
            IJ.log("Powell error: " + e.getMessage());
            if (e instanceof TooManyEvaluationsException) {
                maxEval = (int) (maxEval * 1.5);
            }
        }
        try {
            // Bounded Powell optimiser
            MultivariateFunction fun = getFunction(limits, y, max, maxEval);
            MultivariateFunctionMappingAdapter adapter = new MultivariateFunctionMappingAdapter(fun, lower, upper);
            PointValuePair optimum = o.optimize(new MaxEval(maxEval), new ObjectiveFunction(adapter), GoalType.MINIMIZE, new InitialGuess(adapter.boundedToUnbounded((solution == null) ? startPoint : solution.getPointRef())));
            double[] point = adapter.unboundedToBounded(optimum.getPointRef());
            optimum = new PointValuePair(point, optimum.getValue());
            if (solution == null || optimum.getValue() < solution.getValue()) {
                solution = optimum;
            }
        } catch (Exception e) {
            IJ.log("Bounded Powell error: " + e.getMessage());
            if (e instanceof TooManyEvaluationsException) {
                maxEval = (int) (maxEval * 1.5);
            }
        }
    }
    IJ.showStatus("");
    IJ.showProgress(1);
    if (solution == null) {
        Utils.log("Failed to fit the distribution");
        return;
    }
    double[] point = solution.getPointRef();
    photons = point[0];
    gain = point[1];
    noise = point[2];
    bias = (int) Math.round(point[3]);
    String label = String.format("Fitted bias=%d, gain=%s, noise=%s, photons=%s", (int) bias, Utils.rounded(gain), Utils.rounded(noise), Utils.rounded(photons));
    Utils.log(label);
    if (simulate) {
        Utils.log("Relative Error bias=%s, gain=%s, noise=%s, photons=%s", Utils.rounded(relativeError(bias, _bias)), Utils.rounded(relativeError(gain, _gain)), Utils.rounded(relativeError(noise, _noise)), Utils.rounded(relativeError(photons, _photons)));
    }
    // Show the PoissonGammaGaussian approximation
    double[] f = null;
    if (showApproximation) {
        f = new double[x.length];
        PoissonGammaGaussianFunction fun = new PoissonGammaGaussianFunction(1.0 / gain, noise);
        final double expected = photons * gain;
        for (int i = 0; i < f.length; i++) {
            f[i] = fun.likelihood(x[i] - bias, expected);
        //System.out.printf("x=%d, g=%f, f=%f, error=%f\n", (int) x[i], g[i], f[i],
        //		gdsc.smlm.fitting.utils.DoubleEquality.relativeError(g[i], f[i]));
        }
        yMax = Maths.maxDefault(yMax, f);
    }
    // Replot
    g = pdf(max, photons, gain, noise, (int) bias);
    plot = new Plot2(TITLE, "ADU", "Frequency");
    plot.setLimits(limits[0], limits[1], 0, yMax * 1.05);
    plot.setColor(Color.black);
    plot.addPoints(x, y, Plot2.DOT);
    plot.setColor(Color.red);
    plot.addPoints(x, g, Plot2.LINE);
    plot.addLabel(0, 0, label);
    if (showApproximation) {
        plot.setColor(Color.blue);
        plot.addPoints(x, f, Plot2.LINE);
    }
    Utils.display(TITLE, plot);
}
Also used : MaxEval(org.apache.commons.math3.optim.MaxEval) InitialGuess(org.apache.commons.math3.optim.InitialGuess) ObjectiveFunction(org.apache.commons.math3.optim.nonlinear.scalar.ObjectiveFunction) Plot2(ij.gui.Plot2) Point(java.awt.Point) TooManyEvaluationsException(org.apache.commons.math3.exception.TooManyEvaluationsException) PointValuePair(org.apache.commons.math3.optim.PointValuePair) MultivariateFunction(org.apache.commons.math3.analysis.MultivariateFunction) MultivariateFunctionMappingAdapter(org.apache.commons.math3.optim.nonlinear.scalar.MultivariateFunctionMappingAdapter) TooManyEvaluationsException(org.apache.commons.math3.exception.TooManyEvaluationsException) CustomPowellOptimizer(org.apache.commons.math3.optim.nonlinear.scalar.noderiv.CustomPowellOptimizer) PoissonGammaGaussianFunction(gdsc.smlm.function.PoissonGammaGaussianFunction)

Example 2 with CustomPowellOptimizer

use of uk.ac.sussex.gdsc.smlm.math3.optim.nonlinear.scalar.noderiv.CustomPowellOptimizer in project GDSC-SMLM by aherbert.

the class JumpDistanceAnalysis method doFitJumpDistanceHistogram.

/**
	 * Fit the jump distance histogram using a cumulative sum with the given number of species.
	 * <p>
	 * Results are sorted by the diffusion coefficient ascending.
	 * 
	 * @param jdHistogram
	 *            The cumulative jump distance histogram. X-axis is um^2, Y-axis is cumulative probability. Must be
	 *            monototic ascending.
	 * @param estimatedD
	 *            The estimated diffusion coefficient
	 * @param n
	 *            The number of species in the mixed population
	 * @return Array containing: { D (um^2), Fractions }. Can be null if no fit was made.
	 */
private double[][] doFitJumpDistanceHistogram(double[][] jdHistogram, double estimatedD, int n) {
    calibrated = isCalibrated();
    if (n == 1) {
        // Fit using a single population model
        LevenbergMarquardtOptimizer lvmOptimizer = new LevenbergMarquardtOptimizer();
        try {
            final JumpDistanceCumulFunction function = new JumpDistanceCumulFunction(jdHistogram[0], jdHistogram[1], estimatedD);
            //@formatter:off
            LeastSquaresProblem problem = new LeastSquaresBuilder().maxEvaluations(Integer.MAX_VALUE).maxIterations(3000).start(function.guess()).target(function.getY()).weight(new DiagonalMatrix(function.getWeights())).model(function, new MultivariateMatrixFunction() {

