Custom MATLAB problem libraries

A custom provider can live in any user-owned writable directory. It must define a selector and loader with stable MATLAB function names:

• myproblems_select(options) returns the matching problem names;

• myproblems_load(name) returns a Problem.

A loader may attach a feasible reference fact to the problem it returns through the optional reference field of the Problem input struct: a struct with exactly the fields merit, kind, source and mapping (the token 'feasible_objective/1'), for example a known optimal value read from an offline catalog. The record is validated when the problem is built; validating it evaluates nothing, so it adds no callback call to the load, and no solver runs. A malformed record, including one with the superseded fields fun, maxcv or point or with an unknown mapping, makes the load fail and is never repaired. Problems without a reference load as before, with an unknown reference.

Register the root and canonical functions once:

registration = struct( ...
    'name', 'myproblems', ...
    'root', '/path/to/myproblems', ...
    'select_function', 'myproblems_select', ...
    'load_function', 'myproblems_load');
registerProblemLibrary(registration);

Then select it like a built-in provider:

options.plibs = {'s2mpj', 'myproblems'};
scores = benchmark({@solver1, @solver2}, options);

An identical repeated registration is harmless. The same public name cannot silently be redirected to a different root or different functions; unregister the old provider before registering a replacement.

unregisterProblemLibrary('myproblems');

Unregistering removes the registry and MATLAB path entries but preserves the provider directory and its data. Generic tools may call resolveProblemLibrary('myproblems') to obtain validated function handles. Ordinary benchmark users need only options.plibs.