Hello all,
In reference to Acer's workaround for the operator form of the optimization package, I'm having difficulty applying the method to an optimization problem with 7 variables. I'm hoping you can help me see what I'm doing wrong. This is what I've got coded, where EIG is a fairly long, complex procedure that outputs a float value:
> objf := proc (V::Vector)
EIG(V[1], V[2], V[3], V[4], V[5], V[6], V[7])
end proc;
> objfgradient := proc (X::Vector, G::Vector)
G[1] := fdiff(EIG, [1], [X[1], X[7]]);
G[2] := fdiff(EIG, [2], [X[1], X[7]]);
G[3] := fdiff(EIG, [3], [X[1], X[7]]);
G[4] := fdiff(EIG, [4], [X[1], X[7]]);
G[5] := fdiff(EIG, [5], [X[1], X[7]]);
G[6] := fdiff(EIG, [6], [X[1], X[7]]);
G[7] := fdiff(EIG, [7], [X[1], X[7]]);
NULL;
end proc;
> Optimization:-NLPSolve(7, objf, objectivegradient = objfgradient, initialpoint = Vector([0.5e-3, 0.5e-3, 0.1e-2, 0.5e-3, .3, .7, 0.8e-3]));
NLPSolve: calling NLP solver
SolveUnconstrained: using method=pcg
SolveUnconstrained: number of problem variables 7
PrintSettings: optimality tolerance set to 0.3256082241e-11
PrintSettings: iteration limit set to 50
SolveUnconstrained: trying evalhf mode
SolveUnconstrained: trying evalf mode
Error, (in Optimization:-NLPSolve) could not store fdiff(EIGEXP, [1], [0.500000000000000012e-3, 0.800000000000000040e-3]) in a floating-point rtable
It seems like I'm doing something wrong with the 'objfgradient' procedure, using fdiff. I admittedly don't understand very well how it works, so I appreciate your help and patience!
