[
https://issues.apache.org/jira/browse/MATH-1321?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15138935#comment-15138935
]
Gilles commented on MATH-1321:
------------------------------
bq. \[2.3595947265625, -1.44864501953125\] is not equal to alpha/beta from the
debug output
I had missed the point, sorry.
The "getKey()", or preferrably "getPoint()" since it is more to the point :),
should indeed return the optimal value.
But as you use a {{MultivariateFunctionMappingAdapter}} wrapper, it is expected
that the debug output, which shows values of the original function, is not the
same as the values of the function that is actually being optimized.
In order to retrieve the values in the original range, you must perform the
[inverse
transformation|http://commons.apache.org/proper/commons-math/javadocs/api-3.6/org/apache/commons/math3/optim/nonlinear/scalar/MultivariateFunctionMappingAdapter.html]:
{code}
double[] actual = optimizationFn.unboundedToBounded(nelderResult.getKey());
{code}
which hopefully will return the values shown in the output (modulo the possible
loss of accuracy that could result from the conversions).
> SimplexOptimizer returns wrong values
> -------------------------------------
>
> Key: MATH-1321
> URL: https://issues.apache.org/jira/browse/MATH-1321
> Project: Commons Math
> Issue Type: Bug
> Affects Versions: 3.6
> Environment: Fedora 23,
> java version "1.8.0_51"
> Java(TM) SE Runtime Environment (build 1.8.0_51-b16)
> Java HotSpot(TM) 64-Bit Server VM (build 25.51-b03, mixed mode)
> Reporter: Pavol Loffay
> Priority: Blocker
>
> Hello I'm trying to minimize MSE of double exponential smoothing and get
> optimal parameters alpha and beta.
> https://www.otexts.org/fpp/7/2
> During the minimization the output shows values of alpha and beta which
> differs from alpha, and beta returned from SimplexOptimizer.optimize().
> {code:title=output|borderStyle=solid}
> ...
> Nelder MSE = 0.007226473669598979, alpha=0.896283, beta=0.228161
> Nelder MSE = 0.0069843320509952005, alpha=0.913694, beta=0.190210 # returned
> minimum
> Nelder MSE = 0.008577342645261695, alpha=0.931617, beta=0.131448
> Nelder MSE = 0.00743296945818598, alpha=0.918018, beta=0.166808
> Nelder MSE = 0.007818891499431175, alpha=0.936768, beta=0.136053
> Nelder MSE = 0.007293932014855209, alpha=0.927010, beta=0.155973
> Nelder MSE = 0.007319455298330941, alpha=0.923120, beta=0.178180
> Nelder MSE = 0.007110221641945739, alpha=0.921873, beta=0.175281
> Nelder MSE = 0.007271067724068611, alpha=0.907713, beta=0.212689
> Nelder MSE = 0.007084561548618076, alpha=0.912928, beta=0.197226
> Nelder MSE = 0.007072487763137581, alpha=0.903911, beta=0.213540
> Nelder -> key = [2.3595947265625, -1.44864501953125], fce minimum=
> 0.00698433205099520050
> {code}
> {code:title=Test.java|borderStyle=solid}
> @Test
> public void testOptimization() throws IOException {
> int maxEval = 1000;
> int maxIter = 1000;
> // Nelder-Mead Simplex
> SimplexOptimizer nelderSimplexOptimizer = new
> SimplexOptimizer(0.0001, 0.0001);
> PointValuePair nelderResult = nelderSimplexOptimizer.optimize(
> GoalType.MINIMIZE, new MaxIter(maxIter), new MaxEval(maxEval),
> new InitialGuess(new double[]{0.4, 0.1}), new
> ObjectiveFunction(optimizationFn("Nelder")),
> new NelderMeadSimplex(2));
> System.out.format("\nNelder (%d eval) -> key = %s, fce minimum= %.20f",
> nelderSimplexOptimizer.getEvaluations(),
> Arrays.toString(nelderResult.getKey()),
> nelderResult.getValue());
> }
> private MultivariateFunction optimizationFn(String algorithm) {
> final List<DataPoint> testData = metricData.subList(0, 50);
> // func for minimization
> MultivariateFunction multivariateFunction = point -> {
> double alpha = point[0];
> double beta = point[1];
> DoubleExponentialSmoothing doubleExponentialSmoothing = new
> DoubleExponentialSmoothing(alpha, beta);
> AccuracyStatistics accuracyStatistics =
> doubleExponentialSmoothing.init(testData);
> System.out.format("%s MSE = %s, alpha=%f, beta=%f\n", algorithm,
> accuracyStatistics.getMse(), alpha, beta);
> return accuracyStatistics.getMse();
> };
> MultivariateFunctionMappingAdapter multivariateFunctionMappingAdapter
> =
> new MultivariateFunctionMappingAdapter(multivariateFunction,
> new double[]{0.0, 0.0}, new double[]{1, 1});
> return multivariateFunctionMappingAdapter;
> }
> {code}
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
