The majority of stochastic optimisation methods such as Simulated Annealing (SA), Evolutionary Algorithms (EA), Ant Colony Optimisation (ACO) and Estimation of Distribution Algorithms (EDA) have a range of adjustable parameters like learning rates, crossover probabilities, pheromone evaporation rates and weighting factors. Poor algorithm parameterisation hinders the discovery of good solutions. The parameter values required for optimal algorithm performance are known to be problem-specific, often even specific to the problem instance at hand. Practitioners often apply stochastic methods with parameter values chosen on the basis of few tuning iterations, in which various parameter settings are explored in an attempt to fine-tune the algorithm to their particular problem. Depending on the number of parameters and their plausible value ranges, investigative trials for parameter optimisations can themselves be attempts to solve a combinatorially complex problem. Moreover, it has also been established that some of the parameter values ought to vary during the search process for best algorithm performance. Acknowledging these facts, many researchers have shifted their focus to parameter control methods, where parameter values are optimised based on algorithm performance. This thesis presents an adaptive parameter control method which redefines parameter values repeatedly based on a separate optimisation process that receives its feedback from the primary optimisation algorithm. The feedback is used for a projection of the value performing well in the future. The method uses an evaluation of the recent performance of previously applied parameter values and predicts how likely each of the parameter values is to produce optimal outcomes in the next cycle of the algorithm. The parameter values are sampled from intervals which are adapted dynamically, a method which has proven particularly effective and outperforms all existing adaptive parameter controls significantly. To test the applicability of the approach to a real-world problem, the adaptive parameter control method is applied to a case-study from the automotive industry.
Copyright © 2012 Aldeida Aleti.
A thesis submitted in fulfillment of the requirements of the degree of Doctor of Philosophy, Swinburne University of Technology, 2012.