                public double[][] value(double[] point) throws IllegalArgumentException {
                    return function.jacobian(point);
                }
            }).build();
            //@formatter:on
            Optimum lvmSolution = lvmOptimizer.optimize(problem);
            double[] fitParams = lvmSolution.getPoint().toArray();
            // True for an unweighted fit
            ss = lvmSolution.getResiduals().dotProduct(lvmSolution.getResiduals());
            //ss = calculateSumOfSquares(function.getY(), function.value(fitParams));
            lastIC = ic = Maths.getAkaikeInformationCriterionFromResiduals(ss, function.x.length, 1);
            double[] coefficients = fitParams;
            double[] fractions = new double[] { 1 };
            logger.info("Fit Jump distance (N=1) : %s, SS = %s, IC = %s (%d evaluations)", formatD(fitParams[0]), Maths.rounded(ss, 4), Maths.rounded(ic, 4), lvmSolution.getEvaluations());
            return new double[][] { coefficients, fractions };
        } catch (TooManyIterationsException e) {
            logger.info("LVM optimiser failed to fit (N=1) : Too many iterations : %s", e.getMessage());
        } catch (ConvergenceException e) {
            logger.info("LVM optimiser failed to fit (N=1) : %s", e.getMessage());
        }
    }
    // Uses a weighted sum of n exponential functions, each function models a fraction of the particles.
    // An LVM fit cannot restrict the parameters so the fractions do not go below zero.
    // Use the CustomPowell/CMEASOptimizer which supports bounded fitting.
    MixedJumpDistanceCumulFunctionMultivariate function = new MixedJumpDistanceCumulFunctionMultivariate(jdHistogram[0], jdHistogram[1], estimatedD, n);
    double[] lB = function.getLowerBounds();
    int evaluations = 0;
    PointValuePair constrainedSolution = null;
    MaxEval maxEval = new MaxEval(20000);
    CustomPowellOptimizer powellOptimizer = createCustomPowellOptimizer();
    try {
        // The Powell algorithm can use more general bounds: 0 - Infinity
        constrainedSolution = powellOptimizer.optimize(maxEval, new ObjectiveFunction(function), new InitialGuess(function.guess()), new SimpleBounds(lB, function.getUpperBounds(Double.POSITIVE_INFINITY, Double.POSITIVE_INFINITY)), new CustomPowellOptimizer.BasisStep(function.step()), GoalType.MINIMIZE);
        evaluations = powellOptimizer.getEvaluations();
        logger.debug("Powell optimiser fit (N=%d) : SS = %f (%d evaluations)", n, constrainedSolution.getValue(), evaluations);
    } catch (TooManyEvaluationsException e) {
        logger.info("Powell optimiser failed to fit (N=%d) : Too many evaluations (%d)", n, powellOptimizer.getEvaluations());
    } catch (TooManyIterationsException e) {
        logger.info("Powell optimiser failed to fit (N=%d) : Too many iterations (%d)", n, powellOptimizer.getIterations());
    } catch (ConvergenceException e) {
        logger.info("Powell optimiser failed to fit (N=%d) : %s", n, e.getMessage());
    }
    if (constrainedSolution == null) {
        logger.info("Trying CMAES optimiser with restarts ...");
        double[] uB = function.getUpperBounds();
        SimpleBounds bounds = new SimpleBounds(lB, uB);
        // The sigma determines the search range for the variables. It should be 1/3 of the initial search region.
        double[] s = new double[lB.length];
        for (int i = 0; i < s.length; i++) s[i] = (uB[i] - lB[i]) / 3;
        OptimizationData sigma = new CMAESOptimizer.Sigma(s);
        OptimizationData popSize = new CMAESOptimizer.PopulationSize((int) (4 + Math.floor(3 * Math.log(function.x.length))));
        // Iterate this for stability in the initial guess
        CMAESOptimizer cmaesOptimizer = createCMAESOptimizer();
        for (int i = 0; i <= fitRestarts; i++) {
            // Try from the initial guess
            try {
                PointValuePair solution = cmaesOptimizer.optimize(new InitialGuess(function.guess()), new ObjectiveFunction(function), GoalType.MINIMIZE, bounds, sigma, popSize, maxEval);
                if (constrainedSolution == null || solution.getValue() < constrainedSolution.getValue()) {
                    evaluations = cmaesOptimizer.getEvaluations();
                    constrainedSolution = solution;
                    logger.debug("CMAES optimiser [%da] fit (N=%d) : SS = %f (%d evaluations)", i, n, solution.getValue(), evaluations);
                }
            } catch (TooManyEvaluationsException e) {
            }
            if (constrainedSolution == null)
                continue;
            // Try from the current optimum
            try {
                PointValuePair solution = cmaesOptimizer.optimize(new InitialGuess(constrainedSolution.getPointRef()), new ObjectiveFunction(function), GoalType.MINIMIZE, bounds, sigma, popSize, maxEval);
                if (solution.getValue() < constrainedSolution.getValue()) {
                    evaluations = cmaesOptimizer.getEvaluations();
                    constrainedSolution = solution;
                    logger.debug("CMAES optimiser [%db] fit (N=%d) : SS = %f (%d evaluations)", i, n, solution.getValue(), evaluations);
                }
            } catch (TooManyEvaluationsException e) {
            }
        }
        if (constrainedSolution != null) {
            // Re-optimise with Powell?
            try {
                PointValuePair solution = powellOptimizer.optimize(maxEval, new ObjectiveFunction(function), new InitialGuess(constrainedSolution.getPointRef()), new SimpleBounds(lB, function.getUpperBounds(Double.POSITIVE_INFINITY, Double.POSITIVE_INFINITY)), new CustomPowellOptimizer.BasisStep(function.step()), GoalType.MINIMIZE);
                if (solution.getValue() < constrainedSolution.getValue()) {
                    evaluations = cmaesOptimizer.getEvaluations();
                    constrainedSolution = solution;
                    logger.info("Powell optimiser re-fit (N=%d) : SS = %f (%d evaluations)", n, constrainedSolution.getValue(), evaluations);
                }
            } catch (TooManyEvaluationsException e) {
            } catch (TooManyIterationsException e) {
            } catch (ConvergenceException e) {
            }
        }
    }
    if (constrainedSolution == null) {
        logger.info("Failed to fit N=%d", n);
        return null;
    }
    double[] fitParams = constrainedSolution.getPointRef();
    ss = constrainedSolution.getValue();
    // TODO - Try a bounded BFGS optimiser
    // Try and improve using a LVM fit
    final MixedJumpDistanceCumulFunctionGradient functionGradient = new MixedJumpDistanceCumulFunctionGradient(jdHistogram[0], jdHistogram[1], estimatedD, n);
    Optimum lvmSolution;
    LevenbergMarquardtOptimizer lvmOptimizer = new LevenbergMarquardtOptimizer();
    try {
        //@formatter:off
        LeastSquaresProblem problem = new LeastSquaresBuilder().maxEvaluations(Integer.MAX_VALUE).maxIterations(3000).start(fitParams).target(functionGradient.getY()).weight(new DiagonalMatrix(functionGradient.getWeights())).model(functionGradient, new MultivariateMatrixFunction() {

