Closed sscheele closed 2 years ago
Sorry for not addressing this earlier, I'm going back over open PRs after a period of not actively maintaining this. This is being fixed in the new version where we use LDLt decomposition on the entire KKT system, which should be much more robust. See #31
_projection_solve!
uses a Cholesky factorization to help evaluate constraint violations (pn_methods.jl:98
). However, Cholesky factorization only works on matrices which are positive definite. I don't think the matrix in question is necessarily positive definite (for instance, in the case of goal constraints. I've edited the code to useLinearAlgebra.factorize
instead, which will still perform Cholesky factorization when possible for performance, but will use alternate factorizations when necessary. I'm not certain this is a mathematically correct approach.