search

ericphanson / ConvexTests.jl

Uses the Convex.jl Problem Depot to test various optimization solvers
MIT License
10 stars 5 forks source link

Provide a mechanism for correctly reporting non-solves #3

Open ericphanson opened 4 years ago

ericphanson commented 4 years ago

If an optimizer reaches an iteration limit or time limit etc, it can communicate that via an MOI status. I think failing a tolerance test but correctly reporting an iteration limit is much better than failing the tolerance test but reporting OPTIMAL. So we should have a mechanism for that (e.g. tolerance tests are skipped if a failing status is reported, and have a separate count for # problems optimally solved).

joaquimg commented 4 years ago

We tried to do that as you say in ProxSDP. The failing tests, return TIME/ITERATION_LIMIT, and check if the available solution is feasible or not to return FEASIBLE_POINT or INFEASIBLE_POINT.