            public double[][] value(double[] point) throws IllegalArgumentException {
                return functionGradient.jacobian(point);
            }
        }).build();
        //@formatter:on
        lvmSolution = lvmOptimizer.optimize(problem);
        // True for an unweighted fit
        double ss = lvmSolution.getResiduals().dotProduct(lvmSolution.getResiduals());
        // All fitted parameters must be above zero
        if (ss < this.ss && Maths.min(lvmSolution.getPoint().toArray()) > 0) {
            logger.info("  Re-fitting improved the SS from %s to %s (-%s%%)", Maths.rounded(this.ss, 4), Maths.rounded(ss, 4), Maths.rounded(100 * (this.ss - ss) / this.ss, 4));
            fitParams = lvmSolution.getPoint().toArray();
            this.ss = ss;
            evaluations += lvmSolution.getEvaluations();
        }
    } catch (TooManyIterationsException e) {
        logger.error("Failed to re-fit : Too many iterations : %s", e.getMessage());
    } catch (ConvergenceException e) {
        logger.error("Failed to re-fit : %s", e.getMessage());
    }
    // Since the fractions must sum to one we subtract 1 degree of freedom from the number of parameters
    ic = Maths.getAkaikeInformationCriterionFromResiduals(ss, function.x.length, fitParams.length - 1);
    double[] d = new double[n];
    double[] f = new double[n];
    double sum = 0;
    for (int i = 0; i < d.length; i++) {
        f[i] = fitParams[i * 2];
        sum += f[i];
        d[i] = fitParams[i * 2 + 1];
    }
    for (int i = 0; i < f.length; i++) f[i] /= sum;
    // Sort by coefficient size
    sort(d, f);
    double[] coefficients = d;
    double[] fractions = f;
    logger.info("Fit Jump distance (N=%d) : %s (%s), SS = %s, IC = %s (%d evaluations)", n, formatD(d), format(f), Maths.rounded(ss, 4), Maths.rounded(ic, 4), evaluations);
    if (isValid(d, f)) {
        lastIC = ic;
        return new double[][] { coefficients, fractions };
    }
    return null;
}
Also used : MaxEval(org.apache.commons.math3.optim.MaxEval) InitialGuess(org.apache.commons.math3.optim.InitialGuess) SimpleBounds(org.apache.commons.math3.optim.SimpleBounds) ObjectiveFunction(org.apache.commons.math3.optim.nonlinear.scalar.ObjectiveFunction) LeastSquaresBuilder(org.apache.commons.math3.fitting.leastsquares.LeastSquaresBuilder) PointValuePair(org.apache.commons.math3.optim.PointValuePair) TooManyEvaluationsException(org.apache.commons.math3.exception.TooManyEvaluationsException) DiagonalMatrix(org.apache.commons.math3.linear.DiagonalMatrix) ConvergenceException(org.apache.commons.math3.exception.ConvergenceException) TooManyIterationsException(org.apache.commons.math3.exception.TooManyIterationsException) LeastSquaresProblem(org.apache.commons.math3.fitting.leastsquares.LeastSquaresProblem) MultivariateMatrixFunction(org.apache.commons.math3.analysis.MultivariateMatrixFunction) CMAESOptimizer(org.apache.commons.math3.optim.nonlinear.scalar.noderiv.CMAESOptimizer) Optimum(org.apache.commons.math3.fitting.leastsquares.LeastSquaresOptimizer.Optimum) LevenbergMarquardtOptimizer(org.apache.commons.math3.fitting.leastsquares.LevenbergMarquardtOptimizer) OptimizationData(org.apache.commons.math3.optim.OptimizationData) CustomPowellOptimizer(org.apache.commons.math3.optim.nonlinear.scalar.noderiv.CustomPowellOptimizer)

Example 3 with CustomPowellOptimizer

use of uk.ac.sussex.gdsc.smlm.math3.optim.nonlinear.scalar.noderiv.CustomPowellOptimizer in project GDSC-SMLM by aherbert.

the class JumpDistanceAnalysis method doFitJumpDistancesMLE.

