Limited Memory Steepest Descent Methods for Nonlinear Optimization

This dissertation concerns the development of limited memory steepest descent (LMSD) methods for solving unconstrained nonlinear optimization problems. In particular, we focus on the class of LMSD methods recently proposed by Fletcher, which he has shown to be competitive with well-known quasi-Newton methods such as L-BFGS. However, in the design of such methods, much work remains to be done. First of all, Fletcher only showed a convergence result for LMSD methods when minimizing strongly convex quadratics, but no convergence rate result. In addition, his method mainly focused on minimizing strongly convex quadratics and general convex objectives, while when it comes to nonconvex objectives, open questions remain about how to effectively deal with nonpositive curvature. Furthermore, Fletcher\u27s method relies on having access to exact gradients, which can be a limitation when computing exact gradients is too expensive. The focus of this dissertation is the design and analysis of algorithms intended to solve these issues.In the first part of the new results in this dissertation, a convergence rate result for an LMSD method is proved. For context, we note that a basic LMSD method is an extension of the Barzilai-Borwein ``two-point stepsize\u27\u27 strategy for steepest descent methods for solving unconstrained optimization problems. It is known that the Barzilai-Borwein strategy yields a method with an R-linear rate of convergence when it is employed to minimize a strongly convex quadratic. Our contribution is to extend this analysis for LMSD, also for strongly convex quadratics. In particular, it is shown that, under reasonable assumptions, the method is R-linearly convergent for any choice of the history length parameter. The results of numerical experiments are also provided to illustrate behaviors of the method that are revealed through the theoretical analysis.The second part proposes an LMSD method for solving unconstrained nonconvex optimization problems. As a steepest descent method, the step computation in each iteration only requires the evaluation of a gradient of the objective function and the calculation of a scalar stepsize. When employed to solve certain convex problems, our method reduces to a variant of LMSD method proposed by Fletcher, which means that, when the history length parameter is set to one, it reduces to a steepest descent method inspired by that proposed by Barzilai and Borwein. However, our method is novel in that we propose new algorithmic features for cases when nonpositive curvature is encountered. That is, our method is particularly suited for solving nonconvex problems. With a nonmonotone line search, we ensure global convergence for a variant of our method. We also illustrate with numerical experiments that our approach often yields superior performance when employed to solve nonconvex problems.In the third part, we propose a limited memory stochastic gradient (LMSG) method for solving optimization problems arising in machine learning. As a start, we focus on problems that are strongly convex. When the dataset is too large such that the computation of full gradients is too expensive, our method computes stepsizes and iterates based on (mini-batch) stochastic gradients. Although in stochastic gradient (SG) methods, a best-tuned fixed stepsize or diminishing stepsize is most widely used, it can be inefficient in practice. Our method adopts a cubic model and always guarantees a positive meaningful stepsize, even when nonpositive curvature is encountered (which can happen when using stochastic gradients, even when the problem is convex). Our approach is based on the LMSD method with cubic regularization proposed in the second part of this dissertation. With a projection of stepsizes, we ensure convergence to a neighborhood of the optimal solution when the interval is fixed and convergence to the optimal solution when the interval is diminishing. We also illustrate with numerical experiments that our approach can outperform an SG method with a fixed stepsize