[
https://issues.apache.org/jira/browse/MATH-413?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12904592#action_12904592
]
Gilles commented on MATH-413:
-----------------------------
IMO, the most urgent issue to be dealt with is the design decision to make the
convergence checking criteria independent from the optimization algorithm.
[This has a direct influence on the API and should be stabilized in v3.0.]
Points 1, 2, 6, 10 in the issue description indicate the problems with this
approach.
IMO, while modularization is often a good thing, the attempt produces several
drawbacks in this case:
* The user interface is more difficult than it could be: the user has to
specify a "complex" {{ConvergenceChecker}} object, instead of just {{double}}s
(relative and/or absolute threshold/tolerance/accuracy).
* At the CM level, there is no single implementation of {{ConvergenceChecker}}
that is applicable for every algorithm (see {{BrentOptimizer}} and
{{LevenbergMarquardtOptimizer}}).
* If we don't set a default, the user will get a NPE.
* If we set a default convergence checker, in an attempt to make usage easier,
it will still not be very useful because the user will have to set a new
checker as soon as he needs to change the algorithm accuracy.
* The user cannot rely on the description of convergence given by a
{{ConvergenceChecker}}'s implementation because if he chooses a checker that is
"incompatible" with the optimization algorithm, the behaviour is "undefined".
[I have overridden {{setConvergenceChecker}} to forbid the setting of anything
other than a {{BrentConvergenceChecker}} but this completely voids the intended
flexibility of having a separate convergence checking procedure.]
* If we want to enforce the strict separation between optimization and
convergence checking, we'll have to _modify_ standard algorithm so that they
fit into the mold (i.e. we must ensure that "convergence checking" and
"optimization" parts are indeed independent). This may not even be possible
without removing features of some algorithms (e.g. specific tolerance
definitions in "Levenberg-Marquardt"). And if it is done, then the resulting
algorithm will not be standard anymore!
In light of all these problems, we should seriously re-examine whether the
{{ConvergenceChecker}} approach is the right way to go.
In the end, it is very well possible the best that can be done is to pass
appropriate arguments to the constructor of each optimizer's implementation.
[With the consequence that if one wants to change the convergence checking
behaviour, one has to create a new instance. (Side note: this is consistent
with going towards instances' immutability.)]
> Miscellaneous issues concerning the "optimization" package
> ----------------------------------------------------------
>
> Key: MATH-413
> URL: https://issues.apache.org/jira/browse/MATH-413
> Project: Commons Math
> Issue Type: Bug
> Reporter: Gilles
> Fix For: 3.0
>
>
> Revision 990792 contains changes triggered the following issues:
> * [MATH-394|https://issues.apache.org/jira/browse/MATH-394]
> * [MATH-397|https://issues.apache.org/jira/browse/MATH-397]
> * [MATH-404|https://issues.apache.org/jira/browse/MATH-404]
> This issue collects the currently still unsatisfactory code (not necessarily
> sorted in order of annoyance):
> # "BrentOptimizer": a specific convergence checker must be used.
> "LevenbergMarquardtOptimizer" also has specific convergence checks.
> # Trying to make convergence checking independent of the optimization
> algorithm creates problems (conceptual and practical):
> ** See "BrentOptimizer" and "LevenbergMarquardtOptimizer", the algorithm
> passes "points" to the convergence checker, but the actual meaning of the
> points can very well be different in the caller (optimization algorithm) and
> the callee (convergence checker).
> ** In "PowellOptimizer" the line search ("BrentOptimizer") tolerances depend
> on the tolerances within the main algorithm. Since tolerances come with
> "ConvergenceChecker" and so can be changed at any time, it is awkward to
> adapt the values within the line search optimizer without exposing its
> internals ("BrentOptimizer" field) to the enclosing class ("PowellOptimizer").
> # Given the numerous changes, some Javadoc comments might be out-of-sync,
> although I did try to update them all.
> # Class "DirectSearchOptimizer" (in package "optimization.direct") inherits
> from class "AbstractScalarOptimizer" (in package "optimization.general").
> # Some interfaces are defined in package "optimization" but their base
> implementations (abstract class that contain the boiler-plate code) are in
> package "optimization.general" (e.g.
> "DifferentiableMultivariateVectorialOptimizer" and
> "BaseAbstractVectorialOptimizer").
> # No check is performed to ensure the the convergence checker has been set
> (see e.g. "BrentOptimizer" and "PowellOptimizer"); if it hasn't there will be
> a NPE. The alternative is to initialize a default checker that will never be
> used in case the user had intended to explicitly sets the checker.
> # "NonLinearConjugateGradientOptimizer": Ugly workaround for the checked
> "ConvergenceException".
> # Everywhere, we trail the checked "FunctionEvaluationException" although it
> is never used.
> # There remains some duplicate code (such as the "multi-start loop" in the
> various "MultiStart..." implementations).
> # The "ConvergenceChecker" interface is very general (the "converged" method
> can take any number of "...PointValuePair"). However there remains a
> "semantic" problem: One cannot be sure that the list of points means the same
> thing for the caller of "converged" and within the implementation of the
> "ConvergenceChecker" that was independently set.
> # It is not clear whether it is wise to aggregate the counter of gradient
> evaluations to the function evaluation counter. In
> "LevenbergMarquartdOptimizer" for example, it would be unfair to do so.
> Currently I had to remove all tests referring to gradient and Jacobian
> evaluations.
> # In "AbstractLeastSquaresOptimizer" and "LevenbergMarquardtOptimizer",
> occurences of "OptimizationException" were replaced by the unchecked
> "ConvergenceException" but in some cases it might not be the most appropriate
> one.
> # "MultiStartUnivariateRealOptimizer": in the other classes
> ("MultiStartMultivariate...") similar to this one, the randomization is on
> the firts-guess value while in this class, it is on the search interval. I
> think that here also we should randomly choose the start value (within the
> user-selected interval).
> # The Javadoc utility raises warnings (see output of "mvn site") which I
> couldn't figure out how to correct.
> # Some previously existing classes and interfaces have become no more than a
> specialisation of new "generics" classes; it might be interesting to remove
> them in order to reduce the number of classes and thus limit the potential
> for confusion.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