/**
	 * Fit the jump distances using a maximum likelihood estimation with the given number of species.
	 * | *
	 * <p>
	 * Results are sorted by the diffusion coefficient ascending.
	 * 
	 * @param jumpDistances
	 *            The jump distances (in um^2)
	 * @param estimatedD
	 *            The estimated diffusion coefficient
	 * @param n
	 *            The number of species in the mixed population
	 * @return Array containing: { D (um^2), Fractions }. Can be null if no fit was made.
	 */
private double[][] doFitJumpDistancesMLE(double[] jumpDistances, double estimatedD, int n) {
    MaxEval maxEval = new MaxEval(20000);
    CustomPowellOptimizer powellOptimizer = createCustomPowellOptimizer();
    calibrated = isCalibrated();
    if (n == 1) {
        try {
            final JumpDistanceFunction function = new JumpDistanceFunction(jumpDistances, estimatedD);
            // The Powell algorithm can use more general bounds: 0 - Infinity
            SimpleBounds bounds = new SimpleBounds(function.getLowerBounds(), function.getUpperBounds(Double.POSITIVE_INFINITY, Double.POSITIVE_INFINITY));
            PointValuePair solution = powellOptimizer.optimize(maxEval, new ObjectiveFunction(function), new InitialGuess(function.guess()), bounds, new CustomPowellOptimizer.BasisStep(function.step()), GoalType.MAXIMIZE);
            double[] fitParams = solution.getPointRef();
            ll = solution.getValue();
            lastIC = ic = Maths.getAkaikeInformationCriterion(ll, jumpDistances.length, 1);
            double[] coefficients = fitParams;
            double[] fractions = new double[] { 1 };
            logger.info("Fit Jump distance (N=1) : %s, MLE = %s, IC = %s (%d evaluations)", formatD(fitParams[0]), Maths.rounded(ll, 4), Maths.rounded(ic, 4), powellOptimizer.getEvaluations());
            return new double[][] { coefficients, fractions };
        } catch (TooManyEvaluationsException e) {
            logger.info("Powell optimiser failed to fit (N=1) : Too many evaluation (%d)", powellOptimizer.getEvaluations());
        } catch (TooManyIterationsException e) {
            logger.info("Powell optimiser failed to fit (N=1) : Too many iterations (%d)", powellOptimizer.getIterations());
        } catch (ConvergenceException e) {
            logger.info("Powell optimiser failed to fit (N=1) : %s", e.getMessage());
        }
        return null;
    }
    MixedJumpDistanceFunction function = new MixedJumpDistanceFunction(jumpDistances, estimatedD, n);
    double[] lB = function.getLowerBounds();
    int evaluations = 0;
    PointValuePair constrainedSolution = null;
    try {
        // The Powell algorithm can use more general bounds: 0 - Infinity
        constrainedSolution = powellOptimizer.optimize(maxEval, new ObjectiveFunction(function), new InitialGuess(function.guess()), new SimpleBounds(lB, function.getUpperBounds(Double.POSITIVE_INFINITY, Double.POSITIVE_INFINITY)), new CustomPowellOptimizer.BasisStep(function.step()), GoalType.MAXIMIZE);
        evaluations = powellOptimizer.getEvaluations();
        logger.debug("Powell optimiser fit (N=%d) : MLE = %f (%d evaluations)", n, constrainedSolution.getValue(), powellOptimizer.getEvaluations());
    } catch (TooManyEvaluationsException e) {
        logger.info("Powell optimiser failed to fit (N=%d) : Too many evaluation (%d)", n, powellOptimizer.getEvaluations());
    } catch (TooManyIterationsException e) {
        logger.info("Powell optimiser failed to fit (N=%d) : Too many iterations (%d)", n, powellOptimizer.getIterations());
    } catch (ConvergenceException e) {
        logger.info("Powell optimiser failed to fit (N=%d) : %s", n, e.getMessage());
    }
    if (constrainedSolution == null) {
        logger.info("Trying CMAES optimiser with restarts ...");
        double[] uB = function.getUpperBounds();
        SimpleBounds bounds = new SimpleBounds(lB, uB);
        // Try a bounded CMAES optimiser since the Powell optimiser appears to be 
        // sensitive to the order of the parameters. It is not good when the fast particle
        // is the minority fraction. Could this be due to too low an upper bound?
        // The sigma determines the search range for the variables. It should be 1/3 of the initial search region.
        double[] s = new double[lB.length];
        for (int i = 0; i < s.length; i++) s[i] = (uB[i] - lB[i]) / 3;
        OptimizationData sigma = new CMAESOptimizer.Sigma(s);
        OptimizationData popSize = new CMAESOptimizer.PopulationSize((int) (4 + Math.floor(3 * Math.log(function.x.length))));
        // Iterate this for stability in the initial guess
        CMAESOptimizer cmaesOptimizer = createCMAESOptimizer();
        for (int i = 0; i <= fitRestarts; i++) {
            // Try from the initial guess
            try {
                PointValuePair solution = cmaesOptimizer.optimize(new InitialGuess(function.guess()), new ObjectiveFunction(function), GoalType.MAXIMIZE, bounds, sigma, popSize, maxEval);
                if (constrainedSolution == null || solution.getValue() > constrainedSolution.getValue()) {
                    evaluations = cmaesOptimizer.getEvaluations();
                    constrainedSolution = solution;
                    logger.debug("CMAES optimiser [%da] fit (N=%d) : MLE = %f (%d evaluations)", i, n, solution.getValue(), evaluations);
                }
            } catch (TooManyEvaluationsException e) {
            }
            if (constrainedSolution == null)
                continue;
            // Try from the current optimum
            try {
                PointValuePair solution = cmaesOptimizer.optimize(new InitialGuess(constrainedSolution.getPointRef()), new ObjectiveFunction(function), GoalType.MAXIMIZE, bounds, sigma, popSize, maxEval);
                if (solution.getValue() > constrainedSolution.getValue()) {
                    evaluations = cmaesOptimizer.getEvaluations();
                    constrainedSolution = solution;
                    logger.debug("CMAES optimiser [%db] fit (N=%d) : MLE = %f (%d evaluations)", i, n, solution.getValue(), evaluations);
                }
            } catch (TooManyEvaluationsException e) {
            }
        }
        if (constrainedSolution != null) {
            try {
                // Re-optimise with Powell?
                PointValuePair solution = powellOptimizer.optimize(maxEval, new ObjectiveFunction(function), new InitialGuess(constrainedSolution.getPointRef()), new SimpleBounds(lB, function.getUpperBounds(Double.POSITIVE_INFINITY, Double.POSITIVE_INFINITY)), new CustomPowellOptimizer.BasisStep(function.step()), GoalType.MAXIMIZE);
                if (solution.getValue() > constrainedSolution.getValue()) {
                    evaluations = cmaesOptimizer.getEvaluations();
                    constrainedSolution = solution;
                    logger.info("Powell optimiser re-fit (N=%d) : MLE = %f (%d evaluations)", n, constrainedSolution.getValue(), powellOptimizer.getEvaluations());
                }
            } catch (TooManyEvaluationsException e) {
            } catch (TooManyIterationsException e) {
            } catch (ConvergenceException e) {
            }
        }
    }
    if (constrainedSolution == null) {
        logger.info("Failed to fit N=%d", n);
        return null;
    }
    double[] fitParams = constrainedSolution.getPointRef();
    ll = constrainedSolution.getValue();
    // Since the fractions must sum to one we subtract 1 degree of freedom from the number of parameters
    ic = Maths.getAkaikeInformationCriterion(ll, jumpDistances.length, fitParams.length - 1);
    double[] d = new double[n];
    double[] f = new double[n];
    double sum = 0;
    for (int i = 0; i < d.length; i++) {
        f[i] = fitParams[i * 2];
        sum += f[i];
        d[i] = fitParams[i * 2 + 1];
    }
    for (int i = 0; i < f.length; i++) f[i] /= sum;
    // Sort by coefficient size
    sort(d, f);
    double[] coefficients = d;
    double[] fractions = f;
    logger.info("Fit Jump distance (N=%d) : %s (%s), MLE = %s, IC = %s (%d evaluations)", n, formatD(d), format(f), Maths.rounded(ll, 4), Maths.rounded(ic, 4), evaluations);
    if (isValid(d, f)) {
        lastIC = ic;
        return new double[][] { coefficients, fractions };
    }
    return null;
}
Also used : MaxEval(org.apache.commons.math3.optim.MaxEval) InitialGuess(org.apache.commons.math3.optim.InitialGuess) SimpleBounds(org.apache.commons.math3.optim.SimpleBounds) CMAESOptimizer(org.apache.commons.math3.optim.nonlinear.scalar.noderiv.CMAESOptimizer) ObjectiveFunction(org.apache.commons.math3.optim.nonlinear.scalar.ObjectiveFunction) PointValuePair(org.apache.commons.math3.optim.PointValuePair) TooManyEvaluationsException(org.apache.commons.math3.exception.TooManyEvaluationsException) ConvergenceException(org.apache.commons.math3.exception.ConvergenceException) TooManyIterationsException(org.apache.commons.math3.exception.TooManyIterationsException) OptimizationData(org.apache.commons.math3.optim.OptimizationData) CustomPowellOptimizer(org.apache.commons.math3.optim.nonlinear.scalar.noderiv.CustomPowellOptimizer)

