Bucher, Max (2018)
Optimality Conditions and Numerical Methods for a Continuous Reformulation of Cardinality Constrained Optimization Problems.
Technische Universität Darmstadt
Ph.D. Thesis, Primary publication
|
Text
mbucher_cardinality_reformulation180919.pdf Copyright Information: CC BY-SA 4.0 International - Creative Commons, Attribution ShareAlike. Download (1MB) | Preview |
Item Type: | Ph.D. Thesis | ||||
---|---|---|---|---|---|
Type of entry: | Primary publication | ||||
Title: | Optimality Conditions and Numerical Methods for a Continuous Reformulation of Cardinality Constrained Optimization Problems | ||||
Language: | English | ||||
Referees: | Schwartz, Prof. Dr. Alexandra ; Kanzow, Prof. Dr. Christian | ||||
Date: | 5 April 2018 | ||||
Place of Publication: | Darmstadt | ||||
Date of oral examination: | 1 June 2018 | ||||
Abstract: | Nonlinear constrained optimization problems can be used to model practical and theoretical questions from a vast range of areas in science and industry. In this thesis, we consider a certain class of these problems: cardinality constrained optimization problems. The goal is to minimise a, possibly nonlinear, objective function subject to given constraints, which we also allow to be nonlinear. The cardinality constraint is of our particular interest. It restricts the maximum number of nonzero components of a feasible vector. Cardinality constraints play an important role in a range of applications such as portfolio optimization, compressed sensing as well as in logistics. Since the cardinality constraint is given by a discontinuous function, theoretical results and methods from nonlinear optimization cannot be readily applied. In this thesis, we consider an approach by Burdakov et al. (2016): a reformulation of the cardinality constraint with a continuous auxiliary variable. This reformulation bears strong similarities to a mathematical program with complementarity constraints (MPCC). Similarly, it violates common constraint qualifications which ensure that first order optimality conditions hold at a local minimum. For this reason custom constraint qualifications and first order optimality condition were introduced by Burdakov et al. (2016) and Červinka et al. (2016). In this thesis we follow the approach by Burdakov et. al. (2016). We introduce new second order optimality conditions for the reformulation which hold under custom constraint qualifications. These include a necessary second order optimality condition, a sufficient second order optimality condition and a result on the local uniqueness of stationary points. Additionally, we deduce counterparts of those results for the original cardinality constrained problem. Moreover, the existence of a local error bound for the reformulation is shown. These optimality conditions as well as the result on the existence of a local error bound are subsequently used for the convergence theories of numerical methods. In this thesis, we furthermore place emphasis on numerical methods. We proof exactness of a penalty term using the existence of a local error bound. For the case that non negativity constraints are additionally present, we consider an $\ell^1$-penalty term. This special case occurs, for instance, in portfolio optimization. For this approach, we show that Karush-Kuhn-Tucker points of a penalised problem, for increasing penalty parameters, fulfil a necessary optimality condition for the reformulation in the limit. Furthermore, we consider the application of a sequential quadratic programming (SQP) method to the reformulation by using a piecewise decomposition of the feasible set. Our theoretical results yield an explanation for the results delivered by a (standard) SQP solver applied to the reformulation. Moreover, we consider regularisation methods for the reformulation. We prove convergence of a Scholtes-type regularisation and an exponential regularisation. The former already proved to perform well numerically for MPCCs. Using the second order optimality conditions, we expand the convergence theory for this method. We then show how the approach can be used to expand the convergence theory of a whole class of regularisation methods. Subsequently we present and discuss computational results. We use the reformulation as a model for sparse portfolios, which we construct using historical market data. For various time spans these portfolios yield a better Sharpe-ratio than an equal-weight portfolio. Additionally, we present computational results of the numerical methods for the reformulation. We test the methods for portfolio optimization problems using different risk measures and a range of test cases. The application of the regularisation methods yields better results than a commercial solver for nonlinear optimization problems. Among the penalty methods, the $\ell^1$-penalty method yields the best results. If the objective is to compute a good solution in a short amount of time, the Scholtes-type regularisation compares favourably against a commercial global solver, as further numerical results indicate. |
||||
Alternative Abstract: |
|
||||
URN: | urn:nbn:de:tuda-tuprints-77402 | ||||
Classification DDC: | 500 Science and mathematics > 510 Mathematics | ||||
Divisions: | Exzellenzinitiative > Graduate Schools > Graduate School of Computational Engineering (CE) 04 Department of Mathematics 04 Department of Mathematics > Optimization 04 Department of Mathematics > Optimization > Nonlinear Optimization |
||||
Date Deposited: | 25 Sep 2018 12:24 | ||||
Last Modified: | 09 Jul 2020 02:14 | ||||
URI: | https://tuprints.ulb.tu-darmstadt.de/id/eprint/7740 | ||||
PPN: | 437022536 | ||||
Export: |
View Item |