[Os-project-managers] NaN in xs:double (OSrL)

[Horand Gassmann]

Kipp Martin <kmartin at chicagobooth.edu> wrote:

> Hi Gus:
>>
>> That is not how I remember the discussion from last time (or maybe  
>> the time before that):
>>
>> <initialVariableValues numberOfVar="2">
>>    <var idx="0" value="NaN"/>
>>    <var idx="1"/>
>>
>> should set both x0 and x1 to NaN (this I am clear on) and should  
>> send those values to the solver (this is how I understood the  
>> discussion, but perhaps I misunderstood).
>
> No we don't send those values on to the solver, we send those to the  
> OS solver interface. Consider Ipopt. We never give Ipopt a starting  
> value of NaN for a variable. However, the OS Ipopt solver interface  
> "intercepts" the NaN and replaces it with  1.7171. So we never send  
> NaN to a solver.

Having just looked at OSIpoptSolver.cpp again, I think this is  
incorrect. Let's have a closer look. The parser reads "NaN" in the  
first case and places OSNAN in osoption->...->initVarValue[0]->value.  
In the second case
osoption->...->initVarValue[1]->value is initialed to OSNAN by the  
initVarValue constructor. Either way,  
osoption->...->initVarValue[i]->idx has a legal value, and hence the  
OSNAN is transferred into the x-array in line 255 of OSIpoptSolver,  
and initialed is set to true. Therefore the loop in line 269/270 is  
not executed, and the default_initval is skipped. I think this is as  
it should be.

BTW, I don't think the if statements in lines 241, 246 and 251 are  
very graceful. I believe a quiet reset with a notation in the osrl  
file would be much better.

Cheers

gus