Example 4 with CustomPowellOptimizer

use of uk.ac.sussex.gdsc.smlm.math3.optim.nonlinear.scalar.noderiv.CustomPowellOptimizer in project GDSC-SMLM by aherbert.

the class MaximumLikelihoodFitter method computeFit.

/*
	 * (non-Javadoc)
	 * 
	 * @see gdsc.smlm.fitting.nonlinear.BaseFunctionSolver#computeFit(double[], double[], double[], double[])
	 */
public FitStatus computeFit(double[] y, double[] y_fit, double[] a, double[] a_dev) {
    final int n = y.length;
    LikelihoodWrapper maximumLikelihoodFunction = createLikelihoodWrapper((NonLinearFunction) f, n, y, a);
    @SuppressWarnings("rawtypes") BaseOptimizer baseOptimiser = null;
    try {
        double[] startPoint = getInitialSolution(a);
        PointValuePair optimum = null;
        if (searchMethod == SearchMethod.POWELL || searchMethod == SearchMethod.POWELL_BOUNDED || searchMethod == SearchMethod.POWELL_ADAPTER) {
            // 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);
            baseOptimiser = o;
            OptimizationData maxIterationData = null;
            if (getMaxIterations() > 0)
                maxIterationData = new MaxIter(getMaxIterations());
            if (searchMethod == SearchMethod.POWELL_ADAPTER) {
                // 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());
            } else {
                if (powellFunction == null) {
                    // 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);
                SimpleBounds simpleBounds = null;
                if (powellFunction.isMapped()) {
                    MappedMultivariateLikelihood adapter = (MappedMultivariateLikelihood) powellFunction;
                    if (searchMethod == SearchMethod.POWELL_BOUNDED)
                        simpleBounds = new SimpleBounds(adapter.map(lower), adapter.map(upper));
                    optimum = o.optimize(maxIterationData, new MaxEval(getMaxEvaluations()), new ObjectiveFunction(powellFunction), GoalType.MINIMIZE, new InitialGuess(adapter.map(startPoint)), positionChecker, simpleBounds);
                    double[] solution = adapter.unmap(optimum.getPointRef());
                    optimum = new PointValuePair(solution, optimum.getValue());
                } else {
                    if (searchMethod == SearchMethod.POWELL_BOUNDED)
                        simpleBounds = new SimpleBounds(lower, upper);
                    optimum = o.optimize(maxIterationData, new MaxEval(getMaxEvaluations()), new ObjectiveFunction(powellFunction), GoalType.MINIMIZE, new InitialGuess(startPoint), positionChecker, simpleBounds);
                }
            }
        } 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);
            baseOptimiser = o;
            optimum = o.optimize(new MaxEval(getMaxEvaluations()), new ObjectiveFunction(new MultivariateLikelihood(maximumLikelihoodFunction)), GoalType.MINIMIZE, new InitialGuess(startPoint), new SimpleBounds(lower, upper));
        } 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.NEGATIVE_INFINITY;
            double stopFitness = 0;
            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));
                baseOptimiser = o;
                PointValuePair result = o.optimize(data);
                iterations += o.getIterations();
                evaluations += o.getEvaluations();
                //		o.getEvaluations(), totalEvaluations);
                if (optimum == null || result.getValue() < optimum.getValue()) {
                    optimum = result;
                }
            }
            // Prevent incrementing the iterations again
            baseOptimiser = null;
        } 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();
            baseOptimiser = o;
            // 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));
        } 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));
            baseOptimiser = o;
            // 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));
        //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);
        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.
            FisherInformationMatrix m = new FisherInformationMatrix(maximumLikelihoodFunction.fisherInformation(a));
            setDeviations(a_dev, m.crlb(true));
        }
        // Reverse negative log likelihood for maximum likelihood score
        value = -optimum.getValue();
    } catch (TooManyIterationsException e) {
        //e.printStackTrace();
        return FitStatus.TOO_MANY_ITERATIONS;
    } catch (TooManyEvaluationsException e) {
        //e.printStackTrace();
        return FitStatus.TOO_MANY_EVALUATIONS;
    } catch (ConvergenceException e) {
        //System.out.printf("Singular non linear model = %s\n", e.getMessage());
        return FitStatus.SINGULAR_NON_LINEAR_MODEL;
    } catch (BFGSOptimizer.LineSearchRoundoffException e) {
        //e.printStackTrace();
        return FitStatus.FAILED_TO_CONVERGE;
    } catch (Exception e) {
        //System.out.printf("Unknown error = %s\n", e.getMessage());
        e.printStackTrace();
        return FitStatus.UNKNOWN;
    } finally {
        if (baseOptimiser != null) {
            iterations += baseOptimiser.getIterations();
            evaluations += baseOptimiser.getEvaluations();
        }
    }
    // Check this as likelihood functions can go wrong
    if (Double.isInfinite(value) || Double.isNaN(value))
        return FitStatus.INVALID_LIKELIHOOD;
    return FitStatus.OK;
}
Also used : MaxEval(org.apache.commons.math3.optim.MaxEval) InitialGuess(org.apache.commons.math3.optim.InitialGuess) BOBYQAOptimizer(org.apache.commons.math3.optim.nonlinear.scalar.noderiv.BOBYQAOptimizer) SimpleBounds(org.apache.commons.math3.optim.SimpleBounds) ObjectiveFunction(org.apache.commons.math3.optim.nonlinear.scalar.ObjectiveFunction) Well19937c(org.apache.commons.math3.random.Well19937c) SimpleValueChecker(org.apache.commons.math3.optim.SimpleValueChecker) RandomGenerator(org.apache.commons.math3.random.RandomGenerator) BFGSOptimizer(org.apache.commons.math3.optim.nonlinear.scalar.gradient.BFGSOptimizer) PointValuePair(org.apache.commons.math3.optim.PointValuePair) TooManyEvaluationsException(org.apache.commons.math3.exception.TooManyEvaluationsException) Gaussian2DFunction(gdsc.smlm.function.gaussian.Gaussian2DFunction) ConvergenceException(org.apache.commons.math3.exception.ConvergenceException) BoundedNonLinearConjugateGradientOptimizer(org.apache.commons.math3.optim.nonlinear.scalar.gradient.BoundedNonLinearConjugateGradientOptimizer) TooManyIterationsException(org.apache.commons.math3.exception.TooManyIterationsException) BaseOptimizer(org.apache.commons.math3.optim.BaseOptimizer) CMAESOptimizer(org.apache.commons.math3.optim.nonlinear.scalar.noderiv.CMAESOptimizer) FisherInformationMatrix(gdsc.smlm.fitting.FisherInformationMatrix) PoissonGammaGaussianLikelihoodWrapper(gdsc.smlm.function.PoissonGammaGaussianLikelihoodWrapper) PoissonGaussianLikelihoodWrapper(gdsc.smlm.function.PoissonGaussianLikelihoodWrapper) PoissonLikelihoodWrapper(gdsc.smlm.function.PoissonLikelihoodWrapper) LikelihoodWrapper(gdsc.smlm.function.LikelihoodWrapper) ConvergenceException(org.apache.commons.math3.exception.ConvergenceException) TooManyIterationsException(org.apache.commons.math3.exception.TooManyIterationsException) TooManyEvaluationsException(org.apache.commons.math3.exception.TooManyEvaluationsException) ObjectiveFunctionGradient(org.apache.commons.math3.optim.nonlinear.scalar.ObjectiveFunctionGradient) MultivariateFunctionMappingAdapter(org.apache.commons.math3.optim.nonlinear.scalar.MultivariateFunctionMappingAdapter) OptimizationData(org.apache.commons.math3.optim.OptimizationData) CustomPowellOptimizer(org.apache.commons.math3.optim.nonlinear.scalar.noderiv.CustomPowellOptimizer) MaxIter(org.apache.commons.math3.optim.MaxIter)

Example 5 with CustomPowellOptimizer

use of uk.ac.sussex.gdsc.smlm.math3.optim.nonlinear.scalar.noderiv.CustomPowellOptimizer in project GDSC-SMLM by aherbert.

the class MaximumLikelihoodFitter method computeFit.

