Passing constraint values to tell

[nikohansen]

Should we allow using tell like (based on a suggestion by @brockho)

tell(solutions, objective_values, constraints_values=constraints_values)

Implementation thoughts: the above call would probably instantiate a AugmentedLagrangian instance and call set_coefficients. The instance call takes a g-value list and returns a penalties list. The penalties sum is meant to be added to the objective value similar to penalty boundary handling. update can be called with the best solution afterwards.

See also ConstrainedFitnessAL which uses an AugmentedLagrangian instance and cma.fmin_con2 which uses the former.

Next steps:

• outline/implement a AugmentedLagrangian ask-and-tell use case scenario similar to the one with ConstrainedFitnessAL found in this notebook trying to be less opaque than in these docs
• evaluate the advantages/disadvantages to incorporate this functionality directly into tell.

[nikohansen]

Attempt of a simplified ask-and-tell use case:

def F_penalized(al, F, G):
    """update `al` and return penalized `F`"""
    al.set_coefficients(F, G)
    F1 = [f + sum(al(g)) for f, g in zip(F, G)]
    i = F1.index(min(F1))  # argmin
    al.update(F[i], G[i])
    return [f + sum(al(g)) for f, g in zip(F, G)]

es = cma.CMAEvolutionStrategy(...)
al = AugmentedLagrangian(es.N)
while not es.stop():
    X = es.ask()
    F = [objective(x) for x in X]
    G = [constraints(x) for x in X]
    es.tell(X, F_penalized(al, F, G))

This looks like it could be easily implemented as an F-penalization within tell which creates an AugmentedLagrangian instance (only) when constraints are passed.