Resume or extend a run

generate_mga_files writes every solve to output_dir as it goes, and reads those files back when it starts. To resume an interrupted run, or to continue a finished one with a tighter tolerance or a higher iteration cap, call it again with the same network, dimensions and output_dir:

generate_mga_files(n, dimensions, "mga-output", norm="linf", slack_levels=[0.05],
                   solver_name="highs", tol=2, max_iter=300, max_parallel=4)

On restart:

  • The optimum and the initial box-direction solves are read back instead of solved again.

  • Every earlier nearest-point result is replayed into the approximations, so neither its point nor its cuts have to be found again. Only new iterations solve the model.

Things to keep in mind:

  • Keep the network and dimensions unchanged. The files are assumed to belong to the same problem. Results from a different network or different dimensions would describe a different set; use a new output_dir for those.

  • Adding a slack level to slack_levels is fine: it gets its own slack_<ε>/ directory and starts from scratch, while existing levels resume.

  • max_iter counts the iterations of the current call, so max_iter=300 adds up to 300 iterations to what is already there.

The run summary

When a slack level finishes, generate_mga_files writes slack_<ε>/run.json: the level’s settings and result.

{
  "slack": 0.05, "norm": "linf", "normalize": true,
  "tol": 5.0, "tol_unit": "percent",
  "vertex_method": "bb", "objective_cuts": true,
  "iterations": 31, "gap": 4.45, "converged": true
}

tol and gap are in the unit tol was given in, a percentage when normalized. iterations counts all nearest-point solves of the level, including those of earlier calls that were resumed. A level that did not converge still gets a summary, with "converged": false; a level without one was interrupted.

Exporting the results as a near-optimal set needs a summary for every slack level: they become the set’s slack_levels.

A finished run without summaries, from a version that did not write them yet, gets them by resuming it with max_iter=0 and the run’s original settings (norm, normalize, tol, sep_tol, vertex_method, use_objective_cuts). With max_iter=0 no new nearest-point problem is solved: the results are read back, the gap is measured on them, and the summary is written. Check the log: it should report every box direction as already cached. If it starts solving box directions instead, the files do not match these settings.