## Regularized interior point methods for convex programming

##### View/Open

##### Date

23/03/2022##### Author

Pougkakiotis, Spyros

##### Metadata

##### Abstract

Interior point methods (IPMs) constitute one of the most important classes of optimization methods, due to their unparalleled robustness, as well as their generality. It is well known that a very large class of convex optimization problems can be solved by means of IPMs, in a polynomial number of iterations. As a result, IPMs are being used to solve problems arising in a plethora of fields, ranging from physics, engineering, and mathematics, to the social sciences, to name just a few. Nevertheless, there remain certain numerical issues that have not yet been addressed. More specifically, the main drawback of IPMs is that the linear algebra task involved is inherently ill-conditioned. At every iteration of the method, one has to solve a (possibly large-scale) linear system of equations (also known as the Newton system), the conditioning of which deteriorates as the IPM converges to an optimal solution. If these linear systems are of very large dimension, prohibiting the use of direct factorization, then iterative schemes may have to be employed. Such schemes are significantly affected by the inherent ill-conditioning within IPMs.
One common approach for improving the aforementioned numerical issues, is to employ regularized IPM variants. Such methods tend to be more robust and numerically stable in practice. Over the last two decades, the theory behind regularization has been significantly advanced. In particular, it is well known that regularized IPM variants can be interpreted as hybrid approaches combining IPMs with the proximal point method. However, it remained unknown whether regularized IPMs retain the polynomial complexity of their non-regularized counterparts. Furthermore, the very important issue of tuning the regularization parameters appropriately, which is also crucial in augmented Lagrangian methods, was not addressed.
In this thesis, we focus on addressing the previous open questions, as well as on creating robust implementations that solve various convex optimization problems. We discuss in detail the effect of regularization, and derive two different regularization strategies; one based on the proximal method of multipliers, and another one based on a Bregman proximal point method. The latter tends to be more efficient, while the former is more robust and has better convergence guarantees. In addition, we discuss the use of iterative linear algebra within the presented algorithms, by proposing some general purpose preconditioning strategies (used to accelerate the iterative schemes) that take advantage of the regularized nature of the systems being solved.
In Chapter 2 we present a dynamic non-diagonal regularization for IPMs. The non-diagonal aspect of this regularization is implicit, since all the off-diagonal elements of the regularization matrices are cancelled out by those elements present in the Newton system, which do not contribute important information in the computation of the Newton direction. Such a regularization, which can be interpreted as the application of a Bregman proximal point method, has multiple goals. The obvious one is to improve the spectral properties of the Newton system solved at each IPM iteration. On the other hand, the regularization matrices introduce sparsity to the aforementioned linear system, allowing for more efficient factorizations. We propose a rule for tuning the regularization dynamically based on the properties of the problem, such that sufficiently large eigenvalues of the non-regularized system are perturbed insignificantly. This alleviates the need of finding specific regularization values through experimentation, which is the most common approach in the literature. We provide perturbation bounds for the eigenvalues of the non-regularized system matrix, and then discuss the spectral properties of the regularized matrix. Finally, we demonstrate the efficiency of the method applied to solve standard small- and medium-scale linear and convex quadratic programming test problems.
In Chapter 3 we combine an IPM with the proximal method of multipliers (PMM). The resulting algorithm (IP-PMM) is interpreted as a primal-dual regularized IPM, suitable for solving linearly constrained convex quadratic programming problems. We apply few iterations of the interior point method to each sub-problem of the proximal method of multipliers. Once a satisfactory solution of the PMM sub-problem is found, we update the PMM parameters, form a new IPM neighbourhood, and repeat this process. Given this framework, we prove polynomial complexity of the algorithm, under standard assumptions. To our knowledge, this is the first polynomial complexity result for a primal-dual regularized IPM. The algorithm is guided by the use of a single penalty parameter; that of the logarithmic barrier. In other words, we show that IP-PMM inherits the polynomial complexity of IPMs, as well as the strong convexity of the PMM sub-problems. The updates of the penalty parameter are controlled by IPM, and hence are well-tuned, and do not depend on the problem solved. Furthermore, we study the behavior of the method when it is applied to an infeasible problem, and identify a necessary condition for infeasibility. The latter is used to construct an infeasibility detection mechanism. Subsequently, we provide a robust implementation of the presented algorithm and test it over a set of small to large scale linear and convex quadratic programming problems, demonstrating the benefits of using regularization in IPMs as well as the reliability of the approach.
In Chapter 4 we extend IP-PMM to the case of linear semi-definite programming (SDP) problems. In particular, we prove polynomial complexity of the algorithm, under mild assumptions, and without requiring exact computations for the Newton directions. We furthermore provide a necessary condition for lack of strong duality, which can be used as a basis for constructing detection mechanisms for identifying pathological cases within IP-PMM.
In Chapter 5 we present general-purpose preconditioners for regularized Newton systems arising within regularized interior point methods. We discuss positive definite preconditioners, suitable for iterative schemes like the conjugate gradient (CG), or the minimal residual (MINRES) method. We study the spectral properties of the preconditioned systems, and discuss the use of each presented approach, depending on the properties of the problem under consideration. All preconditioning strategies are numerically tested on various medium- to large-scale problems coming from standard test sets, as well as problems arising from partial differential equation (PDE) optimization.
In Chapter 6 we apply specialized regularized IPM variants to problems arising from portfolio optimization, machine learning, image processing, and statistics. Such problems are usually solved by specialized first-order approaches. The efficiency of the proposed regularized IPM variants is confirmed by comparing them against problem-specific state--of--the--art first-order alternatives given in the literature.
Finally, in Chapter 7 we present some conclusions as well as open questions, and possible future research directions.