This issue refers to version 1.0.0 (last stable version), though it might occur also in later versions (I haven't checked that). This is the issue:
The example nonnegls.cpp crashes in the very first call to LbfgsbSolver::SubspaceMinimization.
I think the reason is that for the first iteration matrix W still has zero columns (as there are no previous steps yet). I think that the problem will be fixed if, for the first iteration,
du(i) = -g(i) , if bounds are not active at i, or they are active but g(i) points outward
du(i) = 0, otherwise
Obviously, if at every position bounds are active with the gradient pointing inward, an optimum has been reached
This issue refers to version 1.0.0 (last stable version), though it might occur also in later versions (I haven't checked that). This is the issue:
The example nonnegls.cpp crashes in the very first call to LbfgsbSolver::SubspaceMinimization.
I think the reason is that for the first iteration matrix W still has zero columns (as there are no previous steps yet). I think that the problem will be fixed if, for the first iteration,
du(i) = -g(i) , if bounds are not active at i, or they are active but g(i) points outward du(i) = 0, otherwise
Obviously, if at every position bounds are active with the gradient pointing inward, an optimum has been reached