@Override
public FitStatus computeFit(double[] y, double[] fx, double[] a, double[] parametersVariance) {
    final int n = y.length;
    final LikelihoodWrapper maximumLikelihoodFunction = createLikelihoodWrapper((NonLinearFunction) function, n, y, a);
    @SuppressWarnings("rawtypes") BaseOptimizer baseOptimiser = null;
    try {
        final double[] startPoint = getInitialSolution(a);
        PointValuePair optimum = null;
        if (searchMethod == SearchMethod.POWELL || searchMethod == SearchMethod.POWELL_BOUNDED || searchMethod == SearchMethod.POWELL_ADAPTER) {
            // Non-differentiable version using Powell Optimiser
            // Background: see Numerical Recipes 10.5 (Direction Set (Powell's) method).
            // The optimiser could be extended to 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 = Math.sqrt(relativeThreshold);
            // final double lineAbs = Math.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;
            final CustomPowellOptimizer o = new CustomPowellOptimizer(relativeThreshold, absoluteThreshold, lineRel, lineAbs, null, basisConvergence);
            baseOptimiser = o;
            OptimizationData maxIterationData = null;
            if (getMaxIterations() > 0) {
                maxIterationData = new MaxIter(getMaxIterations());
            }
            if (searchMethod == SearchMethod.POWELL_ADAPTER) {
                // Try using the mapping adapter for a bounded Powell search
                final 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)));
                final double[] solution = adapter.unboundedToBounded(optimum.getPointRef());
                optimum = new PointValuePair(solution, optimum.getValue());
            } else {
                if (powellFunction == null) {
                    powellFunction = new MultivariateLikelihood(maximumLikelihoodFunction);
                }
                // Update the maximum likelihood function in the Powell function wrapper
                powellFunction.fun = maximumLikelihoodFunction;
                final OptimizationData positionChecker = null;
                // new org.apache.commons.math3.optim.PositionChecker(relativeThreshold,
                // absoluteThreshold);
                SimpleBounds simpleBounds = null;
                if (powellFunction.isMapped()) {
                    final MappedMultivariateLikelihood adapter = (MappedMultivariateLikelihood) powellFunction;
                    if (searchMethod == SearchMethod.POWELL_BOUNDED) {
                        simpleBounds = new SimpleBounds(adapter.map(lower), adapter.map(upper));
                    }
                    optimum = o.optimize(maxIterationData, new MaxEval(getMaxEvaluations()), new ObjectiveFunction(powellFunction), GoalType.MINIMIZE, new InitialGuess(adapter.map(startPoint)), positionChecker, simpleBounds);
                    final double[] solution = adapter.unmap(optimum.getPointRef());
                    optimum = new PointValuePair(solution, optimum.getValue());
                } else {
                    if (searchMethod == SearchMethod.POWELL_BOUNDED) {
                        simpleBounds = new SimpleBounds(lower, upper);
                    }
                    optimum = o.optimize(maxIterationData, new MaxEval(getMaxEvaluations()), new ObjectiveFunction(powellFunction), GoalType.MINIMIZE, new InitialGuess(startPoint), positionChecker, simpleBounds);
                }
            }
        } 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.
            final int numberOfInterpolationpoints = this.getNumberOfFittedParameters() + 2;
            final BOBYQAOptimizer o = new BOBYQAOptimizer(numberOfInterpolationpoints);
            baseOptimiser = o;
            optimum = o.optimize(new MaxEval(getMaxEvaluations()), new ObjectiveFunction(new MultivariateLikelihood(maximumLikelihoodFunction)), GoalType.MINIMIZE, new InitialGuess(startPoint), new SimpleBounds(lower, upper));
        } 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.
            // The CMAESOptimiser is based on Matlab code:
            // https://www.lri.fr/~hansen/cmaes.m
            // Take the defaults from the Matlab documentation
            final double stopFitness = 0;
            final boolean isActiveCma = true;
            final int diagonalOnly = 0;
            final int checkFeasableCount = 1;
            final RandomGenerator random = new RandomGeneratorAdapter(UniformRandomProviders.create());
            final boolean generateStatistics = false;
            // The sigma determines the search range for the variables. It should be 1/3 of the initial
            // search region.
            final 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 = Math.min(sigma.length * sigma.length * 30, getMaxEvaluations() / 2);
            evaluations = 0;
            final 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
                    if (optimum != null) {
                        data[0] = new InitialGuess(optimum.getPointRef());
                    }
                    popSize *= 2;
                    data[1] = new CMAESOptimizer.PopulationSize(popSize);
                }
                final CMAESOptimizer o = new CMAESOptimizer(getMaxIterations(), stopFitness, isActiveCma, diagonalOnly, checkFeasableCount, random, generateStatistics, new SimpleValueChecker(relativeThreshold, absoluteThreshold));
