Global Optimization with Hybrid Evolutionary Computation

Abstract

An investigation has been made into hybrid systems which include stochastic and deterministic optimization. This thesis aims to provide new and relevant insights into the design of the nature-inspired hybrid optimization paradigms. It combines evolutionary and gradient-based methods. These hybrid evolutionary methods yield improved performance when applied to complex global optimization tasks and recent research has shown many of such hybridization policies.The thesis has three broad contributions. Firstly, by examination of stochastic optimization, supported by case studies, we utilised the Price's theorem to formulate a new population evolvability measure which assesses the dynamical characteristics of evolutionary operators. This leads to the development of a new convergence assessment method. A novel diversity control mechanism that uses heuristic initialisation and convergence detection mechanism is then proposed. Empirical support is provided to explicitly analyse the benefits of effective diversity control for continuous optimization.Secondly, this study utilised research relevance trees to evolve hybrid systems which combine various evolutionary computation (EC) models with the sequential quadratic programming (SQP) algorithm in a collaborative manner. We reviewed the convergence characteristics of various numerical optimization methods, and the concept of automatic differentiation is applied to design a vectorised forward derivative accumulation technique; this enables provision of accurate derivatives to the SQP algorithm. The SQP serves as a local optimizer in the deterministic phase of the hybrid models. Through benchmarking on stationary and dynamic problems, results showed that the proposed models achieved sufficient diversity control, which suggests improved exploration-exploitation balance. Thirdly, to mitigate the challenges of 'inappropriate' parameter settings, this thesis proposes closed-loop adaptive mechanisms which dynamically evolve effective step sizes for the evolutionary operators. It then examines the effect of incorporating a derivative-free algorithm which extends the hybrid model to a flexible and reusable algorithmic framework.

Date of Award	1 Aug 2014
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Richard Neville (Supervisor)