[
https://issues.apache.org/jira/browse/MATH-823?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13414111#comment-13414111
]
Dennis Hendriks commented on MATH-823:
--------------------------------------
It seems to me that this overlaps with MATH-764. Both issues discuss sampling
of distributions with user provided random generators (or RandomDataImpl).
> Distribution classes should allow randomData to be configured via constructor
> or setter
> ---------------------------------------------------------------------------------------
>
> Key: MATH-823
> URL: https://issues.apache.org/jira/browse/MATH-823
> Project: Commons Math
> Issue Type: Improvement
> Affects Versions: 3.0
> Reporter: Steve Phelps
> Priority: Minor
>
> In my application I would like to represent several different probability
> distributions using polymorphism - see example below. At compile-time the
> distributions are declared of type RealDistribution. These distributions are
> then configured with concrete classes at run-time using, e.g. dependency
> injection. However, because there is no way to configure the underlying PRNG
> -- ie the randomData field of these distributions, there is a danger of
> introducing correlations into the random number streams through poor seed
> management. If I have dozens of distributions, each configured automatically
> with a new RandomDataImpl() then there is the possibility that some of these
> streams will share the same seed value introducing spurious correlations.
> One solution to this would be to explicitly manage seed values for every
> distribution in my simulation by calling reseedRandomGenerator(). However,
> this introduces unnecessary complexity and dependencies throughout my code.
> A far simpler solution for my use case would be to configure all the
> distributions in my simulation to use the same singleton instance of
> RandomDataImpl for their randomData field (eg using dependency injection),
> thus ensuring truly independent distributions - however currently the API
> does not support this.
>
> Could the Distribution classes be refactored to allow the randomData field to
> be configured, eg through setters or via the constructor? This would be
> similar to the design of the distribution classes provided by the CERN colt
> library, which I believe offers more flexibility in this respect. It would
> also be ideal, IMHO, if the Distribution objects were POJOs with zero-arg
> constructors, which would then allow them to be configured using a
> dependency-injection framework.
> public class Simulation {
> RealDistribution interArrivalTimeDistribution;
>
> RealDistribution sizeDistribution;
>
> double t;
>
> public void run() {
> t = 0;
> while (true) {
> t += interArrivalTimeDistribution.sample();
> double magnitude = sizeDistribution.sample();
> System.out.println(t + "\t" + magnitude);
> }
> }
> public RealDistribution getInterArrivalTimeDistribution() {
> return interArrivalTimeDistribution;
> }
> public void setInterArrivalTimeDistribution(RealDistribution
> interArrivalTimes) {
> this.interArrivalTimeDistribution = interArrivalTimes;
> }
> public RealDistribution getSizeDistribution() {
> return sizeDistribution;
> }
> public void setSizeDistribution(RealDistribution xsizeDistribution) {
> this.sizeDistribution = sizeDistribution;
> }
>
>
> }
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
