Performance compared to modde

The purpose of this package is enable the user to find D-optimal designs in the most general setting. doe can be used for larger models (e.g. this example) with more general constraints like NChooseK constraints (see here) or multiple mixture constraints (see here). Unfortunately, the performance is decreased compared to modde in cases where modde can actually find a design.

To see this, one can determine models for the following problems using doe.

Problem 1

problem = opti.Problem(
    inputs=[opti.Continuous(f"x{i+1}", [0, 1]) for i in range(12)]
    + [opti.Continuous(f"p{i+1}", [-1, 1]) for i in range(3)],
    outputs=[opti.Continuous("y")],
    constraints=[opti.LinearEquality(names=[f"x{i+1}" for i in range(12)], rhs=1)],
)

Problem 2

problem = opti.Problem(
    inputs=[opti.Continuous(f"x{i+1}", [0, 1]) for i in range(12)]
    + [opti.Continuous(f"p{i+1}", [-1, 1]) for i in range(3)],
    outputs=[opti.Continuous("y")],
    constraints=[
        opti.LinearEquality(names=[f"x{i+1}" for i in range(12)], rhs=1),
        opti.LinearInequality(names=["x1", "x2", "x3"], rhs=0.7),
        opti.LinearInequality(names=["x1", "x2", "x3"], lhs=-1, rhs=-0.3),
        opti.LinearInequality(names=["x10", "x11", "x12"], rhs=0.8),
        opti.LinearInequality(names=["x10", "x11", "x12"], lhs=-1, rhs=-0.2),
        opti.LinearInequality(names=["x5", "x6"], lhs=[-1, 0.5], rhs=0),
        opti.LinearInequality(names=["x5", "x6"], lhs=[-1, 2], rhs=0),
    ],
)

The diagram below shows the D optimality for the modde result, the doe result and a random design (as a reference).