13.07.2015 Views

MPEC Problem Formulations in Chemical Engineering Applications

MPEC Problem Formulations in Chemical Engineering Applications

MPEC Problem Formulations in Chemical Engineering Applications

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

RegEq(ǫ): m<strong>in</strong> f(w) (19a)s.t. h(w) = 0 (19b)g(w) ≥ 0(19c)x, y ≥ 0 (19d)x i y i = ǫ ∀i (19e)PF(ρ): m<strong>in</strong> f(w) + ρx T y (20a)s.t. h(w) = 0 (20b)g(w) ≥ 0(20c)x, y ≥ 0 (20d)Complementarity conditions may be relaxed and the problem reformulated as<strong>in</strong> the Reg(ǫ), RegComp(ǫ), or RegEq(ǫ) with a positive relaxation parameterǫ. The solution of the MPCC, w ∗ , can be obta<strong>in</strong>ed by solv<strong>in</strong>g a series of relaxedsolutions, w(ǫ), as ǫ approaches zero. These solution strategies are generally attractedto strongly stationary po<strong>in</strong>ts, but with some exceptions (see [1]). Underassumptions of <strong>MPEC</strong>-MFCQ and sufficient second order <strong>MPEC</strong> conditions [24],Reg(ǫ) and RegComp(ǫ) exhibit local convergence of ‖w(ǫ) − w ∗ ‖ ≤ O ( ǫ 1/2) .(These properties can be strengthened to O (ǫ) if we can assume strict complementarityof the bound multipliers for i ∈ I X ∪ I Y .) Reg(ǫ) has been implementedand tested <strong>in</strong> a version of IPOPT called IPOPT-C [22]. On the otherhand, RegEq(ǫ) will exhibit slower convergence with ‖w(ǫ) − w ∗ ‖ ≤ O ( ǫ 1/4)under the same assumptions.In a related development, Nonl<strong>in</strong>ear Complementarity <strong>Problem</strong> (NCP) functionsand smooth<strong>in</strong>g functions have also been used extensively to solve MPCCs[3, 10, 14, 28]. NCP functions replace the complementarity condition <strong>in</strong> theproblem with an equivalent nonl<strong>in</strong>ear equation. One widely studied NCP functionis the Fischer-Burmeister function:φ(x, y) = x + y − √ x 2 + y 2 (21)As this function is non-differentiable at x = y = 0, it is usually smoothed to theform:φ(x, y) = x + y − √ x 2 + y 2 + ǫ (22)for some small ǫ > 0. The solution to the orig<strong>in</strong>al problem is then recovered bysolv<strong>in</strong>g a series of problems as ǫ approaches zero. An equivalence can be madebetween this method and RegEq(ǫ). Accord<strong>in</strong>gly, the problem will converge atthe same slow convergence rate.In contrast to the regularized formulations, we also consider the exact l 1penalization shown <strong>in</strong> PF(ρ). Here the complementarity can be moved fromthe constra<strong>in</strong>ts to the objective function and the result<strong>in</strong>g problem is solvedfor a particular value of ρ. If ρ ≥ ρ c , where ρ c is the critical value of thepenalty parameter, then the complementarity constra<strong>in</strong>ts will be satisfied at8

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!