Introduction

Obtaining accurate numerical solutions to SDEs has been the focus of numerous studies owing to their relevance in many fields of engineering and science [Platen 1987, Kloeden and Platen 1992, Milstein 1995, Higham et al. 2002]. In the context of stochastic filtering (Chapters 6 and 7), it was seen that the process model is often a set of non-linear SDEs and imprecise integration techniques for these SDEs may precipitate significant numerical errors in the predicted particle locations leading to possible degradation in the filter performance. Determining solutions—strong or weak—of stochastically driven non-linear oscillators by direct numerical integration of the associated SDEs has been dealt with in Chapters 4 and 5. In particular, a universal framework for integration schemes is provided in Chapter 5 through an MC approach. There it has been shown that the Ito–Taylor expansion, which is based on an iterated Ito's formula, helps to construct integration schemes for SDEs. The possibility of developing higher order numerical integration schemes, e.g., the Milstein method [Milstein 1995], numeric–analytical techniques of LTL (locally transversal linearization) type [Roy 2000, 2001, 2004] and the stochastic Newmark method [Roy 2006], has been demonstrated along with the estimation of the order of accuracy of the higher order numerical integration schemes. However, unlike ordinary DEs, deriving higher order numerical schemes for SDEs is generally hindered by the difficulty of computing higher order MSIs. On the other hand, avoidance of the higher order MSIs, which implies retaining fewer terms in the hierarchical stochastic Taylor's approximation used to construct the integration scheme, naturally achieves relatively lower order accuracy. Most of the lower order explicit schemes (for instance, the explicit version of the EM scheme) may lose stability, for instance in the case of stiff SDEs, thus requiring impracticably low time steps to get stable solutions. Thanks to its computational expedience and ease of implementation, a lower order scheme would be ideal, were it not for a loss of integration accuracy.

Introduction

The efficacy of the concept of change of measures was demonstrated in the last few chapters in the context of non-linear stochastic filtering—a tool that also has considerable scientific usefulness in developing numerical schemes for system identification problems. This chapter also concerns an application of the same notion leading to a paradigm [Sarkar et al. 2014] on global optimization problems, wherein solutions are guided mainly through derivative-free directional information computable from the sample statistical moments of the design (state) variables within a MC setup. Before the ideas on this approach are presented in some detail, it is advisable to first focus on some of the available methodologies/strategies for solving such optimization problems.

In most cases of practical interest, the cost or objective functional, whose extremization solves the optimization problem, could be non-convex, non-separable and even non-smooth. Here separability means that the cost function can be additively split in terms of the component functions and the optimization problem may actually be split into a set of sub-problems. An optimization problem is convex if it involves minimization of a convex function (or maximization of a concave function) where the admissible state variables are in a convex set. For a convex problem, a fundamental result is that a locally optimal solution is also globally optimal. The classical methods [Fletcher and Reeves 1964, Fox 1971, Rao 2009] that mostly use directional derivatives are particularly useful in solving convex problems (Fig. 9.1). Non-convex problems, on the other hand, may have many local optima, and choosing the best one (i.e., the global extremum) could be an extremely hard task. In global optimization, we seek, in the design or state or parameter space, the extremal locations of nonconvex functions subject to (possibly) nonconvex constraints. Here the objective functional could be multivariate, multimodal and even non-differentiable, which together precludes applying a gradient-based Newton–step whilst solving the optimization problem.

Introduction

In the last chapter, we have laid down a procedure to solve an optimization problem by first posing it as a martingale problem (see Section 4.12, Chapter 4), whose solution may lead to a local extremization of the cost functional. The stochastic search is next guided to reach global maximum by random perturbation strategies—coalescence and scrambling—specifically devised for the purpose. To realize a single reliable scheme that satisfies the diverse and conflicting needs of an optimization problem defined in terms of multi-cost functions under prescribed constraints is a tough task [Fonseca and Fleming 1995, Deb 2001]. This chapter addresses precisely this issue and considers some modifications to the skeletal optimization approach considered in the last chapter so as to impart greater flexibility with which the innovation process may be designed in the presence of conflicting demands en route to the detection of the global extremum. The efficiency of the global search basically relies upon the ability of the algorithm to explore the search space whilst preserving some directionality that helps in quickly resolving the nearest extremum. The development of the modified setup, referred to as COMBEO (Change Of Measure Based Evolutionary Optimization), recognizes the near impossibility of a specific optimization scheme performing uniformly well across a large class of problems. Recognition of this fact had earlier [Hart et al. 2005, Vrugt and Robinson 2007] led to a proposal of an evolutionary scheme that simultaneously ran different optimization methods for a given problem with some communications built amongst the updates by the different methods. We herein similarly aim at combining a few of the basic ideas for global search used with some well-known optimization schemes under a single unified framework propped up by a sound probabilistic basis.

In a way to be explained in the sections to follow, the ideas (or their possible generalizations) behind some of the existing optimization methods may sometimes be readily included in the present setting—COMBEO—often by just tweaking the innovation process and attempting to incorporate the best practices of some of the available stochastic search methods.