                baseOptimiser = o;
                final PointValuePair result = o.optimize(data);
                iterations += o.getIterations();
                evaluations += o.getEvaluations();
                if (optimum == null || result.getValue() < optimum.getValue()) {
                    optimum = result;
                }
            }
            // Prevent incrementing the iterations again
            baseOptimiser = null;
        } 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.
            final BoundedNonLinearConjugateGradientOptimizer o = new BoundedNonLinearConjugateGradientOptimizer((searchMethod == SearchMethod.CONJUGATE_GRADIENT_FR) ? Formula.FLETCHER_REEVES : Formula.POLAK_RIBIERE, new SimpleValueChecker(relativeThreshold, absoluteThreshold));
            baseOptimiser = o;
            // 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 length = 0;
            for (final double d : gradient) {
                length += d * d;
            }
            final double bracketingStep = Math.min(0.001, ((length > 1) ? 1.0 / length : 1));
            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));
        }
        if (optimum == null) {
            return FitStatus.FAILED_TO_CONVERGE;
        }
        final double[] solution = optimum.getPointRef();
        setSolution(a, solution);
        if (parametersVariance != null) {
            // Compute assuming a Poisson process.
            // Note:
            // If using a Poisson-Gamma-Gaussian model then these will be incorrect.
            // However the variance for the position estimates of a 2D PSF can be
            // scaled by a factor of 2 as in Mortensen, et al (2010) Nature Methods 7, 377-383, SI 4.3.
            // Since the type of function is unknown this cannot be done here.
            final FisherInformationMatrix m = new FisherInformationMatrix(maximumLikelihoodFunction.fisherInformation(solution));
            setDeviations(parametersVariance, m);
        }
        // Reverse negative log likelihood for maximum likelihood score
        value = -optimum.getValue();
    } catch (final TooManyIterationsException ex) {
        return FitStatus.TOO_MANY_ITERATIONS;
    } catch (final TooManyEvaluationsException ex) {
        return FitStatus.TOO_MANY_EVALUATIONS;
    } catch (final ConvergenceException ex) {
        // Occurs when QR decomposition fails - mark as a singular non-linear model (no solution)
        return FitStatus.SINGULAR_NON_LINEAR_MODEL;
    } catch (final Exception ex) {
        Logger.getLogger(getClass().getName()).log(Level.SEVERE, "Failed to fit", ex);
        return FitStatus.UNKNOWN;
    } finally {
        if (baseOptimiser != null) {
            iterations += baseOptimiser.getIterations();
            evaluations += baseOptimiser.getEvaluations();
        }
    }
    // Check this as likelihood functions can go wrong
    if (Double.isInfinite(value) || Double.isNaN(value)) {
        return FitStatus.INVALID_LIKELIHOOD;
    }
    return FitStatus.OK;
}
Also used : MaxEval(org.apache.commons.math3.optim.MaxEval) InitialGuess(org.apache.commons.math3.optim.InitialGuess) BOBYQAOptimizer(org.apache.commons.math3.optim.nonlinear.scalar.noderiv.BOBYQAOptimizer) SimpleBounds(org.apache.commons.math3.optim.SimpleBounds) ObjectiveFunction(org.apache.commons.math3.optim.nonlinear.scalar.ObjectiveFunction) SimpleValueChecker(org.apache.commons.math3.optim.SimpleValueChecker) RandomGenerator(org.apache.commons.math3.random.RandomGenerator) PointValuePair(org.apache.commons.math3.optim.PointValuePair) TooManyEvaluationsException(org.apache.commons.math3.exception.TooManyEvaluationsException) ConvergenceException(org.apache.commons.math3.exception.ConvergenceException) BoundedNonLinearConjugateGradientOptimizer(uk.ac.sussex.gdsc.smlm.math3.optim.nonlinear.scalar.gradient.BoundedNonLinearConjugateGradientOptimizer) TooManyIterationsException(org.apache.commons.math3.exception.TooManyIterationsException) RandomGeneratorAdapter(uk.ac.sussex.gdsc.core.utils.rng.RandomGeneratorAdapter) BaseOptimizer(org.apache.commons.math3.optim.BaseOptimizer) CMAESOptimizer(org.apache.commons.math3.optim.nonlinear.scalar.noderiv.CMAESOptimizer) FisherInformationMatrix(uk.ac.sussex.gdsc.smlm.fitting.FisherInformationMatrix) LikelihoodWrapper(uk.ac.sussex.gdsc.smlm.function.LikelihoodWrapper) PoissonLikelihoodWrapper(uk.ac.sussex.gdsc.smlm.function.PoissonLikelihoodWrapper) PoissonGammaGaussianLikelihoodWrapper(uk.ac.sussex.gdsc.smlm.function.PoissonGammaGaussianLikelihoodWrapper) PoissonGaussianLikelihoodWrapper(uk.ac.sussex.gdsc.smlm.function.PoissonGaussianLikelihoodWrapper) ConvergenceException(org.apache.commons.math3.exception.ConvergenceException) TooManyIterationsException(org.apache.commons.math3.exception.TooManyIterationsException) TooManyEvaluationsException(org.apache.commons.math3.exception.TooManyEvaluationsException) ObjectiveFunctionGradient(org.apache.commons.math3.optim.nonlinear.scalar.ObjectiveFunctionGradient) MultivariateFunctionMappingAdapter(org.apache.commons.math3.optim.nonlinear.scalar.MultivariateFunctionMappingAdapter) OptimizationData(org.apache.commons.math3.optim.OptimizationData) CustomPowellOptimizer(uk.ac.sussex.gdsc.smlm.math3.optim.nonlinear.scalar.noderiv.CustomPowellOptimizer) MaxIter(org.apache.commons.math3.optim.MaxIter)

