Population-based optimization on Riemannian manifolds (Q2140530)

The book is organized into two parts. The first part investigates the Information Geometrical structure of statistical models over Riemannian manifolds. This establishes a geometrical framework to construct Extended Riemannian Stochastic Derivative-Free Optimization Optimization (RSDFO) incorporating both the statistical geometry of the decision space and the Riemannian geometry of the search space. The second part of the book describes Extended RSDFO, a principled population-based meta-algorithm that uses existing manifold optimization algorithms as it's local module. The construction of Extended RSDFO is detailed rigorously from a geometrical perspective, and it's properties derived from first principles. Chapter 1 overviews the recent advancements in geometry and optimization theories, and outlines how both the manifold of data and the manifold of statistical models can be combined to construct optimization algorithms from geometrical principles. Chapter 2 establishes the domain of discourse of the book and formally introduces the essential foundations of Differential Geometry with a flavour towards numerical computation. Fundamentals of statistical manifolds studied in the field of Information Geometry are discussed in Chapter 3. In Chapter 4 the authors survey notions of volume form and intrinsic probability distributions on manifolds in the literature, which can be roughly classified into a ``geometrical'' and ``statistical'' approach. In Chapter 5 the authors begins by discussing how neither the ``geometrical'' nor ``statistical'' approach described in Chapter 4 is suitable for their purpose of establishing a geometrical framework for stochastic optimization on manifolds. In Chapter 6 the authors extend the notion of pameterized probability densities over manifolds beyond the confines of a single normal neighbourhood, overcoming the locality of the ``statistical'' approach described in Chapter 4. In Chapter 7 the authors survey the geometric aspects of two contemporary branches of optimization theories. First they review adaptations of optimization algorithms from Euclidean spaces to Riemannian manifolds. They then discuss the information geometric interpretation of population-based stochastic optimization algorithms on Euclidean spaces. In Chapter 8 the main algorithm of the book, Extended Riemannian Stochastic Derivative-Free optimization (RSDFO), is described. Chapter 9 consists of several illustrative examples comparing Extended RSDFO with state-of-the-art manifold optimization algorithms such as Riemannian Trust-Region method, Riemannian CMA-ES and Riemannian Particle Swarm optimization on the \(n\)-sphere, Grassmannian manifold, and Jacob's ladder.