On Thu, Jan 17, 2013 at 3:52 PM, Gilles Sadowski < gil...@harfang.homelinux.org> wrote:
> On Thu, Jan 17, 2013 at 02:28:56PM +0100, Thomas Neidhart wrote: > > Hi, > > > > by investigating a recent issue (MATH-930), I discovered that the newly > > introduced LinearConstraintSet stores the constraints internally in a > > HashSet. > > This leads to undeterministic behavior as the iteration order of > > constraints is not fixed (especially between different JRE versions). > > > > TBH, I never reviewed this change and I am a bit surprised to see it. > Imho > > there is no need to have a Set here, and would prefer to change this to a > > list implementation. > > -1 > Unless the order of iteration is a requirement of the mathematical > description of the algorithm. > If so: > * This requirement should appear prominently in the Javadoc.[1] > * Please discard the rest of this mail. > If not, then the order is only an implementation detail (IOW, with infinite > precision, the solution would not change with different iteration orders), > I would be cautious to rely on a specific implementation of the container > used to store input data. > > Maybe that the behaviour described in MATH-930 indicates an inherent > weakness of the algorithm (or too stringent requirements); if so, the > better > solution is probably not to rely on whether side-effects are visible or > not. > > Also, I note that in the (deprecated) package > "o.a.c.m.optimization.linear", > the constraints are passed to "optimize" as a > Collection<LinearConstraint> > IIUC, the same behaviour could thus be produced by a user who'd call > optimize several times, passing the same set of constraints but stored in a > different order in the Collection. > Hi Gilles, yes I understand your concerns, and the problem is two-fold: * make the algorithm behave deterministic in all environments (I do not want to spend each time 1 day debugging why it works in one case or another) * fix the obvious numerical stability problem with certain ordering of constraints Although before the parameter was passed as a Collection, most people actually used a List or something similar. At least you had control over it, now it is just a HashSet, inside a CM class which you can not control. So I would really prefer to at least change it to a LinkedHashSet with deterministic iteration order, and investigate further the other problem. Thomas