Aggregations

TooManyEvaluationsException (org.apache.commons.math3.exception.TooManyEvaluationsException)8 InitialGuess (org.apache.commons.math3.optim.InitialGuess)8 MaxEval (org.apache.commons.math3.optim.MaxEval)8 PointValuePair (org.apache.commons.math3.optim.PointValuePair)8 ObjectiveFunction (org.apache.commons.math3.optim.nonlinear.scalar.ObjectiveFunction)8 ConvergenceException (org.apache.commons.math3.exception.ConvergenceException)6 TooManyIterationsException (org.apache.commons.math3.exception.TooManyIterationsException)6 OptimizationData (org.apache.commons.math3.optim.OptimizationData)6 SimpleBounds (org.apache.commons.math3.optim.SimpleBounds)6 CMAESOptimizer (org.apache.commons.math3.optim.nonlinear.scalar.noderiv.CMAESOptimizer)6 MultivariateFunctionMappingAdapter (org.apache.commons.math3.optim.nonlinear.scalar.MultivariateFunctionMappingAdapter)4 CustomPowellOptimizer (org.apache.commons.math3.optim.nonlinear.scalar.noderiv.CustomPowellOptimizer)4 CustomPowellOptimizer (uk.ac.sussex.gdsc.smlm.math3.optim.nonlinear.scalar.noderiv.CustomPowellOptimizer)4 Point (java.awt.Point)2 MultivariateFunction (org.apache.commons.math3.analysis.MultivariateFunction)2 LeastSquaresBuilder (org.apache.commons.math3.fitting.leastsquares.LeastSquaresBuilder)2 Optimum (org.apache.commons.math3.fitting.leastsquares.LeastSquaresOptimizer.Optimum)2 LeastSquaresProblem (org.apache.commons.math3.fitting.leastsquares.LeastSquaresProblem)2 LevenbergMarquardtOptimizer (org.apache.commons.math3.fitting.leastsquares.LevenbergMarquardtOptimizer)2 DiagonalMatrix (org.apache.commons.math3.linear.DiagonalMatrix)2